A combined convolutional and recurrent neural network for enhanced glaucoma detection

Abstract

Glaucoma, a leading cause of blindness, is a multifaceted disease with several patho-physiological features manifesting in single fundus images (e.g., optic nerve cupping) as well as fundus videos (e.g., vascular pulsatility index). Current convolutional neural networks (CNNs) developed to detect glaucoma are all based on spatial features embedded in an image. We developed a combined CNN and recurrent neural network (RNN) that not only extracts the spatial features in a fundus image but also the temporal features embedded in a fundus video (i.e., sequential images). A total of 1810 fundus images and 295 fundus videos were used to train a CNN and a combined CNN and Long Short-Term Memory RNN. The combined CNN/RNN model reached an average F-measure of 96.2% in separating glaucoma from healthy eyes. In contrast, the base CNN model reached an average F-measure of only 79.2%. This proof-of-concept study demonstrates that extracting spatial and temporal features from fundus videos using a combined CNN and RNN, can markedly enhance the accuracy of glaucoma detection.

A Statistical Robust Glaucoma Detection Framework Combining Retinex, CNN, and DOE Using Fundus Images

“…Figures appear as a hyphen in C5 and C6 if the associated information are not clearly shown in the cited paper. All images, except those used in [9], are independent. Images adopted in [9] are sequential images…”

“…Unlike CNN, GLN, VGG, and RN152, RNN is designed for sequential data instead of independent data. For example, images adopted in [9] are sequential images, therefore a combination of CNN and RNN performs better than CNN alone.…”

“…Beginning in 2006, earlier works include [3], [14], [18], [2], [25], [5], [10], [15], [28], [24], [19], and [27]. Since 2018, more recent investigations include [7], [26], [8], [29], [9], [30], [32], and [4].…”

Image-based Glaucoma Classification Using Fundus Images and Deep Learning

“…GIST y PHOG (SVM) [8] Accuracy: 83.4 ROC-AUC: 88 VGG16 + LSTM [25] Sensitivity: 95 Specificity. 96 VGG19 + Transfer learning [26] ROC-AUC: 94.2 Sensitivity: 87 Specificity: 89.01…”

“…Gheisari et al [25] implement two architectures, the VGG16 and ResNet, concatenating LSTM blocks. To determine the best one, they carry out several experiments varying the number of epochs and learning rate.…”