CorNet: Autonomous feature learning in raw Corvis ST data for keratoconus diagnosis via residual CNN approach

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Keratoconus (KC) is a progressive corneal ectasia requiring accurate early diagnosis.
• The Corvis ST provides in vivo corneal biomechanical measurements, but diagnostic parameters can be improved.
• Current diagnostic methods may not fully leverage the rich data from corneal deformation analysis.

Purpose of the Study:

• To evaluate the efficacy of an end-to-end Convolutional Neural Network (CNN), termed CorNet, in diagnosing KC.
• To determine if integrating raw Corvis ST data with CNN can enhance KC detection compared to existing parameters.
• To investigate the diagnostic utility of raw corneal biomechanical data.

Main Methods:

• A dataset of 1786 Corvis ST raw data points (corneal surface elevation during dynamic deformation) from 1112 normal and 674 KC eyes was used.
• An end-to-end CNN (CorNet) architecture, inspired by ResNet, was developed and trained on this dataset.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was used for visualization, and performance was assessed using AUC, sensitivity, specificity, and F1 score.

Main Results:

• CorNet achieved an AUC of 0.971, outperforming the Corvis Biomechanical Index (CBI) AUC of 0.947 and cCBI AUC of 0.963.
• CorNet demonstrated high sensitivity (92.49%) and specificity (91.54%) in distinguishing KC from normal eyes.
• Grad-CAM analysis indicated that corneal deformation during the loading phase is more critical for KC diagnosis than during the unloading phase.

Conclusions:

• The proposed CorNet model effectively utilizes raw Corvis ST biomechanical data for KC detection.
• CorNet shows comparable or superior diagnostic performance to existing Corvis ST parameters.
• This AI-driven approach holds promise for advancing the diagnosis of keratoconus in clinical ophthalmology.