A Hybrid‐Mamba network with dual level attention fusion for multimodal COPD diagnosis

BACKGROUND: Chronic obstructive pulmonary disease (COPD) remains difficult to diagnose reliably due to limitations of conventional spirometry and CT interpretation. Although deep learning has shown promise, CNNs are constrained by local receptive fields, and ViTs are constrained by high computational cost, highlighting the relevance of multimodal integration of CT and clinical data for improving COPD diagnostic accuracy. PURPOSE: In this study, we propose a Hybrid-Mamba Network with Dual Level Attention Fusion for multimodal COPD diagnosis. METHODS: We retrospectively enrolled 381 participants (184 with COPD and 197 healthy controls). Clinical data encompassed basic information, respiratory symptoms, blood gas analysis, pulmonary function tests, and blood routine tests. The framework employs a Hybrid-Mamba architecture for efficient CT feature representation, leverages a tailored Hybrid-DWConv-AAS Block for enhanced feature integration, and incorporates a Dual Level Attention Fusion Block to adaptively integrate CT and clinical data. RESULTS: On the test set, our proposed network achieved an AUC of 0.985 and an accuracy of 0.947, with an average per-patient inference time of 97.54 ms, while maintaining robust diagnostic performance under simulated perturbations. CONCLUSIONS: These findings indicate that the framework provides an efficient and robust approach for COPD diagnosis.