H-UDMIR: A Hybrid Unsupervised Deep Learning Framework for Deformable Medical Image Registration

The study presents H-UDMIR, a novel hybrid medical unsupervised deep learning (DL) framework for deformable image registration (DIR) that integrates U-Net architecture and Spatial Transformer Networks (STNs). The U-Net initially predicts coarse deformation fields, which are then enhanced by cascaded STNs to dynamically register local image regions while tolerating a loss function to effectively balance deformation smoothness and image fidelity. This hybrid approach exhibits enhanced registration performance and resilience without the necessity for annotated training data. Experimental results using H-UDMIR highlight its precise registration capabilities across varied anatomical structures and imaging modalities: kidney computed tomography (CT) scans and brain magnetic resonance imaging (MRI) datasets. Notably, H-UDMIR achieves a Dice similarity coefficient of 92.72% and a Hausdorff distance of 881.9244 on low- and high-dose kidney CT scans, surpassing VoxelMorph’s 84.04% Dice score. A comparative analysis against state-of-the-art approaches such as VoxelMorph underscores H-UDMIR’s superior performance.