Deep-learning-based AI for evaluating estimated nonperfusion areas requiring further examination in ultra-widefield fundus images

Abstract

AbstractWe herein propose a PraNet-based deep-learning model for estimating the size of non-perfusion area (NPA) in pseudo-color fundus photos from an ultra-wide-field (UWF) image. We trained the model with focal loss and weighted binary cross-entropy loss to deal with the class-imbalanced dataset, and optimized hyperparameters in order to minimize validation loss. As expected, the resultant PraNet-based deep-learning model outperformed previously published methods. For verification, we used UWF fundus images with NPA and used Bland–Altman plots to compare estimated NPA with the ground truth in FA, which demonstrated that bias between the eNPA and ground truth was smaller than 10% of the confidence limits zone and that the number of outliers was less than 10% of observed paired images. The accuracy of the model was also tested on an external dataset from another institution, which confirmed the generalization of the model. For validation, we employed a contingency table for ROC analysis to judge the sensitivity and specificity of the estimated-NPA (eNPA). The results demonstrated that the sensitivity and specificity ranged from 83.3–87.0% and 79.3–85.7%, respectively. In conclusion, we developed an AI model capable of estimating NPA size from only an UWF image without angiography using PraNet-based deep learning. This is a potentially useful tool in monitoring eyes with ischemic retinal diseases.