Using 3D Convolutional Neural Network and Corvis ST Corneal Dynamic Video for Detecting Forme Fruste Keratoconus

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Forme fruste keratoconus (FFKC) represents an early stage of keratoconus, often challenging to diagnose with standard methods.
• Early detection of FFKC is crucial for timely intervention and preventing disease progression.

Purpose of the Study:

• To evaluate the diagnostic performance of a three-dimensional convolutional neural network (3D CNN) for detecting forme fruste keratoconus (FFKC).

Main Methods:

• A dataset of 415 corneal dynamic videos was analyzed, including 150 FFKC cases and 265 normal controls.
• A 3D CNN model was developed and trained using Corvis ST (Oculus Optikgeräte GmbH) videos.
• Model performance was assessed using accuracy, AUC, F1 score, sensitivity, and specificity; Grad-CAM visualized model attention.

Main Results:

• The 3D CNN model achieved an accuracy of 87.95% in identifying FFKC.
• The area under the receiver operating characteristic curve (AUC) was 0.95, with an F1 score of 0.85.
• Sensitivity reached 83.33% and specificity was 90.57%.

Conclusions:

• The integration of 3D CNN with Corvis ST videos presents a novel method for differentiating FFKC from normal corneas.
• This AI-driven approach may provide valuable clinical insights for FFKC detection.
• Further external validation is necessary to confirm the model's generalizability for clinical implementation.