This project will address the segmentation of MRI brain volumes into labelled regions such that each cerebral voxel has one or more anatomical labels, each with an associated label probability index. Software tools will be developed to allow manual labelling of voxels in two types of image volumes (a) cryomacrotome data from human cadaver brains and (b) MRI volumes from normal subjects. These tools will be applied to a subset of both data types to generate reference data for subsequent automated procedures. Second, we will continue existing projects for computerized segmentation with a minimum of investigator involvement. The overall segmentation program, manual and automated, will develop at successively finer scales of anatomic detail by identifying major zones, e.g., frontal lobe, toward specific structures, e.g., thalamus. Finer labelling will be generated manually. Each new level of anatomical scale will require the development of more refined algorithms and more detailed models. At the completion of each cycle, we will conduct a re- evaluation of the accuracy and precision of the computerized techniques as compared with manually-derived results from a subset of the acquired data. Phantom data will be generated to assess the robustness of the results against injected image noise, variations in contrast level between tissues and slice thickness. The project has the following specific aims: a) Establish a hierarchical classification of neuroanatomy from gross to fine structure. b) Collect axial high-resolution, high- contrast MRI data from a homogeneous population. Combine these data with UCLA cryomacrotome data in stereotaxic space by application of 3-D image warping software. c) Use existing and develop further tools for computer-aided manual labelling of every voxel in a volume, using the above hierarchy for regional labelling. Apply these techniques to a subset of both cryomacrotome data and MRI data to generate labelled volumes. The labelled data will be used both for comparison with and as a starting model for automated techniques. d) Extend our existing automated segmentation algorithms in the areas of gross tissue classification, model-based regional parcellation and cortical labelling. e) Conduct validation experiments for tissue classification algorithms with geometric and brain-like phantoms. f) Apply automated segmentation techniques to sub-sample of data volumes and conduct statistical comparisons with results with manual methods.
Showing the most recent 10 out of 36 publications