Latest publications from Digital Signal Processing and Image Analysis (DSB)

• Tan, Martine Hjelkrem; Aasan, Marius; Chakraborty, Rwiddhi; Arteaga, Gabriel Yanci; Choi, Changkyu & Rivera, Adín Ramírez (2026). Suppressing Non-Semantic Noise in Masked Image Modeling Representations. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 19319–19329. doi: https:/openaccess.thecvf.com/content/CVPR2026/papers/Hjelkrem-Tan_Suppressing_Non-Semantic_Noise_in_Masked_Image_Modeling_Representations_CVPR_2026_paper.pdf. Full text in Research Archive Show summary

  Masked Image Modeling (MIM) has become a ubiquitous self-supervised vision paradigm. In this work, we show that MIM objectives cause the learned representations to retain non-semantic information, which ultimately hurts performance during inference. We introduce a model-agnostic score for semantic invariance using Principal Component Analysis (PCA) on real and synthetic non-semantic images. Based on this score, we propose a simple method, Semantically Orthogonal Artifact Projection (SOAP), to directly suppress non-semantic information in patch representations, leading to consistent improvements in zero-shot performance across various MIM-based models. SOAP is a post-hoc suppression method, requires zero training, and can be attached to any model as a single linear head.

• Basnet, Laxman; Szorkovszky, Alex; Lind, Pedro; Yazidi, Anis & Bhandari, Shailendra (2026). Diffusion models for eye-gaze trajectory generation using position and velocity representations. arXiv. doi: 10.48550/ARXIV.2609.05522. Full text in Research Archive Show summary

  Eye-tracking data are expensive to collect, requiring specialized hardware and controlled laboratory conditions, and difficult to share due to privacy regulations, which together limit the quantity and diversity of gaze data available for research. We address this with two complementary denoising diffusion probabilistic models (DDPMs) for unconditional generation of continuous eye-gaze dynamics from the same visual-search eye-tracking dataset. Both use an identical FiLM-conditioned one-dimensional U-Net with self-attention (19.35\,M parameters) trained on data from visual searches performed by 28 participants, with sliding-window samples 8\,s in length. The first model generates raw two-dimensional gaze-position sequences directly, and the second generates the corresponding two-component velocity sequences; the two formulations use representation-specific preprocessing, training configurations, data partitions, and evaluation protocols. Each model is evaluated across three independent training seeds over its own fixed train/validation split; seed-aggregated metrics are reported as mean\,$\pm$\,SD while single-reference-run analyses are identified separately. The position-space model achieves a mean Jensen-Shannon (JS) divergence of $0.016\pm0.004$ over nine kinematic features, with the highest feature-wise mean below $0.030$, fixation duration within 2\% of real data, and a Fr\'echet Gaze Distance more than an order of magnitude below statistical and Markovian baselines. Under a Train-on-Synthetic-Test-on-Real protocol, synthetic-only training achieves $R^2=0.66\pm0.02$, corresponding to 82.7\% of the real-data $R^2$ point estimate. The velocity-space model achieves a mean JS divergence of $0.0065$ across velocity-component, speed, log-speed, and turning angle distributions, with a maximum of $0.015\pm0.005$. Reconstructed path length is less accurately reproduced ($0.21\pm0.02$ versus $0.03\pm0.01$ in position space), although the validation protocols differ. Taken together, the two representations indicate that unconditional diffusion reliably captures the local kinematic structure of human gaze motion, velocity-component marginals, scalar speed, and short-range temporal and directional statistics, while integrated, long-range structure such as saccade counts and cumulative path geometry remains the principal target for future stimulus- or participant-conditioned models.

• Turquet, Antoine Leo; Wüstefeld, Andreas; Baird, Alan F.; Iranpour, Kamran; Dichiarante, Anna Maria & Stokkan, Sindre [Show all 11 contributors for this article] (2026). NORFOX: An Experimental Fibre-Optic DAS Array for Seismo-Acoustic Monitoring in Norway. Pure and Applied Geophysics (PAGEOPH). ISSN 0033-4553. doi: 10.1007/s00024-026-04076-y. Full text in Research Archive Show summary

  Distributed Acoustic Sensing (DAS) is a rapidly developing technology providing spatially dense data of great value for seismic and acoustic monitoring. This paper presents NORFOX, a dedicated DAS installation in southeastern Norway. NORFOX comprises a dedicated fibre-optic array with a geometry designed for earthquake and explosion monitoring. A defining feature of the site is its co-location with the NORES seismic and infrasound arrays, enabling direct comparison of DAS observations with conventional seismic and acoustic measurements. This multi-sensor configuration provides a controlled environment for benchmarking DAS data, and it paves the way for advanced, multi-physics wavefield modelling. NORFOX records various natural and anthropogenic seismo-acoustic signals, including earthquakes, explosions, aircraft, and thunder. The ultra-dense spatial sampling provided by DAS can complement traditional monitoring networks, including those used for nuclear-test monitoring. NORFOX helps address key challenges in small-yield event detection, phase discrimination, and event classification. We describe the DAS array layout, the design rationale, the co-located instrumentation, and provide an initial characterisation of fibre performance and background noise behaviour. We also discuss practical aspects of operating the site and outline current limitations. The site is further complemented by an all-sky camera and weather station to record visual events such as meteoroids and lightning activity, and to measure environmental changes; both of which can support the interpretation of the DAS measurements. Using representative examples, we demonstrate the capability of NORFOX for DAS-based array analysis and benchmarking against conventional arrays. NORFOX therefore provides a valuable test-bed for developing, validating and comparing future DAS monitoring approaches in seismology and seismo-acoustic sensing.

• Mulani, Sufayan; Ziksari, Mahsa Sotoodeh; Austeng, Andreas & Näsholm, Sven Peter (2026). Rectified wavenumber algorithms for efficient ultrasound imaging. Ultrasonics. ISSN 0041-624X. 168. doi: 10.1016/j.ultras.2026.108227. Full text in Research Archive Show summary

  Wavenumber-domain beamforming offers substantial computational advantage over conventional time-domain algorithms in ultrasound image formation. However, these methods are found to introduce distortions at large angles and greater depths, making it nontrivial to replace conventional processing pipelines with wavenumber-domain approaches. This paper presents a wavenumber-domain framework that produces images consistent with conventional delay-and-sum (DAS) beamforming, while retaining the computational efficiency of Fourier-domain processing. Building on recent work on DAS-consistent imaging for multistatic acquisition data, the key utility of this novel algorithm is its applicability in plane-wave, focused-transmit, and diverging-wave imaging. The approach compensates for the implicit spatial-frequency filtering of earlier wavenumber-domain formulations through a modified Fourier-domain weighting and an axial scaling to the reconstructed image. Using data from simulations, phantom experiments, and in-vivo liver imaging, we demonstrate that the proposed method preserves DAS-equivalent amplitude, speckle statistics, resolution, and contrast. Quantitative image comparison confirms a close agreement between DAS and the proposed method with Structural Similarity (SSIM) values exceeding 0.99, whereas SSIM between earlier wavenumber-domain beamformers and DAS is lower, ranging from 0.67 to 0.92 depending on the transmit scheme. The proposed method yields approximately one order-of-magnitude reduction in computational cost compared to conventional DAS, making it well-suited for real-time and resource-constrained ultrasound systems.

• Kjetil, Indrehus,; Duric, Adrian; Choi, Changkyu & Ramezani-Kebrya, Ali (2026). Towards Self-Explainable Document Visual Question Answering with Chain-of-Explanation Predictions. arXiv. 2605.06058. doi: https:/arxiv.org/pdf/2605.06058. Full text in Research Archive Show summary

  Document Visual Question Answering (DocVQA) requires vision-language models to reason not only about what information in a document is relevant to a question, but also where the answer is grounded on the page. Existing DocVQA models entangle question-relevant evidence and answer localization and operate largely as black boxes, offering limited means to verify how predictions depend on visual evidence. We propose CoExVQA, a self-explainable DocVQA framework with a grounded reasoning process through a chain-of-explanation design. CoExVQA first identifies question-relevant evidence, then explicitly localizes the answer region, and finally decodes the answer exclusively from the grounded region. Prediction via CoExVQA's chain-of-explanation enables direct inspection and verification of the reasoning process across modalities. Empirical results show that restricting decoding to grounded evidence achieves SotA explainable DocVQA performance on PFL-DocVQA, improving ANLS by 12% over the current explainable baselines while providing transparent and verifiable predictions.

• Suyoung, Bae,; Jaehoon, Lee,; Choi, Changkyu; YunSeok, Choi, & Jee-Hyong, Lee, (2026). Remember Your Trace: Memory-Guided Long-Horizon Agentic Framework for Consistent and Hierarchical Repository-Level Code Documentation. Advances in Neural Information Processing Systems. ISSN 1049-5258. doi: https:/arxiv.org/pdf/2605.14563. Full text in Research Archive Show summary

  Automated code documentation is essential for modern software development, providing the contextual grounding that both human developers and coding agents rely on to navigate large codebases. Existing repository-level approaches process components independently, causing redundant retrieval and conflicting descriptions across documents while producing outputs that lack hierarchical structure. Therefore, we propose MemDocAgent, a long-horizon agentic framework that generates documentation within a single, integrated context spanning the entire repository. It combines two components: (i) Dependency-Aware Traversal Guiding that predetermines a traversal order respecting dependency and granularity hierarchies; (ii) Memory-Guided Agentic Interaction, in which the agent interacts with RepoMemory, a shared memory accumulating prior work traces through read, write, and verify operations. Through an in-depth multi-criteria evaluation, MemDocAgent achieves the best performance over both open and closed-source baselines and demonstrates practical applicability in real software development workflows.

• Eilertsen, Brage; Bjørgfinsdóttir, Røskva; Vargas, Francielle & Ramezani-Kebrya, Ali (2026). Aligning Attention with Human Rationales for Self-Explaining Hate Speech Detection. Proceedings of the AAAI Conference on Artificial Intelligence. ISSN 2159-5399. 40(44), p. 37378–37378. doi: 10.1609/aaai.v40i44.41069. Full text in Research Archive Show summary

  The opaque nature of deep learning models presents significant challenges for the ethical deployment of hate speech detection systems. To address this limitation, we introduce Supervised Rational Attention (SRA), a framework that explicitly aligns model attention with human rationales, improving both interpretability and fairness in hate speech classification. SRA integrates a supervised attention mechanism into transformer-based classifiers, optimizing a joint objective that combines standard classification loss with an alignment loss term that minimizes the discrepancy between attention weights and human-annotated rationales. We evaluated SRA on hate speech benchmarks in English (HateXplain) and Portuguese (HateBRXplain) with rationale annotations. Empirically, SRA achieves 2.4× better explainability compared to current baselines, and produces token-level explanations that are more faithful and human-aligned. In terms of fairness, SRA achieves competitive fairness across all measures, with second-best performance in detecting toxic posts targeting identity groups, while maintaining comparable results on other metrics. These findings demonstrate that incorporating human rationales into attention mechanisms can enhance interpretability and faithfulness without compromising fairness.

• Jahanmard, Vahidreza; Hordoir, Robinson; Ramezani-Kebrya, Ali; Ellmann, Artu; Delpeche‐Ellmann, Nicole & Delpeche-Ellmann, Nicole (2026). Principled Fourier Neural Operators for High‐Resolution Regional Ocean Modeling. Journal of Geophysical Research (JGR): Machine Learning and Computation. 3(2). doi: 10.1029/2025jh001131. Full text in Research Archive Show summary

  Abstract Reliable and computationally efficient ocean state forecasts are essential for climate resilience, maritime safety, and science‐to‐decision applications. Growing demand has stimulated interest in machine learning approaches as scalable complements to traditional numerical models. Neural operators have emerged as promising tools for learning solution operators of governing partial differential equations in ocean and weather prediction. Despite encouraging progress, challenges remain in maintaining long‐term stability and spectral fidelity, particularly for high‐frequency processes. In this study, we investigate the impact of incorporating temporal structure into operator learning on autoregressive stability in high‐resolution regional ocean emulation. We compare the standard Fourier Neural Operator (FNO) with a modified variant, FNOtD, which incorporates joint spatial–temporal Fourier parameterization to better represent spatiotemporal spectral interactions relevant to unsteady ocean dynamics. Within a controlled regional configuration, FNOtD exhibits substantially improved long‐horizon stability and reduced error growth compared to the standard FNO. Furthermore, training with a multi‐lead‐time objective enhances predictive accuracy and spectral agreement. In the evaluated case study, the modified operator achieves predictive skill comparable to a state‐of‐the‐art high‐resolution numerical ocean model while operating at a fraction of the computational cost. These results demonstrate improved robustness and generalization of the operator learning in a regional setting, despite being trained on a relatively short time span and coarse‐resolution data set. However, further validation across diverse oceanic regimes is required to assess the transferability, generalizability, and operational robustness of FNOtD models.

• Han, Jiin; Bae, Suyoung; Choi, Yerim; Lee, Sangyun; Kim, Kyutae & Kim, Ryunho [Show all 9 contributors for this article] (2026). PubTables-QA: A Benchmark Toward Cross-Page Table Reasoning in Table Visual Question Answering. ACM Transactions on Knowledge Discovery from Data. ISSN 1556-4681. doi: https:/respmultimodal2026.github.io. Full text in Research Archive Show summary

  Tables are central to scientific communication, yet their interpretation depends critically on layout, continuity, and surrounding context. While existing multi-page and multi-table benchmarks formally require document-level reasoning, models can often succeed through stochastic shortcut search—locating local answer-bearing regions without navigating table structure. To address this gap, we introduce PubTables-QA, a multi-page TableVQA benchmark designed to audit responsible cross-page visual reasoning in academic documents. Built on PubTables-v2’s structural annotations, PubTables-QA provides 2,106 QA pairs that are unanswerable from any single page, compelling models to reconstruct table continuity across pages or integrate evidence across multiple tables. We further introduce a three-level QA taxonomy spanning Document, Table, and Cell/column levels. Experiments on recent MLLMs, including GPT-4o, Gemini-2.5-Pro, and Gemma-4, show that even the strongest model reaches only 32.6% accuracy. Oracle-setting analysis further shows that performance depends strongly on access to the correct page, table region, and cell-level evidence, highlighting persistent failures in faithful cross-page evidence integration.

