Relationship between number of annotations and accuracy in segmentation of the erector spinae muscle using Bayesian U-Net in torso CT images

摘要

Supervised learning for image segmentation requires annotated images. However, image annotation has the problem that it is time-consuming. This problem is particularly significant in the erector spinae muscle segmentation due to the large size of the muscle. Therefore, this study considers the relationship between the number of annotated images used for training and segmentation accuracy of the erector spinae muscle in torso CT images. We use Bayesian U-Net, which has shown high accuracy in thigh muscle segmentation, for the segmentation of the erector spinae muscle. In the network training, we limit the number of slices for each case and the number of cases to 100%, 50%, 25%, and 10%. In the experiment, we use 30 torso CT images, including 6 cases for the test dataset. Experimental results are evaluated by the mean Dice value of the test dataset. Using 100% of the slices per case, the segmentation accuracy with 100%, 50%, 25%, and 10% of the cases was 0.934, 0.927, 0.926, and 0.890, respectively. On the other hand, using 100% of the cases, the segmentation accuracy with 100%, 50%, 25%, and 10% of the slices per case was 0.934, 0.934, 0.933, and 0.931, respectively. Furthermore, the segmentation accuracy with 100% of the cases and 10% of the slices per case was higher than that of the previous method. We showed that it is feasible to achieve high segmentation accuracy with a limited number of annotated images by selecting several slices from a limited number of cases for training.