Abstract
Purpose :
The accurate preoperative estimation of the intraocular lens (IOL) position implanted in cataract surgery is important to correctly select the IOL power and thus for accurate refractive outcomes. We used geometrical features obtained from preoperative IOL Master700 images and a machine learning algorithm (Gaussian process regression, GPR) for improving the estimation of the IOL position.
Methods :
OCT images from IOLMaster700 (ZEISS, SS-OCT, 6 radial images per measurement) were obtained pre- and post -cataract surgery (Flaum Eye Institute, Rochester, NY) on 41 eyes from 21 subjects (67±7 y/o; -11 D to 4.5 D pre-op spherical error) implanted with three different IOLs (AcrsySof (n=27) & Clareon (n=12) by Alcon; enVista (n=2) by B&L).
Surface segmentation, distortion correction and registration were performed using custom algorithms to obtain 3-D models of the eye, including cornea, crystalline lens/IOL, iris and retina. Four sets of geometrical features were analyzed from pre-op: F_A: Radius of curvature of anterior cornea (RAC), standard axial length (AL) and IOL model implanted; F_B: RAC, AL using custom indices of refraction for the different structures of the eye and IOL model; F_C: F_B + axial measurements (anterior chamber and vitreous chamber depths; cornea and crystalline lens thickness); F_D: F_C + radius of curvature of anterior and posterior crystalline lens. Post-op actual IOL position was quantified and used as ground truth. A total of 1000 independent experiments were run. For each experiment, a GPR model (exponential kernel) was trained and evaluated with a 10-fold cross validation. Standard SRK/T results were also obtained for comparison.
Results :
The mean of absolute estimation errors (MAE) were 296±262 µm (standard SRK/T), 191±120 µm (F_A), 186±121 µm (F_B), 120±81 µm (F_C) and 117±88 µm (F_D). MAE using F_D and F_C was significantly lower than with SRK/T, F_A and F_B (ANOVA p<0.05, Bonferroni). Maximum error across subjects were 1.30 mm (SRK/T), 588 µm (F_A), 568 µm (F_B), 350 µm (F_C) and 369 µm (F_D). The error was higher than 200 µm in 23 eyes for SRK/T, in 15 eyes for F_A and F_B, in 5 eyes for F_C and in 6 eyes for F_D.
Conclusions :
Preoperative geometrical features quantified from OCT images in combination with machine learning algorithms improved the accuracy of the IOL position estimation in cataract surgery, which is critical for improving refractive outcomes.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.