Dual-branch Pneumoconiosis Staging Model Based on Dynamic Convolution and Mamba
WANG Yifan
SU Shuzhi
YIN Xinle
ZHAO Chenqi
YANG Fan
Abstract：To address the issues of sparse lesion distribution,variable morphology,and small inter-class differences in pneumoconiosis X-ray chest films,a dual-branch pneumoconiosis staging model based on dynamic convolution and Mamba(DC-Mamba)was proposed.First,the model enhanced the extraction of local features of small fibrotic lesions through the adaptive kernel parameter generation strategy of the dynamic convolution branch.Meanwhile,the global spatial dependencies of multi-regional lesions were captured by leveraging the sequential modeling capability of the Mamba branch.Second,a feature fusion module with dual-path attention collaboration mechanism was designed to integrate local details and global contextual information.The model was validated on 1760 real anonymized patient X-ray chest films.The results showed that the accuracy of DC-Mamba model reached 78.3%,the recall was 79.0%,and the F1 score was 77.6%,all of which were superior to contrast models.DC-Mamba significantly improved the model's understanding of the overall distribution of pulmonary lesions and the detection of subtle lesions,thereby enhancing the early screening and precise staging capacities of pneumoconiosis.
Keywords：pneumoconiosis stagingX-ray chest filmsdynamic convolutionMambafeature fusionmulti-regional lesions
Publication Date：2025-09-20
Online Publishing Date：2025-09-24(First online date of this platform, not the publication date of the document)
Pages：6( 346-350,356 )