Intermediate-image-guided unsupervised deformable image registration using a latent diffusion model

Deformable image registration aims to estimate a geometric transformation that accurately aligns a pair of images, which is crucial for various medical image analysis tasks, such as disease biomarker extraction, image fusion, and region of interest (ROI) segmentation. The U-Net architecture, known for its effectiveness in medical image segmentation, has become a fundamental network in deformable image registration. However, current U-Net-based deformable registration frameworks have a limited capacity to estimate large-displacement deformations between image pairs. To address this limitation, various advanced frameworks have been proposed. Nonetheless, a significant challenge in these advanced registration frameworks is that predicting a single deformation field to handle large deformations will disrupt image topology. Therefore, there is an imperative need for further research to enhance registration accuracy while maintaining the topology-preservation ability. To mitigate the issue of topological distortion in deformable image registration and to improve registration accuracy, we focused on the foundational U-Net architecture. We proposed a novel intermediate image-guided, unsupervised deformation registration method for 2D medical images, leveraging the feature representation power of the latent diffusion model (LDM), one of the most advanced models for image synthesis. Unlike previous architecture improvement methods, our approach first synthesized an intermediate image through latent space interpolation using a pre-trained LDM. The intermediate image decomposed the single transformation flow between the image pair into two flows. This strategy effectively captured large deformations between image pairs, enhancing registration accuracy without further deteriorating the topological integrity.