Foundation Models for Ovarian Cancer Subtype Classification: Jack Breen, 04/11/24

TIA Centre Seminar Series: Jack Breen

Full Title: Foundation Models and Multiple Instance Learning Methods for Ovarian Cancer Subtype Classification

Abstract: Ovarian cancer histological subtyping is a vital diagnostic task as the subtypes represent vastly different diseases with varied treatment options and prognoses. It has previously been an underrepresented task in AI subtyping research. In this talk, I will discuss my thesis "Artificial Intelligence for Ovarian Cancer Diagnosis from Digital Pathology Slides", which focuses on the robust validation of multiple instance learning (MIL) subtyping models, and explores factors limiting the clinical utility of these models. Histopathology foundation models show great promise across many tasks, but analyses have typically been limited by arbitrary hyperparameters that were not tuned to the specific task - we recently reported the most rigorous single-task validation of histopathology foundation models to date (https://arxiv.org/abs/2405.09990). We have also explored the effects of replacing standard MIL methods with a multi-resolution graph network (https://arxiv.org/abs/2407.18105), alongside earlier studies focused on the computational efficiency of classification. Our other work has included investigating the effects of chemotherapy on morphological subtyping, predicting treatment response effectiveness, and the automated detection of metastases in the lymph nodes and omentum. The modern classifiers have drastically improved diagnostic performance, though some vital hurdles still need to be cleared for these models to achieve clinical utility.

Paper Link: https://arxiv.org/abs/2405.09990