Detection of Abnormalities using Fusion of Brain CT and MR Images

Abstract

The rapid technological development in instrumentation has resulted in many medical imaging techniques to capture and represent the visual information of different parts of the human body. The imaging modalities such as magnetic resonance imaging (MRI) and computed tomography (CT) represent structural details of tissues or organs. The radiologists often advise multimodal imaging to identify the diseased tissues of a particular organ as many a time a single modality is insufficient in conveying the complete information of a tissue or organ. The multimodal imaging may consist of a CT image which represents bone information precisely and MRI image which provides clear soft-tissue information. The intensity and texture information of different tissues represented by CT and MRI are fused to visualize hard and soft tissues clearly in a single image. The details of abnormalities such as edges and textures can be accurately demarcated from the surrounding tissues by the fusion image. The fused images can be clustered for the detection of abnormalities. The fusion can enhance the clustering results and classification accuracy. The existing fusion schemes fail to meet the real-time requirements of the radiologists such as the computational speed of the algorithm, contrast, clarity and edge information of the fused image. Hence, there is a need to develop new fusion methods that can meet the real-time requirements of the radiologists. Based on the above research gaps in this field, research objectives have been defined. The main aims are to design and develop faster fusion algorithms with better contrast, clarity, brightness and edges which can help in the detection of abnormalities. To achieve the goals multi-resolution transforms based fusion methods with different fusion rules are developed and clustering schemes are applied for the detection of abnormalities in the fusion images. In order to validate the proposed methods image pairs are acquired from various sources. The real-time dataset consisting of 40 image pairs of five patients suffering from various diseases are acquired from Postgraduate Institute of Medical Education and Research (PGIMER), Chandigarh, India. Another standard dataset consisting of ten image pairs of ten patients suffering from different brain diseases are acquired from internet repositories. To evaluate the performance of the proposed fusion and clustering methods qualitative analysis is done by the author as well as by the expert radiologists to know their clinical application. To validate the visual results quantitative analysis is also done using fusion and clustering quality assessment parameters. iv Three different fusion schemes are developed which are based on non-subsampled contourlet transform (NSCT), stationary wavelet transform (SWT) and non-subsampled shearlet transform (NSST). These three different techniques provide sufficient coverage of the applicability of the multi-resolution transformation domain in image fusion. The first scheme is an NSST based scheme in which a novel fusion rule is proposed. It is based on the morphological gradient motivated pulse-coupled neural network. The scheme is found to be faster and better in terms of contrast, clarity, brightness and edge information as compared to state of the art schemes. The second developed scheme is a hybrid multiresolution scheme which is a combination of NSCT and SWT. The SWT is used in the form of a novel LF sub-band fusion rule. This scheme carries better texture, contrast, clarity and edges in the fused image as compared to state-of-the-art NSST based fusion methods. The third scheme is also based on NSST and a novel fusion rule. This fusion rule is based on a human visual system motivated operator called the smallest uni-value segment assimilating nucleus. This scheme is also found to be a faster option when compared to state of the art schemes while preserving the structural details of CT and MR images. These schemes serve as an automatic tool to the radiologists while obtaining the diagnostic information during skull injury, to mark the disease orientation as regards to bone and also to decide the point and depth of insertion of surgical instruments during brain surgery. All these schemes are useful for the clinical applications according to the radiologists. The fourth scheme is a clustering scheme that segments the tissues of fusion images into different classes. Four different clustering schemes are applied to segment the fusion images in this method. The clustered fusion images are visually and quantitatively compared for the detection of diseases. A visual analysis of clustered fusion images is also done along with the clustered MR images to compare the information gained in the clustered fusion images. It is observed that the clustered fusion images show the edema and tumor as two separate classes whereas, clustered MR images show these tissues as a single class. The clustered fusion images show the intensity inhomogeneous regions of the same tissues as a single class whereas, the clustered MR images show these as separate classes. This proves that the fusion information aids in clustering for the detection and separation of abnormalities. The developed methods help the radiologists in precise localization, diagnosis, interpretation and treatment planning of brain diseases.