Self-Attention-Based Xception Convolutional Neural Network for Keratoconus Stage Classification Using 2-D RGB Corneal Images
IEEE Access · 2026 · 14 · 104264-104280
DOI: 10.1109/access.2026.3712107 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Muhammed Sideeq Anwar und 1 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Keratoconus is an eye disease that can cause significant vision loss if not detected early. The traditional diagnostic procedure uses corneal tomography images, which provide multi-colour-coded maps and quantitative indices describing the rich structural measurements of the cornea. However, obtaining corneal tomography images requires expensive devices, such as the OCULUS Pentacam system. To increase diagnostic accessibility in low-income regions and reduce diagnostic costs, we first proposed using two-dimensional (2D) RGB colour images captured with a smartphone which is a low-cost imaging device. To detect the keratoconus stage using 2D RGB colour images and overcome the limitations of traditional convolutional neural networks, in this study, we developed an attention-based convolutional neural network derived by modifying the Xception architecture for automatic keratoconus stage classification. The proposed network incorporates query-key-value attention modules, an extended middle flow, and depthwise convolutional blocks to better preserve spatial information and enhance channel-wise feature interactions while maintaining a reasonable computational cost. Experiments were conducted on a novel collected labelled dataset comprising normal, mild, and advanced keratoconus cases to validate the proposed attention-enhanced Xception architecture. The proposed model achieved an overall classification accuracy of 92.86% and reported macro-average precision, recall, and F1 scores above 92%, demonstrating consistent performance across the three disease stage classes. Analysis of the confusion matrix demonstrated strong stage-based discrimination, with perfect identification of advanced cases and minimal misclassification between mild and normal stages. These findings demonstrate that the proposed attention-enhanced Xception architecture provides an effective and generalisable solution for early keratoconus screening, particularly in resource-constrained environments where access to advanced ophthalmic imaging systems is limited.
Zitieren
APA
Anwar, M. S., & Özbílge, E. (2026). Self-Attention-Based Xception Convolutional Neural Network for Keratoconus Stage Classification Using 2-D RGB Corneal Images. IEEE Access, 14, 104264-104280. https://doi.org/10.1109/access.2026.3712107
Vancouver
Anwar MS, Özbílge E. Self-Attention-Based Xception Convolutional Neural Network for Keratoconus Stage Classification Using 2-D RGB Corneal Images. IEEE Access. 2026;14:104264-104280. doi:10.1109/access.2026.3712107
BibTeX
@article{Anwar2026,
author = {Anwar, Muhammed Sideeq and Özbílge, Emre},
title = {Self-Attention-Based Xception Convolutional Neural Network for Keratoconus Stage Classification Using 2-D RGB Corneal Images},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {104264--104280},
doi = {10.1109/access.2026.3712107}
}
EUVIMED