PNNL

Abstract

SIGN Fracture Care International partners with surgeons in low-resource hospitals worldwide to provide access to effective orthopedic care by donating educational materials and innovatively designed surgical implants. SIGN reaches across 52 countries and interacts with over 5,000 surgeons, but expanding their care has led to an overwhelming amount of medical data. Physicians at SIGN pour through hundreds of radiographs daily, and SIGN's Online Surgical Database (SOSD) contains over 500,000 images spanning two decades. To continue expanding their reach and effectively helping surgeons and patients, we apply machine learning tools to the SOSD to improve the throughput of radiograph analysis. One challenge in working with hospitals worldwide is that both the radiographs uploaded to the SOSD and the data entry accompanying the uploads vary greatly in quality. In order to develop tools to analyze the radiographs, we first need to apply database cleanup tools. The first tool we developed sorts radiographs from other images, achieving near-perfect classification, only misclassifying 1 of 794 test images. We then apply object detection methods to determine the number and placement of nail, screw, and plate implants visible in a radiograph. Active learning was used to generate a training set containing 2,510 radiographs with screws, nails, and plates labeled by bounding boxes. Training a model to recognize all three classes gave a low average precision (AP) for the plate class, likely due to the low number of plates in our training set and the large variety of surgical plates used by SIGN-partnered surgeons. Applying standard image augmentation techniques to increase the count of plates in our training set did not appreciably increase the AP of plate detection. We, therefore, trained one model to detect nails and screws and a separate model to detect plates by redrawing the bounding boxes to account for correlations between the screw and plate classes. This strategy increased the AP of plate detection by 75.3 percentage points. The AP of each class was 78% for screws, 93% for nails, and 89% for plates; meanwhile, the sensitivity was 90% for screws, 85% for nails, and 78% for plates. These two tools were then applied to the full SOSD to correct erroneous entries.