Multi-head consistent semi-supervised learning for lumbar CT segmentation

Segmentation of lumbar spine from computed tomography (CT) images is a necessary prerequisite for the intelligent diagnosis of many lumbar diseases. Currently, deep learning based methods have obtained encouraging experimental results with amount of labeled medical images. However, image labeling is time and effort consuming and lots of unlabeled data cannot be fully used. In this paper, a Multi-Head Consistent learning Network (MHC-Net) is designed to effectively utilize unlabeled data for semi-supervised medical image segmentation. The proposed MHC-Net method is based on the advanced teacher–student structure. The method raises two novel modules, multi-head decomposition (MHD) and cross supervision module (CSM), to enhance the precision of lumbar vertebral segmentation. The MHD promotes feature diversity by extending output layers of the encoder–decoder network, while the CSM learns generalized feature representations by computing the similarity between multi-head decompositions. The study conducts numerous experiments on two widely-used medical databases for the spine. The proposed MHC-Net approach still achieves competitive performance with only 20% labeled data, which is significant to reduce data labeling burden and assist intelligent spinal diagnosis.