Quality assessment of a semi-automated spinal disc volume segmentation method for use in lumbar herniated disc studies, following patients over five years. Therefore we demonstrate the reliability of an interpolation method which relies on a reduced number of delineated regions of interest (ROI) to reduce time and effort.

For a total of 21 patients T1 weighted anatomical MRI lumbar spine data is obtained at three different sites, using 1.5 T scanners (2x Philips Acheiva, 1x Siemens Sonata) with standard spine array coils. Therefor a 2D turbo spin echo sequence with a sagittal in-plane resolution of 1.2mm x 0.6mm, 25 slices with a thickness of 2.5mm plus 0.2mm slice gap and three averages is used. Image post-processing and interpolation of spinal disc volume assessment is done in MeVisLab (1). The spinal disc volume is calculated as follows and illustrated in Fig 1.:

1. Ground truth: The spinal disc is marked in every single sagittal slice (approx. 17-23 slices).

2. Optimized number of ROIs: Five sagittal, one transversal and one or two coronal ROIs are marked.

3. Retest: as in 2., independently marked

The ground truth volume of the total disc is the sum of single ROI volumes, considering the slice gap as well. For 2. and 3., total disc volume is interpolated by computing a 3D implicit function that describes a surface, and this surface is scanned using a recursive marching-cubes algorithm (2). The quality of the interpolation is verified optically to assure a sufficient approximation. Examples of the calculated volume masks are shown in Fig. 2.

Fig. 3 shows the correlation of the spinal disk volume of the ground truth with the interpolated spinal disk volume of the optimized number of ROIs. The coefficient of determination and slope are R²=0.96, s=1.01+/-0.05. As can be seen in Fig. 3 the interpolated volume is in general a bit bigger as the volume of the ground truth, leading to a constant offset. For test-retest evaluation, R² is 0.95, with a mean volume difference of 6%.

The disc volume can be robustly approximated by applying the interpolation algorithm on the reduced number of ROIs, as indicated by R² and a slope not significantly deviating from unity. Test and re-test values correlate with equally high R². The volume difference between test and retest reads is 6%, which needs to be regarded as uncertainty in longitudinal evaluations. The fact that the interpolated volume is a bit bigger as the ground truth is due to the fact that the outline of a ROI is rather considered to be in the volume than outside. But when comparing the change in disc volume for patients over time with the interpolation method this constant offset can be negligible.

As a resulting advantage of the semi-automated segmentation only one third of the ROIs are required, resulting in substantial time saving of 65%, especially when the method is applied to high number study data.