Segmentation of the Quadratus Lumborum Muscle Using Statistical Shape Modeling

Purpose: To compare automated segmentation of the quadratus lumborum (QL) based on statistical shape modeling (SSM) with conventional manual processing of magnetic resonance (MR) images for segmentation of this paraspinal muscle. Materials and Methods: The automated SSM scheme for QL segmentation was developed using an MR database of 7 mm axial images of the lumbar region from 20 subjects (cricket fast bowlers and athletic controls). Specifically, a hierarchical 3D-SSM scheme for segmentation of the QL, and surrounding psoas major (PS) and erector spinae+multifidus (ES+MT) musculature, was implemented after image preprocessing (bias field correction, partial volume interpolation) followed by image registration procedures to develop average and probabilistic MR atlases for initializing and constraining the SSM segmentation of the QL. The automated and manual QL segmentations were compared using spatial overlap and average surface distance metrics. Results: The spatial overlap between the automated SSM and manual segmentations had a median Dice similarity metric of 0.87 (mean = 0.86, SD = 0.08) and mean average surface distance of 1.26 mm (SD = 0.61) and 1.32 mm (SD = 0.60) for the right and left QL muscles, respectively. Conclusion: The current SSM scheme represents a promising approach for future automated morphometric analyses of the QL and other paraspinal muscles from MR images.