Towards an Interpretable Chest X-ray Classifier through Optimal Transport Regularization

Chih-Chieh Chen, Chang-Fu Kuo

Keywords: Chest X-ray Classification, Kernel Optimal Transport, Weakly Supervised localization

TL;DR: We designed a chest X-ray classifier that provides patch-level outputs while training solely on class labels. Therefore, we can directly generate heatmaps from the patch outputs.

Abstract:

In addition to model performance, interpretability is essential for integrating artificial
intelligence into clinical settings. In this study, we designed a chest X-ray classifier that
provides patch-level outputs while training solely on class labels. To align the patch-level
outputs with the locations of diseases, we introduced an optimal transport-based regularization
into our architecture. We present results and observations to demonstrate the effectiveness
of our approach.

Submission Number: 3