Sarina Thomas

• Utseth, Ingrid; Vedal, Amund Hansen; Thomas, Sarina & Eikvil, Line (2026). Comparing Foundation Models for Medical Images: A Study on Limited Data and Generalization. Proceedings of Machine Learning Research (PMLR). 307, p. 439–447. doi: https:/proceedings.mlr.press/v307/utseth26a.html. 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.

• 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.

• Singh, Durgesh Kumar; Cao, Qing; Thomas, Sarina; Boubekki, Ahcène; Jenssen, Robert & Kampffmeyer, Michael (2025). WiseLVAM: A Novel Framework For Left Ventricle Automatic Measurements. Lecture Notes in Computer Science (LNCS). ISSN 0302-9743. 16165, p. 218–227. doi: 10.1007/978-3-032-06329-8_21. Full text in Research Archive Show summary

  Clinical guidelines recommend performing left ventricular (LV) linear measurements in B-mode echocardiographic images at the basal level—typically at the mitral valve leaflet tips—and aligned perpendicular to the LV long axis along a virtual scanline (SL). However, most automated methods estimate landmarks directly from B-mode images for the measurement task, where even small shifts in predicted points along the LV walls can lead to significant measurement errors, reducing their clinical reliability. A recent semi-automatic method, EnLVAM, addresses this limitation by constraining landmark prediction to a clinician-defined SL and training on generated Anatomical Motion Mode (AMM) images to predict LV landmarks along the same. To enable full automation, a contour-aware SL placement approach is proposed in this work, in which the LV contour is estimated using a weakly supervised B-mode landmark detector. SL placement is then performed by inferring the LV long axis and the basal level—mimicking clinical guidelines. Building on this foundation, we introduce WiseLVAM—a novel framework for fully automated yet manually adaptable framework for automatically placing the SL and then automatically performing the LV linear measurements in the AMM mode. WiseLVAM utilizes the structure-awareness from B-mode images and the motion-awareness from AMM mode to enhance robustness and accuracy with the potential to provide a practical solution for the routine clinical application. The source code is publicly available at https://github.com/SFI-Visual-Intelligence/wiselvam.git.

• Andreassen, Børge Solli; Thomas, Sarina; Solberg, Anne H Schistad; Samset, Eigil & Volgyes, David (2024). Introducing Anatomical Constraints in Mitral Annulus Segmentation in Transesophageal Echocardiography. In Gomez, Alberto; Khanal, Bishesh; King, Andrew & Namburete, Ana (Ed.), Simplifying Medical Ultrasound: 5th International Workshop, ASMUS 2024, Held in Conjunction with MICCAI 2024, Marrakesh, Morocco, October 6, 2024, Proceedings. Springer Nature Switzerland. ISSN 9783031736476. p. 122–131. doi: 10.1007/978-3-031-73647-6_12. Full text in Research Archive Show summary

  The morphology of the mitral annulus plays an important role in diagnosing and treating mitral valve disorders. Automated segmentation has the promise to be time-saving and improve consistency in clinical practice. In the past years, segmentation has been dominated by methods based on deep learning. Deep learning-based segmentation methods have shown good results, but their consistency and robustness are still subjects of active research. In this work, we introduce a method that combines Graph Convolutional Networks with a 3D CNN model to integrate an anatomical shape template for the predictions. Our method leverages the feature extraction capability of CNN models to provide input features to the graph neural networks. The proposed method leverages strengths from a shape model approach with the strengths of deep learning. Further, we propose loss functions for the CNN designed to guide the graph model training. The CNN was trained with transfer learning, using a limited number of labeled transesophageal echocardiography volumes to adapt to the mitral annulus segmentation task. When comparing the segmentation of the mitral annulus achieved by the proposed method with the test set annotations, the method showed a high degree of accuracy, achieving a curve-to-curve error of 2.00±0.81 mm and a relative perimeter error of 4.42±3.33%. Our results show that the proposed method is a promising new approach for introducing anatomical template structures in medical segmentation tasks.

• Engelhardt, Sandy; Dar, Salman Ul Hussan; Sharan, Lalith; André, Florian; Nagel, Eike & Thomas, Sarina (2024). Artificial intelligence in cardiovascular imaging and intervention. Herz. ISSN 0340-9937. 49, p. 327–334. doi: 10.1007/s00059-024-05264-z. Full text in Research Archive
• Vyver, Gilles van de; Thomas, Sarina; Ben-Yosef, Guy; Olaisen, Sindre Hellum; Dalen, Håvard & Løvstakken, Lasse [Show all 7 contributors for this article] (2024). Toward Robust Cardiac Segmentation Using Graph Convolutional Networks. IEEE Access. 12, p. 33876–33888. doi: 10.1109/ACCESS.2024.3373046. Full text in Research Archive Show summary

  Fully automatic cardiac segmentation can be a fast and reproducible method to extract clinical measurements from an echocardiography examination. The U-Net architecture is the current state-of-the-art deep learning architecture for medical segmentation and can segment cardiac structures in real-time with average errors comparable to inter-observer variability. However, this architecture still generates large outliers that are often anatomically incorrect. This work uses the concept of graph convolutional neural networks that predict the contour points of the structures of interest instead of labeling each pixel. We propose a graph architecture that uses two convolutional rings based on cardiac anatomy. While this architecture does not improve performance on classical measures like Dice score and Hausdorff distance, it does eliminate anatomical incorrect segmentations. Additionally, we propose to use the inter-model agreement of the U-Net and the graph network as a predictor of both the input and segmentation quality in real-time. The results show that from the 100 high agreement samples, 93 were in distribution, while from the 100 low agreement samples, only 7 were in distribution. Finally, this work contributes with an ablation study of the graph convolutional architecture on the publicly available CAMUS dataset and an evaluation of clinical measurements on the clinical HUNT4 dataset. Source code is available online: https://github.com/gillesvntnu/GCN_multistructure

