MTSegNetSemi-supervised Abdominal Organ Segmentation in CT

Shiman Li, Siqi Yin

Abstract: Multi-organ segmentation from CT scan is useful in clinical applications. However, difficulties in data annotation impede its practical usage. In this work, we propose MTSegNet for multi-organ segmentation task in semi-supervised way. Total number of 13 organs in chest and abdomen are included. For network architecture, Attention U-Net serves as basic structure to guarantee segmentation performance and usage of context information. For those unlabeled data, Mean Teacher Model, which is a commonly used semi-supervised structure, is added to the pipeline to facilitate better use of unlabeled data. Besides, class-aware weight and post-process are used as auxiliary methods to further improve performance of model. Experiments on validation set and test set got averaged Dice Similarity Coefficient (DSC) of 0.6743 and 0.7034, respectively.