AutoLMISeg: A Fully Automated Pipeline for 3D Lumbar Muscle Volumetry and Densitometry on CT - Technical Validation Against an Operator-Driven Workflow.

Opportunistic body composition analysis using single-slice 2D computed tomography scans is limited by anatomical variations and field-of-view constraints. Although 3D volumetry is more comprehensive, it is not widely adopted due to the time-consuming nature of manual segmentation. The objective is to technically validate AutoLMISeg (Automated Lumbar Muscle Index Segmentation), a fully automated "zero-click" deep learning pipeline for 3D L1-L5 lumbar muscle volumetry and densitometry. Sixty-nine CT scans were analyzed retrospectively. AutoLMISeg (GPU-accelerated) was compared to a human-in-the-loop reference standard (operator-driven segmentation). Accuracy was assessed using the Dice similarity coefficient (DSC), and volumetric agreement was evaluated using the concordance correlation coefficient (CCC) and Bland-Altman analysis. Processing times were compared using the Wilcoxon signed-rank test. AutoLMISeg showed excellent spatial agreement, with a mean DSC of 0.928 ± 0.019. Volumetric concordance was nearly perfect (CCC = 0.971), with a negligible mean bias of + 16.32 cm3 (approximately 2%). The automated pipeline significantly reduced processing time, completing the task in an average of 31.2 ± 7.2 s, compared to an average of 111.7 ± 25.9 s for the manual reference (p < 0.001). AutoLMISeg provides highly accurate and reliable 3D muscle biomarkers in less than a minute, making large-scale opportunistic screening in routine practice possible.