Statistical Analysis of Deep Learning Models for Diabetic Macular Edema Classification using OCT Images

Abstract

Diabetic macular edema (DME) is a potentially blinding complication of Diabetic retinopathy (DR) and indeed the main cause of visual impairment in diabetic patients. DME can indeed be diagnosed in varying levels of severity by employing Optical Coherence Tomography (OCT), which is a standard imaging modality to capture the 3D view of the retina. Computerized detection of DME is beneficial, and automated identification can assist doctors in their daily activities. Deep Learning (DL), a widely recognized method in this regard, has contributed to improving the effectiveness of classification algorithms. The focus of this research is to use a standard OCT dataset to test and analyze two DL models, Optic Net and DenseNet for DME classification. A statistical analysis of the accuracy measures collected during the experiments is performed to evaluate the performance of the two models. The statistical findings suggest that the model Optic Net (Accuracy-98%, Specificity-100%) outperforms DenseNet (Accuracy-94%, Specificity-96%) in terms of accuracy, and the results could be used to choose an optimal model for DME detection.