Edge-enhanced Instance Segmentation of Wrist CT via a Semi-Automatic Annotation Database Construction Method

Instance segmentation of the wrist in Computed Tomography (CT) is crucial for clinical analysis. Existing works mainly focus on extracting bones from CT data and neglect the other parts such as muscle. However, muscle is vital to provide health-related information, such as body fat and body mass index. This paper proposes a deep learning-based instance segmentation model to segment all components in the wrist CT, including bones, muscle, cast, and background. The major challenges of applying deep learning-based approach in wrist CT instance segmentation are: 1) the lack of annotation data, 2) to design a model for accurate segmentation. We propose a semi-automatic annotation database construction method via the Otsu-based and U-net-based self-training semi-supervised learning model. The proposed annotation method greatly reduces the annotation time and workload compared with the time-consuming and laborious manual annotation. To further design an accurate wrist instance segmentation model, an edge-enhanced U-net model has been proposed. The mean Intersection-Over-Union result on 504 testing slices is 0.9868. This is the first work on the instance segmentation of wrist CT to the best of our knowledge.