Radiologist expertise is in high demand, and the queue for quality interpretation is consistently full. Ellis et al. sought to ease the load by using artificial intelligence (AI) to classify X-rays in a binary fashion – normal or abnormal – while maintaining accuracy and sensitivity. Confidently identifying and removing normal X-rays from the queue would allow radiologists to focus on abnormal cases that require subsequent attention.

Duygu Tosun-Turgut and Michael Hope, along with their co-authors, trained CNNs (convolutional neural networks) to enable AI to classify 7,000 retrospective clinical two-view chest radiographs as “normal” or “abnormal.” The goal was to optimize attention maps toward a biological, meaningful determination of abnormal radiographs. Five different training frameworks were tested, and they found that a hybrid supervision model is accurate, sensitive, and minimizes the number of false negative cases. 

This paper was published in Computers in Biology and Medicine. The use of this practical clinical application may help reduce backlogs and interpretation delays and help to provide swift, quality care.

Additional authors include Ryan Ellis, Erik Ellestad (SFVAMC), and Brett Elicker (UCSF).