LKG-Net: lightweight keratoconus grading network based on corneal topography

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Keratoconus (KC) is a progressive corneal ectatic disease.
• Accurate KC severity grading is crucial for effective treatment.
• Current automated KC detection methods lack focus on severity grading.

Purpose of the Study:

• To develop a lightweight deep learning model for 4-level keratoconus grading (Normal, Mild, Moderate, Severe).
• To improve the accuracy and efficiency of automated KC severity assessment using corneal topography.

Main Methods:

• Proposed a novel Lightweight KC Grading Network (LKG-Net) utilizing depth-wise separable convolution and self-attention mechanisms for feature extraction.
• Implemented a multi-level feature fusion module to enhance feature representation.
• Evaluated LKG-Net on 488 eyes using 4-fold cross-validation.

Main Results:

• LKG-Net achieved high performance metrics: 89.55% weighted recall, 89.98% weighted precision, and 89.50% weighted F1 score.
• The model demonstrated a Kappa score of 94.38%, indicating strong agreement.
• LKG-Net also showed effectiveness in keratoconus screening tasks.

Conclusions:

• The proposed LKG-Net offers an effective and efficient solution for automated keratoconus severity grading.
• This AI-driven approach has the potential to significantly aid clinicians in managing KC patients.
• The model's lightweight design makes it suitable for practical clinical applications.