Autism is a developmental disorder that severely disrupts the social, cognitive and communicative skills of an individual. A host of brain structures are probably involved in its pathophysiology, including the cerebellum, brain stem, cortex and basal ganglia. Recent Magnetic Resonance Imaging (MRI) studies have produced mixed results, perhaps because of sample size limitations, etiologic heterogeneity and difficulties associated with accurately quantifying the regions of interest. The accurate, robust and efficient segmentation of these neuroanatomical structures from MRIs is a difficult problem. A variety of segmentation methods have been attempted to quantitate those brain structures. Success, however, has been limited, mainly because a single technique is likely to be unsuccessful over a wide range of structures. To circumvent that, we propose an integrated approach that attempts to capitalize on both region homogeneity and boundary information allowing us to use the wider source of naturally available information. This combination of information sources is achieved using a rational form of decision making which besides being technically more general than the commonly used single objective optimization approach, is computationally far less burdensome as well. The resulting robust method not only has the ability to handle a wide collection of neuroanatomical structures, but also is capable of suitably addressing MR imaging artifacts such as noise, poor image contrast, partial volume effect and inhomogeneous gain field due to appropriate integration of the aforementioned information sources. Our initial effort would be to extend and adapt our method which presently works on two dimensional images to more accurately capture the subtleties of the present problem. Also, we plan to extend it to three dimensional images to be better able to capture the actual structures involved. Validation for accuracy and reproducibility would be performed using electronically-generated phantoms as well as post mortem data available from the NIH Visible Human project. Finally, we intend to compare the algorithm-generated results with those found by human expert tracing of the same data, using MRI studies from a separately funded NIH project studying high functioning autism, Asperger's syndrome and matched controls. The automated approach will also allow us to investigate additional structures and will illustrate the ability of our approach to reproducibly detect subtle changes in structure.
Showing the most recent 10 out of 36 publications
