Image Information Measures for Predicting Image Registration Performance on iThemba LABS Image Registration System

Abstract

Image information measures such as mutual information and entropy quotient can do more than just giving a quantitative impression on how good an image is. In recent times, they have found application in radiotherapy, especially in treatment planning. Thinking about maximizing information in an image, image registration has been found to be handy, though computationally expensive to achieve. In this work, we describe an experiment to help establish a relationship between registration performance and the extent of misalignment that made registration necessary using data from a CT scan of the body. We present ways in which information measures can be used to decide whether registering images is a necessary operation. We visualize the experimental result and define functions that fits the distribution by making an educated guess and then optimized the functions to arrive at as minimum parameter as possible. We finally screen the various models to arrive at the optimally performing ones. We have found these models to perform very well in predicting registration performance pre-operationally, explaining between ~62.94% to ~99.96% of the effect of rotating around or translating along x, y, z on the performance of registration output should it be carried out, thereby saving more computer power required in image registration, time and boredom on the part of the patient.