Tissue segmentation is performed on multispectral image data via an EM (expectation maximization) maximized hidden Markov random field (HMRF) model with an outlier detection and de-weighting function. Tissue classes are parametrically modeled as multivariate Gaussian distributions.


An HMRF-EM algorithm is applied to source MRI (possibly multispectral) and partial volume is estimated at tissue boundaries. Voxel-wise tissue type maps are created. For brain extraction and tissue segmentation, SIENA/X cross-sectional and longitudinal brain atrophy analyses methods are used. The Brain Extraction Tool (BET) removes all nonbrain, noncerebrospinal fluid (non-CSF) tissue from the image, and identified the outer surface of the skull. Brain and skull images are then used to perform a scaling-constrained registration to a standard brain and skull image set to determine a subject-specific normalization factor. The de-skulled image is then processed with an automated image segmentation tool. Compartment-specific absolute volumes are then quantified (taking partial volume estimates into account) and multiplied by the subject-specific normalization factor to obtain normalized tissue volume measures.


As atrophy measurements are becoming more accepted as primary and secondary endpoints of clinical studies, it is becoming increasingly important to determine and validate the best methodology for obtaining these measures. The BNAC has invested considerable time and effort in this area, researching many different semi-automated and automated global atrophy measurement techniques, as well as a number of regional techniques. Quantitative analyses were performed using “gold-standard” expert segmentations to determine error rates of individual techniques.

The BNAC makes extensive use of the SIENA method for atrophy analysis, and is intimately familiar with all algorithm details and options.



While many methods have been proposed for the study of regionally specific brain areas, nearly all have been either extremely time-consuming or suffered from poor reproducibility. In order to address this problem, the Buffalo Neuroimaging Analysis Center (BNAC) has implemented a rapid and reliable technique for the parcellation of brain regions called Semi-Automated Brain Region Extraction (SABRE). The approach is a melding of automated and manual techniques, using a small number of user-defined landmarks and regions of interest to automatically parcellate the brain into 26 distinct regions. These regions are then subdivided into grey matter, white matter, and CSF compartments to provide a very detailed regional analysis of brain atrophy.

Further, the NeuroSTREAM analysis tool has been substantially updated with a multi-atlas algorithm to provide more accurate and robust atrophy measures. To build a comprehensive normative database, it has been run on over 14,000 baseline and follow-up MRI scans from 2,140 people with MS, 200 people with CIS and 381 healthy individuals. This normative data is in the process of being incorporated into an interactive, web-based predictive modeling system. In the meantime, the NeuroSTREAM method has also been applied to a prospectively collected real-world multi-center study of clinical routine imaging in MS (MS-MRIUS). In this large, heterogeneous dataset, NeuroSTREAM was the only atrophy measure that produced significant and clinically-correlated results
Additionally, the BNAC also utilizes a variety of other region-specific parenchyma methods for both cross-sectional and longitudinal atrophy analyses. These include FreeSurfer and voxel-based morphometry (VBM) techniques. BNAC is also capable to perform measurement of linear regional measures including third ventricle width, bicaudate ratio, lateral ventricle volume, and callosal atrophy.