Abstract / Summary
Abstract Differential diagnosis between Non-Small Cell Lung Cancer (NSCLC) and Small Cell Lung Cancer (SCLC) represents a critical decision boundary in thoracic oncology, dictating immediate therapeutic pathways. However, conventional single-modality strategies often lack concurrent anatomical and functional comprehension, while traditional multi-modal fusion networks are limited by pixel-level signal suppression and highly disruptive slice-level labeling noise. This paper introduces a unified Hybrid Multiple Instance Learning (MIL) Fusion Framework designed explicitly for automated, patient-level lung cancer differentiation using paired structural Computed Tomography (CT) and metabolic 18F-FDG Positron Emission Tomography (PET) volumes. The proposed architecture employs independent, parallel EfficientNet-B0 encoders to preserve native modality topologies, utilizing a parameterized multi-head cross-attention mechanism to dynamically map complex non-linear inter-modality dependencies. To bypass the historic constraint of tedious manual slice-by-slice segmentations, an attention-gated MIL pooling bottleneck is integrated to aggregate 3D volumetric slice sequences into cohesive patient-level bags, optimizing learning parameters directly from high-level clinical diagnoses. Rigorous evaluation on the open-source Lung-PET-CT-Dx dataset across a verified, highly stratified cohort of 355 patients (284 training, 71 validation) demonstrates that the proposed framework significantly outperforms baseline paradigms. The optimized hybrid network achieves a prominent patient-level accuracy of 0.7887, a Balanced Accuracy of 0.5536, an Area Under the ROC Curve (AUC) of 0.6607, and a Macro F_1-score of 0.5443. Crucially, architecture actively overcomes severe epidemiological class skew to recover the aggressive minority target, establishing a non-zero SCLC recall of 37.50% while operating at a rapid inference latency of 1.71 ms per slice equivalent. These empirical outcomes demonstrate the system's potential viability as an explainable, high-throughput radiology decision-support tool for accelerated clinical triage.