The overall goal of this project is the integration of advanced image-processing, data-analysis, and data-management techniques into a brain-image database (BRAID). The integration of these components has greatly aided our collaborators'management and analysis of image-based clinical trials (IBCTs) for the elucidation of structure- function associations in the human brain. In the previous cycle, we extended our segmentation algorithm to incorporate more complex spatial and multispectral signal-intensity information;extended BRAID and its image-processing pipeline to accommodate acute-stroke data;implemented our Bayesian approach to morphometry;constructed a probabilistic atlas of acute-stroke lesions;and re-implemented BRAID using open-source components, improving the user interface and performance in the process. Although these results, in conjunction with BRAID's visualization and other statistical tools, have enabled our collaborators and us to contribute to the peer-reviewed clinical and engineering literature, our experience has demonstrated the need for extensions to BRAID. First, our current data-mining approaches are designed to generate Bayesian networks that model a structure-function relationship;that is, these models are descriptive. Increasingly, clinical neuroscientists are attempting to construct predictive models, which they could apply to new subjects, or even patients, to predict group membership based on image data, or vice versa. Such predictive models could guide early therapy for Alzheimer disease or stroke, among other diseases, and thus have immense potential. Second, although our Bayesian morphometry algorithm has shown great promise for mining cross-sectional data, many morphometry studies center on degenerative diseases, and therefore require longitudinal analysis, which takes into account temporal changes as the disease progresses, or responds to therapy. Third, given the rapid advances in modality development, and based on our experience developing successful approaches to mining voxel-wise lesion and volumetric data, we believe that generalizing our Bayesian data-mining approach to accommodate arbitrary statistical and spatial models would result in a widely applicable structure-function analysis library. Fourth, the rapid expansion of modalities available to clinical neuroscientists has also led to the acquisition of large, multimodality image sets, including diffusion-tensor data, functional MR images, lesion-deficit data, and voxel-wise volumetry. Such data sets are becoming the standard for determining the pathophysiologic mechanisms of neurodegenerative disease. For example, several Alzheimer research groups are collecting fMR and structural MR data, in order to determine whether regional interactions between morphology and activation predict the development of Alzheimer disease better than either modality alone. Such researchers would benefit greatly from software that would allow them to analyze multispectral data. Finally, although our open-source reimplementation of BRAID has already greatly extended the range of statistical models available on-line to our collaborators, we could further extend BRAID's utility by augmenting this statistics datablade to support SQL- based access to all of the data-mining algorithms listed above. Toward these ends, we propose five specific aims to further extend BRAID's functionality: development of a Bayesian method for generating robust, scalable classifiers for brain-image data;implementation of a Bayesian longitudinal morphometry algorithm;development of a class library for modality-independent structure-function data mining;develop data-mining support for multimodality image data;and augmentation of BRAID's statistics datablade to integrate these data-mining algorithms, and to provide on-line access to these tools. We will test these image-analysis, segmentation, and statistical extensions to BRAID, using data from 6 IBCTs.
The goal of this project is to develop software that simplifies the analysis of brain-image data. We will make all of these software components available from a web-accessible image database. We expect that, as we implement more components of this software suite, neuroscientists will find it much easier to perform complex multivariate analyses of their data relating structure and function of the human brain.
Showing the most recent 10 out of 30 publications