• Jung, Kangsoo; Souibgui, Mohamed Ali; Choi, Changkyu & Palamidessi, Catuscia (2026). Selective Disclosure: Controlling Information Leakage in DocVQA Explanations. International Conference on Learning Representations (ICLR). doi: https:/iclr.cc/virtual/2026/10019246. Full text in Research Archive Show summary

  Explainable document visual question answering systems improve transparency by visualizing document regions relevant to a query, but they can unintentionally expose sensitive information through explanation outputs, even when textual answers are restricted. We analyze privacy risks arising from coarse explanations and adversarial jailbreaking queries. To mitigate this, we propose a policy-aware visual explanation sanitization framework based on a Role-centric Attribute-Based Access Control (RABAC) model, combining document structure analysis and line-level localization to mask restricted regions. Our method is model-agnostic, supports dynamic policies, and significantly reduces sensitive information leakage while preserving interpretability.

• Gorosito, Martin; Yazidi, Anis; Hermansen, Åsmund; Berg, Bjørnar; Øiestad, Britt Elin & Grotle, Margreth [Show all 8 contributors for this article] (2026). Exploring risk factors for long-term sickness absence during emerging adulthood: Continuous and discrete time models using Young-HUNT data on psychological distress and chronic pain. International Journal of Medical Informatics. ISSN 1386-5056. 219. doi: 10.1016/j.ijmedinf.2026.106551. Full text in Research Archive Show summary

  Introduction: Long-term sickness absence (LTSA) in young adults has important consequences for labour market participation and future work disability. Chronic pain and psychological distress are key risk factors and frequently co-occur, yet their combined impact during adolescence on later LTSA remains insufficiently understood. This study aims to explore factors that influence adolescents’ and young people’s risk of receiving LTSA benefits during emerging adulthood. Methods: This longitudinal study used data from the Young-HUNT1 (1995–1997; n = 8736) and Young-HUNT3 (2006–2008; n = 7935) cohorts linked to Norwegian registry data and followed into early adulthood. The outcome was time to LTSA (90 or 180 days). Associations were examined using Cox proportional hazards models and Kaplan–Meier analyses. Continuous- and discrete-time models were developed and evaluated using the concordance index, time-dependent AUC, and integrated Brier score. Risk factors were analysed using SurvSHAP, SHAP, and regression-based methods. Results: Chronic pain and co-occurring pain and psychological distress were consistently associated with increased LTSA risk (adjusted HRs between 1.3 and 1.5 for pain and between 1.6 and 1.7 for co-occurrence). In contrast, psychological distress alone showed no consistent association. Model performance was moderate and similar across approaches (C-index between 0.63 and 0.67). Key predictors included female sex, low parental education, chronic pain, poor perceived health, and indicators of early health problems. Conclusion: Adolescent chronic pain, particularly when co-occurring with psychological distress, is an important predictor of LTSA in early adulthood. While absolute LTSA levels may vary across cohorts, underlying risk patterns remain stable. More complex models did not outperform traditional approaches. These findings highlight the importance of early-life conditions and support early identification and intervention to reduce later work absence.

• Vrålstad, Anders Emil; Fosodeder, Peter; Deibele, Karin Ulrike; Nyrnes, Siri Ann; Rindal, Ole Marius Hoel & Skoura-Torvik, Vibeke [Show all 8 contributors for this article] (2026). Coherence Based Sound Speed Aberration Correction — with clinical validation in fetal ultrasound. IEEE Transactions on Medical Imaging. ISSN 0278-0062. doi: 10.1109/TMI.2026.3691415. Full text in Research Archive
• Wu, Zhiyuan; Choi, Changkyu; Yu, Shujian; Jenssen, Robert & Ramezani-Kebrya, Ali (2026). Mitigating Embedding Leakage via Latent Disruption with Controlled Reconstruction. Transactions on Machine Learning Research (TMLR). doi: https:/openreview.net/forum?id=nZWBrxJyrS. Full text in Research Archive Show summary

  Pre-trained encoders produce semantically rich latent embeddings, which, however, may expose unintended information through malicious inference or exploitation. We propose SEAL, a framework that mitigates embedding leakage by disrupting latent representations based on information-theoretic principles. It reduces the risk of potential misuse while enabling controlled reconstruction for trusted users. SEAL learns to encode controlled perturbations by minimizing the Matrix Norm-based Quadratic Mutual Information (MQMI) functional between original and perturbed embeddings within a hyperspherical latent space. Meanwhile, a private decoder, jointly trained with the SEAL encoder, is trained to reconstruct the original data that is accessible only to authorized users under an access-controlled setting. Extensive experiments on vision and text datasets demonstrate that SEAL reduces latent leakage, weakens the effectiveness of evaluated inference attacks, and preserves reconstruction under the considered setting.

• Boukraa, Mohamed Aziz; Karabiyik, Yücel; Austeng, Andreas; Holm, Sverre & Näsholm, Sven Peter (2026). Proof of concept for full-waveform inversion in ultrasound time-harmonic shear-wave elastography. Physics in Medicine and Biology. ISSN 0031-9155. 71(10), p. 1–22. doi: 10.1088/1361-6560/ae6af4. Full text in Research Archive Show summary

  Objective. In ultrasound time-harmonic elastography, tissue elasticity is estimated from the ultrasound-measured shear-wave displacement fields, generated by an external vibrator. Commonly used techniques to map from the ultrasound pulse-echo data to tissue elasticity maps, such as the phase-gradient and local-frequency estimation methods, base their estimates on a set of displacement field data and do not involve direct numerical modeling of the underlying shear-wave propagation. In the current proof-of-concept study, the objective is to develop a Full-Waveform Inversion (FWI) approach to ultrasound-based time-harmonic elastography, providing a physics-driven framework for improved tissue elasticity estimation. Approach. FWI is a methodology commonly applied in geophysics. It relies on simulating the wave propagation and scattering in heterogeneous media to model the intricate relationship between the observed wavefield and the propagation medium properties. Our FWI-based tissue-harmonic elastography approach is based on minimizing a cross-correlation based objective function that quantifies phase mismatches between measured and simulated shear waves. Results. Experiments on synthetic phantom data obtained from the fractional Kelvin-Voigt wave model demonstrate that our approach accurately reconstructs the shear wave speed of the medium, even in the presence of significant noise, and without relying on pre-filtering steps such as directional filtering, which selects waves based on their propagation direction. An improvement of approximately 0.4 m.s -1 within the inclusion region, compared to an established method, is observed for a noisy wavefield with decreasing signal-to-noise ratio, ranging from 20 to -10 dB from top to bottom of the ultrasound image. In addition, the reconstructed shear-wave speed map achieved a 50% reduction in standard deviation. Significance. This work demonstrates the potential for enhanced elasticity mapping using FWI-based reconstruction in time-harmonic elastography with prospective impacts for future clinical applications.

• Brissaud, Quentin; Kong, Chenxi; Froment, Marouchka; Wilding, John; Gülcher, Anna J. P. & Jackson, Jennifer M. [Show all 10 contributors for this article] (2026). Aerial and Space-borne Seismology on Venus: Viability and Design Implications for Future Missions. Earth and Space Science. 13(6). doi: 10.1029/2025EA004541. Full text in Research Archive Show summary

  Venus' evolution remains a mystery because of the lack of in situ geophysical data to constrain its interior structure. Recently-selected planetary missions VERITAS (NASA), DAVINCI+ (NASA), and EnVision (ESA) will investigate the planet's interior, surface, and atmospheric chemistry. However, none of these missions includes sensors capable of accurately probing Venus' crustal and mantle properties. Seismometer deployments are challenging on Venus due to high surface temperature and pressure. Acoustic balloon measurements and airglow observations—that monitor Venus' upper atmosphere glow caused by chemical and radiative processes—have been suggested as alternatives to surface deployments. However, it is critical to assess the potential of such missions under realistic conditions of geology, atmospheric states, network geometry, and seismicity using physics-based modeling. We employ a probabilistic framework to investigate detection probabilities as a function of Signal-to-Noise Ratio (SNR) for airglow and acoustic balloon missions using wave simulations, thermodynamically-consistent seismic velocity models, and realistic seismicity estimates. Our results demonstrate that the probability of detecting a single venusquake at SNR > 1 over 6 months is around 65% across an entire 3-balloon network of about 5,000 km extent. Probabilities using dayglow imager data are below 60% and below 10% using nightglow data. Seismo-volcanic sequences enhance detectability but only if high seismic activity occurs at multiple volcanoes. Long-duration missions with both airglow and balloon-borne sensors could allow seismic wave measurements over a broad range of frequencies. Our results are highly dependent on seismic velocities, attenuation, seismicity, noise levels, mission duration, and airglow-coupling efficiency which should be the focus of future studies.

• Zhou, Hao; Xu, Wentao; Zhang, Xuan; Yazidi, Anis & Yu, Ke (2026). Temporal Conditional Score Network for Multivariate Time Series Anomaly Detection. IEEE Internet of Things Journal. ISSN 2327-4662. 13(13), p. 28165–28180. doi: 10.1109/JIOT.2026.3681780. Full text in Research Archive Show summary

  Time series anomaly detection is a critical technique for secure monitoring in cyber-physical systems (CPSs). Most existing methods learn a normal data profile and identify anomalies as deviations from this profile. Recently, score-based models, which estimate the gradient of the log data probability density, have shown promise in improving data distribution estimation due to their strong generative capabilities. However, their application to time series anomaly detection in CPSs remains underexplored, with key challenges in network architecture design and the accuracy-efficiency trade-off yet to be addressed. In this paper, we propose the Temporal Conditional Score Network (TCSN), a novel framework that jointly estimates conditional scores and detects anomalies based on reconstruction error. Our approach enhances a U-Net backbone with temporal attention skip connections, preserving detailed spatio-temporal features at both variable and time-step levels. Additionally, signature matrix embedding is introduced to improve sensitivity to anomalous patterns. We further establish connections between score-based and reconstruction-based anomaly measurements, leading to a more efficient and accurate detection strategy. To ensure practical applicability, we design a full-cycle deployment pipeline that integrates data sanitization, denoising score matching–based training, hyperparameter tuning, and proactive model updates, enhancing scalability and long-term stability. Extensive experiments on public datasets demonstrate that TCSN consistently outperforms state-of-the-art baselines, achieving up to 0.2155 improvement in composition F1 score while maintaining strong robustness under varying noise levels.

• Julbø, Frida Marie Ihle; Henriksen, Audun Ljone; Pradhan, Manohar; Lindstrøm, Erika Kristina; Kostolomov, Ilyá & Oukrif, Dahmane [Show all 21 contributors for this article] (2026). Reliable classification of polyps based on artificial intelligence: a development and validation study. EClinicalMedicine. 93. doi: 10.1016/j.eclinm.2026.103826. Full text in Research Archive Show summary

  Background The shortage of pathologists presents a significant bottleneck in delivering timely and accurate diagnoses. To address this challenge, we developed the POLyp Artificial Intelligence-based RISk classifier (POLARIS), a prescreening tool designed to assist pathologists in handling the increasing volume of colorectal biopsies. Methods Using 15,079 whole-slide images (WSIs) from 2993 patients in the UK bowel cancer screening program between 2014 and 2018 and the histological diagnoses as ground truth, POLARIS was developed applying the open-source foundation model H-optimus-0, multiple instance learning, and a training scheme designed to provide robust models. Each WSI was categorised into one of five classes based on the diagnosis, reflecting increasing risk of malignancy, and regrouped into two broader categories for clinical interpretation. One category includes only samples not requiring pathological review, specifically those classified as normal tissue or tubular adenomas with low-grade dysplasia (LGD). The second category includes all other types of polyps, which are polyps recommended for review by a pathologist. A prespecified validation protocol defined how a specific model should be evaluated on a geographically external dataset comprising 10,842 WSIs from Cheltenham General Hospital between 2008 and 2019 and acquired using two scanners, Leica Aperio AT2 and Hamamatsu NanoZoomer XR. After validation, three experienced pathologists independently assessed the cases in which model prediction differed most from clinical diagnosis and, blinded to model output, reached a consensus. Findings In the external validation, POLARIS correctly identified 98.94% (95% CI 96.92–99.78%) of polyps with high-grade dysplasia (HGD) and adenocarcinoma as POLARIS-positive, and correctly classified 83.04% (81.66–84.36%) of normal and tubular adenomas with LGD as POLARIS-negative. The overall balanced accuracy was 86.65% for this prespecified primary analysis. In the tuning set, balanced accuracy was 87.64%, the sensitivity for polyps with HGD and adenocarcinoma was 100% (95% CI 97.14–100), and the specificity for normal and tubular adenomas with LGD was 81.89% (79.94–83.70). POLARIS achieved an area under the receiver operating characteristic curve (AUROC) in external validation of 0.9449 (95% CI 0.9384–0.9507) for distinguishing normal and tubular adenomas with LGD from polyps recommended for review by a pathologist and 0.9788 (0.9718–0.9844) for polyps with HGD and adenocarcinoma cases specifically. The model predicted the same class for both scanners in 97.93% (97.49–98.32%) of the cases. In the review of selected cases, the pathologists agreed with the model and not the clinical diagnosis in 92.5% of the cases. Heatmaps highlighting regions where the model indicates high-risk features closely correlated with areas annotated by expert pathologists as high-risk. Interpretation POLARIS has the potential to enhance diagnostic workflows in colorectal pathology by reliably identifying high-risk lesions and highlighting high-risk regions while substantially reducing the number of slides requiring pathologist review.

