Scholarlink

<p>To address the issues of insufficient feature extraction and difficulty in preserving the original topological structure of aligned images in existing 3D medical image registration networks, this paper proposes a 3D medical image registration network based on a bidirectional recursive pyramid (BRP-Net). First, a multi-scale large-kernel feature extraction module is designed to expand the model’s receptive field, enhancing its global perception and fine-grained structural modeling capabilities. Second, a bidirectional pyramid registration network structure is constructed, introducing forward and backward deformation fields during the registration process to impose mutual constraints, thereby enhancing the model’s ability to preserve structural consistency and topological integrity in the registration results. Subsequently, a recursive strategy is proposed to progressively fuse multi-scale deformation fields from low to high resolution, improving the continuity and accuracy of the deformation fields. Finally, a bidirectional consistency loss function is designed to jointly constrain the model training with bidirectional registration results, thereby improving the smoothness of the deformation fields. Experimental results demonstrate that the proposed BRP-Net achieves Dice coefficient, HD95 value, and negative Jacobian value of 0.746, 8.06, and 0.135, respectively, on the LPBA40 dataset. This model not only effectively preserves the original image topology but also outperforms existing mainstream methods in overall registration performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Bidirectional Recursive Pyramid Network for Medical Image Registration

  • 期刊论文

  • 发表日期：

    2025年12月30日

查看全文
  • Jing Peng,
  • Jiarong Yan,
  • Ziyi Wei,
  • Jiaying Liu,
  • Yahong Deng,
  • Jiale Yang

摘要

To address the issues of insufficient feature extraction and difficulty in preserving the original topological structure of aligned images in existing 3D medical image registration networks, this paper proposes a 3D medical image registration network based on a bidirectional recursive pyramid (BRP-Net). First, a multi-scale large-kernel feature extraction module is designed to expand the model’s receptive field, enhancing its global perception and fine-grained structural modeling capabilities. Second, a bidirectional pyramid registration network structure is constructed, introducing forward and backward deformation fields during the registration process to impose mutual constraints, thereby enhancing the model’s ability to preserve structural consistency and topological integrity in the registration results. Subsequently, a recursive strategy is proposed to progressively fuse multi-scale deformation fields from low to high resolution, improving the continuity and accuracy of the deformation fields. Finally, a bidirectional consistency loss function is designed to jointly constrain the model training with bidirectional registration results, thereby improving the smoothness of the deformation fields. Experimental results demonstrate that the proposed BRP-Net achieves Dice coefficient, HD95 value, and negative Jacobian value of 0.746, 8.06, and 0.135, respectively, on the LPBA40 dataset. This model not only effectively preserves the original image topology but also outperforms existing mainstream methods in overall registration performance.