Speaker
Anja Witte
(Institute of Medical Systems Bioinformatics, Center for Biomedical AI (bAIome), Center for Molecular Neurobiology Hamburg (ZMNH), University Medical Center Hamburg-Eppendorf, Hamburg, Germany)
Description
Prostate cancer relapse prediction is a challenging task within computational pathology as tissue preparation and digitalization is not standardized. The different protocols lead to domain shifts, against which a deep learning model must be robust and focus on biological information rather than variations in appearance. We address this challenge through the usage of vision foundation models that are pre-trained on large and diverse pathology datasets. Six different models are fine-tuned and evaluated on an internal dataset that includes multi-domain prostate cancer images on patient-level. The comparison shows that larger models are in general superior and that robustness depends on the vision foundation models and the domains.
Author
Anja Witte
(Institute of Medical Systems Bioinformatics, Center for Biomedical AI (bAIome), Center for Molecular Neurobiology Hamburg (ZMNH), University Medical Center Hamburg-Eppendorf, Hamburg, Germany)