Abstract: Traditional single-stream deep registration models generally rely on implicit feature matching between im-age pairs. While Transformers address the locality constraints of convolutional neural networks (CNNs) by capturing long-range contextual dependencies, they often underrepresent fine-grained local details. Con-sequently, dual-stream architectures, whether CNN- or Transformer-based, struggle to effectively model explicit feature correspondences. To overcome these limitations, we propose CADReg, a triple-stream un-supervised deformable registration framework for multi-resolution of medical imaging, leveraging cross-neighborhood attention embedded displacement fusion. CADReg explicitly establishes mappings between spatial point features in the fixed image and corresponding neighborhood features within the de-formed moving image via cross-neighborhood attention. This strengthens the synergy between attention mechanisms and deformation estimation, enhancing both registration accuracy and efficiency. A subse-quent decoupling module integrates three complementary elements: current-resolution features, deep mu-tual information, and features encoding long-range dependencies. This fusion enables comprehensive ex-traction of inter-feature and intra-feature contextual information, simultaneously aggregating low-level lo-cal details and propagating high-resolution structural cues to progressively refine deformation fields in a coarse-to-fine manner across resolutions. Extensive evaluations on three public 3D brain MRI datasets (LONI LPBA40, IXI and OASIS) demonstrate that CADReg consistently outperforms mainstream registra-tion methods, achieving Dice similarity coefficient of 72.6%±1.3%, 76.8%±2.4%, and 89.6%±1.4%, and structure similarity index of 0.979±0.003, 0.954±0.003, and 0.980±0.003, respectively. These results con-firm CADReg’s enhanced ability to identify and interpret clinically relevant regions of interest.