MSRT: Multi-Scale Spatial Regularization Transformer For Multi-Label Classification in Calcaneus Radiograph

Calcaneus fracture is one of the most common fractures which affect daily life quality. However, calcaneus fracture subtype classification is a challenging task due to the nature of multi-label as well as limited annotated data. In this paper, an augmentation strategy called GridDropIn&Out (GDIO) is proposed to increase the uncertainty of the rough input mask and enlarge the dataset. A spatial regularization transformer (SRT) is designed to capture labels' spatial information, while a multi-scale attention SRT (MSRT) is built to synthesize spatial features from different levels. Our final proposal achieves an mAP of 87.54% in classifying six calcaneus fracture types.