MFTAnet: Two-step Aggregation Net of Multiscale Features for Pneumoconiosis Screening

Pneumoconiosis is an occupational disease among mine and tunnel workers. Early screening is important in prevention and treatment of pneumoconiosis. In response to lack of capability in early screening pneumoconiosis by doctors in primary hospitals, a deep learning model, named Two-step Aggregation Networks of Multiscale Features (MFTAnet), for pneumoconiosis screening was proposed in this paper. The multi-scale features are extracted by a four-stage branch of MFTAnet composed by convolutions, atrous convolutions and layer normalizations (LN). A two-step iterative feature integration is used in each stage. Iterative normal projection and cross-stage smooth projection carried on the intermediate-scale are proposed to initial fuse intra- and inter-scale features respectively by point convolutions and LN while iterative fine integration based on attention is composed by a Max Pooling Channel Attention (MPCA) block and an Inverted Residual MLP (IRMLP) block to further integrate the initial fused features. The integrated features of stage 4 are fed into classifier to screening early-stage pneumoconiosis. The MFTAnet is compared with state-of-art methods and proves that the proposed method can extract complex features in chestX images more effectively and achieve higher accuracy of primary screening of pneumoconiosis than state-of-art methods. The accuracy, F1 score, precision and recall of proposed method achieve 92.72%, 92.48%, 92.05% and 92.90% respectively, indicating the proposed method can be used in pneumoconiosis screening to help primary doctors’ early screening pneumoconiosis.