Journal of applied clinical medical physics · 2026
FS-Mamba: Feature-wise scanning Mamba UNet for automatic image segmentation in liver tumor radiotherapy.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Yin Peijun, Zhang Xueren, Liu Xin, Jiang Zekun, Qiu Qingtao, Yin Yong, Li Zhenjiang
Аннотация
BACKGROUND: Medical image segmentation is fundamental to radiotherapy planning, yet accurate delineation of organs at risk and tumor targets remain challenging due to anatomical variability and low soft-tissue contrast in CT images.
PURPOSE: To develop a lightweight, high-precision automatic segmentation network that meets the dual clinical requirements of accuracy and computational efficiency for online adaptive liver cancer radiotherapy.
METHODS: We propose FS-Mamba, a U-shaped encoder21-decoder network built upon a novel Frequency-Long SSM Block that integrates three innovations: (1) a flip-selective scanning strategy (FS-Scanning) that dynamically reorders feature sequences to prioritize salient information; (2) a frequency-domain modeling branch using Fast Fourier Transform to enhance boundary discrimination; and (3) a long-memory state space model (LSSM) that augments the A-matrix to strengthen historical state retention. The model was validated on three datasets: the public CT-ORG (140 CT volumes, 6 organs) and Synapse (30 cases, 8 organs) benchmarks, and an in-house liver cancer dataset (50 patients, 12 structures including PTV and GTV). Comparisons were made against UNet, nnUNet, TransUNet, Swin-UMamba, and the clinically mainstream TotalSegmentator using Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), and subjective physicist scoring.
RESULTS: FS-Mamba achieved mean DSC/HD95 of 92.81%/10.74 on CT-ORG, 81.28%/14.32 on Synapse30, and 92.68%/10.17 on the liver cancer dataset, outperforming all compared methods with statistical significance (p < 0.05). It received the highest mean subjective clinical score of 4.69/5.0 from 10 blinded medical physicists, approaching the ground-truth reference (4.94/5.0). Despite its leading accuracy, FS-Mamba maintains the smallest model size (20.57 M parameters, 31.94 GFLOPs) and fastest inference (20.16 ms per case), representing a 4.5x parameter reduction and 6.5x speedup compared to TotalSegmentator.
CONCLUSIONS: FS-Mamba provides a clinically viable solution for automated multi-organ and tumor target segmentation that satisfies the stringent time constraints of online adaptive radiotherapy while maintaining expert-level accuracy.
PURPOSE: To develop a lightweight, high-precision automatic segmentation network that meets the dual clinical requirements of accuracy and computational efficiency for online adaptive liver cancer radiotherapy.
METHODS: We propose FS-Mamba, a U-shaped encoder21-decoder network built upon a novel Frequency-Long SSM Block that integrates three innovations: (1) a flip-selective scanning strategy (FS-Scanning) that dynamically reorders feature sequences to prioritize salient information; (2) a frequency-domain modeling branch using Fast Fourier Transform to enhance boundary discrimination; and (3) a long-memory state space model (LSSM) that augments the A-matrix to strengthen historical state retention. The model was validated on three datasets: the public CT-ORG (140 CT volumes, 6 organs) and Synapse (30 cases, 8 organs) benchmarks, and an in-house liver cancer dataset (50 patients, 12 structures including PTV and GTV). Comparisons were made against UNet, nnUNet, TransUNet, Swin-UMamba, and the clinically mainstream TotalSegmentator using Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), and subjective physicist scoring.
RESULTS: FS-Mamba achieved mean DSC/HD95 of 92.81%/10.74 on CT-ORG, 81.28%/14.32 on Synapse30, and 92.68%/10.17 on the liver cancer dataset, outperforming all compared methods with statistical significance (p < 0.05). It received the highest mean subjective clinical score of 4.69/5.0 from 10 blinded medical physicists, approaching the ground-truth reference (4.94/5.0). Despite its leading accuracy, FS-Mamba maintains the smallest model size (20.57 M parameters, 31.94 GFLOPs) and fastest inference (20.16 ms per case), representing a 4.5x parameter reduction and 6.5x speedup compared to TotalSegmentator.
CONCLUSIONS: FS-Mamba provides a clinically viable solution for automated multi-organ and tumor target segmentation that satisfies the stringent time constraints of online adaptive radiotherapy while maintaining expert-level accuracy.