KEY POINTS
- Model development and validation study of FS-Mamba, a lightweight automatic segmentation network for computed tomography-based liver tumor radiotherapy.
- The model was tested on 3 datasets: public CT-ORG (140 computed tomography volumes), Synapse (30 cases), and an in-house liver cancer dataset (50 patients, 12 structures including planning target volume and gross tumor volume).
- FS-Mamba achieved mean Dice similarity coefficient / 95th percentile Hausdorff distance of 92.81% / 10.74 on CT-ORG, 81.28% / 14.32 on Synapse, and 92.68% / 10.17 on the liver cancer dataset.
- It outperformed UNet, nnUNet, TransUNet, Swin-UMamba, and TotalSegmentator with statistical significance (p < 0.05).
- FS-Mamba had 20.57 million parameters, 31.94 giga floating-point operations, and 20.16 milliseconds inference time per case; subjective scoring by 10 blinded medical physicists was 4.69 / 5.0, compared with 4.94 / 5.0 for ground-truth contours.
CLINICAL TAKEAWAY
FS-Mamba looks technically promising for automated organ-at-risk and target segmentation in liver radiotherapy, especially where online adaptive workflows require fast contour generation. The model combines strong segmentation metrics with low computational burden, but the evidence is still model-validation evidence. It needs broader multi-institutional testing, prospective workflow validation, and evaluation beyond computed tomography before being treated as clinically mature.