Predicting visual improvement after macular hole surgery: a combined model using deep learning and clinical features

Abstract

Purpose : Analysis of preoperative optical coherence tomography (OCT) in patients with macular hole (MH) can provide insight into surgical success and postoperative visual acuity (VA). The aim of this study was to assess the feasibility of deep learning (DL) methods to improve the prediction of VA improvement after MH surgery from a combined model using DL on high-definition (HD) OCT B-scans and clinical features.

Methods : This was a retrospective single-center study (CHU de Québec – Université Laval (Canada)). A DL convolutional neural network (CNN) trained using preoperative HD-OCT B-scans (ZEISS, Dublin, CA) of the macula and combined with a logistic regression model of preoperative clinical features was designed to predict VA increase ≥15 Early Treatment of Diabetic Retinopathy Study letters at 6 months post-vitrectomy in closed MH. A total of 121 MH with 242 HD-OCT B-scans and 484 clinical data points were used to train, validate and test the model. These were randomly split into a training set of 83 eyes (69%), a validation set of 21 eyes (17%), and a held-out test set of 17 eyes (14%). Prediction of VA increase was evaluated using the area under the receiver operating characteristic curve (AUROC) and F1 scores. We also extracted the weight of each input feature in the regression-based hybrid model.

Results : All performances are reported on the held-out test set. Using a regression on clinical features, the AUROC was 80.63, with a F1 score of 79.71. For the CNN, relying solely on the HD-OCT B-scans, the AUROC was 72.83 ± 14.57, with a F1 score of 61.47 ± 23.72. For our hybrid regression model using clinical features and CNN prediction, the AUROC was 81.94 ± 5.22, with a F1 score of 80.35 ± 7.68. In the hybrid model, the baseline VA was the most important feature (59.11 ± 6.94% of the model’s weight) while HD-OCT prediction was 9.59 ± 4.17%.

Conclusions : Both the clinical data and HD-OCT models had good discriminative performances. Combining both into a hybrid model did not significantly improve performance.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Overview of our proposed hybrid model. (Top) Illustration of the extraction of the OCT-based prediction from the trained DL model. (Bottom) Flow-chart representing the combination of clinical data and OCT-based data to predict the clinical ground truth. CVA: Corrected-visual acuity.