A Hybrid Transformers-based Convolutional Neural Network Model for Keratoconus Detection in Scheimpflug-based Dynamic Corneal Deformation Videos

Hazem Abdelmotaal ¹, Rossen Mihaylov Hazarbasanov ^2,3, Ramin Salouti ⁴

¹Department of Ophthalmology, Assiut University, Assiut, Egypt.
²Hospital de Olhos-CRO, Guarulhos, SP, Brazil.
³Department of Ophthalmology and Visual Science, Paulista School of Medicine, Federal University of S ao Paulo, SP, Brazil.
⁴Poostchi Ophthalmology Research Center, Shiraz University of Medical Sciences, Shiraz, Iran.
⁵Ruhr University, Bochum, Germany.
⁶Zentrum für Refraktive Chirurgie, Muenster, Germany.
⁷Biomedical Engineering Department, Al-Khwarizmi College of Engineering, University of Baghdad, Iraq.
⁸Computers, Electronics and Automation Department, Stefan cel Mare University of Suceava, Romania.
⁹Department of Ophthalmology, Jichi Medical University, Tochigi, Japan.
¹⁰Department of Ophthalmology, University of Tennessee Health Science Center, Memphis, TN, USA.
¹¹Department of Genetics, Genomics, and Informatics, University of Tennessee Health Science Center, Memphis, TN, USA.

⁰Department of Ophthalmology, Assiut University, Assiut, Egypt.

Journal of Ophthalmic & Vision Research +

|

July 21, 2025

View abstract on PubMed

Summary

A new hybrid Transformer-CNN model accurately detects keratoconus from dynamic corneal deformation videos (DCDVs). This AI tool shows high sensitivity and specificity, offering potential for clinical use in diagnosing this eye condition.

Area Of Science

Ophthalmology
Artificial Intelligence
Medical Imaging

Background

Keratoconus is a progressive eye condition affecting corneal shape.
Accurate and early detection of keratoconus is crucial for effective management.
Dynamic corneal deformation videos (DCDVs) offer rich data for corneal analysis.

Purpose Of The Study

To evaluate a hybrid Transformer-based convolutional neural network (CNN) model for automated keratoconus detection.
To assess the model's performance using stand-alone Scheimpflug-based DCDVs.
To determine the clinical utility of AI in diagnosing keratoconus.

Main Methods

Utilized transfer learning for feature extraction from DCDVs.
Incorporated self-attention mechanisms to capture long-range dependencies in feature maps.
Classified DCDVs to directly identify keratoconus.
Validated model performance on two independent cohorts (275 and 546 subjects).

Main Results

Achieved 93% sensitivity and 84% specificity in keratoconus detection.
Obtained an Area Under the Curve (AUC) of 0.97 for the keratoconus probability score on external validation data.
Demonstrated high accuracy in discriminating between normal and keratoconic corneas.

Conclusions

The hybrid Transformer-CNN model exhibits high sensitivity and specificity for keratoconus detection using DCDVs.
The model's performance suggests significant potential for integration into clinical practice.
AI-driven analysis of DCDVs can enhance the diagnostic capabilities for keratoconus.

Keywords:

Corneal Imaging Deep Learning Keratoconus Scheimpflug Imaging Artificial Intelligence