Foundation model-guided multi-view semi-supervised CT segmentation of liver tumors in resource-constrained settings.
Authors
Affiliations (7)
Affiliations (7)
- Department of Gastroenterology, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, Jiangxi, China. [email protected].
- Department of General Surgery, The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, Jiangxi, China.
- Department of Gastroenterology, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, Jiangxi, China.
- Department of Medical Oncology, The Second Affiliated Hospital, School of Medicine, South China University of Technology, Guangzhou, Guangdong, China.
- Department of Hepatobiliary and Pancreatic Surgery, The First Affiliated Hospital, Jinan University, Guangzhou, Guangdong, China.
- Department of Oncology, The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, Jiangxi, China. [email protected].
- Department of General Surgery, The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, Jiangxi, China. [email protected].
Abstract
We present a label-efficient pipeline for CT auto-segmentation in resource-constrained settings. The framework couples a semi-supervised segmentation backbone with a foundation model-guided regularizer to strengthen the learning from scarce annotations. To better exploit volumetric context, we introduce a multi-view collaborative learning procedure that performs view-specific inference to form a unified supervision signal that suppresses view-dependent noise and improves mask fidelity. We evaluate on a public CT benchmark with varying numbers of labeled scans. In the highly label-limited regime, the approach yields strong accuracy with average Dice 83.79% for the liver and 60.08% for the tumor using 20 labeled cases, outperforms existing segmentation methods. By reducing contouring from hours to seconds, improving small-structure recovery and boundary fidelity, and requiring no interactive prompts, the method offers a plug-and-play path to deployment and a reliable basis for downstream radiomics and longitudinal monitoring.