EUVIMEDEUVIMED entdecken
Crossref

DHIT-Net: Multi-scale interactive perception with hybrid context attention for deformable medical image registration

Xiangchao Liu, Shuoshuo Wang, Yonglin Chen, Chihui Long, Liangliang Su, Yalong Yang, Huabin Wang

Biomedical Signal Processing and Control · 2027-02 · 130 · 111565

DOI: 10.1016/j.bspc.2026.111565 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Biomedical Signal Processing and Control, 2027-02. Verfasst von Xiangchao Liu und 6 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Gefördert von: National Key Research and Development Program of China, Natural Science Foundation of Anhui Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Deformable medical image registration requires the joint modeling of long-range anatomical correspondence and fine-grained local structure. However, many existing CNN- and Transformer-based methods either concatenate moving and fixed images at the input or extract their shallow features independently before introducing cross-image interaction at deeper stages. In both cases, correspondence modeling is not explicitly emphasized during early representation learning, which may increase ambiguity when estimating complex deformations. To address this issue, we propose DHIT-Net, a dual-stream registration framework centered on early cross-image interaction. A Dual-Stream Multi-scale Interactive Perception (DMIP) module is introduced near the input stage to extract multi-scale representations and recalibrate each stream through lightweight mutual gated attention, thereby producing correspondence-aware features before hierarchical encoding. The resulting features are processed by Hybrid Context Attention (HCA) blocks, which combine window-based multi-head self-attention with convolutional local refinement to model contextual dependencies while retaining local anatomical details. During decoding, a Spatially Guided Transfer Module (SGTM) recalibrates encoder features before skip-connection fusion, and a sub-pixel reconstruction module predicts the full-resolution displacement field. The brain MRI experiments are conducted under an unsupervised objective, while the abdominal MR–CT experiment additionally uses auxiliary segmentation supervision during training. Experiments on the IXI, OASIS, and Abdomen MR–CT datasets show that DHIT-Net achieves competitive or leading performance among the evaluated methods on anatomical overlap and boundary-distance metrics. The ablation results suggest that the early interaction module provides the largest individual improvement among the evaluated components. The source code is available at https://github.com/lxc523/DHIT-Net .

Förderung

  • National Key Research and Development Program of China
  • Natural Science Foundation of Anhui Province

Zitieren

APA

Liu, X., Wang, S., Chen, Y., Long, C., Su, L., Yang, Y., & Wang, H. (2027). DHIT-Net: Multi-scale interactive perception with hybrid context attention for deformable medical image registration. Biomedical Signal Processing and Control, 130, 111565. https://doi.org/10.1016/j.bspc.2026.111565

Vancouver

Liu X, Wang S, Chen Y, Long C, Su L, Yang Y, et al. DHIT-Net: Multi-scale interactive perception with hybrid context attention for deformable medical image registration. Biomedical Signal Processing and Control. 2027;130:111565. doi:10.1016/j.bspc.2026.111565

BibTeX
@article{Liu2027,
  author = {Liu, Xiangchao and Wang, Shuoshuo and Chen, Yonglin and Long, Chihui and Su, Liangliang and Yang, Yalong and Wang, Huabin},
  title = {DHIT-Net: Multi-scale interactive perception with hybrid context attention for deformable medical image registration},
  journal = {Biomedical Signal Processing and Control},
  year = {2027},
  volume = {130},
  pages = {111565},
  doi = {10.1016/j.bspc.2026.111565}
}
Kostenlos starten →