A Diagnostic Gap Framework for Evaluating Reconstruction Fidelity in Weakly Supervised Mammography

2026-07-22

AI Summary

This work addresses the challenge of evaluating whether clinically relevant information is preserved during the reconstruction stage in weakly supervised medical imaging pipelines, where conventional metrics fall short. The authors propose a “diagnostic gap” framework that, for the first time, quantifies diagnostic information retention through two dimensions—decision consistency and explanation fidelity—and establishes its functional relationship with reconstruction fidelity, moving beyond mere reconstruction error. Using a fidelity ladder built from VQ-VAE-GAN, VAE-GAN, and diffusion-based SDEdit models, they evaluate lesion regions of interest in mammograms via AUC, LPIPS, and attribution maps (HiResCAM/Grad-CAM++). Results show near-perfect diagnostic preservation under autoencoder-level reconstructions (AUC shift ≤±0.005, high attribution agreement), whereas low-fidelity diffusion reconstructions cause significant degradation (AUC drop of 0.253, falling below random performance in some tasks). This framework offers an architecture-agnostic tool for assessing diagnostic consistency across multi-stage medical imaging workflows.

Abstract

Weakly supervised pipelines for medical imaging have become increasingly popular over the years. These systems often include multiple stages and components, such as reconstruction, generation, and localization, yet standard evaluation metrics provide limited insight into whether clinically relevant information is preserved across each stage. We present the diagnostic gap framework, a practical evaluation tool that measures decision preservation and explanation preservation as a function of measured reconstruction fidelity. To isolate the effect of reconstruction from localization, we evaluate on curated lesion ROI crops using a fidelity ladder of three class-conditional reconstructors---VQ-VAE-GAN, VAE-GAN, and diffusion (SDEdit)---spanning a twenty-fold range in perceptual distance (LPIPS 0.029--0.584). At autoencoder fidelity, both decision and explanation are preserved: AUC changes remain within $\pm$0.005 and attribution similarity (HiResCAM, Grad-CAM++) stays high. At diffusion fidelity, both collapse: pooled AUC drops by 0.253 and mass-pathology AUC falls below chance. The diagnostic gap is thus a measurable function of reconstruction fidelity rather than an intrinsic cost of reconstruction, and the framework provides an architecture-agnostic instrument for identifying when and where multi-stage pipelines lose diagnostic signal.

Problem

• weakly supervised learning
• medical imaging
• reconstruction fidelity
• diagnostic gap
• mammography

Innovation

• diagnostic gap
• reconstruction fidelity
• weakly supervised learning
• medical imaging
• explainability preservation