| dc.description.abstract | Microscopy has developed into a very powerful medium for studying the brain. The Knife-Edge Scanning Microscope (KESM), for example, is capable of imaging whole rat and mouse brains in three dimensions, and produces over 1.5 terabytes of images per brain. These data can reveal the structure and organization of the brain's internals including neurons and blood vessels. Neuron count and density strongly influence the behavior of an organism, and measuring their spatial distribution is key to a better understanding of the workings of the brain. This kind of analysis involves identifying neurons in large brain regions, for which fast automated detection methods are necessary. Most of the current automated cell detection techniques require complex preprocessing of images, use heuristics that are time consuming to develop, or do not generalize well to three dimensional data. In this thesis, I propose two methods based on random forests for detecting neuron bodies in the rat and mouse brain KESM data.
The proposed methods require a few hundred cell centers to be manually labeled. Random forests are trained to predict if a voxel is a cell center or not by using these labeled data and features derived from orthogonal image patches. They can then be used to locate cell centers in 3-D in other images, aided by a refinement step whose parameters are determined from the training data. Minimal manual input is required, and random forests provide a good combination of accuracy and speed. This is expected to enable fast counting and density measurements of neurons in brain regions. The detected cell centers should also be valuable as seeds for cell segmentation methods. | en |
