Privacy-Preserving Multimodal Learning for Lung Cancer: FLiGenX-Net Combining CT Radiomics and Genomic Profiles for Robust Prognosis

• Syed Zaheer Ahammed,
• Radhika Baskar,
• G. Nalinipriya,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.004, Download Paper (Free):

• Lung cancer
• Radiogenomics
• Training
• LIDC
• Multimodal
• TCIA

Abstract

Lung cancer remains a leading cause of cancer-related mortality worldwide, necessitating advanced predictive models that integrate heterogeneous data sources for improved diagnosis and prognosis. This study introduces FLiGenX-Net, a multimodal federated deep learning framework that integrates CT-derived radiomics with genomic profiles to enable lung cancer detection, subtype classification, and survival prediction. The architecture employs dual-stream encoders for imaging and genomic modalities, followed by a cross-modal attention transformer to generate a unified latent representation. Prediction heads simultaneously handle cancer classification through a Softmax layer and survival estimation via a Cox proportional hazards model, while explainability modules—Grad-CAM for imaging and SHAP for genomics—provide clinically interpretable insights. Training is conducted in a federated setting to ensure patient data privacy across institutions, addressing critical challenges in real-world deployment. Experimental investigations across internal and external datasets (LIDC-IDRI, RIDER, TCIA Lung1, and NSCLC Radiogenomics) demonstrate that FLiGenX-Net achieves superior performance, attaining a concordance index of 0.682 and consistent improvements over unimodal and conventional fusion baselines. The framework also exhibits strong calibration and generalization across heterogeneous cohorts, confirming its robustness. By unifying multimodal learning, privacy-preserving federated optimization, and explainable prediction, FLiGenX-Net represents a clinically actionable advance toward precision oncology for lung cancer.

Statistics

Cite this article