Research - masters
Bayesian multiple sclerosis lesion classification modeling regional and local spatial information
Supervisor: Professor Tal Arbel
Centre for Intelligent Machines, McGill University
In collaboration with: Dr. Douglas Arnold, Dr. Louis Collins and Simon Francis from the Montreal Neurological Institute
Project Summary: [Abstract]
Multiple sclerosis is an inflammatory demyelinating disease of the central nervous system. The white matter in the brain is mostly affected and is degenerated, resulting in lesions. Tracking lesion size and count helps assess disease progression. The 2 types of lesions (T1-les and T2-les) can be located on Magnetic Resonance Images (MRI) .
Manual identification of lesions by experts is time consuming and subjective. This is particularly true in the posterior fossa, where the variability between manual classifications is greater than in the cerebrum, and where important structures are densely packed. Furthermore, tissues and lesions have varying intensities depending on spatial location in the brain, rendering both manual and automatic classification difficult.
We present a Bayesian approach for the classification of MS lesions. Training the classifier consists of modeling the spatial variability in intensity distributions over the brain by segmenting the brain into distinct anatomical regions and building the likelihood distributions of each tissue class in each region based on multimodal magnetic resonance image (MRI) intensities. The following regions are used:
Local smoothness is ensured by incorporating neighboring voxel information in the prior probability via Markov random fields. Prior probability distributions and Markov random fields parameters are also learned using preclassified training data.
During the classification stage, the posterior probability distribution is used to determine voxel labels for regular tissue as well as T1-hypointense lesions and T2-hyperintense lesions by finding the maximum a posteriori probability (MAP), and to provide experts with a confidence level in the classification using the entropy measure. Some results are shown below in the cerebrum and the posterior fossa. Automatic and silver standard (SS) classification results, the latter derived from five expert labelings, in the cerebrum (top 2 rows) and posterior fossa (bottom 2 rows) are shown. The entropy and posterior probability spectral maps range from purple (0.01) to red (1). The classified labels are Bkg (dark blue), CSF (light blue), GM (green),WM (yellow),T2-les (red) andT1-les (white). The extracted lesions are shown in green.All lesions are successfuly identified.A few false positives are inside circle "A" in the cerebrum. “B” and “C” show successfully classified lesions.
The Markov random field spatially smoothes the classification. Click on the image below to see the classification process followed by iterations of the Markov random fields algorithm.