• Thomas, Sarina; Tiago, Cristiana Ferreira; Andreassen, Børge Solli; Aase, Svein Arne; Šprem, Jurica & Steen, Erik [Show all 8 contributors for this article] (2023). Graph Convolutional Neural Networks for Automated Echocardiography View Recognition: A Holistic Approach. In Kainz, Bernhard; Noble, Alison; Schnabel, Julia; Khanal, Bishesh; Müller, Johanna Paula & Day, Thomas (Ed.), Simplifying Medical Ultrasound: 4th International Workshop, ASMUS 2023, Held in Conjunction with MICCAI 2023, Vancouver, BC, Canada, October 8, 2023, Proceedings. Springer Nature Switzerland. ISSN 9783031445217. p. 44–54. doi: 10.1007/978-3-031-44521-7_5. Full text in Research Archive Show summary

  To facilitate diagnosis on cardiac ultrasound (US), clinical practice has established several standard views of the heart, which serve as reference points for diagnostic measurements and define viewports from which images are acquired. Automatic view recognition involves grouping those images into classes of standard views. Although deep learning techniques have been successful in achieving this, they still struggle with fully verifying the suitability of an image for specific measurements due to factors like the correct location, pose, and potential occlusions of cardiac structures. Our approach goes beyond view classification and incorporates a 3D mesh reconstruction of the heart that enables several more downstream tasks, like segmentation and pose estimation. In this work, we explore learning 3D heart meshes via graph convolutions, using similar techniques to learn 3D meshes in natural images, such as human pose estimation. As the availability of fully annotated 3D images is limited, we generate synthetic US images from 3D meshes by training an adversarial denoising diffusion model. Experiments were conducted on synthetic and clinical cases for view recognition and structure detection. The approach yielded good performance on synthetic images and, despite being exclusively trained on synthetic data, it already showed potential when applied to clinical images. With this proof-of-concept, we aim to demonstrate the benefits of graphs to improve cardiac view recognition that can ultimately lead to better efficiency in cardiac diagnosis.

View all works in NVA

• 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.

• 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.

• Rivera, Adín Ramírez; Thomas, Sarina & Choi, Changkyu (2024). Unsupervised Learning of visual feature hierarchy by contrasting hierarchical topic assignments. Full text in Research Archive
• Engelhardt, Sandy; Dar, Salman Ul Hussan; Sharan, Lalith; André, Florian; Nagel, Eike & Thomas, Sarina (2024). Erratum to: Artificial intelligence in cardiovascular imaging and intervention (Herz, (2024), 49, 5, (327-334), 10.1007/s00059-024-05264-z). Herz. ISSN 0340-9937. 49(5), p. 393–393. doi: 10.1007/s00059-024-05271-0. Full text in Research Archive
• Choi, Changkyu; Thomas, Sarina & Rivera, Adín Ramírez (2024). Unsupervised learning of visual feature hierarchy by contrasting hierarchical topic assignments. Full text in Research Archive
• Vyver, Gilles van de; Thomas, Sarina; Ben-Yosef, Guy; Olaisen, Sindre Hellum; Dalen, Håvard & Løvstakken, Lasse (2023). Towards Robust Cardiac Segmentation using Graph Convolutional Networks. Full text in Research Archive
• Thomas, Sarina; Andreassen, Børge Solli; Tiago, Cristiana Ferreira; Aase, Svein Arne; Steen, Erik Normann & Šprem, Jurica [Show all 8 contributors for this article] (2023). Graph Convolutional Neural Networks for Automated Echocardiography View Recognition: A Holistic Approach. Full text in Research Archive
• Kausch, Lisa; Thomas, Sarina; Kunze, Holger; Barbari, Jan El & Maier-Hain, Klaus (2023). Shape-based pose estimation for automatic standard views of the knee. Full text in Research Archive
• Thomas, Sarina (2023). Image understanding using AI - Segmentation. Full text in Research Archive
• Thomas, Sarina (2023). Graph neural networks for automatic view recognition. Full text in Research Archive
• Tan, Martine Hjelkrem; Rivera, Adín Ramírez & Thomas, Sarina (2023). Hierarchical topic discovery in images; probabilistic modelling of visual data. Full text in Research Archive
• Vyver, Gilles Van De; Smistad, Erik; Løvstakken, Lasse; Olaisen, Sindre Hellum; Dalen, Håvard & Thomas, Sarina [Show all 7 contributors for this article] (2023). Towards Robust Cardiac Segmentation using Graph Convolutional Networks. Full text in Research Archive
• Thomas, Sarina; Gilbert, Andrew & Ben-Yosef, Guy (2022). Light-weight spatio-temporal graphs for segmentation and ejection fraction prediction in cardiac ultrasound. Full text in Research Archive
• Thomas, Sarina (2022). Applying Graph-neural networks to contour detection in medical ultrasound. Full text in Research Archive
• 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