Content-based image analysis techniques and biological validation schemes for images depicting spatial gene expression patterns in developing embryos

Recent advances in high-throughput biological experiments have enabled the acquisition of a large number of images depicting spatial patterns of gene expression in developing embryos. This, in turn, has resulted in an unprecedented growth of gene expression images, available in both on-line journal articles and websites. Developmental biologists often perform a visual inspection of these images in order to make inferences relating to gene interaction and regulation, which is both laborious and time-consuming. As a remedy to this problem, this dissertation focuses on computational techniques based on image analysis of two-dimensional gene expression images, using the model organism, Drosophila melanogaster (fruit fly).; First, this research discusses a novel application of image processing techniques for annotating the imaging view and developmental stage of gene expression images. Results obtained from this research suggest that discriminatory features can be extracted from specific anatomical regions of the embryo, which is consistent with longstanding biological knowledge.; Second, this dissertation presents a novel adaptation of a dynamic expectation-maximization (EM) algorithm on gene expression images for obtaining a Gaussian mixture model (GMM) representation. The results show that this model leads to a better representation of gene expression staining patterns in an image. Also, the GMM representation thus obtained provides superior image matching compared to GMM representations obtained using classical EM algorithms, and to an existing technique based on the binary feature vector (BFV).; Third, this research demonstrates methods by which biological validation could be applied to image retrieval results obtained using gene expression image comparison techniques (BFV and GMM) in a high-throughput fashion. It discusses a novel adaptation of image retrieval performance measures (namely recall and normalized average rank) based on biological relevance. Further, this research proposes a novel measure of the degree of relationship between image similarity and biological connections of genes, and demonstrates that these measures can be successfully employed to identify sets of images that are both spatially similar and biologically related.; The results of this research are expected to provide a significant contribution to future work in the area of image-based analysis of gene expression in developing embryos.