|M.Sc Student||Greenblum Ayala|
|Subject||Statistical Learning Methods for Segmentation Problems in|
the Model Organism C. elegans
|Department||Department of Biomedical Engineering||Supervisor||ASSOCIATE PROF. Josue Sznitman|
|Full Thesis text|
Over the past half century or so, the nematode Caenorhabditis elegans has established itself as a ubiquitous model organism to study fundamental questions in biology. In particular, C. elegans is increasingly used for research pertaining amongst other to reverse genetic approaches, neuro-development and mechanosensation, as well as motility phenotyping. One ongoing bottleneck lies in reliably providing users with automatic segmentations from single images or sequences that feature complex and/or dynamic visual cues. Such computer vision tasks serve as a first and necessary step prior to extracting nematode phenotypes of interest. Here, we propose to tackle these automatic segmentation challenges by introducing a statistical learning method that revolves around the use of combined intensity- and texture-based image features integrated within a probabilistic framework. To begin, we use a number of filter kernels to compute image features and their ensuing statistics from a single or finite set of annotated images. From these statistics, probabilistic models of the appearance of the desired object and the surrounding background(s) are learned. Finally, when a new object must be extracted from a given image, we classify each image pixel using the learned probabilistic models. We illustrate the potential and versatility of our statistical learning methods on two state-of-the-art segmentation tasks: (i) segmenting whole-body nematodes in image sequences featuring diverse and complex motility environments, and (ii) segmenting imaged neuronal dendritic trees of C. elegans from confocal microscopy. We compare our method performances to existing segmentation methods and demonstrate that overall our probabilistic methods provide users with a tangible solution to tackle the growing needs of segmentation challenges in the C. elegans nematode, while unburdening the user from strenuous efforts.