Visual attribution using Adversarial Latent Transformations

Tehseen Zia, Abdul Wahab, David Windridge, Santosh Tirunagari, Nauman Bashir Bhatti

Abstract: Highlights•The article presents a GAN-based approach for medical image visual attribution (VA).•Existing classification-based VA methods may not fully capture all disease effects.•The proposed method optimizes nonlinear transformations in the latent space.•Proposed VA2LT employs a cycle-consistency principle.•VA2LT creates distinct normal images of abnormal images.