Fracture Detection and Localization in Chest X-Rays Using Semi-Supervised Learning with Dynamic Sharpening

In this work, we present a low-cost and efficient method for training a rib and clavicle fracture detection model for chest X-ray (CXR) in a semi-supervised setting where only a small portion of training data with location annotation. Our method leverages the teacher-student model paradigm which forms a consensus prediction of unknown labels using the output under different input augmentation conditions. And most importantly, we develop a dynamic sharpening method to make the pseudo label generated by the teacher model approximate to the true label with low entropy. This dynamic sharpening method adaptively adjusts the sharpening effect according to the performance of the model during the training process, which can effectively cope with the label imbalance problem in the real world, and improve the model sensitivity. The experiment results demonstrate that our method achieves the state-of-the-art fracture detection performance, i.e., an area under receiver operating characteristic curve (AUROC) of 0.9767 and a free-response receiver operating characteristic (FROC) score of 0.9300, significantly outperforming previous approaches by a gap of 1.00% and 3.68% respectively.