Determining associations between the structure and the function of the human brain has been one of the principal goals of the human brain project. There are two, largely complementary, approaches to this research: activation studies, in which measurements are taken during stimulus processing or performance of a task, and lesion-deficit analysis, in which associations are determined among lesion locations and functional deficits. The widespread availability of noninvasive assessment of brain structure has enabled researchers to investigate neuroimaging correlates of normal aging, cerebrovascular disease, and other processes; we designate such studies as image-based clinical trials. Despite a rapid increase in the volume of available image and clinical data from image-based clinical trials, true automated support of these projects requires progress in five areas: image-data management, accurate image registration to a common standard, automatic delineation of lesions, development of functionally relevant atlases, and statistical analysis of spatial image data. The BRAID project supports work in each of these areas, with the ultimate goal being support of large-scale image-based clinical trials for determining structure-function associations in the human brain.
<p>
Image-data management
<p>
BRAID is implemented in Illustra, an object-relational DBMS. BRAID has three components: an image datablade, consisting of image-processing and visualization operators for preliminary examination of image data, and for presentation of results; a statistical datablade, consisting of statistical operators used in lesion-deficit analysis; and a morphologically factored image representation (MFIR). The image-processing and statistical-analysis operators are accessible from BRAID's structured query language (SQL) interface; thus, we can sum lesions across subsets of subjects, color the lesions and superimpose colored atlas structures, or statistically analyze the image data in conjunction with clinical variables, with one SQL statement, as shown in [1]. Our grant of software from Illustra has been extended by Informix (including upgrade to Universal Server?), after that company's acquisition of Illustra.
<p>
In implementing BRAID, we have focused on developing programming interfaces to routines for manipulating and analyzing spatial image data, and linking these routines to methods for statistical analysis of data from epidemiological studies. For example, since lesions and atlas structures are treated in the same way by BRAID, we can submit an SQL statement to BRAID that causes it to return a list of atlas structures intersecting a list of lesions. Because BRAID has a uniform interface for lesion-deficit analysis, the list of intersections is in a standard format that is accepted by the statistical operators for the chi-square or Fisher exact test. In designing BRAID, we have focused on the abstraction of what it means to analyze spatial image data in the context of a clinical trial. In doing so, we have developed a suite of specialized image-analysis and statistical operators, any one of which can be replaced without our having to reconfigure any other part of BRAID.
<p>
Just as image operations are modular within BRAID, so are analysis operators. To perform structure-to-function analysis, one can submit an SQL statement to BRAID that will return all subjects with lesions that intersect a particular structure; thresholding can be applied to the volume of intersection. With this subset of subjects, one can perform any standard SQL operation, such as obtaining the average of their scores on a particular clinical variable. Similarly, for function-to-structure analysis, one can submit an SQL statement to BRAID that will return all subjects with certain clinical characteristics. With this subset of subjects, one can display the distribution of lesions using BRAID's visualization operators.
<p>
To implement lesion-deficit analysis within BRAID, we developed statistical operators on image and clinical data, which work in conjunction with BRAID's image datablade. For example, using the intersection operator of our image datablade, we can determine the intersection of a set of lesions with a set of atlas structures, as specified in an SQL statement; a thresholding operator yields a list of lesioned atlas structures for each subject. The status of a given structure and a given clinical variable across subjects determines a contingency table. As reported in [1], we implemented chi-square analysis of contingency tables; BRAID generates and analyzes a contingency table for each atlas structure and clinical variable specified in an SQL statement, returning a table of results in HTML format.
<p>
A critical part of our abstraction of image-based clinical trials rests on our design of a morphologically factored image representation (MFIR) that separates spatial (i. e., displacement vector fields after registration) from signal-intensity properties of a brain image. This modularity has allowed us to analyze deformation fields for changes among subgroups of subjects in cerebral sulci [2] or the corpus callosum [3]. Initially, the MFIR was based on piecewise-linear registration for transforming a 3D brain image to Talairach space for visualization, morphological analysis or lesion-deficit analysis. Although the results of subsequent lesion-deficit analysis were encouraging [1], we share the widely held view that nonlinear registration is required for adequate registration.
<p>
see web site for more
|