(Saturday, 4/7/2018)
Room: Foyer
Purpose: To develop and assess performance of artificial intelligence (AI) algorithms that perform automatic quality control (QC) checks on chest x-ray images.
Methods: Over 100,000 x-ray images of numerous anatomical exams and views were compliantly collected from five institutions in the USA, Canada, and China. All images had been either QC rejected on the modality or QC accepted and sent to PACS. Variational autoencoder (VAE) deep learning and convolutional neural networks (CNNs) were used to create AI algorithms to automatically perform two QC tasks. The first algorithm, a one-class classifier using VAE and CNNs, trained to detect a single class - frontal (AP/PA) chest x-ray image versus image of other anatomy. DICOM header information was used to provide initial image level labels, followed by manual review and annotation correction. The second algorithm, a binary classifier using CNNs, was trained to determine whether the patient positioning in a frontal chest x-ray was acceptable. This algorithm was trained using x-ray images rejected for patient positioning errors by a technologist as well as those accepted. The technologists’ self-reported reject reasons were used as initial image level labels. Then radiographic technologists manually reviewed and conducted annotation corrections. Performance of the algorithm for correct view and positioning were evaluated using receiver operating characteristic (ROC) analysis.
Results: Both algorithms performed very effectively; each with an ROC area-under-the-curve of 0.99. The accuracies of the algorithms were 0.99 and 0.95 for the frontal chest x-ray detection algorithm and patient positioning algorithm respectively.
Conclusion: This work demonstrates the feasibility of using AI as a virtual QC technologist to determine if incorrect anatomy or view were acquired and whether patient positioning was acceptable for chest x-ray images. These results warrant further development to expand anatomy and view types, and additional image reject reasons.
Funding Support, Disclosures, and Conflict of Interest: All authors are employees of GE Healthcare
ePosters
Contact Email
