Talk / Overview
Clinical medical imaging data suffers from heteogeneity both within institutions (due to patient population heterogeneity, motion, and equiment upgrades) and across institutions (due to vendor, protocol, and population differences). This hinders learning large-capacity models for tasks such as classification, anomaly detection, or segmentation. The problem is further compounded in the federated learning setting, which is attractive in medical imaging due to improved privacy and easier collaboration between institutions, but where standard algorithms such as Federated Average suffer from degradation in heterogeneous data settings. This talk will illustrate the challenges and potential technical solutions for centralised and federated machine learning in specific pathologies such as stroke.