Analysis and optimization of machine learning based medical image registration framework

This thesis describes my research work on analyzing and optimizing an experimental machine learning (ML) based medical image registration (IR) framework. We exploit a system developed previously at Biomedical Image and Data Analysis laboratory of SFSU as our base IR framework, which implements a feature-based image registration and uses artificial neural network (ANN) for feature selection and correspondence detection with RANSAC algorithm computing affine rigid transformation between subject images. There are several improvements to the design and implementation of this framework which this research work introduced: specifically speaking, long execution times of training and correspondence detection stages were reduced significantly by parallelizing these stages; new NN bootstrapping procedure enhanced the quality of NN training procedure; we introduced a new image preprocessing technique which improved the ratio of correct correspondence matches; we performed an in-depth comparative study of matching performances of ML-based interest point detector, SIFT and SURF algorithms which provided valuable points for understanding strengths and weaknesses of NN detector. All these improvements bring this framework a few steps closer to real-world usage