The UK universities’ researchers developed an open-source AI system called X-Raydar. This system is used to detect abnormalities in chest x-ray images and reports, demonstrating the capability to classify common chest x-ray findings from images and free-text reports. The X-Raydar achieved a mean AUC of 0.919, 0.864, and 0.842 on different datasets. X-Raydar-NLP, the natural language processing algorithm, was trained to label chest x-rays using a taxonomy of 37 findings. The X-Raydar system demonstrated good detection of clinically relevant findings and matched the performance of trained radiologists for critical findings.
The researchers also developed web-based tools for real-time chest x-ray interpretation. The X-Raydar online portal allows users to upload DICOM images for automatic classification. In addition, the researchers open-sourced their trained network architectures, providing a foundation model for future research and adaptation.
The researchers have successfully created and evaluated the X-Raydar AI system for comprehensive chest x-ray abnormality detection. This has remarkable implications for the advancement of AI applications in radiology. You can read the full research paper here.