LMC-VMamba: A Lightweight Visual State Space Model with Multi-level Feature Optimization for Classification in Chest X-rays

Although Chest X-ray (CXR) is an indispensable modality for screening respiratory infectious diseases, the diagnostic efficacy of CXR remains highly contingent upon the radiologist's expertise and is frequently hindered by anatomical superimposition and radiologist fatigue. Furthermore, existing deep learning methods struggle to achieve an optimal trade-off between accurate depiction of ill-defined lesions and lightweight deployment requirements. In this study, we propose LMC-VMamba, a novel, lightweight classification architecture derived from the VMamba, which introduces a two-stage progressive feature optimization mechanism. Specifically, it employs frequency-domain channel attention to enhance the perception of ill-defined opacities, alongside cross-layer cross-attention to facilitate robust multi-scale feature fusion. Evaluated on a public dataset comprising 5,228 images, the proposed model achieved a classification accuracy of 95.41%, outperforming existing state-of-the-art lightweight networks. Comprehensive ablation studies substantiate the efficacy of the proposed modules. These results underscore LMC-VMamba's viability as a reliable, rapid, and computationally efficient tool for pulmonary disease screening, particularly within resource-constrained primary healthcare settings.