Y-Unet: A multi-scale hybrid deep learning framework for deformable registration of cardiac CT images

Deformable image registration is a critical yet challenging task in medical image analysis, particularly for organs with large deformations like the heart. Existing deep learning methods, often based on Convolutional Neural Networks (CNNs), are limited in capturing long-range spatial dependencies, causing them to converge to suboptimal local minima when handling significant deformations. To address this, we propose Y-Unet, a novel multi-scale unsupervised registration framework. Y-Unet features three key innovations: a coarse-to-fine pyramid strategy to handle large deformations within a single forward pass; an grayscale difference map (DM) to dynamically focus the network on misaligned regions; and a hybrid U-Net architecture. This architecture’s encoder uses a Swin Transformer to capture global context, while its convolutional decoder reconstructs a detailed deformation field. Experiments on the MM-WHS 2017 cardiac CT dataset show that Y-Unet outperforms five state-of-the-art methods, including VoxelMorph and TransMorph, across multiple metrics for both registration accuracy (DSC, HD95) and deformation plausibility (non-positive Jacobian determinants). These results demonstrate the effectiveness and superiority of our proposed method for complex cardiac image registration.