A Deep Feature Fusion of Improved Suspected Keratoconus Detection with Deep Learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early detection of keratoconus (KCN) is difficult, even for specialists.
• Accurate diagnosis of subclinical KCN is crucial for timely intervention.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for detecting early and established keratoconus (KCN).
• To improve the accuracy and robustness of KCN detection using fused corneal map features.

Main Methods:

• Utilized Xception and InceptionResNetV2 deep learning architectures for feature extraction from corneal maps.
• Fused features from DL models to enhance subclinical KCN detection.
• Trained and validated the model on datasets from Egypt (1371 eyes) and Iraq (213 eyes).

Main Results:

• Achieved an Area Under the ROC Curve (AUC) of 0.99 and 97-100% accuracy in the Egyptian dataset.
• Validated with AUCs of 0.91-0.92 and 88-92% accuracy in the independent Iraqi dataset.
• Demonstrated high performance in distinguishing normal eyes from those with subclinical and established KCN.

Conclusions:

• The proposed DL model shows significant promise for improving the detection of clinical and subclinical keratoconus.
• This AI-driven approach offers a more accurate and robust method for diagnosing KCN.
• The findings suggest a potential advancement in ophthalmological diagnostic tools for corneal diseases.