Atlas based automated segmentation of the quadratus lumborum muscle using non-rigid registration on magnetic resonance images of the thoracolumbar region

Large volume asymmetries of the quadratus lumborum (QL) muscle, determined from time- and expertise-intensive manual segmentation of axial magnetic resonance (MR) images, have been associated with an increased risk of developing pars interarticularis stress lesions in the lumbar spine of cricket fast bowlers. The purpose of the present study was to develop an atlas-based automated segmentation procedure to determine QL volume from MR images. An MR database of axial lumbar spine images from 15 fast bowlers and 6 athletic control subjects was used to generate the atlas-based segmentation procedures. Initially, all images were preprocessed with a bias field correction algorithm and reverse diffusion interpolation algorithm followed by affine and non-rigid registration methods to generate firstly an average shape atlas (AVG), then based on propagation of manually segmented QL data, develop a probability atlas for automated QL segmentation to calculate muscle volume. The Dice similarity metric (DSC) was used to compare between the QL volume data from the manual and automated segmentation procedures. The mean DICE similarity coefficients between the manual and atlas-based automated segmentation values for the right and left QL muscle volumes were 0.75 (sd=0.1) and 0.76 (sd=0.09), respectively. These preliminary results for the automated segmentation of the QL are encouraging. Further development of the atlas-based segmentation procedures will involve incorporating hierarchical probability atlases for adjacent thoracolumbar muscles to improve the robustness and accuracy of the morphometric analyses obtained by this statistical shape modeling approach.