Prediction of postoperative IOL tilt using preoperative 3D OCT imaging and machine learning

Abstract

Purpose : Postoperative tilt can negatively impact the optical performance of intraocular lenses (IOLs). We estimated IOL tilt after cataract surgery using swept-source optical coherence tomography (SS-OCT) imaging of the crystalline lens and machine learning (ML) algorithms. We analyzed preoperative features to predict postoperative IOL tilt.

Methods : Pre/post-operative SS-OCT images (IOLMaster700, Zeiss, 6 meridians) were obtained in 129 eyes (spherical equivalent: -26D to 9D) from 94 patients (mean age: 70±8 years) at two institutions (Rochester, NY; Houston, TX). Measurements with pupil size less than 3 mm were excluded (n=96 eyes remaining). 3D eye models were created following surface segmentation, distortion correction, and registration. Geometrical and clinical features were obtained as candidate input to the ML estimation algorithm, including tilt magnitude (TM) and tilt direction (TD). TM and TD were also obtained from postoperative measurements and used as ground truth. 100 independent experiments were performed, on which a Ridge Regression model was trained and assessed with a 5-fold cross-validation. Relevant preoperative predictive features of postoperative TM and TD were chosen using a sequential feature selection algorithm. For comparison, results were also acquired using only preoperative TM and TD.

Results : Mean pre/post-operative TM were 3.9±1.1 and 5±1.3 degrees, respectively. Mean pre/post-operative TD were 13.6±16 and 16±16 deg, respectively. Postoperative TM and TD correlated with preoperative TM and TD, respectively (Rho=0.64, p<0.05; Rho=0.63, p<0.05). The mean absolute error (MAE) for postoperative TM prediction was 0.66 ± 0.01 deg across all experiments using 5 features (preoperative TM, equatorial position of the crystalline lens, pupil size, radius of curvature of the anterior cornea, and age), compared to 0.73±0.004 deg using only preoperative TM (paired t-test, p<0.05). The MAE for postoperative TD prediction was 9.27±0.1 deg across all experiments using 4 features (preoperative TD, pupil size, estimated full crystalline lens size, and preoperative TM), and 9.84±0.08 deg using only preoperative TD (paired t-test, p<0.05).

Conclusions : Preoperative crystalline lens tilt may reliably predict postsurgical IOL tilt. ML algorithms and features of the full-shape crystalline lens can improve IOL tilt estimation which may be useful in creating pseudophakic eye models and selecting IOLs.

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