Longitudinal VoxelMorph : spatiotemporal modeling of medical images

Abstract

Medical image registration is an important initial step in many downstream clinical tasks. Repeated imaging for diagnostic, therapeutic, or scientific discovery is common. Classical longitudinal image registration systems are too slow to be useful in practice and are often only designed for a highly specific type of image data. Efficient pairwise image registration models are limited by not accounting for the temporal nature of the data. We present Longitudinal VoxelMorph, a novel machine-learning-based model for efficient and scalable spatiotemporal medical image registration. We also dene a new evaluation metric to quantify the temporal smoothness of a longitudinal deformation field. We evaluate the model on cardiac cine-MRI data and echocardiography data, and nd that Longitudinal VoxelMorph is more temporally consistent than state-of-the-art pairwise models, and achieves comparable or improved anatomical accuracy. Longitudinal VoxelMorph has the potential to be incorporated in downstream medical image tasks, such as image prediction and diagnosis, facilitating better clinical outcomes.