Histological image analysis and gland modelling for biopsy classification

The area of computer-aided diagnosis (CAD) has undergone tremendous growth in recent years. In CAD, the computer output is used as a second opinion for cancer diagnosis. Development of cancer is a multiphase process and mutation of genes is involved over the years. Cancer grows out of normal cells in the body and it usually occurs when growth of the cells in the body is out of control. This phenomenon changes the shape and structure of the tissue glands. In this thesis, we have developed three algorithms for classification of colon and prostate biopsy samples. First, we computed morphological and shape based parameters from hyperspectral images of colon samples and used linear and non-linear classifiers for the identification of cancerous regions. To investigate the importance of hyperspectral imagery in histopathology, we selected a single spectral band from its hyperspectral cube and performed an analysis based on texture of the images. Texture refers to an arrangement of the basic constituents of the material and it is represented by the interrelationships between the spatial arrangements of the image pixels. A novel feature selection method based on the quality of clustering is developed to discard redundant information. In the third algorithm, we used Bayesian inference for segmentation of glands in colon and prostate biopsy samples. In this approach, glands in a tissue are represented by polygonal models with variuos number of vertices depending on the size of glands. An appropriate set of proposals for Metropolis- Hastings-Green algorithm is formulated and a series of Markov chain Monte Carlo (MCMC) simulations are run to extract polygonal models for the glands. We demonstrate the performance of 3D spectral and spatial and 2D spatial analyses with over 90% classification accuracies and less than 10% average segmentation error for the polygonal models.