A Clinically Interpretable Multimodal Framework for Severity-Aware Lung Cancer Diagnosis via Bidirectional Fusion, Graph Transformer Networks, and Generative Learning

Abstract

Clinically interpretable lung cancer diagnosis remains a critical challenge in computational oncology, particularly when severity-aware staging demands the integration of heterogeneous imaging and clinical data. Existing deep learning approaches either operate on single modalities with limited clinical context or employ unidirectional fusion strategies that fail to capture the reciprocal cross-modal dependencies essential for advanced-stage diagnosis. This paper proposes BioFuse-GT, a clinically interpretable multimodal framework that integrates (i) bidirectional cross-modal fusion of computed tomography (CT) imaging and clinical data, (ii) a graph transformer network for population-level relational modeling, and (iii) variational adversarial generative learning for stage-balanced data augmentation. The bidirectional fusion module enables reciprocal information exchange through cross-attention gating between the imaging and clinical pathways, while the graph transformer constructs patient-level relational graphs refined by severity-aware multi-head attention. Comprehensive evaluation on the LIDC-IDRI dataset demonstrates that BioFuse-GT achieves a staging accuracy of 95.2%, a multimodal F1-score of 94.9%, an area under the curve (AUC) of 0.978, and a clinical interpretability agreement of 93.7% with expert radiologists, outperforming representative single-modality and unidirectional-fusion baselines by substantial margins. Ablation studies confirm that every architectural component contributes meaningfully to overall performance, with bidirectional fusion alone providing an 8.4 percentage point improvement in staging accuracy over a unidirectional alternative. These results indicate that BioFuse-GT offers a promising and clinically transparent pathway toward trustworthy multimodal decision support in lung cancer staging.