Protostage-NET: A Clinically Guided Foundation Model with Prototypical Learning for Automated Amd Grading from Oct Images

Author First name, Last name, Institution

• Mohamed Elsharkawy, J.B. Speed School of Engineering
• Moumen El-Melegy, J.B. Speed School of Engineering
• Asem Ali, J.B. Speed School of Engineering
• Ali Mahmoud, J.B. Speed School of Engineering
• Mohammed Ghazal, Abu Dhabi University
• Ashraf Khalil, Zayed University, Abu Dhabi CampusFollow
• Wei Wang, University of Louisville School of Medicine
• Ayman El-Baz, J.B. Speed School of Engineering

Document Type

Conference Proceeding

Source of Publication

Proceedings International Symposium on Biomedical Imaging

Publication Date

4-8-2026

Abstract

Age-related macular degeneration (AMD) is a major cause of vision loss, where accurate staging is essential for treatment. However, most recent approaches for OCT images remain limited to binary classification and do not capture the full clinical spectrum of the disease. To address this gap, we propose ProtoStage-Net, an explainable foundation model framework for fine-grained AMD staging. Our method leverages the DINOv2 vision foundation model as a backbone, fully fine-tuned for OCT data, and introduces two clinically inspired attention mechanisms: the Clinically Guided Channel Attention (CGCA) module, which emphasizes diseaserelevant feature channels, and the Lesion Localization Module (LLM), which highlights pathological retinal regions. Classification is performed through a hierarchical prototypical learning strategy at binary, ternary, and detailed six-class levels, providing transparent similarity-based decisions. In addition, we incorporate temporal ordinal constraints to enforce clinically valid AMD progression pathways. We evaluate ProtoStage-Net on a cohort of 954 OCT B-scans acquired from three different imaging devices. Our framework achieves a macro F1-score of 93.49 %, surpassing the baseline foundation model by more than 2 % and outperforming recent state-of-the-art methods. Extensive experiments further demonstrate the clinical explainability of our approach, with saliency maps confirming that ProtoStage-Net accurately localizes lesion regions across different AMD grades.

DOI Link

10.1109/isbi61048.2026.11515452

ISBN

[9798331577636]

ISSN

1945-7928

Publisher

IEEE

Volume

2026-April

Disciplines

Computer Sciences

Keywords

Artificial intelligence (0.52) | Computer science (0.5) | Grading (engineering) (0.4) | Computer vision (0.38) | Medicine (0.38) | Foundation (evidence) (0.36) | Medical imaging (0.34) | Visualization (0.31) | Medical physics (0.26) | Optical coherence tomography (0.26) | Deep learning (0.26)

Scopus ID

105041628405

Recommended Citation

Elsharkawy, Mohamed; El-Melegy, Moumen; Ali, Asem; Mahmoud, Ali; Ghazal, Mohammed; Khalil, Ashraf; Wang, Wei; and El-Baz, Ayman, "Protostage-NET: A Clinically Guided Foundation Model with Prototypical Learning for Automated Amd Grading from Oct Images" (2026). All Works. 8196.
https://zuscholars.zu.ac.ae/works/8196

Indexed in Scopus

yes

Open Access

no