RESUMO
Background: The reported level of lumbar paraspinal intramuscular fat (IMF) in people with low back pain (LBP) varies considerably across studies using conventional T1- and T2-weighted magnetic resonance imaging (MRI) sequences. This may be due to the different thresholding models employed to quantify IMF. In this study we investigated the accuracy and reliability of established (two-component) and novel (three-component) thresholding models to measure lumbar paraspinal IMF from T2-weighted MRI. Methods: In this cross-sectional study, we included MRI scans from 30 people with LBP (50% female; mean (SD) age: 46.3 (15.0) years). Gaussian mixture modelling (GMM) and K-means clustering were used to quantify IMF bilaterally from the lumbar multifidus, erector spinae, and psoas major using two and three-component thresholding approaches (GMM2C; K-means2C; GMM3C; and K-means3C). Dixon fat-water MRI was used as the reference for IMF. Accuracy was measured using Bland-Altman analyses, and reliability was measured using ICC3,1. The mean absolute error between thresholding models was compared using repeated-measures ANOVA and post-hoc paired sample t-tests (α = 0.05). Results: We found poor reliability for K-means2C (ICC3,1 ≤ 0.38), moderate to good reliability for K-means3C (ICC3,1 ≥ 0.68), moderate reliability for GMM2C (ICC3,1 ≥ 0.63) and good reliability for GMM3C (ICC3,1 ≥ 0.77). The GMM (p < .001) and three-component models (p < .001) had smaller mean absolute errors than K-means and two-component models, respectively. None of the investigated models adequately quantified IMF for psoas major (ICC3,1 ≤ 0.01). Conclusions: The performance of automated thresholding models is strongly dependent on the choice of algorithms, number of components, and muscle assessed. Compared to Dixon MRI, the GMM performed better than K-means and three-component performed better than two-component models for quantifying lumbar multifidus and erector spinae IMF. None of the investigated models accurately quantified IMF for psoas major. Future research is needed to investigate the performance of thresholding models in a more heterogeneous clinical dataset and across different sites and vendors.
RESUMO
The size, shape, and composition of paraspinal muscles have been widely reported in disorders of the cervical and lumbar spine. Measures of size, shape, and composition have required time-consuming and rater-dependent manual segmentation techniques. Convolutional neural networks (CNNs) provide alternate timesaving, state-of-the-art performance measures, which could realise clinical translation. Here we trained a CNN for the automatic segmentation of lumbar paraspinal muscles and determined the impact of CNN architecture and training choices on segmentation performance. T2-weighted MRI axial images from 76 participants (46 female; age (SD): 45.6 (12.8) years) with low back pain were used to train CNN models to segment the multifidus, erector spinae, and psoas major muscles (left and right segmented separately). Using cross-validation, we compared 2D and 3D CNNs with and without data augmentation. Segmentation accuracy was compared between the models using the Sørensen-Dice index as the primary outcome measure. The effect of increasing network depth on segmentation accuracy was also investigated. Each model showed high segmentation accuracy (Sørensen-Dice index ≥ 0.885) and excellent reliability (ICC2,1 ≥ 0.941). Overall, across all muscles, 2D models performed better than 3D models (p = 0.012), and training without data augmentation outperformed training with data augmentation (p < 0.001). The 2D model trained without data augmentation demonstrated the highest average segmentation accuracy. Increasing network depth did not improve accuracy (p = 0.771). All trained CNN models demonstrated high accuracy and excellent reliability for segmenting lumbar paraspinal muscles. CNNs can be used to efficiently and accurately extract measures of paraspinal muscle health from MRI.
