Consistent regularization and proxy label based bonesemi-supervised point cloud segmentation method

Abstract

In deep learning, supervised learning methods consume enormous amount of labeled data. Exceptionally, bone labeling needs to be finished by specialty physicians. In addition, bone data is very difficult to acquire and expensive to annotate. It is desirable to utilize unlabeled data effectively. This paper proposed a semi-supervised method for bone semantic segmentation by combining both advantages of consistent regularization and proxy label. Based on a teacher-student mutual learning framework, proxy labels of the unlabeled data select from the softmax output of student network, and the student network is compared with the teacher output using consistency cost to get trained. Then the teacher network is supervised using proxy label generated from the student network. The teacher network parameters are passed by the student network through a sliding exponential average. The experiments show that compared with the supervised network our method using 10% labeled data achieved similar performance. For bone piece surface extraction and femur surface segmentation, the prediction accuracy was improved by 4.25% after optimizing the graph-cutting algorithm. The accuracy of extracting femur surface segmentation reached 94.4%, and the bone piece outer surface reached 84.3%, this method improves the segmentation accuracy and efficiency, overcomes labeling difficulty.

Ähnliche Arbeiten