Nonrigid Image Registration using 3D Convolutional Neural Network with Application to Brain MR images

Abstract

Image registration is a vital tool in medical image analysis with a large number of applications assisting the medical experts. Currently, conventional image registration approach with predefined dissimilarity metric and iterative optimization, is widely used. In this thesis, we proposed a method to solve medical image registration problem using a deep learning approach. A convolutional neural network architecture is proposed, named as "RegNet", applied on monomodality image registration problem. The proposed RegNet does not require any dissimilarity metric and is capable of directly estimating the spatial relationship between two images. The training is based on the pseudo-real-world displacement vector field, created by the conventional registration tool and artificial deformation simulation, resembling a deformation similar with the real-world deformation case. Multi-stage framework is also implemented to increase the capture range of RegNet. This thesis evaluates the performance of "RegNet" for an intrasubject magnetic resonance brain images registration problem.