• Kildal, Wanja; Pradhan, Manohar; Kristensen, Gunnar S Balle; Vlatkovic, Ljiljana; Tobin, Kari Anne Risan & Hveem, Tarjei Sveinsgjerd [Show all 9 contributors for this article] (2026). Prognostic role of estrogen receptor, L1 cell adhesion molecule, and tumor grade in patients with endometrial carcinoma of no specific molecular profile. International Journal of Gynecological Cancer. ISSN 1048-891X. 36(5). doi: 10.1016/j.ijgc.2026.104669. Full text in Research Archive
• Acharya, Sushil; Fabian, Karl; Yazidi, Anis & Westeng, Kjetil (2026). Semi-automated bed boundary detection and log alignment using fast Walsh–Hadamard transform. Frontiers in Earth Science. 14. doi: 10.3389/feart.2026.1736164. Full text in Research Archive Show summary

  Depth mismatches between well logs acquired in different runs (e.g., electrical wireline logging (EWL) versus logging while drilling (LWD)) impede stratigraphic correlation and bias reservoir property estimation. Existing alignment methods, such as maximum cross-correlation (MaxCC), are fast but generally assume near-uniform shifts and can be noise-sensitive, whereas dynamic time warping (DTW) can accommodate nonlinear discrepancies but is computationally heavier than global-shift methods and less interpretable. We present a semi-automated Fast Walsh–Hadamard Transform (FWHT) workflow that enhances step-like responses associated with lithological transitions, quantifies boundary evidence via a Walsh Boundary Index (WBI), and uses matched boundary pairs as control points to construct a monotone, piecewise-linear depth-warping function that maps LWD depths onto an EWL reference.“Semi-automated” denotes that, once the log channels and the user-specified cutoff C and WBI threshold τ are defined, all remaining steps are automatic and deterministic. The method is applied to normalized EWL–LWD datasets containing gamma ray (GR), bulk density (RHOB), neutron porosity (NPHI), and compressional sonic slowness (DTC) logs. FWHT increases the slice-averaged Pearson correlation r (mean across logs) from 0.079 to 0.574 and reduces the Euclidean distance e from 1.863 to 0.964; the MaxCC baseline achieves a similar mean correlation (0.595) but a higher distance (1.008), consistent with its global-shift assumption. A constrained global DTW baseline achieves higher similarity ( r = 0.878; e = 0.632), but does so via dense pointwise warping that is less interpretable and can introduce locally nonuniform distortions. Scalability is evaluated on 89 GR log pairs from the Norwegian Continental Shelf: mean Pearson correlation r increases from 0.367 (unaligned) to 0.851 with FWHT (excluding 5 NA pairs), compared with 0.860 using MaxCC. With a uniform threshold ( τ = 0.08 ) , MaxCC performs better in 45 cases, FWHT in 29, and 15 are comparable. Across the evaluated wells, FWHT shows the largest performance gains relative to global-shift methods in intervals exhibiting depth-dependent (nonlinear) misalignment. Overall, the FWHT framework provides an interpretable and computationally efficient approach for boundary-controlled, nonlinear well-log depth alignment.

• Tan, Martine Hjelkrem; Aasan, Marius; Arteaga, Gabriel Yanci & Rivera, Adín Ramírez (2026). SPoT: Subpixel Placement of Tokens in Vision Transformers. Transactions on Machine Learning Research (TMLR). 3(26). Full text in Research Archive Show summary

  Vision Transformers naturally accommodate sparsity, yet standard tokenization methods confine features to discrete patch grids. This constraint prevents models from fully exploiting sparse regimes, forcing awkward compromises. We propose Subpixel Placement of Tokens (SPoT), a novel tokenization strategy that positions tokens continuously within images, effectively sidestepping grid-based limitations. With our proposed oracle-guided search, we uncover substantial performance gains achievable with ideal subpixel token positioning, drastically reducing the number of tokens necessary for accurate predictions during inference. SPoT provides a new direction for flexible, efficient, and interpretable ViT architectures, redefining sparsity as a strategic advantage rather than an imposed limitation.

• Jensen, Are Charles; Ziksari, Mahsa Sotoodeh; Austeng, Andreas & Näsholm, Sven Peter (2026). A Coherence-Restoring Subspace Projection for Adaptive Array Spectral Estimation. IEEE Access. 14, p. 37062–37071. doi: 10.1109/access.2026.3671486. Full text in Research Archive Show summary

  Wide-beam or single-transmit acquisitions often reduce local spatial coherence, breaking the narrowband model assumed by high-resolution array spectral estimators such as IAA and Capon and thereby degrading performance. We propose a discrete prolate spheroidal sequence (DPSS) subspace projection of delay-focused aperture data. This projection suppresses incoherent off-angle energy and restores local spatial coherence, enabling scanline-wise adaptive spectral estimation under severe model mismatch. Each delay-focused aperture vector is projected onto a DPSS subspace spanned by the first K eigenvectors corresponding to a small angular bandwidth. The approach is lightweight, with precomputation and a per-point complexity of O(MK), and integrates naturally into standard delay-focused processing pipelines. Frequency–angle plots reveal how the projection reconstructs coherent ridge structures that are otherwise obscured by wide‑beam incoherence. Simulations in both plane-wave and diverging-wave ultrasound scenarios demonstrate improved resolution and contrast in single-transmit wide-beam imaging. Qualitative results on recorded channel data from the public PICMUS dataset provide an experimental sanity check and validation, indicating that the same coherence-restoration behavior is observed in real recordings. All experimental validation in this work is confined to ultrasound imaging; assessment of other array-processing applications is left for future work.

• Amezua, Javier; Averbuch, Gil; Näsholm, Sven Peter & Arrowsmith, Stephen (2026). Listening to the Atmosphere: Using Infrasound Observations to Infer Atmospheric Conditions. Journal of Geophysical Research (JGR): Atmospheres. ISSN 2169-897X. 131. doi: 10.1029/2025JD044841. Full text in Research Archive Show summary

  The stratosphere and mesosphere are important regions for the prediction of weather at the Earth's surface for medium- and long-range forecasts. The availability of observations in these layers is lower than that of the troposphere, especially for the dynamics. While seeking new observational sources is important, there are existing infrasound data sets that provide indirect observations of these layers to be exploited. Infrasound waves generated at the Earth's surface travel horizontally and vertically through the atmosphere, and can be detected by sensor arrays at ranges of hundreds or thousands of kilometers. These waves are affected by the atmospheric conditions they encounter during propagation, and the integrated contributions can be observed in the recorded observations. Inverse problem techniques can be readily used to extract information from these integrated observations and provide valuable data related to the atmospheric conditions. We exploit observations from controlled ammunition explosions in Oklahoma, which generate waves traveling to 30–50 km altitude before being refracted back to the surface and detected 256 km from the explosion site. As model background we use the 10-member European Reanalysis product, valid 1 hr before the explosions. We use the Modulated Ensemble Transform Kalman Filter to combine these two sources of information and obtain updated atmospheric profiles. The assimilated observations bring the atmospheric profiles closer to those obtained by solely interpolating the reanalysis product to the time of the explosions. The most benefited altitudes are those close to the refraction heights of the infrasound waves, 35–55 km.

• Mulani, Sufayan Ikabal; Ziksari, Mahsa Sotoodeh; Austeng, Andreas & Näsholm, Sven Peter (2026). Delay-and-Sum Consistent Wavenumber Algorithm. IEEE Transactions on Ultrasonics. 73(3), p. 379–390. doi: 10.1109/tuson.2026.3667456. Full text in Research Archive Show summary

  Fourier-domain beamforming methods offer a computationally efficient alternative to time domain methods in ultrasound imaging. Among these, the wavenumber algorithm (WA) achieves fast image reconstruction from full-matrix capture (FMC) data but introduces amplitude distortions due to its inherent weighting. In this study, we present a modified Fourier-domain beamforming method that operates on FMC data and reproduces the amplitude response of the conventional delay-and-sum (DAS) method while maintaining the computational efficiency of WA. This is achieved through a derivation of a correction factor that restores DAS-like amplitude characteristics in WA. The correction is implemented in two stages, consisting of a matrix multiplication to the data in the Fourier domain and a subsequent scaling with the z coordinate in the reconstructed image. Validation on simulated point targets, phantom experiment, and in-vivo imaging demonstrates that the proposed method produces images nearly indistinguishable from DAS but with significantly reduced computational burden. We also investigate the role of zero-padding, showing how interpolation accuracy affects reconstruction fidelity and computational efficiency. The proposed approach can, for example, facilitate the development of ultrasound systems with reduced power consumption.

• Skrede, Ole-Johan; Pradhan, Manohar; Isaksen, Maria; Hveem, Tarjei Sveinsgjerd; Vlatkovic, Ljiljana & Nesbakken, Arild [Show all 23 contributors for this article] (2026). Generalisation of automatic tumour segmentation in histopathological whole-slide images across multiple cancer types. NPJ precision oncology. 10. doi: 10.1038/s41698-026-01311-6. Full text in Research Archive Show summary

  Deep learning is expected to aid pathologists in tasks such as tumour segmentation. We developed a general tumour segmentation model for histopathological images and examined its performance in different cancer types. The model was developed using over 20,000 whole-slide images from over 4000 patients with colorectal, endometrial, lung, or prostate carcinoma. Performance was validated in pre-planned analyses on external cohorts with over 3000 patients across six cancer types. Exploratory analyses included over 1500 additional patients from The Cancer Genome Atlas. Average Dice coefficient was over 80% in all validation cohorts with en bloc resection specimens and in The Cancer Genome Atlas cohorts. No performance loss was observed when comparing the general model with single-cancer models specialised in cancer types from the development set. In conclusion, extensive and rigorous evaluations demonstrate that generic tumour segmentation by a single model is possible across cancer types, patient populations, sample preparations and slide scanners.

• Galliamov, Karim; Kazmi, Syed M Ahsan; Khan, Adil Mehmood & Rivera, Adín Ramírez (2026). Concepts' Information Bottleneck Models. International Conference on Learning Representations (ICLR). doi: https:/openreview.net/forum?id=JGIYfwaNpT. Full text in Research Archive Show summary

  Concept Bottleneck Models (CBMs) aim to deliver interpretable predictions by routing decisions through a human-understandable concept layer, yet they often suffer reduced accuracy and concept leakage that undermines faithfulness. We introduce an explicit Information Bottleneck regularizer on the concept layer that penalizes I(X; C) while preserving task-relevant information in I(C; Y ), encouraging minimal-sufficient concept representations. We derive two practical variants (a variational objective and an entropy-based surrogate) and integrate them into standard CBM training without architectural changes or additional supervision. Evaluated across six CBM families and three benchmarks, the IB-regularized models consistently outperform their vanilla counterparts. Information-plane analyses further corroborate the intended behavior. These results indicate that enforcing a minimal-sufficient concept bottleneck improves both predictive performance and the reliability of concept-level interventions. The proposed regularizer offers a theoretic-grounded, architecture-agnostic path to more faithful and intervenable CBMs, resolving prior evaluation inconsistencies by aligning training protocols and demonstrating robust gains across model families and datasets.

• Arteaga, Gabriel Yanci; Aasan, Marius; Chakraborty, Rwiddhi; Tan, Martine Hjelkrem; Silva, Thalles & Kampffmeyer, Michael [Show all 7 contributors for this article] (2026). Why Prototypes Collapse: Diagnosing and Preventing Partial Collapse in Prototypical Self-Supervised Learning. International Conference on Learning Representations (ICLR). doi: https:/openreview.net/forum?id=fVJEWdwvLO. Full text in Research Archive Show summary

  Prototypical self-supervised learning methods consistently suffer from partial prototype collapse, where multiple prototypes converge to nearly identical representations. This undermines their central purpose—providing diverse and informative targets to guide encoders toward rich representations—and has led practitioners to over-parameterize prototype sets or add ad-hoc regularizers, which mitigate symptoms rather than address the root cause. We empirically trace the collapse to the joint optimization of encoders and prototypes, which encourages a type of shortcut learning: early in training prototypes drift toward redundant representations that minimize loss without necessarily enhancing representation diversity. To break the joint optimization, we introduce a fully decoupled training strategy that learns prototypes and encoders under separate objectives. Concretely, we model prototypes as a Gaussian mixture updated with an online EM-style procedure, independent of the encoder’s loss. This simple yet principled decoupling eliminates prototype collapse without explicit regularization and yields consistently diverse prototypes, which in several settings translate to improved downstream performance.

