LungStage-XMAF: An eXplainable Attention Fusion Framework for Fine-Grained TNM Staging of Lung Cancer Using PET and CT with Clinical Metadata

Background/Objectives Accurate TNM (Tumour, Node, Metastasis) staging is critical for lung cancer diagnosis and treatment planning; however, current methods are limited by the subjectivity in interpreting multimodal data. This study presents a new approach, an eXplainable Multimodal Attention Fusion framework (LungStage-XMAF), designed to integrate positron emission tomography-computed tomography (PET-CT) images with clinical metadata (age, sex, weight, smoking history, histopathological grade) to facilitate automated, fine-grained TNM staging. Unlike conventional single-modality or static-fusion approaches, this model dynamically adapts to the patient-specific clinical context. Methods We developed a LungStage-XMAF, an explainable multimodal fusion framework that integrates PET-CT scans with clinical metadata using a dynamic cross-modal attention mechanism. This mechanism adaptively fuses imaging features according to individual patients’ clinical variables (e.g., age, sex, smoking history, etc.) via learned attention weights. The model was trained on 133 pairs of PET and CT scans along with corresponding clinical data from the Lung PET-CT-Dx dataset, using a dual-stream ResNet-50 for feature extractor and a fine-tuned EfficientNet-B0 classifier for fine-grained TNM substage prediction. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to model output to highlight anatomical and metabolic regions influencing predictions, ensuring clinical interpretability. Performance was evaluated through five-fold cross-validation Results The Model achieved an average accuracy of 98% across full TNM substage classification (e.g., T1, T1a, T1b, T1c), matching the granularity required in clinical staging protocol. Conclusion : Despite evaluation on a limited single-centre cohort, LungStage-XMAF demonstrates a feasible and clinically aligned AI system with strong potential to support consistent, explainable and efficient TNM staging in routine clinical practice.