Deformable medical image registration using Deep Learning

Abstract

Medical image registration involves aligning two or more images of the same subject or different subjects with various imaging modalities. It is a fundamental task in medical image analysis, with applications in disease monitoring, treatment planning, motion tracking, population analysis and many more. However, it is a challenging task due to a variety of factors, such as differences in imaging modalities, anatomical variations, and deformation caused by organ motion or surgical intervention. Deep Learning (DL) has shown great promise in addressing some of the challenges in medical image registration. Advanced deep neural networks are able to learn complex features from data and make predictions of non-linear (deformable) transformations from input images to solve registration problems with significantly higher computational efficiency than non-DL registration methods. The works presented in this thesis focus on investigating and improving aspects of deformable DL registration. First, a study was conducted to investigate the effectiveness of supervised and unsupervised training of DL registration for cardiac motion using Magnetic Resonance Imaging (MRI) images. Second, a DL registration method is proposed which uses a differentiable Mutual Information (MI) loss and diffeomorphic free-form deformation (FFD), enabling accurate and well-regularised registration of medical images in different modalities. Finally, an accurate, data-efficient and robust DL registration method is developed by embedding variational optimisation in the learning-based framework.