Optimizing IOL Calculators with Deep Learning Prediction of Total Corneal Astigmatism.

Area of Science:

• Ophthalmology
• Corneal topography
• Refractive surgery

Background:

• Accurate prediction of total corneal astigmatism (TCA) is crucial for refractive surgery planning.
• Current methods for predicting TCA from anterior corneal astigmatism (ACA) have limitations.
• Optimizing TCA prediction can enhance surgical outcomes and intraocular lens (IOL) selection.

Purpose of the Study:

• To identify the most accurate regression model for predicting TCA from ACA.
• To fine-tune the architecture of the best-performing model for enhanced predictive accuracy.
• To compare the performance of deep learning models against established TCA prediction formulas.

Main Methods:

• Retrospective analysis of 19,468 eyes using Pentacam HR corneal topography data.
• Evaluation of various regression learners, with a focus on deep neural networks (DNNs).
• Refinement of DNN architecture and comparison with a leading TCA prediction formula.

Main Results:

• The optimized deep learning model demonstrated superior performance in predicting TCA magnitude (R² = 0.9740, RMSE = 0.0963 D) compared to the leading formula (R² = 0.8590, RMSE = 0.2257 D).
• Axis prediction error was significantly reduced with the deep learning approach (average error = 4.74° vs. 12.8°).
• Deep learning models consistently showed smaller errors and better data clustering than traditional methods.

Conclusions:

• Deep learning techniques offer significantly improved accuracy for TCA prediction over traditional methods.
• This advanced predictive capability can lead to more precise corneal astigmatism calculations.
• Enhanced TCA prediction accuracy holds potential for improved IOL selection and overall surgical success in refractive procedures.