Deep Learning Based on Ensemble to Diagnose of Retinal Disease using Optical Coherence Tomography

Abstract

Optical coherence tomography (OCT) image plays an essential modality role in medical image analysis to diagnose various retinal diseases. Through the retinal cross-sectional OCT images, ophthalmologists need much time and effort in manual analysis to extract diagnostic features. In this study, we propose a method for the automatic diagnosis of five retinal diseases based on an ensemble of two transfer learning models to analyze OCT images. The individual predicted class probabilities of MobileNetV3Large are fused with the predicted class probabilities of ResNet50 to ensure robustness in the prediction. An image processing technique such as contrast limited adaptive histogram equalization (CLAHE) is applied to enhance the quality of OCT images before feeding them to neural network models. In addition, 5-fold cross-validation and EarlyStopping techniques are also used to prevent the overfitting of the training model. To conduct the experiment, a dataset containing 1999 OCT images is constructed.