• Kaiser, Daniel; Frigessi, Arnoldo; Ramezani-Kebrya, Ali & Ricaud, Benjamin (2026). CogniLoad: A Synthetic Natural Language Reasoning Benchmark With Tunable Length, Intrinsic Difficulty, and Distractor Density. International Conference on Learning Representations (ICLR). doi: https:/openreview.net/forum?id=0Sex2H5Jnn. Full text in Research Archive Show summary

  Current benchmarks for long-context reasoning in Large Language Models (LLMs) often blur critical factors like intrinsic task complexity, distractor interference, and task length. To enable more precise failure analysis, we introduce CogniLoad, a novel synthetic benchmark grounded in Cognitive Load Theory (CLT). CogniLoad generates natural-language logic puzzles with independently tunable parameters that reflect CLT's core dimensions: intrinsic difficulty ($d$) controls intrinsic load; distractor-to-signal ratio ($\rho$) regulates extraneous load; and task length ($N$) serves as an operational proxy for conditions demanding germane load. Evaluating 22 SotA reasoning LLMs, CogniLoad reveals distinct performance sensitivities, identifying task length as a dominant constraint and uncovering varied tolerances to intrinsic complexity and U-shaped responses to distractor ratios. By offering systematic, factorial control over these cognitive load dimensions, CogniLoad provides a reproducible, scalable, and diagnostically rich tool for dissecting LLM reasoning limitations and guiding future model development.

• Mathema, Rujeena; Nav, Shamimeh Mohajeri; Bhandari, Shailendra; Regmi, Manoj; Lind, Pedro & Yazidi, Anis [Show all 7 contributors for this article] (2026). Comprehensive dataset of features describing eye-gaze dynamics across multiple tasks. Scientific Data. doi: 10.1038/s41597-026-06754-x. Full text in Research Archive Show summary

  We present a comprehensive dataset collected from 251 participants, which includes diverse features characterizing eye-gaze dynamics. Participants performed several tasks per session, typical in eye-tracking research: vanishing saccade, cued saccade, flickering cross, rotating ball, and free viewing. Covering these experimental paradigms, the dataset enables analysis of human eye phenomena such as oculomotor control and perceptual processing. As features, the dataset includes timestamped gaze coordinates, pupil sizes, and event classifications, labelled as fixations, saccades, or blinks. All the data were recorded using the EyeLink Portable Duo eye-tracker hardware at 1000 Hz, and processed from raw EyeLink Data Format (EDF) files into structured files via an automated pipeline. The dataset was anonymized according to the Norwegian Agency for Shared Services in Education and Research (SIKT) ethical standards. Possible applications of these datasets are the study of visual attention, cognitive science, and assistive technology.

• Bhoodoo, Preetraj; Thomas, Sarina; Wetzer, Elisabeth; Solberg, Anne H Schistad & Ben-Yosef, Guy (2026). Spatio-Temporal Landmark Detection via Selective Fine-Tuning of Echocardiography Foundation Models. Proceedings of Machine Learning Research (PMLR). 307. doi: https:/openreview.net/forum?id=91igqpzV9M&noteId=91igqpzV9M. Full text in Research Archive Show summary

  Foundation models (FMs) have shown remarkable capabilities across computer vision tasks, yet their effectiveness for complex medical downstream tasks remains underexplored. This work investigates whether state–of–the–art video–based FMs for echocardiography can perform precise spatio–temporal landmark detection without extensive fine–tuning. We evaluate two recent powerful FMs, namely EchoPrime, and PanEcho, pre–trained on few millions of echocardiographic video–text pairs, for left ventricular contour detection at end–diastole (ED) and end–systole (ES) on EchoNet–Dynamic. We compare encoder regimes (frozen, partially frozen, fully trainable) and decoder heads (multilayer perceptron (MLP) vs. graph convolutional network (GCN)), and benchmark against strong non–FM backbones (ResNet–18 2D/3D, ViT–Base, MViTv2–Small). Frozen encoders perform poorly and variably (≈78.00 Dice, ED), whereas selectively unfreezing two blocks with GCN+augmentation yields a large jump (91.71 ±3.49 Dice, ED), recovering most of the improvement. Fully trainable EchoPrime (GCN+augmentation) achieves 93.13 ±3.11/90.95 ±3.71 Dice (ED/ES), which is SOTA for regression-based models on EchoNet. Deploying separate, fully fine–tuned models for each task quickly becomes impractical in resource–constrained settings. Our results suggest that partially fine-tuning the FM is a resource-efficient strategy that recovers most of the performance benefits of end-to-end training, while avoiding the overhead of maintaining a separate model for each task. The code is available at https://github.com/preetrajb/EchoVLMLandmarks.

• Djenouri, Youcef; Belmecheri, Nassim; Michalak, Tomasz; Dubiński, Jan; Belbachir, Nabil & Yazidi, Anis (2025). Learning Graph Representation of Agent Diffusers. In y Das, Sanmay; Nowé, Ann & Vorobeychik, Yevgeniy (Ed.), AAMAS '25: Proceedings of the 24th International Conference on Autonomous Agents and Multiagent Systems. Association for Computing Machinery (ACM). ISSN 9798400714269. p. 620–629. doi: https:/dl.acm.org/doi/10.5555/3709347.3743578. Full text in Research Archive Show summary

  Diffusion-based generative models have significantly advanced text-to-image synthesis, demonstrating impressive text comprehension and zero-shot generalization. These models refine images from random noise based on textual prompts, with initial reliance on text input shifting towards enhanced visual fidelity over time. This transition suggests that static model parameters might not optimally address the distinct phases of generation. We introduce LGR-AD (Learning Graph Representation of Agent Diffusers), a novel multi-agent system designed to improve adaptability in dynamic computer vision tasks. LGR-AD models the generation process as a distributed system of interacting agents, each representing an expert sub-model. These agents dynamically adapt to varying conditions and collaborate through a graph neural network that encodes their relationships and performance metrics. Our approach employs a coordination mechanism based on top-k maximum spanning trees, optimizing the generation process. Each agent's decision-making is guided by a meta-model that minimizes a novel loss function, balancing accuracy and diversity. Theoretical analysis and extensive empirical evaluations show that LGR-AD outperforms traditional diffusion models across various benchmarks, highlighting its potential for scalable and flexible solutions in complex image generation tasks. Code can be found at: https://github.com/YousIA/LGR_AD.

• Silva, Thalles; Pedrini, Helio & Rivera, Adín Ramírez (2025). Self-Organizing Visual Prototypes for Non-Parametric Representation Learning. Proceedings of Machine Learning Research (PMLR). 267, p. 55614–55630. Full text in Research Archive
• Bøe, Vilde Schulerud; Kleppe, Andreas; Foersch, Sebastian; Wagner, Daniel-Christoph; Busund, Lill-Tove Rasmussen & Rivera, Adín Ramírez (2025). Low-Rank Adaptations for increased Generalization in Foundation Model features. Proceedings of Machine Learning Research (PMLR). doi: https:/openreview.net/pdf?id=0BJTRUVDf4. Full text in Research Archive Show summary

  For foundation models (FMs) to truly advance computational pathology, they must deliver consistent and reliable predictions under diverse, unseen test conditions. Without such robustness, clinical trust and widespread adoption remain out of reach. Although many FMs for histopathology now exist, they have to our knowledge not been systematically tested for robustness by external researchers on independent datasets. In this study, we evaluate the robustness of foundation model features on three separate histopathology datasets and find that their performance drops on external data. Our analysis also reveals that these models often encode dataset-specific information, limiting their generalizability. To address this issue, we train a Weight-Decomposed Low-Rank Adaptation (DoRA) with strong data augmentations to improve feature robustness. Our experiments show that models trained with this adapter exhibit fewer signs of dataset-specific information and may generate more robust features across domains. These results highlight the need for robustness testing and encourage incorporating robustness considerations into the development, training, and tuning of FMs for histopathology. The code for this work will be available at https://github.com/dsb-ifi/DoRA-for-FM-robustness

• Helgesen, Sander Elias Magnussen; Manet, Anthony; Cyll, Karolina; Tobin, Kari Anne Risan; Kjæreng, Marna Lill & Kostolomov, Ilya [Show all 15 contributors for this article] (2025). Reliable and Efficient Tissue Segmentation in Whole-Slide Images. Journal of machine learning research. ISSN 1532-4435. 2, p. 1–11. doi: https:/openreview.net/pdf?id=3gsNUjFi1q. Full text in Research Archive Show summary

  Whole-slide images in digital pathology often contain large regions of irrelevant background, making tissue segmentation an important preprocessing step in many applications. Traditional rule-based approaches to tissue segmentation often work quite well, but it is difficult to create general rules that cover all instances. We here apply an unmodified nnU-Net v2 training setup on downsampled whole-slide to develop and test an efficient and robust tissue segmentation model. The dataset contained nearly 30,000 images from slides with different tissue types, imaged using different scanners, and annotated using a semi-automatic workflow so that all annotations have been verified or made by human experts. This large, diverse dataset enables the training of a tissue segmentation model that generalizes well across different scanners and tissue types. We observed that our proposed model achieves similar or better accuracy than other deep learning models, while offering better robustness than simpler rule-based methods. The best compromise between inference speed and accuracy was observed using images at 10 \textmu m per pixel. Our approach can be used as an efficient and well-suited preprocessing step for computational pathology. Source code, Dockerfiles, and model weights are made publicly available at: \url{https://github.com/icgi/Reliable-and-Efficient-Tissue-Segmentation-in-Whole-Slide-Images}.

• Kaiser, Daniel; Ramezani-Kebrya, Ali; Frigessi, Arnoldo & Ricaud, Benjamin (2025). Decomposing Reasoning Efficiency in Large Language Models. International Conference on Learning Representations (ICLR). doi: https:/openreview.net/forum?id=mym1D6mIZR. Full text in Research Archive Show summary

  While large language models (LLMs) are typically evaluated on accuracy metrics alone, real-world deployment requires careful control of computational efficiency. Building on the CogniLoad benchmark, we introduce a unified efficiency metric (i.e., correct answers per 1'000 output tokens) enabling direct cross‑model comparison. We further provide an interpretable factorization into context robustness, logic robustness, and token appetite. Evaluating 15 state‑of‑the‑art reasoning LLMs on CogniLoad, we find that some models fail due to logic errors, others consume excessive tokens, and a few hit context limits. Tokens are an imperfect but practical proxy for computation load, permitting consistent comparisons across closed and open models. By decomposing overall efficiency into actionable components, our framework identifies concrete targets for improving LLM reasoning efficiency.

• Presacan, Oriana; Ojha, Jaya; Yazidi, Anis; Monteiro, Eric & Lind, Pedro G. (2025). A Comprehensive Review of Explainable AI in Deep Learning Algorithms for EEG Analysis. ACM Transactions on Computing for Healthcare (HEALTH). ISSN 2691-1957. doi: 10.1145/3785409. Full text in Research Archive Show summary

  While deep learning techniques are nowadays a powerful field to automatically learn and perform accurate EEG data classification in the clinical context, they still lack wide acceptance within the medical and health research community. This lack of trust is associated with the high complexity deep learning algorithms typically have which contributes to a low level of interpretability of the outcome and predictions. This survey aims to provide a comprehensive discussion of the latest advancements in deep learning models applied to EEG analysis, while also emphasizing research in explainable AI within this domain. It explores commonly used algorithms in EEG analysis, their main application areas, and the insights provided by XAI. Moreover, the survey addresses current limitations, such as the evaluation of XAI methods and the need for clinical validation, as well as ongoing challenges in this field. One critical insight from this review is the relative paucity of clinical evaluations of the rich stock of proposed techniques and methods for AI explainability. By showcasing various applications and breakthroughs in EEG analysis facilitated by XAI, the survey underscores the potential of these technologies to revolutionize neurological diagnosis and treatment, paving the way for wider acceptance and implementation in clinical settings.

• Kasius, J. C.; Kildal, Wanja; Vrede, S. W.; Askautrud, Hanne Arenberg; Pradhan, Manohar & Altena, A. M. van [Show all 29 contributors for this article] (2025). Surgical stage in the era of molecular profiling of endometrial cancer. European Journal of Cancer. ISSN 0959-8049. 233. doi: 10.1016/j.ejca.2025.116164. Full text in Research Archive
• Gorosito, Martin; Berg, Bjørnar; Yazidi, Anis; Hermansen, Åsmund; Øiestad, Britt Elin & Grotle, Margreth [Show all 8 contributors for this article] (2025). Disability pension during emerging adulthood: Insights from the young-HUNT study on psychological distress, chronic pain, and policy reform. International Journal of Medical Informatics. ISSN 1386-5056. 207. doi: 10.1016/j.ijmedinf.2025.106216. Full text in Research Archive Show summary

  Introduction: Disability pension (DP) among young adults has steadily increased in Norway, with chronic pain and psychological distress among the main causes. This study examined the association between co-occurring psychological distress and chronic pain in adolescence and the risk of DP during emerging adulthood, before and after the Norwegian Labour and Welfare Administration (NAV) reform. Machine learning survival models were also developed to predict early DP and identify consistent risk factors across cohorts. Methods: Data from two general population surveys in 1995–1997 (Young-HUNT1) and 2006–2008 (Young-HUNT3) were linked to public registry data for DP, with a mean follow-up of 12 years. These represent populations pre- and post-NAV reform, respectively. Crude and adjusted Cox proportional hazards models tested associations between chronic pain, psychological distress and their co-occurrence with early DP. Machine learning survival models were developed using common predictors across cohorts and validated temporally. SurvSHAP(t) and hazard ratios were used for explainable risk factor analysis. Results: Co-occurring psychological distress and chronic pain were associated with increased early DP risk in Young-HUNT3 (HR = 2.25, [95 % CI 1.62 to 3.14]) but not in Young-HUNT1. Tree-based machine learning models showed strong predictive performance and generalizability over time, with ExtraSurvivalTrees achieving a validation c-index of 0.75 [95 % CI 0.74–0.76]. Consistent risk factors included low social integration, limited physical activity, and low parental education in both cohorts, while additional factors post-reform included low self-esteem, poor self-rated health, and negative school experiences. Conclusion: Adolescents with co-occurring psychological distress and chronic pain face an elevated risk of early DP. Early identification and support are crucial, as risk factors extend beyond medical conditions. Findings also suggest that NAV policy changes may have influenced DP receipt rates among young adults.

