Automatic extraction of multiple-study X-ray images

In medical radiology standard practice, radiographs of different projections or studies are occasionally merged into a single image for convenience. Before the radiographs can be utilized for further automatic processing, e.g., for detecting and/or localizing specific injuries, such merged projections or studies need to be separated. Doing this manually can be tiresome; therefore, we decided to investigate the possibility of automating this process using contemporary image processing and machine learning techniques. We implemented two independent automatic solutions for this problem: (1) a manually-tailored image-splitting method using morphological operations, image binarization, and boundary-box detection, and (2) an instance of the YOLOv4 algorithm. The manually-tailored algorithm and YOLOv4 model were trained on the training data set of 4,000 images and validated on the data set of 250 images. Comparing both on a disjoint test set, consisting of 250 images, the YOLOv4 model noticeably outperformed the manually-tailored method with an accuracy of 0.992, F1 score of 0.996, and the intersection-overunion of 0.8978±0.079. The results suggest that YOLOv4 can efficiently extract large portions of embedded projections or studies from medical radiographs.