Robust unsupervised medical image registration using a recursive deformable pyramid network

Medical image registration is an important task in medicine for providing an accurate anatomical correspondence between multimodal images, which is essential for diagnosis, treatment planning, and longitudinal follow-up. Nevertheless, the natural variability of biological shapes as well as the absence of annotated data in unsupervised settings make conventional registration-related algorithms extremely challenging to apply. In this work, we introduce a new Recursive Deformable Pyramid Network (RDPN) for the task of unsupervised medical image registration that aims to model both global and local deformation fields by hierarchically incorporating multi-scale feature representations. The adopted network consists of a deformable convolutional backbone that is used recursively on the pyramid level, where such instantiation allows us to estimate the adaptive spatial transformation without inheriting ground truth correspondences. Our method is tested on a synthetic brain MRI dataset which was meticulously generated to simulate inter-patient and intra-patient anatomical variability and a real abdominal CT dataset. Experiments show that RDPN achieves a robust performance gain in comparison with the state-of-the-art methods in the aspects of Dice similarity coefficient, target registration error and deformation smoothness. The pyramid mechanism in recursive level is especially effective for the purpose of aligning fine-grained anatomical structures with global structure consistency. Moreover, a thorough ablation study demonstrates the importance of recursive feature fusion and deformable modeling in learning a robust unsupervised registration solution. This work contributes by presenting a pragmatic approach to scalable registration for difficult registration problems in medical imaging, which could elevate the quality of the clinical workflow downstream.