Design and Development of Content Based Image Retrieval System for Analysis of Liver Tumor On Ct Image

Abstract

Liver cancer is a serious threat to human life due to low survival rates. Generally, newlineimages of the tumor are acquired from Computed Tomography (CT) and Magnetic newlineResonance Imaging (MRI) techniques. These images are stored in the hospital newlinedatabase for further analysis. However, manual image analysis by radiologists is newlinesubjective and time consuming. Further, the biopsy is the reliable method for liver newlinetumor diagnosis. But, this method is invasive and hence not recommended. To newlineaddress these problems, there is need to develop computer aided diagnosis system for newlineanalysis of liver tumors. Therefore, this thesis proposes an effective Content-Based newlineMedical Image Retrieval (CBMIR) system for diagnosis of liver tumors. It assists newlinephysicians in making diagnostic decisions by retrieving relevant cases of liver tumor. newlineThus, improves the accuracy and efficiency of the liver tumor diagnosis process. newlineThis research work proposes methods for tumor segmentation, classification and newlineindexing. In the first step, the acquired CT image is enhanced using Contrast Limited newlineAdaptive Histogram Equalization(CLAHE) algorithm. Next, the effective newlinesegmentation technique is proposed based on Kernalized Fuzzy C-Means Clustering newline(KFCM) and Hybrid Bee Colony and Cuckoo Search (HBCCS) techniques to detect newlineliver tumor on the CT image. The LTP (Local Ternary Patterns) and SURF (SpeededUp Robust Features) feature set is extracted from the detected tumor. Then the newlinerelevant features are selected using Modified ReliefF algorithm. Next, the newlineclassification task is carried out using proposed Modified LSTM (Long Short-Term newlineMemory) classifier to distinguish between benign and malignant tumors. These labels newlinealong with feature set are stored in the database. Similar steps are carried out on the newlinequery image and then the database is indexed using the proposed modified K-means newlineclustering algorithm and KD tree techniques. Similarity matching is carried out using newlineEuclidian distance and the most relevant images are retrieved and displayed to the newlinephysician. newlineExperiments a