Research On Statistical Tissue Classification Of Biomedical Images

Vast number of biomedical images are produced routinely everyday. Chinese Digital Human Project is an ongoing research plan in which two high-resolution and high-quality data sets including male and female have been acquired already. To the post-processing of the huge and diverse source of data sets, quantitative analysis and three dimensional reconstructions of normal or abnormal tissues and organs are essential. Image segmentation is the prerequisite of quantitative analysis and building high-quality geometric models.In general, the statistical based medical image segmentation method is also called statistical tissue classification. The tissue classification methods are different due to image modality, characteristics and quality etc. In this thesis, tissue classification is studied based on three different data sets: nerve, blood vessels and brain tissues. An improved K-means clustering algorithm was given which can be used to large size images.This dissertation proposed a method by which nerve micro bundles are extracted and classified automatically from the scanned images of the brachial plexus serial tissue sections. Weighted distance matrix based on the Euclidean distance of the centroid and the shortest distance of the nerve micro bundles was used in hierarchical clustering. A minimum squared error function was used to judge the number of clusters. Experiments showed that this classification method is feasible.This dissertation proposed a maximum intensity projection (MIP) preprocessing mask technique based finite mixture Gaussian model segmentation method for extracting blood vessels from brain magnetic resonance angiography (MRA) dataset. The voxels whose intensity is high in the dataset belong to blood vessels or subcutaneous fat, which may bias the adjustment of the blood vessels. MIP of the dataset in the Z axis direction was computed and segmented as a mask. The masked MRA dataset was segmented by a low threshold and the remanent voxels were modeled by two normal distributions. Expectation-Maximization (EM) algorithm was used to estimate the model parameters. The initial values of the parameters were setup close to the true value by full width half height (FWHM) and the area under the histogram curve, by which the computation iteration number is lessen, the computation is convergence and does not convergent to local maxima. The result showed that this method is feasible for vessel extraction from MRA dataset.This dissertation proposed random forest for segmentation of multi-channel images, which uses bootstrap resampling from the training samples to produce many sample set to train the same number of classification tree in which feature was randomly selected to construct random forests for multi-channel MR image segmentation. Experiments showed that this method has good segmentation performance.