ICLR2025

Time-to-Event Pretraining for 3D Medical Imaging

Zepeng Frazier Huo, Jason Alan Fries, Alejandro Lozano, Jeya Maria Jose Valanarasu, Ethan Steinberg, Louis Blankemeier, Akshay S. Chaudhari, Curtis P. Langlotz, Nigam Shah

Abstract

What's the motivation: -Medical vision foundation models (VFMs) should help to better clinical decision making -In medicine, prognosis can help long-term treatment planning, in addition to diagnosis Clinical use case: Use CT scan to predict probability of cardiomegaly (enlarged heart) in the next 12 month E.g. VFMs should learn to identify biomarkers through self-supervised pretraining ✅Ideal Use GradCAM to probe models * Example from paper CT input ❌ Not ideal Why fail?