The primary goal of this project is to enable identification and segmentation of specific structures of interest from imaging data (such as DTI MR brain images). These regions may be small and inconspicuous with poor contrast, therefore, direct application of classical unsupervised segmentation (designed to extract "salient" regions) is problematic. The alternative pursued here is a system to leverage expert-like high level advice within the image segmentation process: to do this, the underlying engine is endowed with global constraints encoding (a) effort already expended by the user in segmenting similar images in the past, as well as (b) aggregate knowledge from a cohort of similar images. The key algorithmic component is the design of mechanisms to translate such constraints (as best as possible) to a combinatorial framework so that the resultant models can be optimized efficiently for high resolution 3-D imaging data. This research produces the methodology and accompanying software for this important image analysis task.
The project has broad scientific impact. Wide distribution of code produced from this research can enable improvements in various computer vision and medical imaging problems where image segmentation is a key step. Additionally, the algorithms developed here have applications in other problems such as object recognition and image categorization. The project is also well suited to involve undergraduate and graduate students from a diverse spectrum of backgrounds in cutting edge inter-disciplinary computer vision and image processing research.