Semi-Supervised Abdominal Organ and Pan-cancer Segmentation with Efficient nnU-Net

Ziran Chen, Taiyu Han, Xueqiang Zeng, Guangtao Huang, Huihui Yang, Yan Kang

Keywords: Segmentation, Semi-supervised learning, Pseudo-labels.

TL;DR: Semi-Supervised Segmentation with Efficient nnU-Net

Abstract: Abdominal organs serve as frequent sites for the manifestation of cancer, however, a prevailing gap exists in the availability of a widely accessible and precise segmentation model tailored to these organs and associated tumors. While nnU-Net has become a powerful baseline for medical image segmentation in recent years, its default configuration lacks the ability to leverage unlabeled data and falls short in terms of inference efficiency. To surmount these inherent constraints, we propose an improved approach based on nnU-Net. Our proposed method incorporates a semi-supervised algorithm that utilizes pseudo-labeling to effectively process unlabeled data within the nnU-Net framework. We improve the utilization of unlabeled data by generating high quality pseudo-labels with the default nnU-Net. Additionally, we reduce the network complexity of 3D U-Net and train a lightweight student model using a combination of labeled and pseudo-labeled data. In terms of performance, our lightweight student model achieved promising results on the validation set. The method yielded the average DSC of 0.8856 and NSD of 0.9451 in the process of segmenting 13 abdominal organs. For tumor segmentation, the average DSC and NSD were computed as 0.4258 and 0.3513, respectively. The average running time per case is 29s and the average GPU memory is 25411MB. In conclusion, our approach effectively addresses the limitations of nnU-Net, improving both inference efficiency and the utilization of unlabeled data. The encouraging results obtained in the FLARE 2023 challenge underscore the potential of our method to advance practical clinical applications in the field of medical image segmentation.

Submission Number: 1