Pneumoconiosis screening and classification using deep learning models.

OBJECTIVES: To develop and evaluate different deep learning (DL) models for clinically relevant screening and classification of pneumoconiosis based on chest radiographs. METHODS: This retrospective study used B-reader labelled pneumoconiosis lung images from US workers. After pre-processing the images, DL models were trained and evaluated for four classification tasks: (1) parenchymal abnormalities present versus absent, (2) parenchymal and/or pleural abnormalities present versus absent, (3) small parenchymal opacities grade 1 versus grade 0 and (4) progressive massive fibrosis versus simple pneumoconiosis versus no pneumoconiosis. RESULTS: A total of 2968 lung images were divided as follows: 2294 for training, 287 for validation, 287 for independent in-distribution testing and 100 for out-of-distribution testing from a separate scanner manufacturer dataset. On the in-distribution testing data set, the DL models achieved an accuracy of 0.93 (SD=0.01), 0.92 (SD=0.02) and 0.91 (SD=0.02) for task 1, 2 and 3, respectively. DL had comparable or better performance compared with certified B-readers who had average accuracies of 0.9 (SD=0.11), 0.91 (SD=0.10) and 0.77 (SD=0.21). For task 4, the DL model achieved an accuracy of 0.83 (SD=0.02), which is similar to B-readers' performance (average accuracy of 0.86 (SD=0.12)). Performance was markedly lower on the out-of-distribution test set, with reduced accuracy and AUC. CONCLUSIONS: On independent, in-distribution test images, DL models performed comparably to certified B-readers for tasks 1-4, supporting their potential to augment pneumoconiosis screening and classification. Reduced performance on challenging out-of-distribution cases warrants further study to improve generalisation.