XPA-FedCBR: A Federated Case-Based Reasoning System for Explainable and Privacy-Aware Lung Cancer Diagnosis
TL;DR
The proposed XPA-FedCBR combines Federated Learning, which trains a shared model across institutions without moving raw data, with Case-Based Reasoning, which supports each prediction by retrieving clinically similar past cases and, for uncertain predictions, using them to refine the decision.
Abstract
Lung cancer is among the deadliest cancers worldwide, largely because it is usually caught too late. Building AI tools for earlier, stage-aware diagnosis is hard in practice: patient scans are scattered across hospitals that cannot share them freely, and clinicians are reluctant to trust models that cannot explain their reasoning. This study presents XPA-FedCBR, a framework that addresses both concerns. It combines Federated Learning, which trains a shared model across institutions without moving raw data, with Case-Based Reasoning, which supports each prediction by retrieving clinically similar past cases and, for uncertain predictions, using them to refine the decision. We evaluate it on the LIDC-IDRI dataset (roughly 1,018 CT scans from 1,010 patients), split across five simulated institutions under non-IID conditions, with annotations mapped to stages I–IV. Over four independent runs, the proposed model reaches a macro-F1 of 0.915, significantly ahead of FedAvg without CBR (0.875), a centralized CNN (0.895), and a standalone CNN (0.844), with its advantage largest under domain shift. Its case-based explanations, evaluated quantitatively and against gradient-based saliency, give clinicians evidence they can directly inspect.