• Lindemann, Kristina Yvonne Kathe; Kildal, Wanja; Kleppe, Andreas; Tobin, Kari Anne Risan; Pradhan, Manohar & Mascialino, Barbara [Show all 11 contributors for this article] (2025). Real-world outcomes in molecular subgroups for patients with advanced or recurrent endometrial cancer treated with platinum-based chemotherapy. International Journal of Gynecological Cancer. ISSN 1048-891X. 35(5). doi: 10.1016/j.ijgc.2024.101618. Full text in Research Archive Show summary

  Objective There is scarce real-world evidence on patients with advanced/recurrent endometrial cancer treated with platinum-based chemotherapy. We assessed the oncological outcome in groups by molecular classification. Methods This retrospective cohort study included patients with advanced/recurrent endometrial cancer treated with platinum-based chemotherapy after hysterectomy at The Norwegian Radium Hospital, Oslo University Hospital, Norway, between January 2006 and December 2017. Patients were molecularly classified as pathogenic POLE mutated, mismatch repair deficient, p53 abnormal, or no specific molecular profile. Time-to-recurrence and cancer-specific survival were calculated. Results We identified 264 advanced-stage patients (stage III/IV) and 96 patients with recurrent disease. The molecular classification was prognostic for time-to-recurrence (p < .0001) and cancer-specific survival (p < .0001) in patients with advanced disease, but the outcome did not differ significantly by molecular groups in recurrent patients. In all molecular groups, patients with stage III disease had longer time-to-recurrence and cancer-specific survival compared to patients with stage IV disease. The worst outcome was observed in patients with p53 abnormal tumors with an HR of 1.57 (95% CI 1.07 to 2.30) for time-to-recurrence and HR of 1.78 (95% CI 1.19 to 2.65) for cancer-specific survival in stage III/IV disease and an HR of 1.45 (95% CI 0.83 to 2.52) for time-to-recurrence and HR of 1.60 (95% CI 0.99 to 2.68) for cancer-specific survival in patients with recurrent disease. The few patients with POLE mutated tumors had favorable outcomes despite the advanced/recurrent disease status. Conclusions Oncological outcomes differ by molecular groups, in particular among patients with advanced disease. Patients with p53 abnormal tumors have the worst outcome, while patients with POLE mutated tumors have favorable outcomes even with recurrent disease. Implementation of the addition of immunotherapy to chemotherapy is expected to lead to substantial improvement of outcome, particularly in patients with mismatch repair deficient advanced/recurrent disease. There is still a high unmet need in advanced/recurrent patients with p53 abnormal and no specific molecular profile tumors.

• Acharya, Sushil; Fabian, Karl; Yazidi, Anis & Westeng, Kjetil (2025). Deep Hierarchical Graph Correlation: A Two-Stage Approach to Well-Log Alignment Using CNNs and Dynamic Programming. Journal of Marine Science and Engineering (JMSE). 14(1), p. 66–66. doi: 10.3390/jmse14010066. Full text in Research Archive Show summary

  Precise depth alignment of well logs is essential for reliable subsurface characterization, enabling accurate correlation of geological features across multiple wells. This study presents the Deep Hierarchical Graph Correlator (DHGC), a two-stage deep learning framework for scalable and automated well-log depth alignment. DHGC aligns a target log to a reference log by comparing fixed-size windows extracted from both signals. In the first stage, a one-dimensional convolutional neural network (1D CNN) trained on 177,026 triplets using triplet-margin loss learns discriminative embeddings of gamma-ray (GR) log windows from eight Norwegian North Sea wells. In the second stage, a feedforward scoring network evaluates embedded window pairs to estimate local similarity. Dynamic programming then computes the optimal nonlinear warping path from the resulting cost matrix. The feature extractor achieved 99.6% triplet accuracy, and the scoring network achieved 98.93% classification accuracy with an ROC-AUC of 0.9971. Evaluation on 89 unseen GR log pairs demonstrated that DHGC improves the mean Pearson correlation coefficient from 0.35 to 0.91, with successful alignment in 88 cases (98.9%). DHGC achieved an 8.2× speedup over DTW (3.16 s versus 25.83 s per log pair). While DTW achieves a higher mean correlation (0.96 versus 0.91), DHGC avoids singularity artifacts and exhibits lower variability in distance metrics than CC, suggesting improved robustness and scalability for well-log synchronization.

• Geréb, Gábor & Sándor, András (2025). Effective Representation and Fast Computing With Polyarc Bounded Intervals. Acta Cybernetica. ISSN 0324-721X. p. 1–22. doi: 10.14232/actacyb.314364. Full text in Research Archive Show summary

  Complex interval arithmetic is a powerful tool for the analysis of computational errors. The naturally arising rectangular, polar, and circular interval types yield overly relaxed bounds. The later introduced polygonal type allows for arbitrarily precise representation for a higher computational cost. We propose the polyarcular interval type as an effective generalization of the above-mentioned types. The polyarcular interval can represent all types and most of their arithmetic combinations precisely and has a better approximation capability with that of the polygonal interval. In particular, in specific cases of antenna tolerance analysis and robot localization it can achieve perfect accuracy for lower computational cost then the polygonal type, which we show in a relevant case study.

• Aasan, Marius; Tan, Martine Hjelkrem; Catalano, Nico; Choi, Changkyu & Rivera, Adín Ramírez (2025). Differentiable Hierarchical Visual Tokenization. Advances in Neural Information Processing Systems. ISSN 1049-5258. 38. doi: 10.52202/085713-2930. Full text in Research Archive Show summary

  Vision Transformers rely on fixed patch tokens that ignore the spatial and semantic structure of images. In this work, we introduce an end-to-end differentiable tokenizer that adapts to image content with pixel-level granularity while remaining backward-compatible with existing architectures for retrofitting pretrained models. Our method uses hierarchical model selection with information criteria to provide competitive performance in both image-level classification and dense-prediction tasks, and even supports out-of-the-box raster-to-vector conversion.

• Gorosito, Martin; Haugerud, Hårek; Berg, Bjørnar; Yazidi, Anis; Øiestad, Britt Elin & Sivertsen, Børge [Show all 8 contributors for this article] (2025). Chronic pain, psychological distress, and their co-occurrence in Norwegian young adults: Insights from machine learning and explainable AI. International Journal of Medical Informatics. ISSN 1386-5056. 205. doi: 10.1016/j.ijmedinf.2025.106108. Full text in Research Archive Show summary

  Introduction Chronic pain and psychological distress significantly impact young adults' health, yet little is known about their associated factors when they co-occur. This study applied machine learning classification models and explainable AI to uncover associations across diverse factors with chronic pain, psychological distress, and their co-occurrence in young adults. It also examined how associations vary when these conditions occur individually versus when they co-occur. Variable clusters were also obtained to identify latent subgroups. Methods Cross-sectional data from the Norwegian Students' Health and Well-being Study (SHoT2018) were utilised. Three classification models (Logistic Regression, Tsetlin Machine, CatBoost) were developed separately for three mutually exclusive groups (chronic pain only, psychological distress only, and their co-occurrence), against a healthy group. Model performance was assessed using Area Under the Curve (AUC) and internal-external cross-validation using geographical regions as folds. Shapley values and odds ratios were applied to identify factors associated with conditions, with associations compared across them. Clause rules were obtained to search for variable clusters. Results A total of 34,469 participants (mean age 22.9 years; 67.4% female) were included. Chronic pain, psychological distress, and co-occurrence were reported by 27.2%, 12.7%, and 26.7%, respectively. Models performed consistently across regions (AUC 0.70–0.72 for chronic pain, 0.92–0.93 for psychological distress, and 0.94–0.96 for co-occurrence. Explainable AI analyses identified females, perceived health status, and sleep disorders as factors associated with all conditions. Associations shifted when conditions co-occurred versus individually, with some negating each other while others combined, revealing distinct patterns for the co-occurring group. Conclusion The findings underscore the importance of studying these conditions both individually and simultaneously. Machine learning and explainable AI revealed unique associations that can help with early identification of subjects at risk.

• Nguyen, Anh Duc; Markov, Ilia; Wu, Frank Zhengqing; Ramezani-Kebrya, Ali; Antonakopoulos, Kimon & Alistarh, Dan [Show all 7 contributors for this article] (2025). Layer-wise Quantization for Quantized Optimistic Dual Averaging. International Conference on Machine Learning (ICML). doi: https:/openreview.net/forum?id=J6LYjEOxbz&noteId=XPlechyXOj. Full text in Research Archive
• Windecker, Stephan; Fraser, Alan G; Szymanski, Piotr; Gilard, Martine; Lüscher, Thomas F & Abid, Leila [Show all 27 contributors for this article] (2025). Priorities for medical device regulatory approval: a report from the European Society of Cardiology Cardiovascular Round Table. European Heart Journal. ISSN 0195-668X. 46(16), p. 1469–1479. doi: 10.1093/eurheartj/ehaf069. Full text in Research Archive Show summary

  Abstract The European Union (EU) Medical Device Regulation increased regulatory scrutiny to improve the safety and performance of new medical devices. An equally important goal is providing timely access to innovative devices to benefit patient care. The European Society of Cardiology strongly advocates for the evolution of the Medical Device Regulation system to facilitate priority access for innovative devices for unmet needs and orphan cardiovascular (CV) medical devices in EU countries. Although device approval is currently executed by Notified Bodies in the EU, it will be advantageous in the mid-term to consider a single EU regulatory agency for devices. In the short term, steps can be taken to transform the current system into a more efficient, predictable, cost-effective, and user-friendly service. Key strategies include the following: enhancing predictability of the approval process through use of early scientific advice from regulators; establishing unique regulatory pathways for CV orphan, paediatric, and innovative devices; promoting more efficient (re)certification of essential legacy CV devices; improving transparency of sponsor interactions with Notified Bodies; expanding the roles of the Expert Panels to assist in the approval of CV devices; promoting global regulatory harmonization, considering streamlined authorization of CV medical technologies across selected jurisdictions; developing an efficient system to monitor device safety; and ensuring funding for data collection platforms. Some strategies that could help include considering a pilot programme for joint approval processes of selected devices in partnership with other regions (i.e. US Food and Drug Administration); developing priority pathways for accelerated access to innovative or orphan devices; and increasing recognition of the importance of early feasibility studies in the EU.

• Thygesen, Sander Bøe; Sirevaag, Tore Lie & Näsholm, Sven Peter (2025). Estimating annulus sealing properties using the flexural wave spectrum in pitch-catch well-logging. Ultrasonics. ISSN 0041-624X. 146. doi: 10.1016/j.ultras.2024.107491. Full text in Research Archive Show summary

  Safe oil and gas well operations require appropriate sealing of the annulus casing. Pitch-catch ultrasound logging measurements can be used for well-barrier inspection. In the analysis of such data, an important aspect is to determine whether there is cement or mud behind the casing. This paper presents a data processing approach to differentiate between fluid and solid behind the casing from pitch-catch datasets. It is based on the spectral signature of the casing flexural wave, in which a notch-like dip might be observed. This dip is understood to occur in most solid annulus scenarios. However, when the annulus velocity exceeds a certain limit or is below a threshold, a dip is not produced. The frequency where the dip occurs is associated with an overlap between the flexural phase-velocity and the annulus P-wave velocity. This is exploited by picking the notch frequency, and then its value is used to estimate the annulus material P-wave velocity. On the basis of this insight, a method for distinguishing solids from fluids is presented. The outcome is a binary algorithm that detects a dip (or no dip), and which in addition differentiates between annulus materials using the estimated velocity. In addition, we analyze the accuracy of the velocity estimation. It is straightforward to adopt this in an operational setting. This paper demonstrates the performance and accuracy of the algorithm for both simulated data and field recordings.

• Gereb, Gabor; Lønmo, Tor Inge Birkenes; Hansen, Roy Edgar; Näsholm, Sven Peter & Austeng, Andreas (2025). Clairvoyant performance bounds for adaptive beamforming in pulse-echo imaging. Journal of the Acoustical Society of America. ISSN 0001-4966. 157(1), p. 459–470. doi: 10.1121/10.0034836. Full text in Research Archive Show summary

  In adaptive beamforming, the array signal processing adjusts its sensor delays and weights based on the incoming data. In conventional beamforming, these parameters are instead given from a predefined model. Adaptive beamformers can improve measurement precision by dynamically rejecting spatial interference. While an established theory is available on the behavior of adaptive beamformers in textbook scenarios, their expected performance on realistic pulse-echo imaging scenes is still mostly uncharted. Imaging performance can be evaluated by individual pixel precision and aggregated metrics such as resolution and contrast. The achievable gain is strongly related to the sparsity of the scene and the availability of data to appropriately estimate the spatial covariance matrix. In pulse-echo measurements, the nonstationary interference poses a special problem for adaptive beamforming, which is a current research question of academic and industrial interest. The current work establishes a performance bound for adaptive beamforming in simulated realistic pulse-echo scenarios. This is derived and numerically implemented as the clairvoyant minimum variance distortionless response beamformer. The proposed framework allows for an a priori assessment of the applicability of adaptive beamforming, for a given scenario. The performance of the implemented algorithms can be directly compared with the theoretical limit in a simulated environment.

