Assessment of Medical Foundation Models for Survival Prediction with Whole Slide Images

Survival prediction with whole slide images (WSIs) can provide guidance for a better patient care and treatment selection but it is a challenging computer vision task with its particularities. Despite the great results showed by the recent survival analysis models with WSIs, the collection of the large annotated WSI datasets for survival analysis could be hindered by disease rareness or clinical trials constraints and be infeasible in the real-life medical practice. To overcome these limitations we propose to assess the performance of the digital pathology foundation models for prediction of survival outcomes on the small size ovarian cancer datasets. Our experimental results demonstrate that these models show promising results, their improved performance open the possibility to investigate the mechanisms of response to a particular therapy and in general could accelerate the adoption of machine learning models in medical practice.