Abstract
Purpose :
To evaluate the accuracy of different deep learning models for detecting glaucoma from fundus photographs.
Methods :
We used a dataset with 1707 fundus photographs from 919 normal eyes and 788 eyes with glaucoma. We developed five different deep learning architectures in Google Colaboratory to detect glaucoma based on fundus photographs. We randomly selected 80% of the fundus photographs for development of models and 20% of the fundus photographs for final testing and evaluation of models. We developed a customized convolutional neural network (CNN) with 13 layers and compared it against four of the existing CNN architectures including AlexNet, VGG16, DensNet121, and ResNet50. We used different accuracy metrics to compare the performance of developed models based on the testing subset (Table 1).
Results :
Table 1 shows the accuracy of different models in detecting glaucoma based on different CNN architectures using development and testing subsets. The customized CNN architecture achieved an accuracy of 87.2% and 92.6% on testing and development subsets, respectively. Figure 1 shows the receiver operating characteristics (ROC) curves of all CNN architectures. The area under the receiver operating characteristic curve (AUC) of the customized CNN was 81.8% and 96.5% based on testing and development subsets, respectively.
Conclusions :
We performed a pilot study to show that it is feasible to detect glaucoma from fundus photographs based on different deep learning architectures. However, two factors are critical in robustly evaluating deep learning models. The datasets would need to be representative of underlying decease characteristics. The model would need to be generalizable. Validating the customized CNN architecture based on larger and independent datasets is desirable and may augment clinical practice in glaucoma assessment.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.