Automatic calcaneus fracture identification and segmentation using a multi-task U-Net

Calcaneus is the bone in the foot that bears most of the body weight and calcaneus fracture is the most common type of tarsal bone fractures. Plain radiograph examination is usually the first step of calcaneus fracture diagnosis because of its convenience and low cost. A multi-task U-Net is proposed in this paper to develop a computer aided calcaneus fracture diagnosis system. Our approach is an end-to-end CNN for identification and segmentation of calcaneus fracture, which uses regularization of the two tasks for mutual performance enhancement. First, a novel radiograph normalization method to obtain scale rotation invariance under different monochrome type is employed. Second, a classification header with feature from decoder and encoder is added to U-Net for multitask. Finally, a conditional dice-loss which can promote model performance under rough-ground-truth supervision is adopted in training. Experiments show that the network predicts fracture regions more precise than the rough ground-truth and identifies fracture with sensitivity of 99.53% and specificity of 98.59%.