Deep learning-based keratoconus detection from Scheimpflug images.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Keratoconus is a progressive corneal ectasia.
• Early detection, especially forme fruste (FF) keratoconus, is crucial for timely intervention.
• Conventional topographic assessments may miss subtle, preclinical changes.

Purpose of the Study:

• To evaluate deep learning models for keratoconus detection using raw Scheimpflug images.
• To specifically assess the model's capability in identifying forme fruste keratoconus.
• To determine the diagnostic performance of AI in early keratoconus detection.

Main Methods:

• A deep learning model utilizing transfer learning with a VGG16 architecture was developed.
• The model was trained on a dataset of 22,750 raw Scheimpflug corneal images from 910 eyes.
• Preprocessing and data augmentation techniques were incorporated to enhance model robustness.

Main Results:

• The model achieved 90.70% accuracy for FF keratoconus classification (AUC 0.89).
• Sensitivity and specificity for FF keratoconus were 80.57% and 80.56%, respectively.
• For clinical keratoconus, the model demonstrated high performance with 93.28% sensitivity, 99.40% specificity, and an AUC of 1.00.

Conclusions:

• Deep learning applied to raw Scheimpflug images is effective for keratoconus detection.
• The approach shows significant potential for identifying early-stage structural changes characteristic of FF keratoconus.
• This AI-driven method may enhance the detection of preclinical keratoconus compared to traditional methods.