Introduction

Spinal Cord Injury (SCI) is a devastating injury affecting the central nervous system. Accurate detection and segmentation of (pre and post-operative) spinal cord tissue is of high importance in assessing primary spinal cord lesions and secondary tissue damage in spinal cord injury (SCI) and a prerequisite for longitudinal follow-up procedures. In multiple sclerosis (MS) patients, an automatic segmentation of the spinal cord and MS lesions has been published recently¹ using convolutional neural networks. The aim of this preliminary study is to assess the automatic level-wise spinal cord segmentation in spinal cord injury.

Methods

The dataset includes ten chronic spinal cord injury patients. Images were acquired on a 3T system (Philips Achieva, Best, The Netherlands) using a 3D T2-weighted sequence. Segmentation and data processing was done using IntelliSpace Discovery (Philips, Best, The Netherlands). Manual spinal level segmentation was performed by an experienced radiological technician (MA, 23 years of experience) and reviewed by a radiologist (MFB, 28 years of experience) until consensus was achieved. Automatic segmentation was done using two segmentation algorithms offered by the spinal cord toolbox²: PropSeg³ and DeepSeg¹. Pairwise comparison of the segmentation results of the manual and automatic results is done by calculating the Dice Similarity Coefficient (DSC)⁴. The manual segmentation is regarded as gold standard. Boxplots show the median and quartiles of the DSC of all spinal levels for the comparison of 1) manual vs PropSeg 2) manual vs DeepSeg and 3) DeepSeg vs PropSeg.

Results

The demographics of the study participants is shown in Fig.1. We included seven male and three female tetraplegic patients (age: median=57.3 years, range: 28.8-74.7, years since injury: median=16.4 years, range: 1.3-53.5 years), five with complete and five with incomplete tetraplegia. The algorithms provided segmentation results in 9/10 (propSeg) and 8/10 (deepSeg) patients. Fig. 2, 3 and 4 show the segmentation results in patient cases presenting the advantage and pitfalls of automatic segmentation application in tetraplegic spinal cord injury patients. Fig.5 shows the boxplot of the DSC for the three comparisons of all spinal levels where segmentation was possible. Not surprisingly, there is a "dip" of the DSC between C4 and C7, the areas affected by the injury.

Discussion

The main findings of this preliminary studies are that automatic segmentation results in spinal cord injury can evoke three scenarios: 1) Although lesion site affects the image quality, both algorithms results in meaningful labelling (Fig.2). 2) Both algorithms perform well in the unaffected levels, but struggle at the lesion site (Fig.3). 3) Automatic segmentation fails because the starting point is wrongly chosen by the algorithm leading to incorrect following levels (Fig.4). Future studies might include a larger number of subjects to train a convolutional neural network to better assess the injury site.

Conclusion

This study presents preliminary data of automatic spinal levels segmentation using the spinal cord toolbox in tetraplegic spinal cord injury patients. It is a promising tool for future applications and has great potential for longitudinal follow-up assessments.

Acknowledgements

The authors thank the Swiss Paraplegic Foundation for support and J.Cohen-Adad and his group developing the spinal cord toolbox for providing the segmentation routines.