Exploiting Pseudo-Labeling and nnU-Netv2 Inference Acceleration for Abdominal Multi-Organ and Pan-Cancer Segmentation

Ziyan Huang, Jin Ye, Haoyu Wang, Zhongying Deng, Tianbin Li, Junjun He

Keywords: Medical Image Segmentation, Computational Efficiency, Abdominal Tumor

Abstract: Deep-learning based models offer powerful tools for the automatic segmentation of abdominal organs and tumors in CT scans, yet they face challenges such as limited datasets and high computational costs. The FLARE23 challenge addresses these by providing a large-scale dataset featuring both partially and fully annotated data, and by prioritizing both segmentation accuracy and computational efficiency. In this study, we adapt the winning FLARE22 strategy to FLARE23 by utilizing a two-step pseudo-labeling approach. Initially, a large model trained on datasets with complete organ annotations generates pseudo-labels for datasets that originally contain only tumor annotations. These labels are then integrated to create a comprehensive training dataset. A smaller, more efficient model is subsequently trained on this enriched dataset for deployment, targeting both tumors and organs. Our approach, utilizing the FLARE23 dataset, has achieved notable results. On the online validation leaderboard, it reached an average DSC of 89.63\% for organs and 46.07\% for lesions, with an average processing time of 16.1 seconds for 20 selected validation cases. In the final testing set, our model demonstrated improved performance, achieving an organ DSC of 89.98\% and lesion DSC of 62.61\%, while reducing the average processing time to 12.02 seconds. The code and model are publicly available at https://github.com/Ziyan-Huang/FLARE23.

Submission Number: 23