Neural network segmentation of disc volume from magnetic resonance images and the effect of degeneration and spinal level

Abstract Background Magnetic resonance imaging (MRI) noninvasively quantifies disc structure but requires segmentation that is both time intensive and susceptible to human error. Recent advances in neural networks can improve on manual segmentation. The aim of this study was to establish a method for automatic slice‐wise segmentation of 3D disc volumes from subjects with a wide range of age and degrees of disc degeneration. A U‐Net convolutional neural network was trained to segment 3D T1‐weighted spine MRI. Methods Lumbar spine MRIs were acquired from 43 subjects (23–83 years old) and manually segmented. A U‐Net architecture was trained using the TensorFlow framework. Two rounds of model tuning were performed. The performance of the model was measured using a validation set that did not cross over from the training set. The model version with the best Dice similarity coefficient (DSC) was selected in each tuning round. After model development was complete and a final U‐Net model was selected, performance of this model was compared between disc levels and degeneration grades. Results Performance of the final model was equivalent to manual segmentation, with a mean DSC = 0.935 ± 0.014 for degeneration grades I–IV. Neither the manual segmentation nor the U‐Net model performed as well for grade V disc segmentation. Compared with the baseline model at the beginning of round 1, the best model had fewer filters/parameters (75%), was trained using only slices with at least one disc‐labeled pixel, applied contrast stretching to its input images, and used a greater dropout rate. Conclusion This study successfully trained a U‐Net model for automatic slice‐wise segmentation of 3D disc volumes from populations with a wide range of ages and disc degeneration. The final trained model is available to support scientific use.