Vertebral Region Segmentation for CT Images via Semi-supervised Learning

Bone mineral density (BMD) is a key indicator of bone mass in the human body and has received increased attention in recent years. Quantitative computerized scanning (QCT) is commonly used in clinical detection to get the CT slices of human, and then using the bone densitometry to get the bone densitometer. In addition, conventional QCT detection requires the lumbar spine area of interest to be checked out of the CT image first. The checking process may be affected by operating errors and other factors, resulting in excessive or incorrect checking of non-lumbar vertebrae areas, which will affect the evaluation of bone density value. However, the well-trained neural network can accurately segment the lumbar vertebral ROI area and avoid the judgment of bone density value due to the non-related areas around the vertebrae or the hard bone edge. While the lumbar scan images of each patient have large differences, and the data with lumbar vertebrae region of interest labels are less so that training the lumbar vertebrae ROI segmentation network is more difficult. Therefore, this paper proposes a semi-supervised network called SAT-UNet. In SAT-UNet, some channel attention blocks are fused into the U-Net to make the network more focused on the important feature channels. In particular, this paper evaluates the pseudo labels by confidence and selects better quality pseudo-labels for network training. Finally, the network performance was evaluated by using three Metrics compared with the U-Net and the semi-supervised CARes-UNet, the results showed that SAT-UNet increased by about 8% in the case of 180 labeled data, which proved that SAT-Net has an advantage in the task of lumbar vertebral ROI segmentation.