Abstract
Purpose :
Age-related macular degeneration (AMD) is a progressive retinal disease and a common cause of blindness in people over age 50. AMD pathology can be detected by optical coherence tomography (OCT) and categorized into advanced and non-advanced AMD based on presence of geographic atrophy and/or neovascularization. Here we propose a stack-based ensemble deep learning method and demonstrate that it can improve non-advanced and advanced AMD detection using retina OCT scans.
Methods :
Images from 150 subjects and 278 eyes diagnosed as non-advanced or advanced AMD obtained by Zeiss Cirrus OCT (Carl Zeiss Meditec, Inc., Dublin, CA, USA) were used in this study. Low-quality images were excluded before processing. The cleaned dataset had 767 images labeled as non-advanced and 663 images labeled as advanced AMD. The quality of retina images improved by contrast limited adaptive histogram equalization. Each image was divided into 4 patches without overlap between each other. Two center patches near the fovea were used, resulting in 2860 patch images in total. The processed images were split into 80% for training, 15% for validation, and 5% for testing. Base learners include 3 custom sequential models with 15, 23, and 25 hidden layers trained by stochastic gradient descent (SGD), rmsprop, and Adam optimizers, respectively. The weights calculated from each base model were used as an input feature for a meta model. Figure 1 shows the block diagram of the proposed model.
Results :
Classification results for the base and ensemble models are shown in Table 1. Model 3 trained by Adam achieved slightly higher accuracy (87%) than model 1 (85%) and model 2 (84%) using SGD and rmsprop optimizers. By stacking base learners, the ensemble model can improve the accuracy, specificity, sensitivity up to 91%, 92%, 90.9%, respectively, on the test set.
Conclusions :
Stack-based ensemble deep learning can improve the detection of non-advanced and advanced AMD.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.