Atlas-Free Cervical Spinal Cord Segmentation on Midsagittal T2-Weighted Magnetic Resonance Images

Abstract

An automatic atlas-free method for segmenting the cervical spinal cord on midsagittal T2-weighted magnetic resonance images (MRI) is presented. Pertinent anatomical knowledge is transformed into constraints employed at different stages of the algorithm. After picking up the midsagittal image, the spinal cord is detected using expectation maximization and dynamic programming (DP). Using DP, the anterior and posterior edges of the spinal canal and the vertebral column are detected. The vertebral bodies and the intervertebral disks are then segmented using region growing. Then, the anterior and posterior edges of the spinal cord are detected using median filtering followed by DP. We applied this method to 79 noncontrast MRI studies over a 3-month period. The spinal cords were detected in all cases, and the vertebral bodies were successfully labeled in 67 (85%) of them. Our algorithm had very good performance. Compared to manual segmentation results, the Jaccard indices ranged from 0.937 to 1, with a mean of 0.980 ± 0.014. The Hausdorff distances between the automatically detected and manually delineated anterior and posterior spinal cord edges were both 1.0 ± 0.5 mm. Used alone or in combination, our method lays a foundation for computer-aided diagnosis of spinal diseases, particularly cervical spondylotic myelopathy.

Figure 1

The flowchart of our algorithm.

Figure 2

A midsagittal T2-weighted MR image (left) and our schematic drawing showing the spinal cord and its surrounding structures (right). The ligaments, including the anterior longitudinal ligament (ALL), the posterior longitudinal ligament (PLL), and the ligamentum flavum (LF), are deliberately thinned, and the internal architectures of intervertebral disks are neglected.

Figure 3

An example of classifying pixels on the histogram (left) after fitting with four Gaussian distributions (right). Horizontal dotted lines denote relative frequency.

Figure 4

Spinal cord detection using dynamic programming. Near the black line indicating the best path containing the largest number of isointense pixels, the white line denoting the best path traversing the region having the most homogeneous signal intensity is detected.

Figure 5

Detecting the edges of different structures using dynamic programming. From left to right: B ⁵, B ⁶, B ³, B ⁷, B ⁸, and B ⁴. The horizontal lines denote the superior and inferior edges of the spinal canal. The original MR image is shown in the left part of Figure 7.

Figure 6

Vertebral body and intervertebral disk detection and labeling in a midsagittal MR image.

Figure 7

Segmentation results in nearly normal and severely degenerated cervical spines. The left half of each image is the original image containing the cord (gray) and its surrounding cerebrospinal fluid region (white). The right half of each image is the segmentation result. Erroneously classified pixels are shown in white. Components of the epidural space are also shown in some nonstenotic areas, but they are clinically irrelevant.