ChebyRegNet: An Unsupervised Deep Learning Technique for Deformable Medical Image Registration

Image registration supports a variety of clinical applications such as treatment planning, disease monitoring, and interventions. Deep learning-based deformable image registration techniques, including VoxelMorph and HyperMorph, are achieving state-of-the-art results but still exhibit considerable errors in multimodality registrations, such as structural to functional image registration. To address these challenges, this paper proposes ChebyRegNet, a deep learning framework that leverages Chebyshev polynomials and hypernetworks, to enhance registration accuracy. Moreover, a Feature Extraction Network is added that produces deep features for the Chebyshev HyperNetwork, which is responsible for generating parameters for the Image Registration Network. This structure allows ChebyRegNet to successfully handle challenging registration tasks. Experiments show that ChebyRegNet greatly improves registration in terms of target registration errors compared to VoxelMorph and HyperMorph, achieving 0.1 versus 0.2 and 0.19 for HyperMorph and VoxelMorph, respectively. Moreover, ChebyRegNet outperforms the other two techniques in terms of the Dice Score and achieves competitive results in terms of the Jaccard Coefficient. The publicly available code for our work can be accessed at https://github.com/udanish50/ChebyRegNet/tree/main.