Patient Aware Active Learning for Stable Medical Image Classification

Personnel: Yash-yee Logan, Ryan Benket, Ahmad Mustafa, Mohit Prabhushankar, Ghassan AlRegib

Goal: To overcome the elasticity-plasticity dilemma on medical imagery

Challenges: Elasticity-plasticity dilemma has long plagued machine learning applications. However, in order for safety critical applications to become deployable, effective ways of minimizing this problem need to be designed. 

Our Work: We show that patient aware active learning paired with continual learning optimization schemes aids in the minimization of negative flip rate [1]. We describe a continual learning setup that takes advantage of the samples selected using clinically relevant meta-information. Current active learning approaches do not have any impact on retaining prior knowledge after acquiring new information. This work is central to exploiting auxiliary medical data to limiting the elasticity-plasticity dilemma.

References:

  1. Y. Logan, R. Benkert, A. Mustafa, G. Kwon, G. AlRegib, “Patient Aware Stable Active Learning for Medical Image Classification” Under Review at 2022 IEEE Journal of Biomedical and Health Informatics (J-BHI).