Multiple sclerosis lesion detection using constrained GMM and curve evolution
Journal of Biomedical Imaging
MR contrast synthesis for lesion segmentation
ISBI'10 Proceedings of the 2010 IEEE international conference on Biomedical imaging: from nano to Macro
| Hi-index | 0.01 |
A fully automatic Bayesian framework for multiple sclerosis (MS) lesion classification is presented, using posterior probability distributions and entropy values to classify normal and lesion tissue. Spatial variability in intensities of multimodal MR images over the brain is explicitly modeled by building region-specific multivariate likelihood distributions. Local smoothness is ensured by incorporating neighboring voxel tissue information using Markov Random fields. A probabilistic measure of confidence for the classification is then presented, which can also be used to assess disease burden. The method was tested on 10 patients with MS by comparing automatically classified lesions, with and without regional information, to manual classifications by five expert raters using volume count and overlap. Results improve with the incorporation of spatial information, and are comparable to manual classifications. This method also enables a more accurate classification in the posterior fossa, where no other method reports success.