Abstract
Purpose :
Macular edema (ME) is a common ocular manifestation of vascular retinal diseases, including diabetic retinopathy and retinal-vein occlusion (RVO), and it is an important cause of visual deterioration. Optical Coherence Tomography (OCT) is a high-resolution, noncontact, non-invasive imaging technique and it is the preferred imaging modality for diagnosis of ME. Currently, ME diagnosis depends on the subjective evaluation of OCT and the clinical experience of ophthalmologists. Because of the similarity of diabetic ME (DME) and RVO edema in OCT, some cases could be incorrectly classified. The lack of research on deep learning (DL) diagnosis of RVO edema with OCT images led us to propose a convolution Neural Network (CNN) model to differentiate between DME and RVO.
Methods :
After image preprocessing, the Kermany dataset was used to finetune the VGG-19 network, pre-trained on the ImageNet dataset, to classify the OCT images between 4 classes: CNV (choroid neo-vascularization), Drusen, DME and Normal. This network was used to classify the OCT images into three classes: Normal, RVO and DME. A total of 4035 oct images from 1023 patients from Centro Hospitalar Universitário de São João were selected, with 1605 images of normal eyes, 1128 images of eyes with RVO and 1302 eyes with DME. After image preprocessing, 3291 images were used for training and validation and 744 images were used for testing. Then, accuracy, precision, recall, specificity, f-beta score, f-score, receiver-operating characteristic (ROC) and AUROC were calculated. Data augmentation was applied, and the effect of different data augmentation methods and combinations was compared.
Results :
The proposed architecture provides an accuracy of 82.60%, precision of 82.36%, recall of 82.60%, f-beta score of 82.27%, f-score of 82.48% and AUROC of 92.03%. The ROC curve and the confusion matrix are shown in Figure 1 and 2, respectively.
Conclusions :
DL could distinguish DME from RVO edema, and it might be useful in clinical practice and retinal screening. In the future, new studies are needed to improve the accuracy of these models and to extend the application of machine-learning algorithms in the diagnosis of other macular diseases.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.