Abstract
Purpose :
An external validation of an automated Artificial Intelligence AI- and color fundus image-based model for detection of disc hemorrhage, a significant risk factor in glaucoma progression.
Methods :
A deep machine learning architecture named “EfficientNet B5”, pre-trained with the “ImageNet” dataset, was implemented to train a network to detect the disc hemorrhages in fundus images. For this, 150 images with disc hemorrhages and 650 normal or without disc hemorrhages were used to train and test the model. Full-color fundus images were first cropped using automated AI to get only the optic disc area of the retina. These images were then resized to 100x100 pixels. This helps in reducing the number of retinal features that the network encounters, given the relatively small size of training data. The image sets were randomly augmented at each epoch for variation with rotation, translation and sheering with noise addition that resulted in newly generated images upto 35 times the original number. An early stopping mechanism was employed wherein the training is stopped if no improvement in training loss is seen in 25 consecutive epochs. The network weights with the best training loss were saved. For external validation, we have used another dataset obtained from the Department of Ophthalmology at Icahn School of Medicine at Mount Sinai. This dataset consisted of 144 images with disc hemorrhage and 831 normal or without disc hemorrhages.
Results :
For detection of disc hemorrhage on the external validation dataset, we achieved 93.13% accuracy (95% CI: 91.35% to 94.64%) with a sensitivity of 71.53% (95% CI: 63.42% to 78.73%), a specificity of 96.87% (95% CI: 95.45% to 97.95%) and a kappa score of 0.71 (95% CI: 0.65 to 0.78).
Conclusions :
We have performed an external validation study of the automated disc hemorrhage detection from a color fundus image. In the future, we aim to improve the sensitivity by adding more data and study the tool prospectively with deployment in clinical settings.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.