Segmenting degenerated lumbar intervertebral discs from MR images

Michopoulou, Sofia, Costaridou, Lena, Panagiotopoulos, Elias, Speller, Robert and Todd-Pokropek, Andrew (2009) Segmenting degenerated lumbar intervertebral discs from MR images. In, 2008 IEEE Nuclear Science Symposium Conference Record. (IEEE Nuclear Science Symposium Conference Record) IEEE, pp. 4536-4539. (doi:10.1109/NSSMIC.2008.4774298).

Abstract

Magnetic Resonance Imaging is the modality of reference for diagnosing intervertebral disc degeneration, a condition related to chronic back pain. Segmentation of intervertebral discs is a prerequisite for computer aided diagnosis, while it could also serve in computer based surgery planning. A small number of studies report on disc segmentation methods applied on normal discs, while segmentation of degenerated discs remains an open issue. In the present study, a testing sample of 26 normal and 49 degenerated discs from T2-weighted midsaggital MR images of the lumbar spine was utilized to investigate the performance of two different segmentation methods. The first, which is based on the Fuzzy C-Means (FCM) algorithm utilizing three tissue classes (disc, bone and cerebrospinal fluid), suffered from severe leakage of disc border due to overlapping grey-level values between disc and surrounding tissues. To overcome this problem a combined method was developed, utilizing a probabilistic atlas of the disc along with the FCM algorithm. The probabilistic atlas was designed on the basis of an additional sample of 40 manually segmented normal intervertebral discs and was rigidly registered to images of the testing sample by means of a landmark based registration technique. The combined method resulted in reducing border leakage and achieved statistically significantly improved performance in comparison to the FCM method for all metrics tested (p«O.Ol). Specifically, sample overlap accuracies (mean standard deviation) of the combined method were O.81±O.06 for normal and O.77±O.07 for degenerated discs, and of the FCM method O.059±O.19 and O.63±O.15 respectively. Sample root mean square border distances of the combined method were J.53 and J.91 pixels for normal and degenerated discs. Concluding, incorporation of prior anatomical knowledge to the FCM method resulted in significantly nhanced performance, demonstrating increased potential in degenerated disc segmentation. © 2008 IEEE.

This record has no associated files available for download.

Contributors

Download statistics