Human‑centered contrastive explanations for medical imaging using VAE‑AC‑WGAN

Abstract

We introduce a novel contrastive explanation framework, the Variational Autoencoder Auxiliary Classifier Wasserstein Generative Adversarial Network (VAE-AC-WGAN), designed to generate synthetic medical images with and without pathological features, thereby enhancing interpretability and enabling clearer insights into model behavior. Unlike traditional attribution methods such as Gradient-weighted Class Activation Mapping (Grad-CAM), which often fall short in delivering meaningful explanations in critical fields like medical imaging, VAE-AC-WGAN supports more faithful reconstructions and targeted perturbations of key image features. Building upon and improving medXGAN, our approach addresses limitations in image quality and computational efficiency. We evaluate our method qualitatively and quantitatively on brain MRI and the Lung Image Database Consortium (LIDC) datasets and present comparisons with prior techniques, thus contributing to more transparent and trustworthy AI-assisted clinical decision-making systems.

Ähnliche Arbeiten