• Maia, Julia S.; Plesa, Ana-Catalina; Zelst, Iris Van; Ghail, Richard; Gülcher, Anna J. P. & Panning, Mark P [Show all 13 contributors for this article] (2025). The Seismogenic Thickness of Venus. Journal of Geophysical Research (JGR): Planets. ISSN 2169-9097. 130(7). doi: 10.1029/2025JE009065. Full text in Research Archive Show summary

  Growing evidence that volcanism is currently ongoing on Venus suggests that the sister planet of the Earth may also be seismically active. Given the success of seismic measurements on Mars and the Moon to reveal the interior structure of these bodies, seismic investigations on Venus are a natural next step. The potential for seismic activity is closely linked to the thickness of the so-called seismogenic layer, that is, the region where rocks behave in a brittle manner and quakes can nucleate. On Earth, the seismogenic thickness is correlated with the thermal structure of the lithosphere, and is typically associated with the depth of the 600°C isotherm. Here, we combine geophysical constraints with thermal evolution models to estimate the thermal structure of Venus' lithosphere and determine the corresponding seismogenic thickness. Taking all estimates into account, our results show that the seismogenic thickness overall varies from 2 to 35 km. The lowest values are associated with areas that probably correspond to local thermal anomalies associated with magmatic processes. This interpretation is corroborated by geodynamic models, which show that intrusive magmatism can largely increase the temperature within the lithosphere at local scales. The seismogenic layer is thickest at volcanic plains which are commonly associated with regions of mantle downwellings. In these regions, the seismogenic layer likely reaches Venus' mantle, while in areas with a thick crust or anomalously high thermal gradients, quakes might be limited to the crust. Our study provides evidence that Venus has a substantial seismic potential.

View all works in NVA

• Alix, Corentin; Boulestreau, Romain; Ribot, Emeline; Couffinhal, Thierry; Karabiyik, Yücel & Salles, Sebastien Andre Roger (2026). In Vivo Detection of Guided Waves in the Carotid Wall Using a skin-Mounted Piezoelectric Patch and Ultrafast Ultrasound Imaging. Full text in Research Archive
• Tran, Thi Thuy Chau; Erdem, Cagri; Yazidi, Anis; Kjølerbakken, Kai Morgan & Griwodz, Carsten (2026). Research Proposal: Non-intrusive Stress Recognition using Multimodality Deep Learning. doi: 10.1145/3793853.3798410. Full text in Research Archive
• Singh, Durgesh Kumar; Thomas, Sarina; Cao, Qing; Kampffmeyer, Michael; Jenssen, Robert & Solberg, Anne H Schistad (2026). Semi-Supervised nnU-Net with Weak–Strong Consistency and Boundary Refinement for Mitral Valve Segmentation. doi: https:/papers.miccai.org/miccai-2026-sat/paper/MVAA_089.pdf. Full text in Research Archive Show summary

  Analysis of the mitral valve begins with delineating the valve in the image, a step that is usually performed manually and is therefore slow and observer-dependent. The MICCAI 2026 Mitral Valve Anatomy Analysis (MVAA) challenge defines three segmentation tasks based on cardiac computed tomography (CT), three-dimensional transesophageal echocardiography (3D TEE), and intraoperative surgical video. The unpaired images differ in dimensionality, appearance, and label formulation. Therefore, a separate model is trained for each task within a common self-configuring nnU-Net framework. Labeled images are supervised with soft Dice loss and cross-entropy. Since mitral valve structures are thin, these region losses are complemented by a scheduled boundary loss averaged over the available foreground classes. Expert labels are scarce for cardiac CT and surgical video, while many unlabeled images are available. These unlabeled images are used through FixMatch-style weak–strong consistency learning, where a weak view without intensity augmentation generates confidence-filtered pseudo-labels for intensity-augmented strong views. On the private test set held by the challenge organizers, the proposed approach achieves Dice similarity coefficients of 0.83 for cardiac CT, 0.81 for 3D TEE, and 0.55 for surgical video. Hausdorff distance and average surface distance are reported as boundary measures. The code is publicly available at https://github.com/Durgesh93/MVAA.git.

• Boukraa, Mohamed Aziz; Karabiyik, Yücel; Austeng, Andreas; Holm, Sverre & Näsholm, Sven Peter (2026). Improved Full-Waveform Inversion in Time-Harmonic Elastography. Full text in Research Archive
• Han, Chaoran; Alix, Corentin; Näsholm, Sven Peter; Austeng, Andreas; Salles, Sébastien & Karabiyik, Yücel (2026). Transverse Oscillation-Based Velocity Estimation in Ultrasound Time-Harmonic Elastography. Full text in Research Archive
• Vedal, Amund Hansen; Eikvil, Line; Solberg, Anne H Schistad; Salberg, Arnt Børre & Malde, Ketil (2026). Hierarchical classification and embedding quality of plankton. Full text in Research Archive Show summary

  Poster

• Ziksari, Mahsa Sotoodeh; Alix, Corentin; Austeng, Andreas; Salles, Sebastien Andre Roger & Karabiyik, Yücel (2026). Directional clutter filter imaging for time harmonic elastography. Full text in Research Archive
• Melteig, Elina; Glad, Ingrid Kristine; Ramezani-Kebrya, Ali; Pérez, Johan Manuel de Aguas; Holmsen, Sigurd & Palomares, Alfonso Diz-Lois [Show all 7 contributors for this article] (2026). Blir KI vår dommedag? Ekspertene strides. [Journal]. Titan. Full text in Research Archive
• Han, Ji-In & Choi, Changkyu (2026). PubTables-QA: a Multi-Page Table Visual Question Answering for Cross-Page Table Reasoning. doi: https:/www.ekc2026.org/page/programme/stdivision/15. Full text in Research Archive Show summary

  Tables are a central medium for communicating quantitative evidence in scientific documents, yet their compact relational form makes robust interpretation depend on layout and context. However, existing multi-page and multi-table benchmarks may still allow stochastic shortcut search, where models locate local answer-bearing regions without systematically navigating the underlying table structure. To address this gap, we introduce PubTables-QA, a multi-page TableVQA benchmark dataset carefully designed to audit cross-page visual reasoning in authentic academic documents. Built upon the extensive structural table annotations of PubTables-v2, PubTables-QA introduces 2,106 table-based QA pairs, where each question is designed to be unanswerable from a single page alone. This design compels models to reconstruct cross-page table continuity across headers, rows, columns, and cells, or to integrate evidence distributed across multiple tables. PubTables-QA further provides a three-level fine-grained QA taxonomy, Document, Table, and Cell/column levels respectively, ranging from relevant evidence localization and table-level structural understanding, to precise cell/column-level value extraction. Experiments on recent MLLMs including GPT-4o, Gemini-2.5-Pro, and Gemma-4 reveal that even the strongest models achieve only 32.6% overall accuracy. Further, our evidence-aware evaluation confirms that, unlike existing multi-page benchmarks susceptible to stochastic shortcuts, PubTables-QA requires genuine cross-page evidence integration, establishing it as a substantial and unsolved challenge.

• Choi, Changkyu; Tan, Martine Hjelkrem; Aasan, Marius & Rivera, Adín Ramírez (2026). AI in Our Daily Lives - Panel Discussion. Full text in Research Archive Show summary

  AURoRA researchers co-hosted a public seminar and panel discussion with the Embassy of the Republic of Korea to Norway, held at Litteraturhuset in Oslo as part of the K-Initiative campaign. The evening opened with the talk "Mirrors of Intelligence" by Adín Ramírez Rivera (University of Oslo). The panel then discussed trust, fairness and the societal impact of AI with the audience.

• Rivera, Adín Ramírez & Choi, Changkyu (2026). Mirrors of Intelligence. Full text in Research Archive Show summary

  The lecture focuses on how AI systems are built, what they reflect from human data, and why their outputs require critical judgment. In addition, it explains the development of LLMs and VLMs through large-scale data collection, pre-training, fine-tuning, and human-feedback-based alignment. It highlights both the strengths of AI, such as summarizing well-established knowledge and generating standard code, and its limitations, including biases, hallucinations, and sycophancy. The main takeaway is to use AI as an augmentative tool rather than a replacement for human judgment, while remaining aware of bias, questioning plausible but false information, and developing AI literacy as an essential skill.

• Choi, Changkyu (2026). Know-Thy-Not: Training-Free Targeted Negation in Dual-Encoder Vision-Language Models. doi: https:/mishmash.no/assets/events/kristiansand2026/mishmash-opening-conference-2026-abstracts.pdf. Full text in Research Archive Show summary

  Dual-encoder vision-language models systematically fail at negation, scoring "a cat without a dog"' nearly identically to "a cat with a dog.'' We show that this failure is inherent to the single-embedding framework. Under contrastive training, single-embedding methods can at best ignore the negated concept (dilution) but never penalize its presence (rejection). We propose Know-Thy-Not (KTN), a training-free method that introduces an image-side rejection dimension to penalize the negated concept directly in the visual representation. We also introduce COCO-Confusion, the first benchmark requiring genuine rejection to succeed. Across 9 backbones, KTN improves COCO-Confusion mAP by up to +25.6 over vanilla VLMs and +9.6 over the strongest existing method.

• Poiesz, Emil & Holmsen, Sigurd (2026). Hva skjer i kunstig intelligens? Siste nytt fra forskningsfronten. doi: https:/www.arendalsuka.no/program/hva-skjer-i-kunstig-intelligens-siste-nytt-fra-forskningsfronten. Full text in Research Archive Show summary

  Presentasjon under Arendalsuka om energibruk til KI modeller

• Bøe, Vilde Schulerud (2026). Foundation Models for Pathology and the need for Robust Representations. doi: https:/www.youtube.com/watch?v=7dN0KKxmXrQ&t=2s. Full text in Research Archive Show summary

  Deep learning has transformed computational pathology, enabling highly accurate models for tasks such as survival analysis and tumor detection. However, despite impressive progress, these models often struggle to maintain performance across domain shifts arising from variations in staining, scanners, institutions, and patient populations. This highlights the challenge of robustness and generalization. In this talk, I will discuss emerging efforts to build pathology models that are not only accurate, but also reliable across diverse real-world settings. Focusing on foundation models and multiple-instance learning (MIL), we will look at how sources of dataset-specific biases become embedded in learned representations and how they affect downstream performance. I will present current approaches for improving robustness, including our work on encouraging consistent representations across scanner variations during MIL training. I aim to highlight the progress that has been made and the challenges that remain in developing pathology AI systems that generalize beyond the conditions in which they were trained.

• Aasan, Marius (2026). Perceptual Grouping and Adaptive Tokenisation. doi: https:/www.youtube.com/watch?v=l6BeD85wRgQ. Full text in Research Archive Show summary

  As vision modelling is pivoting towards transformer based models, the role of tokenization — the way models partition and interpret the data — becomes increasingly important, as it is the central interface between analysis and data, that dictates how AI models synthesise semantic understanding from sensory data. This talk discusses how this interfact can be leveraged for better interpretability, incorporation of prior knowledge, and new opportunities in self-supervised learning signals. We also discuss the challenges involved in how these mechanisms currently work, and what interesting research opportunities that can be derived from tokenization in the broader context of modern representation learning.

• Tan, Martine Hjelkrem (2026). Sparse Tokens and Patch Representations. doi: https:/www.youtube.com/watch?v=3NvHZFei-S0. Full text in Research Archive
• Aasan, Marius; Wetzer, Elisabeth & Chakraborty, Rwiddhi (2026). Visual Intelligence Workshop on Multimodal AI. Full text in Research Archive Show summary

  Visual Intelligence Workshop on Multimodal AI The workshop focuses on three short talks on different aspects of multimodality. The presentations are designed to appeal to a broad target audience and provide an educational and bottom-up approach to the topic. Our goal is to provide understanding and interesting perspectives for Visual Intelligence’s academic and industry partners. The talks seek to explore the following: How multimodality shapes current AI models. ‍ How it affects how we think of data, as well as interact and understand AI models. ‍ How multimodality can help us paint a broader picture of our data by bridging modalities across domains. ‍ What types of benefits this research can bring to the Visual Intelligence consortium as a whole.

• Golfetto, Cristiana; Dahl, Andreas Reason; Mugaas, Terje; Røstum, Jon; Bruaset, Stian & Thygesen, Sander Bøe [Show all 8 contributors for this article] (2026). Multi-sensor data combined with robotics and AI for automated inspection of wastewater networks. Full text in Research Archive
• Wetzer, Elisabeth; Choi, Changkyu; Jenssen, Robert; Handegard, Nils Olav & Ebbesson, Lars O.E. (2026). Artificial Intelligence for Sustainable Fisheries: Methods, Monitoring, and Practice. doi: https:/virtual.oxfordabstracts.com/event/75716/program?date=%222026-5-27%22&g_submission=129&g_session=215457. Full text in Research Archive Show summary

  Arctic and Subarctic fisheries face increasing pressure as traditional practices struggle to keep pace with shifting ocean conditions. These challenges call for intelligent, adaptive systems to support sustainable monitoring and management. This session explores how Artificial Intelligence (AI) can contribute to the future of fisheries science by bridging algorithmic innovation with practical application and fostering interdisciplinary exchange. We invite papers that demonstrate how AI, ranging from machine learning and computer vision to emerging LLM-based agentic systems, can support fisheries research and regulation. Relevant applications include acoustic data interpretation for species identification, improved stock assessments, vessel activity monitoring, and adaptive regulatory strategies. Submissions addressing challenges such as limited computational resources, sparse data availability, or operational constraints in remote polar environments are particularly encouraged. The session also highlights the importance of engaging Indigenous knowledge holders and coastal communities in AI development. We especially welcome co-designed frameworks that promote fairness, transparency, and local agency in contexts where environmental data and governance intersect. Through this session, we aim to bring together diverse perspectives that advance responsible and locally grounded uses of AI for sustainable marine stewardship.

