MS-HIENet: multi-scale hybrid implicit-explicit registration network

Abstract Objective. Medical image registration serves as a cornerstone for precision diagnosis and treatment, particularly for dynamic organs like the lungs, where high deformability and complex motion patterns impose significant challenges. Traditional iterative methods are time-consuming, while existing deep learning approaches often struggle to synergistically optimize large deformation modeling and localized fine structure preservation due to limited receptive fields. This study aims to address the challenge of jointly handling large-scale and localized deformations in pulmonary imaging by developing a novel deep learning framework. Approach. We propose a multi-scale hybrid implicit-explicit registration network (MS-HIENet), a mask-free end-to-end framework that integrates implicit neural representation (INR) and convolutional neural networks (CNN). The method employs a two-fold strategy: First, a multi-scale optimization mechanism where low-resolution layers leverage INR to capture global deformations and high-resolution layers utilize CNN to refine local anatomical structures, enabling coarse-to-fine hierarchical registration. Second, an INR-based coordinate-to-displacement implicit mapping framework is used to directly model continuous deformation fields, eliminating the dependency on mask annotations. Main results. Experimental results on the deformable image registration laboratory dataset demonstrate that MS-HIENet achieves a mean target registration error of 1.00 mm, representing an average reduction of 29.5% compared to state-of-the-art deep learning methods. Ablation studies validate the effectiveness of the multi-scale collaboration and hybrid implicit-explicit representation, with the deformation field folding rate reaching an minimal level (mean: 0.00017). Significance. The proposed method effectively bridges the gap between global deformation consistency and local anatomical precision. By combining the continuous modeling capabilities of INR with the local feature refinement of CNNs, MS-HIENet significantly enhances topological consistency and clinical applicability, offering a robust solution for high-precision lung image analysis.