Estimation of the IOL position using the full shape of the crystalline lens and machine learning algorithms.

Abstract

Purpose : Accurate pre-operative estimation of the post-operative location of the Intraocular Lens (IOL) implanted in cataract surgery (ELP) is crucial to select the IOL power that minimizes refractive errors. We used geometrical features of patient eyes obtained from preoperative OCT images, including the full shape of the crystalline lens, and machine learning (ML) algorithms for improving ELP estimation.

Methods : OCT images from IOLMaster700 (ZEISS) were obtained pre- and post -cataract surgery in 116 eyes (Flaum Eye Institute, Rochester, NY; Baylor College of Medicine, Houston, TX) from 80 patients (70±8 y/o; -11 D to 9 D pre-op spherical equivalent) implanted with five different IOLs (AcrySof n=37, Clareon n=25, and Vivity n=11 by Alcon; enVista n=14 by B&L; Tecnis n=29 by J&J). Measurements with pupils smaller than 3 mm were discarded, remaining 88 eyes (63 patients). Surface segmentation, distortion correction, and registration were performed to generate 3-D models of the eye. The full shape of the crystalline lens was estimated using the eigenlenses method (coefficients a1-a6; Martinez-Enriquez, BOE 2020). A total of 43 features were considered as candidate input to the ML estimation algorithm, including geometrical and clinical features. Post-op actual IOL position was quantified and used as ground truth. 100 independent experiments were run, for each of which a Gaussian Process Regression model was trained and evaluated with a 5-fold cross validation. A sequential feature selection algorithm (SFS) was used to obtain the most relevant features. SRK/T, Haigis, HofferQ, and intersection approach (IntApp) results were also obtained for comparison.

Results : From the whole set of features, the first 4 in the ranking chosen by the SFS were anterior chamber depth, IOL model, a1 coefficient (lens size), and vitreous chamber depth. The ELP mean absolute estimation error (MAE) was 121±105 µm with the proposed method (ELP-AI) using the first 4 features of the ranking, in comparison with 250±227 µm (SRK/T), 218±164 µm (Haigis), 209±136 µm (HofferQ), and 182±145 µm (IntApp). MAE provided by ELP-AI was significantly lower than those of SRK/T, Haigis, HofferQ, and IntApp (ANOVA p<0.05, Bonferroni).

Conclusions : Full shape crystalline lens features quantified from OCT images and ML algorithms improved the accuracy of IOL position estimation, which is important for improving refractive outcomes.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.