Automatic segmentation of volumetric images with applications to medical diagnosis.

摘要

To facilitate medical diagnosis, image segmentation is the key prerequisite toward quantifying the shape and volume of different tissues, which are then utilized for feature analysis and visualization. In this dissertation, we present a novel mixture-based image segmentation algorithm. The algorithm utilizes a hidden Markov random field Gibbs model to integrate local spatial information into the Expectation Maximization (EM) model-fitting algorithm. Compared with the conventional finite mixture model, it considers neighborhood information, thus overcoming the non-uniformity and noise problems for medical images. Given three-dimensional (3D) images, the algorithm models the likelihood of a realization as a finite multivariate function. Each voxel is classified by a mixture-based maximum a posterior segmentation, indicating its probabilities belonging to each class. In other words, each voxel is labeled as a mixel with tissue percentages. Because the EM algorithm converges locally, the choice of initial conditions is important. We propose a fast self-adaptive vector quantization approach, which has a unique property of independence of initial values. It utilizes the principal component analysis on local feature vectors of each voxel and performs classification according to the nearest neighbor rule. Through that, a preliminary segmentation result and initial parameter estimation are obtained. The presented segmentation scheme requires no prior knowledge of the data except a maximum number of distinct groups for classification, which can be set based on anatomical knowledge.; The above mixture-based segmentation algorithm has been developed as an automatic approach to extract colon lumen through colonic material tagging for virtual colonoscopy. As a result of this electronic colon cleansing, routine physical bowel cleansing prior to virtual colonoscopy may not be necessary. We further demonstrated a mixture-based feature analysis scheme in achieving computer-aided diagnosis (CAD) and automatic polyp detection. By extending the mixture-based algorithm to multispectral MR brain images, we successfully developed a mixture-based scheme for quantitative analysis of Multiple Sclerosis. We also integrated the noise analysis and point spread function into our framework. Validation on accuracy and repeatability as well as correlation studies with clinical findings demonstrated the feasibility of our framework as a clinically accepted tool for quantitative analysis of MS.