Abstract
Purpose :
To detect multiple retina diseases, including age-related macular degeneration (AMD), diabetic retinopathy (DR), and glaucoma from fundus images using a probabilistic deep convolutional neural network (CNN) model and to evaluate the generalizability of the model.
Methods :
We collected fundus photographs from patients with AMD, DR, and glaucoma, and healthy subjects from age-related eye disease studies (AREDS) dataset and multiple publicly available datasets. We developed a probabilistic deep learning model to detect four classes based on 12,418 images in the development dataset and validated the model based on two external datasets with 1305 and 600 images, respectively. The performance of the probabilistic model was compared with the conventional deep learning models.
Results :
The probabilistic CNN model was developed using the ResNet152-V2 architecture. The area under the receiver operating characteristic (AUC) of the probabilistic model based on the internal testing, first, and second external datasets were 0.84 (CI: 0.82-0.86), 0.81 (0.79-0.84) and 0.81 (0.80-0.83), respectively. The AUC of the conventional CNN model based on the internal testing, first, and second external datasets were 0.94 (0.92-0.96), 0.85 (0.83-0.87), 0.86 (0.85-0.87), respectively (Fig. 1). The gradient class activation map (CAM) of the probabilistic CNN mode focused more on the macula region for AMD detection, blood vessel and hemorrhages for DR, and optic disc for glaucoma detection (Fig. 2).
Conclusions :
Probabilistic CNN models can detect multiple retina diseases with a high level of generalizability with a slight compromise in performance compared with conventional CNN model. The model focused more on retinal regions corresponding to pathologies related to the underlying disease. The probabilistic model however provides both the likelihood and uncertainty of prediction that may advance the integration of CNN models in clinical settings.
This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.