Automatic Paraspinal Muscle Segmentation in Patients with Lumbar Pathology Using Deep Convolutional Neural Network
Recent evidence suggests an association between low back pain (LBP) and changes in lumbar paraspinal muscle morphology and composition (i.e., fatty infiltration). Quantitative measurements of muscle cross-sectional areas (CSAs) from MRI scans are commonly used to examine the relationship between paraspinal muscle characters and different lumbar conditions. The current investigation primarily uses manual segmentation that is time-consuming, laborious, and can be inconsistent. However, no automatic MRI segmentation algorithms exist for pathological data, likely due to the complex paraspinal muscle anatomy and high variability in muscle composition among patients. We employed deep convolutional neural networks using U-Net+CRF-RNN with multi-data training to automatically segment paraspinal muscles from T2-weighted MRI axial slices at the L4-L5 and L5-S1 spinal levels and achieved averaged Dice score of 93.9\(\%\) and mean boundary distance of 1 mm. We also demonstrate the application using the segmentation results to reveal tissue characteristics of the muscles in relation to age and sex.
KeywordsSegmentation Deep learning Lumbar pathologies MRI
This work was supported by CIHR, CFI, NSERC and BrainsCAN, as well as the Seventh Framework Programme (Health-2007-2013, grant agreement NO: 201626: GENODISC) and Canada Reearch Chairs program. We acknowledge the support of NVIDIA Corporation and thank Dr. Yingli Lu for his help.
- 4.Wei, Y., Xu, B., Tao, X., Qu, J.: Paraspinal muscle segmentation in CT images using a single atlas. In: Proceedings of the PIC, pp. 211–215. IEEE (2015)Google Scholar
- 5.Ronneberger, O., Fischer, P., Brox, T.: U-Net: convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) MICCAI 2015. LNCS, vol. 9351, pp. 234–241. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-24574-4_28CrossRefGoogle Scholar
- 6.Zheng, S., Jayasumana, S., Romera-Paredes, B., et al.: Conditional random fields as recurrent neural networks. In: ICCV, pp. 1529–1537 (2015)Google Scholar
- 9.Isensee, F., Petersen, J., Klein, A., et al.: nnU-Net: self-adapting framework for u-Net-based medical image segmentation. arXiv preprint arXiv:1809.10486 (2018)