Machine Learning Algorithms for In-Silico Virtual Imaging
Forum Details
Date: October 18, 2024 12-1 PM Eastern Time
Speakers: Vahid Tarokh, PhD
Rapid acceleration in medical imaging technologies has increased the complexity of evaluation and optimization of new imaging technologies through clinical imaging trials due to associated expenses, time-requirements, difficulty in accruing subjects, ethical limitations, etc. It seems that if we only rely on simplistic models and subjective perception of image aesthetics, the results may not readily predict clinical utility. In addressing these issues, our colleagues at the Duke Center for Virtual Imaging Trials have been advancing in silico alternatives to clinical trials developing a Virtual Imaging Trial Platform based on computational models of (i) patients, (ii) imaging systems, and (iii) image analysis or interpretation.
In this talk, I will discuss how I think some modern machine learning techniques may be applied to all these three directions. For instance: What is a nearly optimal set of virtual patients that can represent the population encountered in practice? How to use representation learning to generate samples of rare diseases? Can the dynamics of tumor growth be represented by a latent SDE whose drift can be used to make medical conclusions? How to determine representatives of existing scanners? Can we use causal inference to extract actionable information from the image data for virtual readers? Can this actionable information be disentangled into interpretable components? Can we use transfer learning to make better conclusions about outcomes related to real patients/devices/readers from both simulated data, and a much smaller representative of real data?