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.