Ensemble Convolutional Neural Networks for the Classification and Visualization of Retinal Diseases in Optical Coherence Tomography Images

Optical Coherence Tomography (OCT) is a non-invasive imaging technique that uses light waves to capture cross-sectional images of patients' retina layers, allowing for the diagnosis of various retinal diseases. Ophthalmologists use OCT images to decide whether to perform anti-Vascular Endothelial Growth Factor therapy. However, it is time-consuming work to analyze the images since OCT provides several images for each patient. This paper proposes an ensemble learning (EL) model, based on three deep learning models, that categorize patients' OCT images into four categories such as Choroidal neovascularization, Diabetic macular edema, Drusen, and Normal. Four different Convolutional Neural networks (CNNs) are adapted to train the images. Among them, three CNNs are selected for the proposed EL model such as VGG19, ResNet152, and DenseNet121. Two different voting methods (soft and hard) are also used in the proposed EL model. The proposed EL model shows 0.9930 accuracy, 0.9930 sensitivity, and 0.9977 specificity. New heatmap algorithm is also proposed, based on positive and negative heatmaps, to analyze activity of CNN models and estimate regions of interest from the OCT images accurately. The proposed EL model and heatmap algorithms shows relatively good performance compared to other CNN models and heatmap algorithms. The proposed EL shows the potential to work as a second reader for ophthalmologist.