EUVIMEDEUVIMED entdecken
Crossref

3D-AUNet-DS: A Residual Attention U-Net with Deep Supervision and Composite Loss for Multi-Class Brain Tumour Segmentation on BraTS 2020

Yihong Zhang, Ameer Hamza, Md Saifur Rahman, Shijun Sun, Yaoyao Ran, Ali Sajid, Muhammad Abubakr

International Journal of Innovative Science and Research Technology · 21.05.2026 · 796

DOI: 10.38124/ijisrt/26may529 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Innovative Science and Research Technology, online erschienen am 21.05.2026. Verfasst von Yihong Zhang und 6 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate segmentation of brain tumour sub-regions from multi-modal MRI remains clinically challenging, as conventional three-dimensional U-Net architectures are hampered by vanishing gradients, indiscriminate skip-connection propagation, and insufficient multi-scale supervision, collectively limiting robust delineation of heterogeneous tumour components on the BraTS 2020 benchmark. An Enhanced Three-Dimensional Attention U-Net with Deep Supervision is proposed, trained on 128³ isotropic volumes from three MRI modalities (FLAIR, T1ce, T2) using stochastic multitransform augmentation and a composite class-weighted Dice–focal–boundary loss to jointly address class imbalance and imprecise enhancing tumour delineation. The architecture incorporates a fully residual encoder–decoder backbone with graduated spatial dropout (0.10–0.30), soft spatial attention gates at every skip connection to suppress background activation, and a resolution-aware deep supervision scheme weights with morphological post-processing to enforce anatomical plausibility. The proposed method achieves Dice scores of 0.817, 0.811, and 0.846 for enhancing tumour, tumour core, and whole tumour respectively, with HD95 values of 2.95, 3.24, and 3.97 mm, demonstrating superior boundary precision over nnU-Net, Swin UNETR, TransUNet, H2NF-Net, and ACU-Net, confirming the clinical viability of the integrated framework for precise multi-class brain tumour segmentation.

Zitieren

APA

Zhang, Y., Hamza, A., Rahman, M. S., Sun, S., Ran, Y., Sajid, A., & Abubakr, M. (2026). 3D-AUNet-DS: A Residual Attention U-Net with Deep Supervision and Composite Loss for Multi-Class Brain Tumour Segmentation on BraTS 2020. International Journal of Innovative Science and Research Technology, 796. https://doi.org/10.38124/ijisrt/26may529

Vancouver

Zhang Y, Hamza A, Rahman MS, Sun S, Ran Y, Sajid A, et al. 3D-AUNet-DS: A Residual Attention U-Net with Deep Supervision and Composite Loss for Multi-Class Brain Tumour Segmentation on BraTS 2020. International Journal of Innovative Science and Research Technology. 2026:796. doi:10.38124/ijisrt/26may529

BibTeX
@article{Zhang2026,
  author = {Zhang, Yihong and Hamza, Ameer and Rahman, Md Saifur and Sun, Shijun and Ran, Yaoyao and Sajid, Ali and Abubakr, Muhammad},
  title = {3D-AUNet-DS: A Residual Attention U-Net with Deep Supervision and Composite Loss for Multi-Class Brain Tumour Segmentation on BraTS 2020},
  journal = {International Journal of Innovative Science and Research Technology},
  year = {2026},
  pages = {796},
  doi = {10.38124/ijisrt/26may529}
}
Kostenlos starten →