SOME STUDIES ON MEDICAL IMAGE PROCESSING AND ANALYSIS

Abstract

Medical image processing is currently the foremost emerging and challenging area of research within the field of image processing. The images used for processing and analysis in the medical domain are captured through imaging technologies like MRI, CT Scan etc. Because of its broad availability and capacity to create high-resolution pictures, magnetic resonance imaging (MRI) is the most commonly utilized imaging technique. MR imaging is a sophisticated visualisation technology that allows for the safe and non-invasive acquisition of pictures of the interior architecture of the human body. Processing and analysis of human brain images plays a major role in computer aided diagnosis and neuroscience research. It further helps in identification of several neurodegenerative and mental disorder diseases such as tumors, edema, Schizophrenia, Multiple Sclerosis, Alzheimer disease and Parkinson. The segmentation of MRI brain images allows representation of brain image into meaningful segments which helps in analysis and interpretation of image for identifying the area of interest, detecting tumor, and treatment planning. The segmentation of MR brain images is carried out manually within the clinical environment by the radiologist counting on their visual interpretation. This mapping is extremely time consuming tedious task and error prone to human dependency. In order to beat the shortcomings of mapping the tumor manually, the researchers are keen to develop automatic tumor segmentation system. Also, Automatic retrieval of images is of great importance in medical field as it helps in decision making for solving various problems and helps in treatment planning. viii This thesis work tries to cater various aspects of MR brain image classification and segmentation. In order to classify the image into tumorous and non-tumorous we have proposed to use binary patterns (LBP) as features. The LBP provides a label to each 3 X 3 window based on the connection between the centre pixel and adjacent pixels. The histogram of these labels is then sent into the classification step as a feature space. The pictures are categorised as tumorous or non-tumorous using the Minimal Complexity Machine (MCM) method. The accuracy computed is used to assess the results of the proposed approaches. After identification of MR brain image as tumorous, early detection and accurate treatment based on truthful diagnosis are the major concern to cure brain tumor. The precise position, direction, and extent of the aberrant tissues are crucial in the detection of brain tumour. Detection of anatomical brain structure plays a major role in the planning of treatment. This study involves comparative analysis of three image segmentation techniques for detection of brain tumor from sample MRI images of brain. K-Means Clustering, Fuzzy C-Means Clustering, and Region Growing are the picture segmentation algorithms used. A comparative analysis of the above mentioned image segmentation techniques is done and it was found that utilizing solely feature memberships may result in overlapping clusters, emphasising the importance of using both feature and object memberships. To handle this problem in case of MR brain images we introduce the application of Fuzzy Co-Clustering Algorithm for detection of brain tumor. In this algorithm co-clustering is integrated with the Fuzzy approach with a view to obtaining distinct clusters. However, these segmentation algorithms are not able to handle the uncertainty that arises from boundary between different tissues. To handle this type of uncertainty the generalization of fuzzy sets theory is used, known as intuitionistic fuzzy sets (IFS). In order to handle the ambiguity between different tissues we have formulated Intuitionistic Fuzzy Co- Clustering Algorithm. The performance of the ix algorithms is evaluated on the basis of match score, accuracy score, Dice score and Jaccard's similarity coefficient. Also amongst the most essential fields in computer-aided treatment and neuroscience studies is the segmentation of a human brain picture from MRI scans into three brain tissues: cerebrospinal fluid, grey matter, and white matter. Segmentation algorithms such as FCM is very sensitive to noise. To avoid any stuck in local optimal results, ACO technique is used. ACO is used to determine the value of initial cluster centers. The centers thus obtained are fed into the system to perform segmentation. In modified FCM, Mahalanobis distance is used instead of Euclidean distance as Euclidean distance takes into account only the super-spherical shapes about the center of mass for clustering the data points, whereas data points belonging to same cluster may not be located in that area only. Also, the local neighborhood information is also considered as neighboring pixels are more likely to belong to same cluster. Also, we propose one more algorithm for brain image segmentation, a Fuzzy-Gravitational Search Algorithm (GSA) and its application to MRI brain image segmentation. The proposed approach is based on GSA, and uses fuzzy inference rules for controlling the parameter α as search progresses. The results of the system are compared with GSA and recent work on brain image segmentation algorithms for both real and simulated database on the basis of Dice Coefficient values and is found to outperform. Also, we propose an image retrieval algorithm in this thesis. Automatic retrieval of images is of great importance in medical field as it helps in decision making for solving various problems. The feature extraction in content based image retrieval is a crucial stage whose success is determined on the approach used to extract characteristics from pictures. There are two types of visual content descriptors: global and local. To describe an image, a global descriptor describes the visual features of the entire image, whereas a local descriptor represents the visual features of areas or components. x The feature database is built by arranging these as multi-dimensional feature vectors. In this thesis, we have proposed a novel feature descriptor based on local binary pattern for retrieval of biomedical images.