Research on a Dual-Branch Deep Network Based on an Improved CBAM Attention Mechanism and Interaction-Aware Gated Fusion for Auxiliary Diagnosis of Musculoskeletal Injuries

Abstract

To address the strong subjectivity of conventional diagnostic methods for musculoskeletal injuries, the insufficient fusion of heterogeneous multimodal features in existing gait analysis approaches, and their weak ability to focus on critical lesion regions, this paper proposes a novel dual-branch deep learning architecture for the auxiliary diagnosis of musculoskeletal injuries. The model adopts a dual-branch structure that independently extracts the spatial texture of plantar pressure distribution heatmaps and the nonlinear features of quantified gait parameters. At the architectural level, a Dynamic Cross-Channel and Multi-Scale Spatial Fusion Attention mechanism (DCMSF-CBAM) is innovatively constructed to achieve dynamic multi-path fusion along both the spatial and channel dimensions, which markedly strengthens the model’s ability to focus on critical lesion regions such as the heel and the arch. In addition, an Interaction-Aware Gated Fusion module (IAGF) is designed, in which a dynamic gating mechanism driven by interaction features adaptively and weightedly integrates visual representations with quantified numerical values, effectively overcoming the information redundancy and loss caused by conventional simple concatenation. Validated on a four-class clinical gait dataset covering healthy subjects, knee osteoarthritis, anterior cruciate ligament injury, and meniscus injury, the proposed model achieves a classification accuracy of 0.9476, with precision, recall, and F1-score of 0.9456, 0.9557, and 0.9500, respectively, under the macro-averaging strategy. The results indicate that the proposed method can provide a reference for the early screening and auxiliary diagnosis of musculoskeletal injuries.