Cluster analysis based classification of hepatic tumor cells in histopathology images using machine learning techniques

Abstract

In this thesis a novel algorithm for segmentation of histopathological data is proposed. The differentiation of a cluster of nuclei and multi-nucleation is a critical issue in automated diagnosis systems. The similarities between said clusters and malignant nuclei would influence the misclassifications of these regions in the final decision of automated systems. The proposed automated diagnosis system in the first major aspect of this thesis for detecting the nuclei of Hepatocellular Carcinoma (HCC) or noncancerous is formulated as a joint classification and segmentation problem using Area based adaptive Expectation Maximization (EM) technique. This newlinealgorithm detects the nuclei with high sensitivity and gives the better accuracy of 70.44% which is comparable with the state-of-art method, Convolutional Networks. In the face of assessing nucleus and the extracellular nuclei changes in istopathology images by this algorithm, the nuclei cluster includes undesired objects such as blood cells, mucos and uneven dyeing stains. And also it fails to detect the multi-nuclei in images and hence, the second major aspect of the thesis falls on the segmentation of nuclei alone by excluding the aforementioned undesired objects. The proposed method uses Quick hull to define a convex hull of each hepatic cell in the image to newlineimprove the segmentation of single or multiple nuclei regions by discarding the excess stains and noise. Now this system becomes more trustworthy to detect the single as well as multiple nuclei with a low number of false positives and false negatives provides the accuracy of 89.76%. An optimal combination of features have been proposed to create a distinctive model of nuclei which will be used to classify the diseased nuclei and non-cancerous nuclei to get a significant improvement in true positive and reduction in false negative rates with the classification accuracy of 88.24%. newline newline