A Simple Mean-Teacher UNet Model for Efficient Abdominal Organ Segmentation

Zixiao ZHAO, Jiahua Chu

Keywords: Medical Image Segmentation, Abdominal Organ Segmentation, Semi-supervised, UNet, Mean-Teacher

Abstract: One inevitable barrier to deep learning-based medical image segmentation algorithms is that for such tasks requiring high accuracy, all models must be trained using large datasets annotated by experts, and this process is exceptionally time-consuming and laborious. For abdominal organ segmentation, this problem becomes more prominent as the image size becomes larger. To address this problem, we design a classical UNet model using the Mean-Teacher strategy to obtain relatively satisfactory segmentation ($58.93\%$ DSC and $59.54\%$ NSD)results on a semi-supervised abdominal segmentation dataset. The core idea is to use labeled data to improve the segmentation performance of the model itself, while introducing noise on unlabeled data to improve the generalization of the model. Inspired by nnUNet, we use as simple a model structure as possible, thus ensuring the efficiency during training and inference phases (< 2GB VRAM consumption and $\sim$10s inference time)