Unleashing Registration: Diffusion Models for Synthetic Paired 3D Training Data

Deep learning-based medical image analysis often comes with the drawback of a very limited amount of publicly available data due to privacy concerns. For the task of image registration, this is especially relevant, since networks rely on image pairs for training. Augmenting the dataset with synthetic deformations does not suffice for anonymous publication, as patient-specific topologies are kept. To handle this, we propose to leverage vector quantisation- and diffusion-based image generation to synthesise paired data for training of registration networks. Our method introduces a new concept that directly synthesises paired data. We make this method applicable to both image data in the form of 2D brain MRI slices and sparse point cloud data in the form of 3D vessel representations of intra-patient lung scan pairs by volumetric rasterisation. With paired synthetic data, only minimal fine-tuning on privacy-sensitive real data is needed to achieve comparable or even better results to training on a large real dataset. The code is publicly available at https://github.com/multimodallearning/UnleashingRegistration .