Abstract
Purpose :
To evaluate a deep learning classifier for the automated classification of optical coherence tomography angiography images from diabetic retinopathy (DR) eyes with and without macular ischemia (MI).
Methods :
In this study, 241 3x3 mm superficial vascular complex (SVC) OCTA images of DR patients with and without MI, as well as healthy controls were used to train a multi-layer deep convolutional neural network (DenseNet121) to differentiate between healthy versus DR with and without MI. A two-class (DR versus healthy controls) and a three-class classification system (DR without MI, DR with MI, and healthy control) were implemented. In order to evaluate classification performances overall in the dataset, a 5-fold cross-validation method was used. Visualization of the output of the DL classifier on OCTA images was performed using two different methods: Smoothed Saliency Maps and GradCAM++.
Results :
The accuracy of the binary classification (DR versus healthy controls) was 90.87%, along with a sensitivity of 88.96% and a precision of 96.48%. The three-class classification (DR with MI, without MI, and controls) showed an accuracy of 85.89%. Healthy controls and DR with MI obtained the highest sensitivity, of 94.25% and 93.98%, respectively.
Conclusions :
This study describes the use of a deep learning-based model to automatically classify DR with/without MI on central 3x3mm OCTA images. Thus, this model may be a useful screening tool and help improve the clinical management of DR patients.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.