EUVIMEDEUVIMED entdecken
Crossref

A Trustworthy and Clinically Interpretable Framework for Invasive Ductal Carcinoma Grading: Mitigating Automation Bias via Robust Optimization and Faithful XAI

Hossam Magdy Balaha, Nourhan Saad, Ali Mahmoud, Dibson Gondim, Mohammed Ghazal, Ayman El-Baz

Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 1174-1181

DOI: 10.1609/aaaiss.v10i1.43541 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Hossam Magdy Balaha und 5 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate histological grading of Invasive Ductal Carcinoma (IDC) is essential for prognosis but still suffers from inter-observer variability and domain shifts in histopathological imaging. Deep Learning shows promise, but standard first-order optimizers can suffer catastrophic performance collapse under aggressive stain variations, and "black-box" predictions undermine clinical trust and encourage automation bias. We introduce a robust end-to-end framework implementation utilizing the EVA-02 Vision Transformer (ViT), trained with the Sophia 2nd-order optimizer and Class-Balanced Focal Loss. This combination ensures stable and curvature-aware updates, avoiding the gradient instability that plagues AdamW or Lion. Additionally, we incorporate seven Explainable AI (XAI) methods (Integrated Gradients, SmoothGrad, RISE, Feature Ablation, and ViT adaptations of Grad-CAM, XGrad-CAM, and Eigen-CAM) to confirm that the model focuses on morphological hallmarks rather than spurious staining artifacts. Our method achieves state-of-the-art Weighted Accuracy of 98.90% at 20x and 98.78% at 40x on the PathoIDCG dataset without relying on error-prone upstream nuclei segmentation. This framework mitigates the automation bias by delivering transparent, artifact-resistant explanations and offers the clinical workforce a trustworthy human-in-the-loop decision support tool.

Zitieren

APA

Balaha, H. M., Saad, N., Mahmoud, A., Gondim, D., Ghazal, M., & El-Baz, A. (2026). A Trustworthy and Clinically Interpretable Framework for Invasive Ductal Carcinoma Grading: Mitigating Automation Bias via Robust Optimization and Faithful XAI. Proceedings of the AAAI Symposium Series, 10(1), 1174-1181. https://doi.org/10.1609/aaaiss.v10i1.43541

Vancouver

Balaha HM, Saad N, Mahmoud A, Gondim D, Ghazal M, El-Baz A. A Trustworthy and Clinically Interpretable Framework for Invasive Ductal Carcinoma Grading: Mitigating Automation Bias via Robust Optimization and Faithful XAI. Proceedings of the AAAI Symposium Series. 2026;10(1):1174-1181. doi:10.1609/aaaiss.v10i1.43541

BibTeX
@article{Balaha2026,
  author = {Balaha, Hossam Magdy and Saad, Nourhan and Mahmoud, Ali and Gondim, Dibson and Ghazal, Mohammed and El-Baz, Ayman},
  title = {A Trustworthy and Clinically Interpretable Framework for Invasive Ductal Carcinoma Grading: Mitigating Automation Bias via Robust Optimization and Faithful XAI},
  journal = {Proceedings of the AAAI Symposium Series},
  year = {2026},
  volume = {10},
  number = {1},
  pages = {1174--1181},
  doi = {10.1609/aaaiss.v10i1.43541}
}
Kostenlos starten →