Myopia categorization using PINNs

My Master’s thesis, in collaboration with The University of Burgos, develops a computational system to analyze myopia from a structural perspective of the human eye. Using a physical model based on Le Grand’s optical scheme and Physics-Informed Neural Networks (PINNs), the project seeks to categorize refractive errors from ocular biometric parameters with data from real users measured with a biometer in patients undergoing cataract surgery or lens replacement with intraocular lenses.

The project provides two different results:

• The first thing we did was to create a physical model based on Le Grand’s optical scheme and Physics-Informed Neural Networks (PINNs) to categorize refractive errors from ocular biometric parameters. Then we created two different methods to create regression models to predict the refractive errors and give each feature an importance score so we could understand which features were the most important for causing the refractive errors. The first method involved training a Physics-Informed Neural Network (PINN) to predict the refractive errors and the second method used a SHAP value analysis to give each feature an importance score. In both cases, the most important feature causing the refractive errors was the axial length. Other important features were the anterior corneal radius and the anterior chamber depth.

• This project also created a segmentation model that based on the predictions of the PINN model, it identified five different categories of myopia, each with distinct refractive error patterns. This enables the possibility of adapting the treatments for each category. It also predicts several features of the eye, such as the axial length and the length of the lens based on other biometric features, and gives each features an importance score with two different methods, where the axial length is the most important feature.