University of Houston

The quantification and comparison of gene expression data across images plays a key role in understanding the functional network of various genes. To enable studying of relationships between a large number of images, automated methods that can segment gene expression images into distinct anatomical regions/sub-regions are needed. Automated segmentation of mouse brain gene expression images is a challenging problem mainly due to the complexity of gene expression appearance: (i) the lack of visible edge cue for the anatomical regions, (ii) the inhomogeneous of intensity pattern inside each anatomical region, and (iii) the variation of intensity pattern of the same region across images. Therefore, the use of geometric priors and appearance cues can potentially help in accurate segmentation of gene expression images.

The goal of this research is to develop segmentation methods that incorporate shape priors and appearance cues and apply them to gene expression image data. The specific objectives are: (i) to incorporate appearance cues from selected model points into statistical shape models so they can be applied on images that have complex appearance (e.g., gene expression images), (ii) to improve the representation ability of statistical shape models to represent a larger range of shapes, (iii) to incorporate our methods into fitting a subdivision mesh atlas, and (iv) to evaluate our methods for segmentation of gene expression images.