Deep learning matches B-readers for pneumoconiosis screening on chest X-rays

On in-distribution test data, deep learning models achieved accuracies of 0.93, 0.92, and 0.91 for detecting parenchymal abnormalities, any abnormality, and small opacity grade, respectively, comparable to certified B-readers. Performance dropped on out-of-distribution cases.

• For a three-way classification (progressive massive fibrosis vs. simple pneumoconiosis vs. no disease), the DL model accuracy was 0.83, similar to the average B-reader accuracy of 0.86.
• Model performance was markedly lower on an out-of-distribution test set from a different scanner manufacturer, indicating reduced generalizability.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.