A deep learning approach to spine segmentation using a feed-forward chain of pixel-wise convolutional networks

Segmenting vertebrae and disks in medical images using algorithmic methods promises to improve diagnostics and treatment selection for lower back pain. While most existing spinal segmentation methods are model-based, we propose a deep learning approach that uses a series of four pixel-wise segmentation networks. Each network segments MR images at a different scale. Results from one network in the chain are fed as input to the next network in the chain. By using both the original image and results from the previous network as input, each successive network produces an increasingly refined segmentation result. This approach improves vertebrae and disk segmentations results by 1.3% and 4.9%, respectively, when compared to a state-of-the-art deep-learning segmentation method-the U-net. It also provides additional benefits over U-net architectures by simplifying simultaneous multi-class segmentation, streamlining the overall use of the network, and removing the need for post-processing to filter erroneous classifications from the output.