AbstractPurpose To evaluate a data-efficient segmentation approach for quantifying lumbar paraspinal muscle (LPM) at the single-slice L4/L5 level across multi-sequence MRI, minimizing annotation burden while achieving reliable group-level quantification. Materials and methods A total of 943 MRI scans from 486 participants were retrospectively analyzed, with a single axial slice at the L4/L5 disc level selected from each scan. The images included T1-weighted, T2-weighted, and Dixon MRI. A training-free Segment Anything Model 2 (SAM2) was applied for dataset-level LPM segmentation. For each dataset, one, two, or three ground-truth masks were used as prompt labels (4, 8, and 12 images in total across all four datasets), with the remaining images used for testing (939, 935, and 931 images, respectively). Segmentation accuracy was evaluated using the Dice similarity coefficient (DSC). Automated quantifications were assessed using two one-sided tests (TOST), intraclass correlation coefficients (ICC), and Bland–Altman analysis. Results With three-shot prompting, SAM2 achieved the highest segmentation accuracy for the left and right lumbar erector spinae and multifidus muscles (DSC: 0.871 ± 0.128 to 0.944 ± 0.063). Automated measurements were statistically equivalent to manual references within a 5% margin for bilateral muscle area and a 10% margin for bilateral fat ratio, except for the fat ratio of left erector spinae, which required a 14.14% margin. ICCs ranged from 0.822 to 0.856 for muscle area and 0.771–0.927 for fat ratio. Conclusion L4/L5 slice-level LPM segmentation and quantification are achieved using only three labeled examples per dataset, with multicenter evaluation and publicly available code and data.
Zhang et al. (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: