A Neural Network for the Prediction of the Visual Acuity Gained from Vitrectomy and Peeling for Epiretinal Membrane

A significant proportion of patients with epiretinal membrane (ERM) demonstrate improvement in visual acuity (VA) 3 months after pars plana vitrectomy (PPV) and membrane peeling. The identification of these patients before surgery is clinically relevant.

This retrospective study was conducted to establish a neural network to predict improvement using preoperative clinical factors and OCT.

A total of 427 eyes from 423 patients who underwent a PPV for primary idiopathic ERM combined with or without cataract surgery were included.

The data were automatically labeled according to whether an improvement of at least 2 logarithm of the minimum angle of resolution lines was observed. A multilayer perceptron was trained using a set of 7 clinical factors. The images were processed using a convolutional network. The output of both networks was concatenated and presented to a second multilayer perceptron. The dataset was divided into training, validation, and test datasets.

The accuracy of the neural network on an independent test dataset for the prediction of postoperative VA was analyzed. The impact of individual clinical factors and images on performance was assessed using ablation studies and class activation maps.

The clinical factors alone demonstrated the highest accuracy of 0.74, with a sensitivity of 0.82 and a specificity of 0.67. These results were obtained after the exclusion of less significant factors in an ablation study. The inclusion of the factors age, preoperative lens status, preoperative VA, and the distinction between combined phacovitrectomy and vitrectomy yielded the most accurate results. In contrast, the use of ResNet18 as a neural network for image processing alone (0.61) or images combined with clinical factors (0.70) resulted in reduced accuracy. In the class activation map, image regions corresponding to the outer, central, and inner retina appeared to be important for the decision-making process.

Our neural network has yielded favorable results in predicting improvement in VA in approximately 3-quarters of patients. This artificial intelligence-based personalized therapeutic strategy has the potential to aid decision-making. Future studies are to assess the clinical potential and generalizability and improve accuracy by including a more extensive dataset.

The author(s) have no proprietary or commercial interest in any materials discussed in this article.