This AI Paper Unveils X-Raydar: A Groundbreaking Open-Source Deep Neural Networks for Chest X-Ray Abnormality Detection

Researchers from various universities in the UK have developed an open-source artificial intelligence (AI) system, X-Raydar, for comprehensive chest x-ray abnormality detection. Trained on a dataset from six UK hospitals, the system utilizes neural networks, X-Raydar and X-Raydar-NLP, for classifying common chest X-ray findings from images and their free-text reports. The dataset, spanning 13 years, included 2,513,546 chest x-ray studies and 1,940,508 usable free-text radiological reports. A custom-trained natural language processing (NLP) algorithm, X-Raydar-NLP, labeled the chest X-rays using a taxonomy of 37 findings extracted from the reports. The AI algorithms were evaluated on three retrospective datasets, demonstrating similar performance to historical clinical radiologist reporters for various clinically important findings.

The X-Raydar achieved a mean AUC of 0.919 on the auto-labeled set, 0.864 on the consensus set, and 0.842 on the MIMIC-CXR test. Notably, X-Raydar outperformed historical reporters on 27 of 37 findings on the consensus set, showed non-inferiority on nine, and was inferior on one finding, resulting in an average improvement of 13.3%. The system’s performance matched that of trained radiologists for critical findings, including pneumothorax, parenchymal opacification, and parenchymal mass or nodules.

The development included a radiological taxonomy covering eight anatomical areas and non-anatomical structures, facilitating comprehensive labeling. An NLP algorithm, X-Raydar-NLP, was trained on 23,230 manually annotated reports to extract labels. X-Raydar, the computer vision algorithm, used InceptionV3 for feature extraction and achieved optimal results using a custom loss function and class weighting factors.

For testing, a consensus set of 1,427 images annotated by expert radiologists, an auto-labeled set (n=103,328), and an independent dataset, MIMIC-CXR (n=252,374), were employed. X-Raydar-NLP demonstrated good detection of clinically relevant findings in free-text reports, with a mean sensitivity of 0.921 and specificity of 0.994. X-Raydar’s mean AUC across all findings on the consensus set was 0.864, showing a strong performance for critical, urgent, and non-urgent findings.

The researchers also developed web-based tools, allowing public access to the AI models for real-time chest x-ray interpretation. The X-Raydar online portal lets users upload DICOM images for automatic pre-processing and classification. Additionally, the researchers open-sourced their trained network architectures, providing a foundation model for further research and adaptation. The researchers have successfully developed and evaluated an AI system, X-Raydar, for comprehensive chest x-ray abnormality detection. The system demonstrated comparable performance to historical radiologist reporters and is made freely accessible to the research community, contributing to the advancement of AI applications in radiology.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to join our 34k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..