Three-dimensional (3D) deformable image registration plays a critical role in 3D medical image processing. This technique aligns images from different time points, modalities, or individuals in 3D space, enabling the comparison and fusion of anatomical or functional information. To simultaneously capture the local details of anatomical structures and the long-range dependencies in 3D medical images, while reducing the high costs of manual annotations, this paper proposes an unsupervised 3D medical image registration method based on shifted window Transformer and convolutional neural network (CNN), termed Swin Transformer-CNN-hybrid network (STCHnet). In the encoder part, STCHnet uses Swin Transformer and CNN to extract global and local features from 3D images, respectively, and optimizes feature representation through feature fusion. In the decoder part, STCHnet utilizes Swin Transformer to integrate information globally, and CNN to refine local details, reducing the complexity of the deformation field while maintaining registration accuracy. Experiments on the information extraction from images (IXI) and open access series of imaging studies (OASIS) datasets, along with qualitative and quantitative comparisons with existing registration methods, demonstrate that the proposed STCHnet outperforms baseline methods in terms of Dice similarity coefficient (DSC) and standard deviation of the log-Jacobian determinant (SDlogJ), achieving improved 3D medical image registration performance under unsupervised conditions.
In the field of medical image registration, corresponding landmarks can serve as an crucial reference for the quantitative evaluation of registration algorithm errors. However, the corresponding landmarks extracted by existing methods generally lack anatomical correspondence. While some approaches have begun to take bifurcation points of tubular structures as landmarks, the establishment of accurate one-to-one correspondences between these points is hindered by highly similar local features and missing correspondences, thereby impeding effective quantification of registration errors. To address these limitations, this paper proposes a reweighted graph matching-based algorithm for automatic extraction of corresponding landmarks in medical images. First, the bifurcation points of tubular structures with anatomical correspondence are extracted as landmarks. Secondly, to address the issues of highly similar local features and missing correspondences, a reweighted graph matching model is constructed base on a local similarity strategy and a modality independent neighborhood descriptor, which iteratively identifies correct matches and rectifies mismatches to progressively refine the correspondence between landmarks. Finally, the K-nearest neighbor graph of landmarks is dynamically updated, and outlier matches are filtered based on a local similarity strategy, thereby achieving accurate landmark matching under partial correspondence conditions. Experimental results on two public datasets demonstrate that the proposed method can accurately extract corresponding landmarks and effectively evaluate registration error, thereby providing technical support for enhancing the reliability of medical image registration algorithms in clinical applications.