Intraocular lens calculation using machine learning and deep learning methods

Session Details

Session Title: IOL Power Calculations, Post-LASIK & Extreme Eyes

Session Date/Time: Tuesday 25/09/2018 | 14:00-16:00

Paper Time: 14:12

Venue: Room A2

First Author: : G.Debellemanière FRANCE

Co Author(s): :    D. Gatinel   A. Saad                 

Abstract Details

Purpose:

To determine the accuracy of data-based intraocular lens calculation methods in eyes presenting different preoperative condition, and to compare their result with SRK/T and Haigis formulas.

Setting:

Anterior Segment and Refractive Surgery Dept., Rothschild Foundation, Paris, France

Methods:

Preoperative data (including mean keratometry, axial length, anterior chamber depth, white-to-white diameter, central corneal thickness) as well as model and power of implanted intraocular lenses and 1-month objective postoperative refraction after cataract surgery were retrieved for 2500 eyes and used to train different machine learning models (XGBoost, Deep Neural Network, Support Vector Machine). Preoperative data and implanted lens characteristics were used as input to predict objective postoperative refraction as output. Accuracy of the trained models was assessed using a dataset of 1000 eyes leaved untouched until the end as a validation set to prevent overfitting.

Results:

XGBoost was the best-performing model and had a better accuracy in predicting the objective postoperative refraction than SRK/T and Haigis formulas, especially in eyes with previous refractive surgery or radial keratotomy (p < 0.05) and penetrating keratoplasty (p < 0.05).

Conclusions:

Data-based methods could help avoiding refractive surprises after cataract surgery in atypic eyes.

Financial Disclosure:

-