Application of a Semiautomated Contour Segmentation Tool to Identify the Intervertebral Nucleus Pulposus in MR Images

Validation Dataset

To calculate the amount of segmentation error in the CS method, we generated a validation dataset containing a structure with a known area, and the result of the CS was compared with this area. The validation dataset was a computer-constructed image containing a 2D Gaussian structure with a height intensity equal to 500 and the SD equal to 3 in both directions. This structure was intended to mimic the T2WI of the NP because the signal intensity radially drops off from the center of the NP toward the AF. The base diameter of the Gaussian structure spanned 13 pixels, thus the base area was π × 6.5² = 132.66 pixels². This area will be referred to as Area_true. The image size was 49 × 49 pixels with background intensity equal to zero, with the Gaussian structure located in the bottom right corner of the image. The top left corner of the image was used to measure the average intensity of the background to calculate the image SNR as described below.

The base image was generated without the introduction of any noise; thus, it had a noise level equal to 0% (Fig 2 A). Other validation images were generated with noise added to a specific number of pixels chosen at random. For example, the intensity of 960 randomly chosen pixels of 2401 total pixels was randomly changed between 0 and 500, resulting in an image with 40% noise level (Fig 2B). The following noise levels were added to the base image: 0%, 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, and 80%, and 30 images of each noise level were generated, thus creating a total of 300 validation images. These images were used to assess the sensitivity and accuracy of the CS method.

• Download figure
• Open in new tab
• Download powerpoint

Fig 2.

An example of the computer-generated validation image with 0% noise (A) and 40% noise (B). The Gaussian structure is located in the bottom right corner of the image.

Rabbit Dataset

This dataset was composed of midsagittal T2WI ROI from a rabbit puncture model study approved by the Institutional Animal Care and Use Committee of the University of Pittsburgh, and all procedures were in compliance with guidelines of the Public Health Service Policy on Humane Care and Use of Laboratory Animals. A 3T magnet (Siemens, Erlangen, Germany) was used, and the parameters of the T2WI were as follows: TR = 3800 ms, TE = 114 ms, FOV = 184 × 232, number of sections = 20, section thickness = 0.6 mm, section gap = 0.66 mm. This study was similar to that reported previously by Sobajima et al.^6,8 In brief, 4 skeletally mature New Zealand white rabbits underwent the puncture procedure with a 16-ga needle in 3 lumbar IVDs, L2–3, L3–4, and L4–5, thus initiating a degeneration process. We did not operate on 3 additional rabbits; thus, they were used as nonoperative controls. T2WI was performed in all rabbits before the date of surgery and at 6 and 12 weeks postsurgery. Therefore, the 3 IVDs from 7 rabbits across 3 time points resulted in 63 images of lumbar IVDs. This dataset was used to evaluate the ability of the CS method to detect IDD.

Segmentation

The CS method was written in Matlab, Version 7.5 (MathWorks, Natick, Massachusetts) and was applied to both datasets. For the rabbit dataset, the method was applied to the midsagittal section, which was identified by an orthopedic resident. The user drew a box on the image surrounding each IVD of interest (L2–3, L3–4, and L4–5) and the adjacent vertebrae. The intensity isolines of the chosen section were calculated and displayed as a contour image. The contour lines of the image were separated by contour levels of 50. The user was prompted to identify the NP, our ROI, by selecting the isoline that corresponded to the largest closed area between the 2 adjacent vertebrae. Then as a visual confirmation, the image was displayed with the ROI boundary in red to ensure the proper location of the NP. The segmentation was performed twice with a 1-week gap. The same process was performed on the validation dataset, but the selected region was the Gaussian structure instead of the IVD.

Moreover, the NP of the rabbit MR images was manually segmented using ImageJ software (http://rsb.info.nih.gov/ij/) by 3 observers: 2 orthopedic residents and a student, who were trained to recognize the location and shape of the NP on an MR image. The order of the images was randomized for each observer, and they were also segmented twice, with a 1-week gap between each segmentation session. Additionally, the Gaussian structure of the validation dataset was manually segmented once by the student.

Outcome Measures

The total area of the ROI was calculated for both datasets. Additionally, the SNR of each image was calculated according to the following equation⁹:

To assess IDD, the MR imaging index previously established by Sobajima et al^6,8 was calculated for the rabbit NP: where I and A are the intensity and area of the jth pixel respectively, and N is the total number of pixels in the ROI. Also, a Student t test was used to statistically evaluate the differences between the control and punctured rabbits disks. Finally, an ICC was calculated to compare the performance of the CS method with the manual segmentation in addition to inter- and intraobserver variability.

As previously indicated the Area_true of the Gaussian structure in the validation dataset was 132.66 pixels². The error in the segmented area from the CS and manual method was calculated as the percentage difference between the area of the segmented ROI (Area_ROI) and Area_true as shown in equation 3:

Therefore the area error addresses the size difference between the segmented region of interest and the criterion standard represented here by Area_true.