EUVIMEDEUVIMED entdecken
Crossref

AFHAL: Adaptive Federated Hybrid-Attention Learning for privacy-aware medical image segmentation

Ataberk Urfali, Sait Ali Uymaz, Ozgu Can

Biomedical Signal Processing and Control · 2027-02 · 130 · 111587

DOI: 10.1016/j.bspc.2026.111587 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Biomedical Signal Processing and Control, 2027-02. Verfasst von Ataberk Urfali und 2 weiteren Autor:innen. Die Arbeit nennt 68 Literaturangaben. Gefördert von: The Scientific and Technological Research Council of Türkiye.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Developing multi-institutional medical image segmentation models is challenging because privacy regulations prevent centralized data aggregation. Federated learning enables collaborative training without data sharing, but existing methods suffer from performance degradation under non-IID data distributions and from quality-blind aggregation strategies that weight client contributions by dataset size rather than model quality. AFHAL (Adaptive Federated Hybrid-Attention Learning) is proposed with two complementary innovations: (i) an adaptive aggregation mechanism that derives each institution’s weight from its local validation loss through inverse-loss weighting, favoring better-performing local models regardless of dataset size; and (ii) a Hybrid Attention U-Net integrating SE-based channel attention with spatial attention at each encoder level to emphasize diagnostically relevant features. AFHAL is evaluated on twelve public datasets spanning MRI, CT, X-ray, and endoscopy (18 segmentation tasks), treated as pseudo-institutions, with all federated baselines sharing the same backbone and training protocol over three random seeds. AFHAL achieves an average Dice score of 0.8799 with 0.99% loss versus centralized training, a 60% reduction in FedAvg’s performance gap; it ranks first in 14 of 18 tasks among the classical federated baselines and converges 18% faster. It also outperforms contrastive and personalized methods (MOON, pFedMe, Ditto) on average, remaining statistically on par with the strongest personalized baseline. Ablation and additional analyses confirm the complementarity of the two components, robustness under extreme heterogeneity, stable adaptive weights, negligible overhead, and empirical privacy protection. The achieved performance approaches or meets clinical thresholds for radiotherapy planning and volumetric assessment, providing an effective privacy-aware solution for multi-institutional medical image segmentation.

Förderung

  • The Scientific and Technological Research Council of Türkiye

Zitieren

APA

Urfali, A., Uymaz, S. A., & Can, O. (2027). AFHAL: Adaptive Federated Hybrid-Attention Learning for privacy-aware medical image segmentation. Biomedical Signal Processing and Control, 130, 111587. https://doi.org/10.1016/j.bspc.2026.111587

Vancouver

Urfali A, Uymaz SA, Can O. AFHAL: Adaptive Federated Hybrid-Attention Learning for privacy-aware medical image segmentation. Biomedical Signal Processing and Control. 2027;130:111587. doi:10.1016/j.bspc.2026.111587

BibTeX
@article{Urfali2027,
  author = {Urfali, Ataberk and Uymaz, Sait Ali and Can, Ozgu},
  title = {AFHAL: Adaptive Federated Hybrid-Attention Learning for privacy-aware medical image segmentation},
  journal = {Biomedical Signal Processing and Control},
  year = {2027},
  volume = {130},
  pages = {111587},
  doi = {10.1016/j.bspc.2026.111587}
}
Kostenlos starten →