Abstract:To address the clinical challenges in pneumoconiosis staging caused by ambiguous lesion features and weak anatomical interpretability in chest X-ray images, this study proposes a dual-stream fusion model for pneumoconiosis staging diagnosis, named MDC-Net, which integrates multi-scale perception and Transformer-based representation learning. By integrating a multi-scale perception network with a detail-preserving encoder, MDC-Net synergistically incorporates local anatomical features and global spatial correlation analysis to significantly improve diagnostic accuracy. A global grouped attention convolution module is introduced to enhance the extraction of micronodule features, while a reversible residual encoder is employed to preserve texture details of critical signs such as ground-glass opacities, effectively mitigating texture loss. Furthermore, a global-local dual-stream fusion mechanism is designed, in which a hierarchical anatomical Transformer builds spatial correlations between hilar structures and pleural lines, addressing limitations in spatial relationship modeling. Experimental results on 1760 clinical cases from Southwest Medical University Hospital show that MDC-Net achieves AUC values of 0.9877, 0.8449, 0.7987, and 0.9912 for stages I to IV, respectively. Compared to Vision Transformer, the proposed model improves overall accuracy by 5.42% and the F1-score by 4.92%. Particularly for the challenging differentiation between stages I and II, AUC values are increased by 0.0934 and 0.0426, respectively. These results effectively meet clinical demands for pneumoconiosis-specific feature characterization and precise staging.