Corneal Power Estimation From Anterior Corneal Surface Placido Disc Images Using Artificial Intelligence

To develop an automated artificial intelligence (AI) framework using a vision transformer (ViT) model with a multi-output regression layer to predict K1 and K2 values directly from placido  disc images reflected from the anterior corneal surface.

Ophthalmology Department, Rabin Medical Center, Petah Tikva, Israel 

Anterior corneal surface placido disc images from the TOMEY-OA2000 device were collected and  preprocessed at the Rabin Medical Center, Petach Tikva, Israel, to enhance model performance. First, images were cropped to focus on the central corneal region. Next, a masking process was applied to isolate the mire rings. The dataset was partitioned into 5 folds. A ViT model, modified with a multi-output regression head, was trained to simultaneously estimate the K1 and K2 values. The model was trained using a 5-fold cross-validation technique to ensure robustness. Then, Model performance was assessed using mean squared error (MSE), mean absolute error (MAE), and the coefficient of determination (R²).

9875 placido disc images were obtained from the TOMEY-OA2000 device. The K1 values ranged from 32.4 to 53.7 diopters, while K2 values spanned from 34.4 to 58.9 diopters. The model achieved preliminary results with an MSE of 24%, an MAE of 35%, and an R² score of 93% (all results were averaged over 5 folds). 

The proposed AI model provides a promising, non-invasive approach for automated estimation of K1 and K2 values from anterior corneal surface placido disc images. Its predictive accuracy suggests potential clinical utility in streamlining corneal assessments and better controlling suture related corneal astigmatism following penetrating or anterior lamellar keratoplasty by transforming keratoscopy to keratometry.