Automatic segmentation of target volume of breast radiotherapy on CT using MSR-UNet: A multicenter study.
Authors
Affiliations (8)
Affiliations (8)
- School of Computer Science and Software Engineering, Southwest Petroleum University, No. 8 Xindu Avenue, Xindu District, Chengdu, 610599, Sichuan, China; Kashgar Institute of Electronics and Information Industry, No. 121 Shenka Avenue, Kashgar, 84400, Xinjiang, China.
- School of Computer Science and Software Engineering, Southwest Petroleum University, No. 8 Xindu Avenue, Xindu District, Chengdu, 610599, Sichuan, China.
- Zhejiang Cancer Hospital, No. 1 Banshan East Road, Gongshu District, Hangzhou, 310022, Zhejiang, China.
- School of Information and Software Engineering, University of Electronic Science and Technology of China, No. 4, Section 2, Jian She Bei Road, Cheng Hua District, Chengdu, 510120, Sichuan, China.
- The First Affiliated Hospital of Yangtze University, No. 8 Hangkong Road, Shashi District, jingzhou, 434000, hubei, China.
- School of Information and Software Engineering, University of Electronic Science and Technology of China, No. 4, Section 2, Jian She Bei Road, Cheng Hua District, Chengdu, 510120, Sichuan, China; Kashgar Institute of Electronics and Information Industry, No. 121 Shenka Avenue, Kashgar, 84400, Xinjiang, China. Electronic address: [email protected].
- The First Affiliated Hospital of Yangtze University, No. 8 Hangkong Road, Shashi District, jingzhou, 434000, hubei, China. Electronic address: [email protected].
- The First People's Hospital of Kashgar, No. 66 Yingbin Avenue, Kashgar, 84400, Xinjiang, China. Electronic address: [email protected].
Abstract
We developed MSR-UNet(Mamba-based Skip Refinement U-Net) to accurately segment the clinical target volume (CTV) and tumor bed (TB) for breast radiotherapy. MSR-UNet was developed based on a nnU-NetV2 framework with semantic-gated skip refinement and bidirectional Mamba for long-range context modeling. This retrospective multicenter study included 1260 patients (A/B/C: 1,024/122/94). MSR-UNet and three baselines (nnU-NetV2, Attention U-Net, and Swin-UNETR) were trained at Institution A with five-fold cross-validation and directly tested at Institutions B and C without retraining. Ablation studies and multicenter heterogeneity analyses were performed to evaluate module contributions and model stability. Performance was evaluated using DSC, HD95, and inference time. Clinical usability was assessed on an external subset (B/C, 20 cases each; n=40) using a two-physician review-and-edit workflow, measuring manual vs model-assisted contouring time and pre/post-edit changes in DSC and HD95. At Institution A cross-validation, MSR-UNet achieved CTV/TB DSC of 90.15 ± 2.11%/75.45 ± 6.50% and HD95 of 4.85 ± 1.21 mm/7.20 ± 2.29 mm, outperforming nnU-NetV2. Ablation studies demonstrated that each proposed module contributed to the overall segmentation accuracy. Furthermore, multicenter heterogeneity analysis revealed substantial variability across institutions; however, MSR-UNet maintained consistent performance at Centers B and C, validating its superior robustness. In the clinical evaluation, model assistance reduced contouring time by 87% for CTV and 75% for TB , including inference, and improved post-edit agreement versus the raw outputs. MSR-UNet improves postoperative breast radiotherapy target segmentation accuracy on CT and maintains performance under external multicenter testing. The generated contours support efficient clinician editing and substantially reduce contouring time in an external subset.