An intelligent decision support system based on image processing for evaluating of the endoscopic images

ABSTRACTPhD ThesisAN INTELLIGENT DECISION SUPPORT SYSTEM BASED ON IMAGE PROCESSING FOREVALUATING OF THE ENDOSCOPIC IMAGESAbdulkadir ŞENGÜRFirat UniversityGraduate School of Natural and Applied SciencesDepartment of Electrical - Electronics Engineering2006, Page: 133Intelligent recognition systems have gained importance in various areas with the technologicaldevelopments, and pattern recognition constitutes the bases of these systems. Pattern recognition is torecognize unknown patterns by assigning them into a known class or known patterns belonging to aknown class. Pattern recognition includes two steps: Feature extraction and classification. Imagesegmentation is considered at the feature extraction stage of a pattern recognition system. Thus, imagesegmentation is important for subsequent pattern recognition processes. Its performance directlyaffects the performance of the subsequent classification procedure.In this thesis, three various techniques were proposed for gray and color image segmentation.1. A MRF (Markov Random Fields) model which uses Poisson probability density functionfor gray level image segmentation, is proposed.2. Entropy based wavelet packet neural networks architecture which contain featureextraction period and update the features according to the performance of the classifier wasmodified as it can be used on images. Beside entropy features, energy feature wereembedded for robust and efficient system development.3. A new unsupervised system was proposed which uses wavelet transform and oscillatorneural networks architecture. The most important advantage of the proposed system is thatit achieves the desired segmentation process without any training samples as needed in thesupervised pattern recognition systems.The proposed feature extraction mechanisms and the supervised neural networks architecturewere used for detection of normal and abnormal (polyp) formations at the endoscopic video frames.86.2 % sensitivity and 85.7 % specificity values were obtained when wavelet packet entropy andenergy features with the supervised neural networks were used. We also obtained 90.2 % sensitivityand 88.7 % specificity values by using statistical features of the wavelet transform co occurrencematrices and neural networks.Keywords : Pattern recognition, feature extraction, intelligent diagnosis system, wavelet transform,image segmentation techniques, endoscopic images, artificial neural networks, oscillator neuralnetworks, wavelet neural networks.