• Jenssen, Robert; Eikvil, Line; Solberg, Anne H Schistad; Solheim, Inger & Bjørklund, Petter (2026). Visual Intelligence Annual Report 2025. Full text in Research Archive
• Bhoodoo, Preetraj (2026). Echocardiography Foundation Models in Practice: EchoPrime, PanEcho, and Selective Fine-Tuning for Landmarks Detection. doi: https:/www.youtube.com/watch?v=xMb5n4hrqoA&t=484s. Full text in Research Archive
• Melteig, Elina & Näsholm, Sven Peter (2026). Kan påskeværet varsles i januar? [Journal]. Titan.no. doi: https:/www.titan.uio.no/teknologi/2026/kan-paskeveret-varsles-i-januar-fremtidens-vervars.html. Full text in Research Archive
• Utseth, Ingrid; Vedal, Amund Hansen; Thomas, Sarina & Eikvil, Line (2026). Comparing Foundation Models for Medical Images: A Study on Limited Data and Generalization. Full text in Research Archive Show summary

  In this study we have investigated how vision foundation models, pretrained on different domains, compete with a specialized model for classification as a function of the size of the labeled training set of medical images. Furthermore, we have looked into the different models' ability to generalize to difficult cases. Our experiments are conducted for cardiac ultrasound images and the downstream task of view recognition. Still, this classification task is meant to serve as a demonstrative example, where we think that the findings should be transferable to other classification tasks and other domains. Through these experiments we found that the foundation models were able to beat the performance of our task-specific supervised model when labelled training data were limited. This was true even for models trained on natural images and when using the simple linear probing method to create a classifier. We observed that more domain-specific foundation models achieved an even higher performance with limited data. On the other hand, the more general models showed a greater ability to generalize and perform well on difficult, out-of-distribution cases. Still, for typical in-domain cases with sufficient labeled data, a task-specific ResNet model was competitive with the foundation models, while also being both smaller and faster.

• Mylius-Kroken, Johan; Wetzer, Elisabeth; Ramezani-Kebrya, Ali; Jenssen, Robert & Wickstrøm, Kristoffer (2025). geobin: Geometric Binning Estimator. Full text in Research Archive
• Jenssen, Robert; Eikvil, Line; Solberg, Anne H Schistad; Solheim, Inger & Bjørklund, Petter (2025). Visual Intelligence Annual Report 2024. Full text in Research Archive Show summary

  The Visual Intelligence Annual Report 2024 highlights the centre's progress, activities and achieved innovations for 2022. It describes new deep learning methods which address pressing societal needs in the fields of medicine and health, marine science, the energy sector, and earth observation.

• Froment, Marouchka; Brissaud, Quentin; Näsholm, Sven Peter & Schweitzer, Johannes (2025). Behind the paper: Balloon seismology enables subsurface inversion without ground stations. Springer Nature Research Communities. doi: https:/communities.springernature.com/posts/balloon-seismology-enables-subsurface-inversion-without-ground-stations. Full text in Research Archive Show summary

  What if we can’t place a seismometer directly on the ground? Our team has just shown that infrasound waves, recorded by floating balloons, can reveal information about the seismic sources and subsurface. This finding supports the concept of using balloon missions to explore the interior of Venus.

• Bakker, M-C.E.; Franken, I.A.; Laclé, M.M.; Cyll, Karolina; Stathonikos, N. & Raedt, Sepp de [Show all 17 contributors for this article] (2025). 726O Prognostic value of the combined analysis of pathologists and artificial intelligence (CAPAI) in high-risk stage II-III colon cancer treated without chemotherapy: Interim report from a nationwide validation. Annals of Oncology. ISSN 0923-7534. 36(Supplement 2), p. S476–S476. doi: 10.1016/j.annonc.2025.08.1301. Full text in Research Archive Show summary

  Background The recommended treatment of high-risk stage II (pT4) and stage III (pN+) colon cancer (CC) is surgical resection followed by adjuvant chemotherapy (ACT). However, half of patients are cured by surgery alone and may be spared ACT if identified up-front. The Combined Analysis of Pathologists and Artificial Intelligence (CAPAI) was developed and validated to predict cancer-specific survival (CSS) in patients predominantly treated with ACT. The restrictive use of ACT in the Netherlands enables validation of the prognostic value of CAPAI when withholding ACT. Methods We identified all patients from the Netherlands Cancer Registry (NCR) with high-risk stage II and III CC with R0 resection between 2015-2019, without (neo)adjuvant treatment, under 70 years of age and with a good performance status (WHO0-1 or ASA1-2). Diagnostic H&E slides were retrieved from the nationwide database (Palga), centrally scanned, and analyzed by the DoMore-v1 deep learning biomarker. The score was combined with pT/pN stage and lymph node count to classify patients into CAPAI risk groups: low-, intermediate-, or high-risk. The primary endpoint was 3-year CSS, defined as death with known recurrence. Results This interim analysis includes 453 patients from 65 hospitals, with an overall 3-year CSS of 85.9% (95% CI 82.6-89.2%). CAPAI stratified patients in significantly distinct prognostic groups (log rank p<0.0001) with 3-year CSS of 93.7% (90.5-97.1%; n=215 [47%]) for low-, 87.5% (82.3-93.1; n=156 [34%]) for intermediate- and 60.4% (50.4-72.5; n=82 [18%]) for high-risk. In univariable Cox regression, CAPAI provided a hazard ratio for high- versus low-risk of 7.9 (4.1-15.2; p < 0.001) and for intermediate- versus low-risk of 2.1 (1.0-4.3; p=0.039). Conclusions CAPAI successfully stratified patients into distinct prognostic groups, identifying nearly half of patients as low-risk with a 3-year CSS of 93.7% without ACT. A comparison with a balanced cohort of patients who received ACT is ongoing. Prospective use of this readily available and affordable stratification tool could help avoid ACT in low-risk patients and intensify treatment in high-risk patients. Legal entity responsible for the study The authors. Funding CAPAI is developed thanks to funding by Learning from Deep learning, The Norwegian Research Council (#334862) and Improving diagnosis by utilizing Big Data and software-driven automation of pathology, The Norwegian Research Council (#259204). This public private partnership was sponsored by Health∼Holland (LSHM#19027).

• Lujan, G.; Arole, V.; Raedt, Sepp de; Meyers, K.; Pradhan, Manohar & Maupin, J. [Show all 17 contributors for this article] (2025). 52P External validation of Histotype Px® Colorectal for risk stratification in a US patient cohort. doi: https:/www.sciencedirect.com/science/article/pii/S2949820125001407?via%3Dihub. Full text in Research Archive Show summary

  Background Many patients with stage II/III colorectal cancer (CRC) are routinely offered adjuvant chemotherapy (ACT) according to guidelines. However, with significant side-effects and few patients benefitting, better risk stratification is needed. Histotype Px Colorectal is an AI-based digital biomarker that analyses routine H&E whole-slide images (WSI) and combined with pT, pN, and number of lymph nodes sampled, stratifies patients into distinct risk groups. We externally validated its prognostic performance. Methods Patients diagnosed with stage II/III colon adenocarcinoma at Ohio State University from 2011-2024 were included in this retrospective study. Blinded to clinical outcomes, anonymized routine Philips iSyntax WSIs of H&E sections were analysed by Histotype Px. Logrank test and Cox proportional hazards regression were subsequently used to analyse cancer-specific survival (CSS) and time to recurrence (TTR) to determine the prognostic performance. Results A total of 503 patients were included with a median age of 63 years (22-97), 54% male, 45% stage II, 61% right-sided tumors, 23% MSI, 51% received ACT, and 44 months median follow-up. Histotype Px was a significant predictor of CSS with hazard ratio (HR) 8.7 (95% CI 4.4-17.3, p<0.001) for high vs. low risk patients and HR 2.9 (95% CI 1.4-6.0, p=0.005) for intermediate vs. low risk patients. Histotype Px was a significant predictor of TTR with HR 4.5 (2.9-6.9, p<0.001) for high vs. low risk patients and HR 1.8 (1.2-2.9, p=0.008) for intermediate vs. low risk patients. Only 11 of 228 stage II patients had a cancer-related death, corresponding with 96% and 84% 5-year CSS for low and intermediate risk, respectively. For stage III patients, 41% were classified as low, 29% as intermediate, and 30% as high risk. 40 of the 279 stage III patients had a cancer-related death, with 5-year CSS of 92%, 83%, and 63% for low, intermediate, and high risk, respectively. Conclusions Histotype Px Colorectal is an AI-based digital biomarker that stratifies stage II/III patients into distinct risk groups. It identifies patients with exceedingly low risk of cancer-specific death who may potentially forgo ACT, and conversely, patients who may benefit from treatment escalation, thereby enabling more personalized ACT for CRC. Legal entity responsible for the study Ohio State University Comprehensive Cancer Center. Funding Ohio State University Comprehensive Cancer Center.

• Gullbekk, Sophus Bredesen; Froment, Marouchka; Brissaud, Quentin & Näsholm, Sven Peter (2025). Modeling the characteristics of infrasound signals as detected from a balloon platform: Earth and Venus applications. Full text in Research Archive Show summary

  The Strateole2 campaign deployed superpressure balloons for several weeks in the southern stratosphere. On 15 January 2022, their pressure sensors detected infrasound signals (0.01 to 1 Hz) from the major eruption of the Hunga Tonga Hunga Ha'apai (HTHH) volcano. These signals present differences in amplitude and frequency content compared to recording of ground International Monitoring System stations. Using HTHH as a case study, we apply ray-tracing and normal mode modeling techniques, coupled with ERA5 atmospheric models, to understand the particularities of the infrasound wavefield at these elevated platforms. We also propose to generalize these results to different seasons and geographical source-receiver location combinations. A key objective for our future work is to apply such frameworks to assess the possibilities for balloon based seismoacoustic probing on Venus.

• Gullbekk, Sophus Bredesen; Näsholm, Sven Peter; Froment, Marouchka & Brissaud, Quentin (2025). Morphology of infrasound wave propagation from ground to atmospheric balloons. Full text in Research Archive
• Mulani, Sufayan Ikabal; Austeng, Andreas & Näsholm, Sven Peter (2025). Delay-and-sum consistent Fourier beamforming. Full text in Research Archive
• Svoren, Mathias Horvati; Torsnes, Sigurd Kleiven & Näsholm, Sven Peter (2025). Mapping of stratospheric polar vortex from global infrasound data. Full text in Research Archive
• Froment, Marouchka; Näsholm, Sven Peter & Brissaud, Quentin (2025). Seismology using infrasound: Exploring the subsurface of Earth and Venus with balloons. Full text in Research Archive
• Pradhan, Manohar; Kildal, Wanja; Kleppe, Andreas; Tobin, Kari Anne Risan; Nikolaysen, Filip & Mascialino, Barbara [Show all 12 contributors for this article] (2025). Prognostic Value Of Different Stages Of Maturation Of Tertiary Lymphoid Structures In Molecular Subgroups Of Endometrial Carcinoma. International Journal of Gynecological Cancer. ISSN 1048-891X. 35(2), p. 100104–100104. doi: 10.1016/j.ijgc.2024.100104. Full text in Research Archive Show summary

  Introduction/Background Tertiary lymphoid structures (TLSs) are organized clusters of lymphoid cells at various maturation stages with potential prognostic and predictive impact for response to immunotherapy in several cancer types. We assessed the prognostic significance of TLSs in patients with molecularly classified endometrial carcinoma (EC). Methodology This retrospective cohort included 1221 patients with molecularly classified EC treated at Oslo University Hospital 2006-2017. Formalin-fixed, paraffin-embedded tissue sections from hysterectomy specimens were immunohistochemically stained for CD21, Ki-67, CD3 and CD20. TLSs were categorized as mature (CD21+/Ki-67+), immature (CD21+/Ki-67–), lymphoid aggregates (CD21–, CD3+, and CD20+), or negative (CD21–/no CD3/CD20+ clusters). Univariable and multivariable analyses were performed to assess the association between TLS stages and time to recurrence (TTR). Results We identified 517 patients with lymphoid aggregates, 224 with immature TLSs, 138 with mature TLSs, and 342 TLS-negative cases. TLSs were more frequent in POLE-mutated (98%, 94/96) and MMRd (MMR deficient) tumours (87%, 321/368) compared to p53 abnormal (63%, 175/275) and no specific molecular profile (60%, 278/462) tumours (p<0.001). TLS presence of any maturity was significantly associated with longer TTR in univariable (p<0.001) and multivariable analysis (p<0.001). Compared to TLS-negative cases, the hazard ratio was 0.69 (95% confidence interval (95%CI): 0.54-0.89) for lymphoid aggregates, 0.45 (95%CI: 0.31-0.64) for immature TLSs, and 0.27 (95%CI: 0.16-0.44) for mature TLSs in multivariable analyses. In stratified analysis by molecular subgroup, presence of all maturity stages of TLSs remained significantly associated with longer TTR in patients in MMRd and NSMP tumours in multivariable analysis. TLSs is a favourable prognostic marker for EC patients, especially in the MMRd and NSMP subgroups. Conclusion TLSs were predominantly found in patients with POLE-mutated and MMRd tumours. Prognosis improved progressively with TLS maturation stage, with longest TTR in patients with mature TLSs. Evaluating TLSs may provide additional prognostic information across molecular subgroups. Disclosures This supported collaborative study was funded by GSK (OSF.IO/JHEVG” title="https://doi.org/10.17605/OSF.IO/JHEVG">https://doi.org/10.17605/OSF.IO/JHEVG). BM, DS, and HE are full-time employees of GSK and hold stock and/or stock options. HE holds stock/stock options i Abbvie. KL reports the following conflicts of interest outside the submitted work: Participation on data safety monitoring or advisory boards of Eisai, MSD, Nykode, AstraZeneca, GSK and Karyopharm (honoraria paid to institution); speaker fee from MSD; GSK and Astra Zeneca and research funding paid to institution from GSK. All remaining authors have declared no conflicts of interest.

