Semi-supervised Abdominal Multi-Organ and Tumors Segmentation by Cascaded nnUNet

Bochen Wu, Wenli Fu, Mengyao Zhang

Keywords: Semi-supervised learning, Multi-Organ segmentation, Tumor segmentation

Abstract: Abdominal multi-organ and tumors segmentation can provide anatomical structure information for doctors and is an important step in computer-aided diagnosis. However, accurate segmentation of abdominal multi-organ and tumors is still an urgent problem due to partially labeled issue and variable tumor position. To address these problems, we propose a cascaded approach using cascaded nnU-Net to handle the task of multi-organ and tumors segmentation. Since tumors located in different organs have different gray value and textures, we train segmentation models for each tumor to improve the tumor segmentation accuracy. We also combine semi-supervised method while training to makes full use of the unlabeled data. In addition, we postprocess the segmentation results to refine segmentation based on anatomical prior knowledge. We improve the inference speed by replacing the interpolation function and cropping the probability map. We obtain an average DSC of 90.28\% on abdominal multi-organ segmentation and 42.87\% on pan-tumor segmentation, with an average inference time of 23.77s per case on validation set.

Submission Number: 17