4795
Spinal Cord Grey Matter Segmentation using a Light-Weight Off-The-Shelf Neural Network1Physics and Astronomy, University of British Columbia, Vancouver, BC, Canada, 2International Collaboration on Repair Discoveries, Vancouver, BC, Canada, 3School of Biomedical Engineering, University of British Columbia, Vancouver, BC, Canada, 4Radiology, University of British Columbia, Vancouver, BC, Canada, 5Kinesiology, University of British Columbia, Vancouver, BC, Canada, 6Pathology and Laboratory Medicine, University of British Columbia, Vancouver, BC, Canada
Synopsis
Spinal cord grey matter segmentation is typically done manually. Automatic segmentation methods exist but are generally highly customized. We used an off-the-shelf neural network (LinkNet) to segment the grey matter in the spinal cord to assess the performance of a method with a generic architecture, which may be easier to replicate on different machine learning frameworks. Manual segmentation was used as training data. The performance of our trained network was compared to an automatic segmentation method in the Spinal Cord Toolbox (SCT), and both networks produced similar results, demonstrating the viability of the off-the-shelf approach.
Introduction
Recent advances in human spinal cord imaging have been able to resolve the butterfly-shaped grey matter (GM) within the spinal cord1. Although the contrast between GM and white matter (WM) is enough for human eyes to differentiate, automatic segmentation tools do not reach human-level performance. There are many tools for automatic segmentation of GM2 but they are not always accurate; therefore, manual segmentation remains the gold standard for this process. However, manual segmentation is laborious, tedious and the reproducibility is poor. Recent machine learning methods in spinal cord segmentation, like the Spinal Cord Toolbox (SCT)3, have been successfully used but have customized network architectures that may be difficult to replicate for non-experts in machine learning.
Our objective was to create a neural network that could segment spinal cord grey matter using an encoder-decoder machine learning method based on LinkNet4, which has a generic architecture shown to be quite flexible. To do this, we segmented GM on T2*-weighted spinal cord images manually and used this as training data to refine a neural network already pre-trained on brain white matter segmentation. We hypothesized that a carefully selected off-the-shelf network may approach the accuracy of a highly customized method.
Methods
Segmentation Algorithm: An autoencoder convolutional neural network based on the LinkNet architecture was refined and adapted for spinal cord segmentation. Binary cross-entropy was used to calculate loss and trained with Adam optimizer5. Training ground truths were obtained with manual segmentation drawn using FSL6 (workflow is depicted in Figure 1). Images for neural network inputs were upscaled because the neural network initially performed poorly for the very small segmentation area in the original image.
Data and Analysis Pipeline: 3T MRI (multi-echo fast gradient echo (mFFE), 5-echo, TE/TR=8.2/815ms, 16 slices, resolution=0.3x0.3x2.5mm3) was collected using a phased array spine coil (Philips Achieva). The stack was centered between C2 and C3 for all scans7. The network was trained on augmented data from 7 healthy subjects. Data augmentation produced additional samples with rotations from 0° to 180°, translations up to 0.3 times the image width and height, and upscaling up to 0.3 times. 200 randomly generated augmentations were trained for each epoch; the model was trained for 12 epochs. The network was tested on 1 new subject obtained from the same mFFE sequence. Probabilistic output predictions were binarized at a threshold of 0.5 for comparison with manual segmentation and SCT segmentation. Dice score, a measure of similarity between data sets, was calculated between ground truths and network predictions for accuracy. The Dice score produced by the network was compared to that produced by the SCT. The SCT segmentation algorithm (SCT version: 3.2.0) used for comparison was the multi-atlas based method8 with vertebral levels and SCT neural network method.
Results
Neural network segmentation results are shown in Figure 2. Dice score for our neural network with manual segmentation was 0.6332 vs. 0.6281 for SCT neural network segmentation and 0.3980 for SCT multi-atlas based method with manual segmentation. Computational time for all three segmentation methods was similar.
Discussion and Conclusion
Our neural network segmentation demonstrated a similar dice score to SCT’s neural network but higher score against SCT’s multi-atlas based method. Our neural network segmentation was smoother on the edges of the segmentation mask on every slice compared to the ground truths, likely due to the manual segmentation mask being upscaled. Our neural network segmented regions around the grey commissure poorly, but the majority of grey matter was segmented well. Another machine learning segmentation method, Deepseg2, also had challenges segmenting the grey commissure; this may be due to the model not being optimally tuned or trained on enough training examples to learn to distinguish between the low image contrast and find the small structure. Comparing the neural network segmentation with SCT segmentation, the dorsal and lateral horn regions were smaller in our neural network segmentation. Limitations of our current neural network include the small size of training examples. Furthermore, the model may not be robust for segmentation on images generated by other sequences or even different scanners because the model was only trained on one type of image. Nevertheless, our approach, which is trained on very few examples, demonstrates that a light-weight off-the-shelf network can have similar performance to a highly customized one. Moreover, this study adds further evidence that neural network approaches are a promising alternative to atlas-based methods for segmentation. The SCT approach may still perform better in other scenarios, but the LinkNet may be more easily replicated across computing platforms.Acknowledgements
We sincerely thank the study participants and MRI technologists at our centre. Funding support was provided by a seed grant from the International Collaboration On Repair Discoveries and the Natural Sciences and Engineering Research Council Undergraduate Student Research Award.References
1. Taso M, Le Troter A, Sdika M, et al. Construction of an in vivo human spinal cord atlas based on high-resolution MR images at cervical and thoracic levels: Preliminary results. MAGMA. 2014;27(3):257-267. 2. Prados F, Ashburner J, Blaiotta C, et al. Spinal cord grey matter segmentation challenge. Neuroimage. 2017;152:312-329.
3. Leener BD, Lévy S, Dupont SM, et al. SCT: Spinal Cord Toolbox, an open-source software for processing spinal cord MRI data. NeuroImage. 2017;145A:24-43.
4. Chaurasia A, and Culurciello E. LinkNet: Exploiting encoder representations for efficient semantic segmentation. In 2017 IEEE Visual Communications and Image Processing (VCIP). 2017.
5. Kingma DP, and Adam BJ. A method for stochastic optimization. arXiv preprint arXiv:1412.6980. 2014.
6. Jenkinson M, Beckmann CF, Behrens TE, et al. FSL. NeuroImage. 2012;62:782-790.
7. Ljungberg E, Vavasour I, Tam R, et al. Rapid myelin water imaging in human cervical spinal cord. Proc. Intl. Soc. Mag. Reson. Med.
8. Asman AJ, Bryan FW, Smith SA, et al. Groupwise multi-atlas segmentation of the spinal cord’s internal structure. Medical Image Analysis. 2014;18(3):460-471.
Figures
