Abstract
Multiple Sclerosis (MS) is a neurodegenerative disease that is associated with brain tissue damage primarily observed as white matter abnormalities such as lesions. We present a novel, fully automatic segmentation method for MS lesions in brain MRI that combines outlier detection and region partitioning. The method is based on an atlas of healthy subjects and detects lesions as outliers, without requiring the use of training data with segmented lesions. In order to segment lesions as spatially coherent objects and avoid spurious lesion detection, we perform classification on regions (connected groups of voxels) instead of individual voxels. Each voxel location is assigned to a region that would maximize overall relative entropy or Kullback-Leibler divergence between neighboring regions. Our proposed method is fully automatic and does not require manual selection or outlining of specific brain regions. The method can also be adapted to MR images obtained from different scanners and scanning parameters as it requires no training.
Keywords
Source Code and Data
No source code files available for this publication.
Reviews
Simon Warfield
Friday 25 July 2008
This paper describes a fully automated algorithm for the segmentation of brain MRI of patients with MS lesions.
MS lesions are detected by identifying regions which are different in the patient scan from a brain atlas.
A robust estimator is utilized to prevent outliers from influencing the estimation of parameters of a Gaussian mixture model likelihood function.
The results indicate the method has high specificity but lower sensitivity, indicating a conservative lesion segmentation outcome.
Martin Styner
Monday 28 July 2008
This paper presents an MS lesion segmentation approach via outlier detection from an atlas based tissue classification. The classification procedure assign voxels to regions, which in turn are classified in tissue classes, MS lesions detected as outliers and corrected for intensity inhomogeneity within Expectation Maximization loop.
The results clearly indicate the conservative nature of the current approach, sensitivity is quite low, while specificty and false-positive rate are great.
It is not clear to what sense the training cases were used, as the approach seems to be based on an existing prior atlas. Was it used to fine-tune any parameters?
Overall well written paper.