Hierarchical Explainable Radiogenomics for Early Lung Cancer

Abstract

Lung cancer is a leading cause of cancer deaths, demanding screening systems that are both accurate and interpretable. We propose a hierarchical radiogenomic framework that first performs early screening and then molecular subtype classification. In Stage A, a lightweight attention-based CT model distinguishes non-cancer from cancer cases. Only cases predicted as high-risk or uncertain are forwarded to Stage B, where a tri-modal classifier integrates chest CT, structured clinical variables, and genomic profiles to separate LUAD from LUSC using a 3D ResNet encoder, MLP, and denoising autoencoder with attention-guided fusion. Interpretability is achieved via Integrated Gradients (CT), DeepSHAP (clinical/genomic), ConceptSHAP (pathway-level factors), and DiCE counterfactuals. An uncertainty-aware router with probability calibration governs escalation from screening to tri-modal analysis. Rather than introducing a new single network component, the framework contributes a clinically motivated end-to-end design that unifies calibrated screening, radiogenomic subtype analysis, and multi-level explanation within one workflow. Evaluated on NLST and NSCLC-Radiogenomics, and additionally examined under a simulated federated learning setting with non-IID client partitions, the framework outperforms unimodal and naive fusion baselines while maintaining performance close to the centralized setting.