Hierarchical cross-attention guided deformable registration with multi-level feature fusion for medical images

Deformable medical image registration is fundamental for longitudinal diagnosis, surgical planning, and quantitative anatomical analysis, yet achieving accurate and robust alignment remains challenging. Traditional optimization-based methods are computationally intensive and highly sensitive to parameter settings. In contrast, many recent deep learning approaches fail to effectively capture spatial correspondences across diverse anatomical structures, struggle with multi-scale feature fusion, and lack interpretability. To address these issues, we propose Hierarchical Cross-Attention Morph (HCA-Morph), a novel unsupervised registration framework that explicitly models spatial alignment and adaptively integrates hierarchical features. HCA-Morph comprises two dedicated modules: a Spatial Correspondence-Aware Module (SCAM), which learns interpretable spatial alignment between moving and fixed images from local to global scales, and a Cross-Scale Attention Module (CSAM), which dynamically fuses multi-level features with built-in interpretability to enhance representation consistency. Experiments on two public brain MRI datasets-the Open Access Series of Imaging Studies (OASIS) and Information eXtraction from Images (IXI)-demonstrate that HCA-Morph achieves Dice scores of 0.845 and 0.834 on the OASIS and IXI datasets, respectively, outperforming current state-of-the-art models, while also achieving better HD95 scores (1.365 on OASIS and 1.725 on IXI). Notably, with only 2.05 MB parameters-the smallest among all compared methods-HCA-Morph achieves comparable or superior accuracy while maintaining the lowest memory consumption (4.08 GB on OASIS, 9.02 GB on IXI) and generating more plausible deformation fields with fewer folding regions (|J_ϕ ≤ 0|: 1.77×10^-4 on OASIS, 5.39×10^-6 on IXI). Ablation studies confirm that SCAM and CSAM provide complementary contributions on OASIS, combining synergistically to yield a total improvement of 0.030. Portability experiments further demonstrate that both modules can be successfully integrated into diverse existing architectures including CNN, Transformer, and Mamba-based models, delivering consistent performance improvements across different backbones. Overall, HCA-Morph achieves a strong balance between accuracy, efficiency, and anatomical plausibility for deformable brain MRI registration. The source code and pretrained models are publicly available at https://github.com/Importes/HCA-Morph.