MBM-Mamba: Multi-Branch Model Based on Mamba for Pneumoconiosis Screening and Staging

Existing auxiliary diagnosis methods based on Convolutional Neural Network (CNN) are difficult to achieve ideal accuracy in pneumoconiosis screening and staging tasks. This paper proposes a multi-branch structure model for pneumoconiosis screening and staging based on chest X-ray images, namely MBM-Mamba, and proposes a new six-direction scanning strategy under the framework of 2D-Selective-Scan (SS2D). The diagonal scanning is used to explicitly capture the two-dimensional local dependency with linear time complexity. Then, a Detail Enhancement Module (Dem) is constructed by integrating prior information into CNN, thus forming a local feature extraction module (CNN-Mamba), which significantly improves the expression ability of subtle lesion information.In addition, the MBM-Mamba model designs a global feature extraction module based on the multi-head self-attention mechanism, which effectively enhances the ability to capture global context. The MBM-Mamba model integrates the local and global feature extraction modules by using multi-path residuals, and realizes the synchronous hierarchical fusion of cross-structural features, so that the model can better understand the overall distribution of lung lesions and the degree of fibrosis. Verified on 1760 chest X-ray images of real anonymous patients, the accuracy of the MBM-Mamba model reaches 0.786, and the F1 score is 0.790. Both indicators are better than the existing models.