Code:
Paper: https://openreview.net/pdf?id=0BJTRUVDf4
At the upcoming COMPAYL workshop at the MICCAI conference, Vilde Schulerud Bøe, PhD candidate at DSB and within Visual Intelligence, will present important findings about foundation models (FMs) in computational pathology.
While these models show promise for analyzing histopathological data, their reliability in real-world settings is still in question.
In this study, we have assessed the performance of several FMs on three different histopathology datasets and found that their accuracy decreases significantly when tested on external data. The models often included dataset-specific details, which limits their ability to generalize effectively.
To tackle this issue, we used a fine-tuning method called Weight-Decomposed Low-Rank Adaptation (DoRA), enhanced with strong data augmentations. The experiments revealed that models using this approach displayed improved robustness and were less reliant on the specifics of the training data.
Our findings highlight the need for rigorous robustness testing of FMs and stress the importance of considering these factors during their development and training.
Please come discuss with our team at COMPAYL workshop to learn more about these new insights in computational pathology.
Authors:
Vilde Schulerud Bøe1, Andreas Kleppe1,2,3, Sebastian Foersch4, Daniel-Christoph Wagner4, Lill-Tove Rasmussen Busund5,6, Adín Ramírez Rivera1
1: DSB group, Department of Informatics, University of Oslo
2: Institute for Cancer Genetics and Informatics, Oslo University Hospital
3: Centre for Research-based Innovation Visual Intelligence, UiT The Arctic University of Norway
4: Institute of Pathology, University Medical Center, Mainz
5: Department of Medical Biology, UiT The Arctic University of Norway, Tromsø
6: Department of Clinical Pathology, University Hospital of North Norway, Tromsø