Towards Case-based Interpretability for Medical Federated Learning

TL;DR

This work tackles the challenge of case-based interpretability in privacy-preserving medical federated learning by training a DenseNet-121 classifier with differential privacy and generating synthetic case explanations via a Medfusion diffusion model. Similarity-based retrieval over client-specific synthetic cases provides explanations for new pleural effusion X-rays, with a radiologist evaluating realism and evidential value. Compared to a centralized upper bound, the federated approach shows lower classification F1 due to data heterogeneity and DP noise, but retrieval explanations align better with radiologist rankings than SSIM, highlighting the potential of synthetic case-based explanations in federated settings. The study lays groundwork for privacy-conscious interpretability in clinical AI, while outlining necessary future work to ensure privacy guarantees and clinical readiness.

Abstract

We explore deep generative models to generate case-based explanations in a medical federated learning setting. Explaining AI model decisions through case-based interpretability is paramount to increasing trust and allowing widespread adoption of AI in clinical practice. However, medical AI training paradigms are shifting towards federated learning settings in order to comply with data protection regulations. In a federated scenario, past data is inaccessible to the current user. Thus, we use a deep generative model to generate synthetic examples that protect privacy and explain decisions. Our proof-of-concept focuses on pleural effusion diagnosis and uses publicly available Chest X-ray data.

Figures (3)

• Figure 1: Case-based Explanations for Federated Learning.
• Figure 2: SHAP saliency maps for true positive test samples.
• Figure 3: Example of a test case and the Top-4 retrieved explanations given by the radiologists (top), our model (DP-FL) and SSIM.