• Cyll, Karolina; Skrede, Ole-Johan & Kleppe, Andreas (2025). Can radiology be first to use prognostic deep learning models for oncological treatment? Annals of Oncology. ISSN 0923-7534. 36(10), p. 1122–1122. doi: 10.1016/j.annonc.2025.07.013. Full text in Research Archive
• Pradhan, Manohar; Kildal, Wanja; Vlatkovic, Ljiljana; Tobin, Kari Anne Risan; Lindemann, Kristina Yvonne Kathe & Kristensen, Gunnar S Balle [Show all 8 contributors for this article] (2025). Response to correspondence on prognostic value of tertiary lymphoid structures in molecular subgroups of endometrial carcinoma by Pradhan et al. International Journal of Gynecological Cancer. ISSN 1048-891X. 35(7), p. 101970–101970. doi: 10.1016/j.ijgc.2025.101970. Full text in Research Archive
• Acharya, Sushil; Fabian, Karl & Yazidi, Anis (2025). Automated Depth Alignment of Well Logs Using Siamese Neural Networks. doi: 10.30632/spwla-2025-0048. Full text in Research Archive Show summary

  Accurate depth alignment of well logs is essential for re-liable subsurface interpretation, influencing petrophysi-cal analysis. Traditional methods like cross-correlation struggle with repetitive lithologies, often leading to false matches and requiring extensive manual corrections(Torres Caceres et al., 2022). While machine learning offers alternatives, supervised approaches demand large labeled datasets, and unsupervised clustering requires extensive post-processing, introducing potential errors(Acharya and Fabian, 2024b; Acharya et al., 2025). This study proposes a Siamese Neural Network (SNN) to au-tomate log alignment by learning geological similarity between reference and unsynchronized logs. The SNN eliminates the need for labeled data, minimizes post-processing, and effectively handles irregular depth er-rors. Tested on gamma-ray (GR) logs from different logging methods within the same well, the proposed method improves Pearson correlation from 0.003 to 0.878 and reduces Euclidean distance by 74.623%, significantly outperforming cross-correlation in non-linear alignment scenarios. A novel similarity score enhances geological plausibility, complementing statistical metrics and pro-viding a valuable validation tool, especially in complex intervals. While the SNN is computationally slower, its superior accuracy makes it a reliable choice for post-drilling workflows, ensuring more precise depth align-ment with minimal manual intervention. By propagat-ing corrections to co-acquired logs, it enhances multi-log consistency for reservoir characterization. This work provides a scalable and reliable solution to depth alignment challenges, with potential applications in other geoscientific fields requiring log matching.

• Halnes, Geir; Javaid, Anam; Solvang, Michael; Nichele, Stefano; Riegler, Michael & Singstad, Bjørn-Jostein [Show all 82 contributors for this article] (2025). Proceedings of NORA’s annual conference 2025. Nordic Machine Intelligence (NMI). 5(3), p. 1–14. doi: 10.5617/nmi.12551. Full text in Research Archive
• Bjørklund, Petter; Singh, Durgesh Kumar; Steen, Erik Normann & Kampffmeyer, Michael (2025). Slik kan KI bidra til å oppdage hjertesykdommer raskere. uit.no. doi: https:/uit.no/nyheter/artikkel?p_document_id=913917. Full text in Research Archive
• Mylius-Kroken, Johan; Wickstrøm, Kristoffer; Wetzer, Elisabeth; Ramezani-Kebrya, Ali & Jenssen, Robert (2025). Can a Convex Partition caused by a CPWL Neural Network be used for Density Estimation? National conference on image processing and machine learnin. Full text in Research Archive
• Froment, Marouchka; Brissaud, Quentin; Näsholm, Sven Peter; Schweitzer, Johannes & Gullbekk, Sophus Bredesen (2025). Revealing the structure of Venus' interiors using planetary rumbles. Full text in Research Archive Show summary

  Seismology has been remarkably successful in probing the interior structure of the Earth, the Moon, and more recently Mars. Could it achieve the same on Venus, and reveal what hides beneath the surface of this mysterious and hellish twin of Earth ? Venus’ surface conditions make the deployments of ground-based seismometers challenging. Instead, several mission concepts have proposed to record venusquakes from above, leveraging the conversion of seismic waves into acoustic waves in the dense atmosphere. Could seismology be conducted from floating balloons, or through satellite imaging ? To explore this possibility, we gather insights from balloon campaigns on Earth and from modeling.

• Boukraa, Mohamed Aziz; Karabiyik, Yücel; Austeng, Andreas & Näsholm, Sven Peter (2025). Full-Waveform Inversion in Ultrasound Time-Harmonic Shear-Wave Elastography. Full text in Research Archive
• Boukraa, Mohamed Aziz; Karabiyik, Yücel; Austeng, Andreas & Näsholm, Sven Peter (2025). Multifrequency Full-Waveform Inversion in Ultrasound Time-Harmonic Shear-Wave Elastography. Full text in Research Archive
• Mulani, Sufayan Ikabal; Ziksari, Mahsa Sotoodeh; Austeng, Andreas & Näsholm, Sven Peter (2025). A Modified Wavenumber Algorithm That Preserves the DAS Image. Full text in Research Archive
• Mulani, Sufayan Ikabal; Ziksari, Mahsa Sotoodeh; Austeng, Andreas & Näsholm, Sven Peter (2025). Efficient Ultrasound Imaging by Integrating Refocus with Fourier Beamforming. Full text in Research Archive
• Ziksari, Mahsa Sotoodeh; Austeng, Andreas; Näsholm, Sven Peter & Jensen, Are Charles (2025). Enhanced Diverging-Wave Iterative Adaptive Approach Beamforming using Spatial Subspace Filtering. Full text in Research Archive
• Han, Chaoran; Boukraa, Mohamed Aziz; Näsholm, Sven Peter; Austeng, Andreas; Ziksari, Mahsa Sotoodeh & Karabiyik, Yücel (2025). A Simulation Framework for Ultrasound Time Harmonic Elastography (THE). Full text in Research Archive
• Rivera, Adín Ramírez (2025). Multimodal Intelligence. Full text in Research Archive
• Yazidi, Anis (2025). Robust RL with Learning Automata: Structure, Timescales, and Mixed Strategies. Full text in Research Archive
• Wu, Zhiyuan (2025). Addressing Label Shift in Distributed Learning via Entropy Regularization. doi: https:/www.youtube.com/watch?v=FRFAxS94Kow. Full text in Research Archive
• Arnstad, August & Holmsen, Sigurd (2025). Hvorfor henger KI etter i den fysiske verden? Full text in Research Archive Show summary

  Innblikk i ulike scenarioer hvor virkelighetsgapet mellom mennesker og kunstig intelligens/robotikk er av varierende omfang.

• Acharya, Sushil; Fabian, Karl & Yazidi, Anis (2026). Dynamic Depth Alignment of Wellbore Measurements Using Machine Learning. Norges teknisk-naturvitenskapelige universitet. ISSN 9788235302786. Full text in Research Archive Show summary

  Well logs record physical properties of rock formations along a borehole as a function of depth. They are essential for evaluating subsurface formation properties in hydrocarbon exploration, reservoir development, groundwater studies, and numerous other geoscientific and engineering applications. Despite their importance, well logs acquired by different logging tools, such as electrical wireline logging (EWL) and logging while drilling (LWD), or even by repeat runs of the same tool, often assign the same geological features to slightly different depths because of mechanical effects, operational conditions, and differences in data-processing workflows. For meaningful analysis, measurements from multiple tools or runs must be integrated and interpreted together. Because each run records depth independently, the resulting logs are not automatically aligned to a common depth. The process of aligning these measurements to a single, consistent depth reference is known as well-log depth matching and remains a major challenge in subsurface characterization. In the oil and gas industry, depth matching is performed manually by experts using tie points and correlation-based methods, sometimes supported by dynamic time warping (DTW). Although effective in many cases, these approaches are time-consuming when applied to large datasets, subjective, and less reliable in intervals with repetitive log patterns or non-uniform depth errors. This thesis develops automated and semi-automated machine learning (ML) and deterministic methods for aligning LWD logs to the EWL depth reference within the same well. The aim is to reduce manual effort and provide geologically reliable results that experts can trust and review. Rather than estimating the exact true along-hole depth, the thesis focuses on aligning LWD and EWL logs so that corresponding geological features appear at the same depth. The proposed methods are designed to handle nonuniform depth mismatches, preserve geological consistency, and integrate with practical depth-alignment workflows. The first contribution is a learning-based global alignment framework, the Deep Hierarchical Graph Correlation (DHGC). DHGC learns window-level representations from pairs of gamma ray (GR) logs and uses them to construct a pairwise similarity matrix. A constrained graph-based algorithm, implemented using Dijkstra’s algorithm, then computes a globally consistent nonlinear alignment path. By preserving monotonicity and controlling slope and step size, the method reduces the one-to-many mappings that often arise in ambiguous intervals with standard DTW. The second contribution is a supervised local alignment model called the Attention-Driven Convolutional Neural Network (AD-CNN). The model predicts depth shifts at the window level. Any extreme or unrealistic shifts are corrected using a separate outlier-detection workflow. After outlier correction, the predicted shifts are used either to shift local log segments or to build a continuous depth correction across the well. The thesis also proposes a semi-automated workflow that focuses on geological boundaries to improve interpretability and practical applicability. Unlike the single-log alignment methods described earlier, this workflow performs alignment jointly using multiple logs. The method enhances step-like lithological transitions using the Fast Walsh-Hadamard Transform (FWHT). Low-pass filtering in the Walsh domain highlights step-like structures that reflect formation layering. From this, a Walsh Boundary Index is computed to identify likely formation boundaries. These detected boundaries are then used as control points to build a monotonic, piecewise-linear depth-warping function. This approach is similar to how practitioners manually tie key points between logs. It produces clear intermediate outputs, such as boundary evidence, selected anchor points, and the resulting depth mapping. The method is useful when depth misalignment varies with depth and cannot be corrected by a single global shift. Building on the same idea of boundary-guided alignment, but using data-driven methods instead of deterministic signal processing, the thesis also explores unsupervised multi-log alignment methods. By combining dimensionality-reduction techniques such as PCA or autoencoders with clustering methods such as K-means, the approach identifies clusters that may represent distinct lithological units. It also identifies the boundaries between these units, which may correspond to bed boundaries. This is achieved without requiring explicit depth-shift labels. The proposed methods are evaluated on well-log datasets from the Norwegian Continental Shelf (NCS), using well-wise hold-out validation and comparisons with established industrial baselines, including correlation-based methods and DTW. Performance is evaluated quantitatively using Pearson correlation, Euclidean distance, Earth Mover’s Distance (for DHGC), and runtime, and qualitatively through visual inspection. For alignment based on a single log type, DHGC provides robust global alignment and is faster than DTW. The AD-CNN provides efficient local alignment with accuracy comparable to established methods. For multiple logs, the FWHT-based workflow improves alignment quality while maintaining interpretability. The unsupervised multi-log methods also improve alignment quality across several log types, with the autoencoder combined with K-means approach performing best among the tested unsupervised methods. These results suggest a practical stepwise workflow for industry use: first, apply an initial global shift; then use anchor-point warping where clear markers are available; and finally apply flexible nonlinear methods only in complex intervals. All depth corrections and processing decisions should be stored as metadata, uncertain intervals should be flagged with a “no-align” option, and final verification should use repeat sections or known markers. This work reduces manual effort, improves consistency, and provides clear results for reliable automated depth matching in large-scale subsurface studies.

• Andreassen, Børge Solli; Samset, Eigil; Solberg, Anne H Schistad; Thomas, Sarina & Volgyes, David (2025). Deep Learning for Mitral Annulus Delineation in 3D Transesophageal Echocardiography. Universitetet i Oslo. Full text in Research Archive Show summary

  Valvular heart disease is a significant health concern, particularly in aging populations, with mitral valve disease being among the most common. Three-dimensional transesophageal echocardiography (TEE) provides high-quality imaging for assessing the mitral valve. Delineating the mitral annulus in 3D images is an important step for accurate quantitative morphological assessment of the valve. However, manual analysis is time-consuming and prone to interobserver variability. Deep learning is a class of machine learning methods that has led to significant progress across an extensive range of applications. In medical imaging, it has enabled improvements across a wide array of tasks, including classification, segmentation, and landmark detection. In this thesis, deep learning is applied to automate the delineation of the mitral annulus from 3D TEE images. The thesis presents a progression of methods across three papers, each building on the last, gradually incorporating more anatomical context and spatial structure. The first paper applies a 2D model with post-processing to enforce annular continuity. The second paper introduces a periodic 3D model using cylindrical coordinates, enabling direct segmentation and prediction of anatomical orientation. The third paper integrates anatomical constraints through a graph-based model to improve the robustness and consistency of the predicted annulus. The methods were evaluated on annotated 3D TEE datasets, showing accurate and consistent delineation of the mitral annulus across multiple subjects. In addition to the specific application, the thesis proposes a novel loss function and coordinate decoding strategy for deep learning with potential applicability in landmark detection tasks, both in medical and natural images. These Developments support the potential for integration into practical workflows, which could assist clinicians in assessing mitral valves while contributing to automation in cardiac imaging.

View all works in NVA