UNSUPERVISED TUMOR EXTRACTION AND CLASSIFICATION

Abstract

This thesis is concerned with the problem of tumor extraction and classification. The process of tumor detection and classification is a complex and time consuming task since it requires a careful assessment of medical images. In the effort to produce more efficient and accurate results, image analysis techniques are frequently being used. Therefore, developing a system which could accurately segment the tumor affected regions and categorize the tumors in different classes is very important. It is of benefit to develop a computer system which assists radiologists and also reduces the subjectivity and human errors involved in the diagnosis. The aim was to develop reliable methods that contribute towards accurate extraction and classification of tumor from medical images. This thesis contributes in all major steps of a computer aided system i.e. image preprocessing, image segmentation, feature extraction and classification. In image pre-processing, two image fusion techniques for multi-modal medical images based on local features and fuzzy logic are presented. In first scheme, local entropy and variance are used to calculate the information in images. The scheme assigns weights to pixels depending upon the amount of information. The main advantage of the proposed fuzzy logic based image fusion scheme is improvement in fused results. The second scheme uses undecimated wavelet, local features, improved guided filter and weighted maps. This scheme offers less spectral distortion and produce better spatial information than the existing techniques. In image segmentation, two segmentation schemes based on weighted fuzzy active contour are presented. vi In these techniques, weights have been assigned in proportion to the information provided by local features, two fuzzy systems (Mamdani inference and Takagi-Sugeno inference) based systems are used for assigning weights. A method is presented for feature extraction and classification by using texture features, invariant moments and perception based features. Optimal feature combination using fuzzy weights and classification using multi-class support vector machines is performed. Simulation results when analyzed visually and quantitatively depict the significance of the proposed schemes compared to existing schemes. The results of all the proposed image fusion schemes are demonstrated through examples of medical images and results of test against conventional schemes