Efficient Algorithms for Multi modal Registration and Enhancement of Medical Images

Abstract

newline Image registration is the process of mapping corresponding points in two images. In the medical imaging domain, registration of images is an indispensable part, and is used for innumerable purposes. It has usage in diagnosis and treatment of diseases, in planning of therapy or during surgery.Hence, development of fast and accurate image registration algorithms is a topic of active research. newlineSeveral registration algorithms exist in the literature. Traditional image registration algorithms used corresponding points or corresponding surfaces in the two images to map one image to the other. A comparatively new class of image registration algorithms use image intensities to map images. The idea behind these algorithms is that there exist some relation between the intensities of corresponding points of two images. An information theory based approach, called mutual information, is widely used for registering images. At the registered position, the mutual information between two images is maximum. Mutual information is computed using entropy and joint entropy of images, which are based on image intensities. Mutual information based registration algorithms are one of the most efficient and robust algorithms till date. They allow fully automatic 3D multi-modal medical image registration. With the development of Deep Neural Networks (DNN), a number of newer algorithms have emerged. However, due to the complicated architecture and high computation cost, DNN based algorithms are still not widely in use. newlineThe goal of this thesis is to design computationally efficient multi-modal medical image registration algorithms. Hence we focus our attention on statistical parameter based algorithms, like mutual information, because of their simpler architecture and low computation cost compared to DNN based algorithms. In an attempt to reduce computation, we proposed efficient formulae for computation of entropy and joint entropy of images. Unlike the traditional formulae, which considers each intensity at a time, the