The goal of this study is to demonstrate the accuracy and reliability of a new method for automatic segmentation of spinal cord gray matter. Recent studies have shown that spinal cord gray matter metrics are useful in the diagnosis and tracking of certain diseases, such as multiple sclerosis¹. However, these metrics are often difficult to obtain since spinal cord gray matter segmentation is a time-consuming process that must be manually performed by a trained expert. Automatic gray matter segmentation in the spinal cord has been largely unsuccessful, due to the poor contrast and low resolution of spinal cord MR images. With the use of phase sensitive inversion recovery (PSIR) imaging, we are now able to acquire better quality images that allow for the use of an active contour segmentation algorithm².

2D PSIR images at 3T of the C2-C3 level of the spinal cord were acquired in 40 healthy controls (aged 46±14; 24 women, 16 men) on a Siemens Skyra scanner. A template was created with the data from 20 subjects and the data of 12 subjects was used to test the automatic segmentation procedure. 8 additional subjects were scanned twice with repositioning to investigate the consistency of the automatic segmentation. All images were interpolated by a factor of 10 to aid in segmentation.

In all subjects, the cord was segmented automatically using the software JIM³ and the gray matter was manually segmented by an experienced neurologist. Templates were generated for both cord shape and spinal cord gray matter shape using affine and nonlinear registration with the masks of 20 subjects.

In the remaining subjects, the cord template was registered to the subject cord mask and the same transformation, was applied to the gray matter template as well. Distance maps were used to aid in accurate registration of the binary masks. This step gives an initial guess for the contour defining the gray matter segmentation (Figure 1).

Finally, the spinal cord gray matter was segmented automatically with a morphological geodesic active contours algorithm⁴. The algorithm is based on a Python implementation that is publicly available on GitHub⁵. This algorithm deforms the initial contour provided by the user in a method driven by three image forces: a smoothing force that controls the smoothness of the contour, a balloon force that inflates or deflates the contour in areas where information is lacking, and an image attraction force, which draws the contour to the maximum gradient areas in the image.

The images from the 12 initial subjects were manually segmented by the same experienced neurologist 3 times. Dice coefficients were computed to compare the automatic segmentations with each of the manual segmentations and the average Dice coefficient was calculated. Additionally, the percent change (the difference divided by the mean) of the areas was computed between the automatic segmentations from the 8 subjects that were scanned twice.

From the 20 automatic segmentations, there were 3 obvious errors that were eliminated when calculating statistics. (Figure 2)

When comparing the automatic segmentations of gray matter with the manual segmentations, the Dice coefficients ranged from .82 to .93 and the average was .88 with a standard deviation of .02. The automatic segmentation areas were consistently slightly lower than the areas from the manual segmentations. (Figures 3 and 4)

For the 8 subjects that were scanned twice, the percent changes ranged from 1% to 7%, with an average of 3% and a standard deviation of 2%. (Figure 5)

This study presents an accurate and robust automatic segmentation technique for segmenting gray matter areas in the spinal cord. An automatic segmentation method will help to further evaluate valuable spinal cord gray matter metrics in larger patient cohorts as markers of disease progression and predictors of disability. A reliable segmentation method also sets up the basis to employ gray matter metrics in multi-center settings and clinical trials. Since the algorithm is based on finding the maximum intensity gradient, it depends only on relative intensity differences of the image. In theory, this algorithm could be applied to different types of images, provided the level of contrast is sufficient. While this method works well for healthy controls, it may behave differently in patients due to the presence of significant lesions or morphological abnormalities. However, with further tuning of the parameters, we can change the balance for how much the algorithm depends on the prior shape information versus the intensity information in the image, which will help to provide more accurate assessments for the gray matter area in difficult cases.