Intraocular Lens Position Estimation From The Full Shape Geometry Of The Crystalline Lens And Machine Learning Algorithms

To improve the pre-operative estimation of the post-operative location of the Intraocular Lens (IOL) implanted in cataract surgery using patient eye’s geometrical features obtained from pre-operative OCT images, including the full shape of the crystalline lens, and machine learning (ML) algorithms (the ELP-AI method). To compare the estimation error obtained with the ELP-AI method with that obtained with other IOL position estimation formulas published in the literature such as SRK/T, Haigis, HofferQ, and the intersection approach (IntAp, that estimates the IOL position from the intersection of the two circles that best fit the anterior and posterior crystalline lens surfaces visible through the pupil).

Instituto de Óptica, Consejo Superior de Investigaciones Científicas, Madrid, Madrid, Spain. Flaum Eye Institute, University of Rochester, Rochester, NY, United States. Department of Ophthalmology, Baylor College of Medicine, Houston, TX, United States. Ophthalmology Department, Fundación Jiménez Díaz University Hospital, Madrid, Spain.  

OCT images from IOL Master700 (ZEISS) were obtained pre- and post-cataract surgery in 164 eyes from 130 patients. Measurements with pupils smaller than 3 mm were discarded, leaving 126 eyes from 103 patients (71±9 yo; -11 D to 9 D pre-op spherical equivalent) implanted with five IOL models (AcrsySof n=28, Clareon n=56, and Acrysof IQ Vivity n=11 by Alcon; enVista n=8 by B&L; Tecnis n=23 by J&J). For each measured eye, 3-D models were obtained, and the full shape of the crystalline lens was estimated using the eigenlenses method (Martinez-Enriquez, BOE2020). Clinical and geometrical features obtained from the 3-D models were considered as candidate input to a ML estimation algorithm. Post-op actual IOL position was used as ground truth.

A feature-selection algorithm was used to obtain the most relevant input features for predicting the IOL position. The first 4 features in the ranking were anterior chamber depth, IOL model, first eigenlens coefficient (related to the general size of the crystalline lens), and axial length. The IOL position mean absolute estimation error (MAE) was calculated with the proposed ELP-AI method trained with those 4 features, with the SRK/T, Haigis, and HofferQ formulas, and with the intersection approach implemented in some commercial OCTs (IntAp).The MAE with ELP-AI (123±102 µm) was statistically significantly lower (ANOVA p<0.05, Bonferroni) than with SRK/T (250±210 µm), Haigis (215±155 µm), HofferQ (214±137 µm), and IntAp (167±134 µm).  

Geometrical features quantified from OCT images, including the full shape of the crystalline lens obtained with the eigenlenses method, were used along with clinical features to train a machine learning algorithm for the estimation of the IOL position after cataract surgery, the ELP-AI method. The general size of the lens, described with the first eigenlens coefficient, was selected as one of the most important features for improving the accuracy of the estimation. The estimation error with the proposed method was significantly lower than with SRK/T, Haigis, HofferQ, and IntAp. Given that IOL position estimation errors are a primary source of residual refraction in IOL implantation, ELP-AI should result in improved refractive outcomes.