A bone segmentation method based on Multi-scale features fuse U2Net and improved dice loss in CT image process

Abstract

During the conventional CT image process, window width is set to extract target. However, different tissues may have the same value of hounsfield unit in CT images, which cause false extraction and noise. In this paper, a computer-assisted target segmentation method based on deep learning is proposed. According to the positional relationship between muscle and bone, an adaptive label softening method based on the change of muscle area is proposed to improve the standard dice loss. At the same time, this paper introduces a label based on distance map to improve the segmentation accuracy at the edge of targets, which can solve the problem of fuzzy bone boundary at the epiphysis. This paper also improves the structure of U2Net and proposes a multi-scale feature fuse U2Net (MFF U2Net). Compared with other U-shaped networks, the method proposed in this paper shows high prediction accuracy (mean 95.244%) and small dispersion of data (variance 0.0008) on the test set. The experiment results show that the proposed segmentation method based on deep learning outperforms the conventional segmentation method significantly.