EUVIMEDEUVIMED entdecken
Crossref

RSMIA‐Net : A Novel Residual Skip‐Enhanced Network With Multi‐Dimensional Interactive Attention for Pulmonary Embolism Detection in CT Pulmonary Angiograms

Hao Ma, Yu Gu, Wei Zhang, Lidong Yang, Baohua Zhang, Juan Hao, Meng Chen, Jianjun Li, Xin Liu, Siyuan Tang, Qun He

International Journal of Imaging Systems and Technology · 2026-11 · 36 (6)

DOI: 10.1002/ima.70476 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Imaging Systems and Technology, online erschienen am 09.11.2026. Verfasst von Hao Ma und 10 weiteren Autor:innen. Die Arbeit nennt 53 Literaturangaben. Gefördert von: Natural Science Foundation of Hebei Province, Natural Science Foundation of Inner Mongolia Autonomous Region, National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Pulmonary embolism (PE) is a life‐threatening condition that seriously endangers human health. To address the semantic gap between high‐ and low‐level features and the lack of spatial context modeling in PE detection, we propose RSMIA‐Net, a Residual Skip‐enhanced Network with Multi‐dimensional Interactive Attention. First, we propose an Attention‐Guided Residual Skip (AGRS) module, which combines a Residual Skip Enhancement (RSE) structure with a Multi‐dimensional Interactive Quadruplet Attention (MIQA) module. The RSE structure introduces top‐down multiplicative guidance and residual fusion into skip connections, enabling deeper semantic information to enhance shallow feature representations. The MIQA module captures interdependencies between channel and spatial dimensions from four complementary perspectives and further improves fine‐grained spatial feature extraction in PE regions via a Spatial Cross‐Interaction Architecture. Then, we propose a Context‐Detail Balanced ASPP (CDB‐ASPP) module that fuses multi‐scale contextual information through five parallel convolutional branches with varying dilation rates, combined with a hybrid global–local enhancement strategy. Additionally, a Dual Attention (DA) Transformer module is incorporated at the bottleneck of the U‐Net to enhance global feature modeling. Finally, elastic deformation‐based data augmentation is employed to improve training efficiency on limited datasets. In this study, PE segmentation denotes voxel‐level delineation of embolic regions, whereas PE detection refers to lesion‐level identification derived from the segmentation outputs through post‐processing. Specifically, the predicted PE probability maps are converted into candidate embolic regions and further evaluated using detection metrics after connected‐component analysis and candidate matching. In the CAD‐PE dataset, RSMIA‐Net achieves competitive performance on the PE detection task, with a sensitivity of 0.904 (95% CI: 0.851–0.962), an FROC score of 0.798 (95% CI: 0.727–0.853), and an AUC of 0.862 (95% CI: 0.799–0.907). Compared with the baseline model, the proposed modules increase the parameter count from 14.906 to 24.991 M and the computational cost from 1.550 to 5.265 G FLOPs, accompanied by improved detection performance. To further assess generalization, RSMIA‐Net was directly evaluated on the external FUMPE dataset without retraining or changing the post‐processing criteria. The model achieved an FROC score of 0.742 (95% CI: 0.708–0.822), an AUC of 0.872 (95% CI: 0.814–0.930), and a sensitivity of 0.928 (95% CI: 0.869–0.968), demonstrating its generalization capability on an independent CTPA cohort. The results show that RSMIA‐Net achieves improvements compared with other state‐of‐the‐art methods in PE detection. All implementation codes and supporting documents of the proposed approach can be accessed at https://github.com/GuYuIMUST/RSMIA‐Net .

Förderung

  • Natural Science Foundation of Hebei Province
  • Natural Science Foundation of Inner Mongolia Autonomous Region
  • National Natural Science Foundation of China

Zitieren

APA

Ma, H., Gu, Y., Zhang, W., Yang, L., Zhang, B., Hao, J., Chen, M., Li, J., Liu, X., Tang, S., & He, Q. (2026). RSMIA‐Net : A Novel Residual Skip‐Enhanced Network With Multi‐Dimensional Interactive Attention for Pulmonary Embolism Detection in CT Pulmonary Angiograms. International Journal of Imaging Systems and Technology, 36(6). https://doi.org/10.1002/ima.70476

Vancouver

Ma H, Gu Y, Zhang W, Yang L, Zhang B, Hao J, et al. RSMIA‐Net : A Novel Residual Skip‐Enhanced Network With Multi‐Dimensional Interactive Attention for Pulmonary Embolism Detection in CT Pulmonary Angiograms. International Journal of Imaging Systems and Technology. 2026;36(6). doi:10.1002/ima.70476

BibTeX
@article{Ma2026,
  author = {Ma, Hao and Gu, Yu and Zhang, Wei and Yang, Lidong and Zhang, Baohua and Hao, Juan and Chen, Meng and Li, Jianjun and Liu, Xin and Tang, Siyuan and He, Qun},
  title = {RSMIA‐Net : A Novel Residual Skip‐Enhanced Network With Multi‐Dimensional Interactive Attention for Pulmonary Embolism Detection in CT Pulmonary Angiograms},
  journal = {International Journal of Imaging Systems and Technology},
  year = {2026},
  volume = {36},
  number = {6},
  doi = {10.1002/ima.70476}
}
Kostenlos starten →