Abstract
Purpose :
The application of artificial intelligence (AI) algorithms in the ophthalmology field has great potential for detection, early intervention, and management of patients with retinal, macular, and choroidal pathologies. This study aims to develop an AI-assisted automated system based on a deep learning architecture for early detection and recognition of OCT images with the presence of macular edema (ME), drusen and subretinal or choroidal neovascularization (CNV).
Methods :
We have utilized the OCT dataset published in Kaggle to train (59,352), validate (12,463) and test (12,669) the model. We have utilized another real-world database (RWD) (1,184) that was collected via several optometry clinics for testing. All the participants were enrolled after informed consent in accordance with an Institutional Review Board/Ethics Committee approved protocol. ResNet with 50 layers that is pre-trained on ImageNet is utilized in order to build our AI OCT model. Images are center-cropped, resized to 512x512 and normalized based on mean-standard deviation transformation method. Stochastic Gradient Descent is utilized as the optimizer for updating weights of the network and the binary crossentropy as the loss function to decay the weights. Our model is a binary classification model where we define an abnormal image with Drusen or/and CNV or/and ME, while a normal image does not have these three anomalies.
Results :
The model achieved the following values for accuracy, sensitivity, specificity, and Kappa on Kaggle and RWD test datasets: 98.93%, 85.22%; 99.23%, 94.52%; 98.33%, 74.31%; and 97.57%, 69.82%, respectively. Table 1 shows that our AI-model can achieve a low false negatives (35) that clearly contributes to the medical decision.
Conclusions :
The good performance of our model suggests that it can be used as a valuable tool for early detection of many retinal anomalies. AI programs using OCT images have a great potential to play a significant role in the early diagnosis and management of ophthalmological diseases.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.