Related Experiment Video
Updated: Apr 12, 2026
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
Is Deep Learning Ready for Abdominal Organ-at-Risk Segmentation in the Foundation Model Era: A Comprehensive Study of
1Department of Radiation Oncology, Precision Radiation in Oncology Key Laboratory of Sichuan Province, Sichuan Clinical Research Center for Cancer, Sichuan Cancer Hospital & Institute, Sichuan Cancer Center, University of Electronic Science and Technology of China, Chengdu, China; School of Mechanical and Electrical Engineering, University of Electronic Science and Technology of China, Chengdu, China.
Purpose:
The application of deep learning-based methods for accurate organ-at-risk segmentation in challenging clinical scenarios remains unexplored. This study aims to evaluate state-of-the-art fully supervised learning (FSL) methods and foundation model (FM)-based methods across challenging clinical scenarios, and to propose an effective solution to improve model robustness and reduce organ hallucination.
Methods And Materials:
We retrospectively collected computed tomography (CT) scans from 413 patients across 2 institutions, divided into 3 cohorts based on treatment strategy. Seven FSL and 6 FM methods were comprehensively evaluated on an internal testing cohort (n = 67, without surgery), external testing cohort 2 (n = 22, partial organ resection surgery), and external testing cohort 3 (n = 74, whole organ resection surgery), as well as 3 public datasets. We further introduced an organ erasure augmentation (OEA) strategy to improve generalization and address hallucinations in missing organs. Quantitative metrics included Dice similarity coefficient, normalized surface Dice (NSD), and hallucination ratio.
Results:
Two of three fine-tuned FM methods failed to produce any segmentation outputs for 5 and 6 of 19 organs, respectively. Prompt-based FM methods using tight bounding box prompts demonstrated stable performance but struggled with complex anatomy such as the intestine. Our proposed OEA method outperformed existing FM-based and FSL methods, achieving mean Dice similarity coefficient and mean normalized surface Dice of 87.29% and 87.84% on the internal testing cohort, 85.15% and 85.01% on the external testing cohort 2, and 82.29% and 81.81% on the external testing cohort 3, respectively. Compared with the best-performing method, our method reduced the mean hallucination ratio from 0.571 to 0.516 and demonstrated superior cross-dataset generalization with less performance degradation.
Conclusions:
Current FM-based and FSL methods remain insufficient for clinical use in cases involving irregular anatomy or significant distribution shifts. The proposed OEA strategy reduces hallucination and enhances segmentation robustness, offering a promising step toward reliable clinical application.
