Automated paraspinal muscle quantification at L3-L4 on lumbar MRI: external validation in a cohort study

To develop and externally validate a workflow for automatically evaluating paraspinal muscle at the L3-L4 level on lumbar Magnetic Resonance Imaging (MRI) without manual slice selection or manual segmentation. In this cross-sectional external validation study, spinal MR images from 152 participants in the Chongqing Osteoporosis Screening Study were screened. After participant selection, 146 eligible cases entered the pipeline, and 30 cases were chosen by age-stratified random sampling for manual reference validation. Localization performance was evaluated using absolute slice error (ASE) and hit rate (HR). Segmentation performance was assessed using the Dice similarity coefficient (Dice) for paraspinal muscle regions of interest. Agreement between automated measurements and measurements derived from the final reviewed GT1 reference was evaluated using intraclass correlation coefficients (ICCs). All the above metrics were derived from a subgroup of 30 age-stratified validation cases by comparing automated segmentations with the final reviewed GT1 reference segmentations. Robustness was further examined across age strata, with exploratory analysis of the degeneration subgroup. In the 146 eligible cases entering the pipeline, 144 were successfully processed, yielding a success rate of 98.6%. In the validation subset, automated localization identified the target level with complete agreement relative to manual localization (ASE = 0.07 ± 0.26; HR = 100%). Automated segmentation showed high overlap with final reviewed GT1 reference segmentation (Dice = 0.934 ± 0.016). Measurements of lean cross-sectional area (lean CSA) and the T2-weighted signal-intensity-based fat fraction (FF) estimate, which differs from chemical-shift-derived proton density fat fraction (PDFF), showed high agreement with the manual workflow (ICC = 0.91–0.98), with more stable consistency for lean CSA. Performance was broadly consistent across age strata. In the exploratory degeneration subgroup ( n = 6), the ICC for the T2-weighted signal-intensity-based FF estimate had a wide confidence interval, indicating substantial uncertainty in the agreement estimate. At the L3-L4 level, the proposed workflow achieved a high processing success rate and showed good localization performance, segmentation overlap, and agreement with measurements derived from the final reviewed GT1 reference. These findings support the feasibility of automated paraspinal muscle quantification at L3-L4 under the evaluated acquisition protocol, although agreement of the T2-weighted signal-intensity-based FF estimate in the exploratory degeneration subgroup should be interpreted cautiously.