A Novel Framework for Medical Learning: Using AI Based Grad-CAM for Improving Otitis Media Diagnosis

Abstract

Otitis Media (OM) and its sub-categories of pathology are the number one pathology in children. Diagnosis is very difficult as it requires visual inspection the tympanic membrane of a child, which is in view for only a few seconds during a clinical exam. Improving diagnosis requires the transfer of visual insights which is a complex learning and training task. A validated method to understand visual task insights has been to use eye-tracking as a surrogate for neural attention. Eye tracking data can be represented in the form of a heat-map or a visual saliency map. Considering the power and benefits of using state-of- the-art Machine Learning techniques in diagnosing visual pathology, our purpose is to derive a heat-map from a Machine Learning algorithm that acts as an "expert", and to provide these heat-maps for medical students with the final aim of understanding if this improves medical learning, specifically for OM. Our results indicate a significant improvement in diagnostic performance when showing medical students heat-maps derived from machine learning models, in conjunction to traditional teaching tutorials when compared to a control group not exposed to the heat-maps. This research provides a simple, cost-effective proof-of- concept framework to enhance the diagnostic accuracy and training speed for medical student as well as contribute in bridging the disparity gap in diagnostic accuracy of otitis media amongst practitioners.