Patch-shuffle-based semi-supervised segmentation of bone computed tomography via consistent learning

Bone segmentation is essential in Computed Tomography (CT) imaging, which assists physicians in diagnosing, planning operations, and evaluating treatment effects. A recent trend is to apply deep neural networks for bone segmentation in CT slices. However, training a feasible segmentation model often requires a large number of annotated CT slices. The annotation procedure is often laborious and time-consuming. This paper proposed a novel semi-supervised learning method for training a bone segmentation model in CT images. A patch-shuffle -based data augmentation method was proposed by leveraging the unique bone structures in CT. To utilize the unlabeled CT slices, we employed a consistent learning loss between the feature of the original CT slice and the corresponding patch-shuffled CT slice. This encouraged the model to generate the same feature on each corresponding pixel in the two kinds of data. We validated the proposed method on three CT datasets of different image qualities and different anatomies: wrist, foot, chest, head, abdomen, and limb. The experiment results on the three bone CT datasets demonstrated the outperformance of our proposed method over the other semi-supervised methods.

Automated summary

This paper proposes a novel semi-supervised learning method for training a bone segmentation model in CT images. It leverages the unique bone structures in CT for data augmentation and utilizes unlabeled CT slices for training. The experiment results demonstrate the superior performance of the proposed method over other methods on different bone CT datasets.