Learning Rigid Image Registration - Utilizing Convolutional Neural Networks for Medical Image Registration

Many traditional computer vision tasks, such as segmentation, have seen large step-changes in accuracy and/or speed with the application of Convolutional Neural Networks (CNNs). Image registration, the alignment of two or more images to a common space, is a fundamental step in many medical imaging workflows. In this paper we investigate whether these techniques can also bring tangible benefits to the registration task. We describe and evaluate the use of convolutional neural networks (CNNs) for both mono- and multi- modality registration and compare their performance to more traditional schemes, namely multi-scale, iterative registration. This paper also investigates incorporating inverse consistency of the learned spatial transformations to impose additional constraints on the network during training and investigate any benefit in accuracy during detection. The approaches are validated with a series of artificial mono-modal registration tasks utilizing T1-weighted MR brain i mages from the Open Access Series of Imaging Studies (OASIS) study and IXI brain development dataset and a series of real multi-modality registration tasks using T1-weighted and T2-weighted MR brain images from the 2015 Ischemia Stroke Lesion segmentation (ISLES) challenge. The results demonstrate that CNNs give excellent performance for both mono- and multi- modality head and neck registration compared to the baseline method with significantly fewer outliers and lower mean errors.