Rib Fracture Diagnosis System on Chest X-Rays with Deep Learning

Abstract

Chest X-Rays is an important imaging modality for diagnosing thorax diseases. However, it is a challenging task to identify rib fractures. The features of rib fractures including various shapes, various sizes, and might overlapped by other organs. It is then time consuming and hard for clinicians to deal with it. Hence, automatically rib fractures recognition is important to clinicians. In this work, we trained two deep convolutional neural networks in order to classify rib fractures in chest x-ray images. One for frontal view images, and another for oblique view images. We present a two-stage method including lung field cropping with U-Net [1] and rib fracture classification with EfficientNet- B0 [2]. We evaluated models’ classification performance using Area Under Receiver Operating Characteristic Curve (AUROC), Area Under Precision Recall Curve (AUPRC), and Accuracy. Class Activation Maps (CAMs) were provided to know the image regions most relevant to the model’s prediction.