DeepOCT: An explainable deep learning architecture to analyze macular edema on OCT images

Abstract

Macular edema (ME) is one of the most common retinal diseases that occur as a result of the detachment of the retinal layers on the macula. This study provides computer-aided identification of ME for even small pathologies on OCT using the advantages of Deep Learning. The study aims to identify ME on OCT images using a lightweight explainable Convolutional neural networks (CNN) architecture by composing significant feature activation maps and reducing the trainable parameters. A CNN is an effective Deep Learning algorithm, which consists of feature learning and classification stages. The proposed model, DeepOCT, focuses on reaching high classification performances as well as popular pre-trained architectures using less feature learning and shallow dense layers in addition to visualizing the most responsible regions and pathology on feature activation maps. The DeepOCT encapsulates the block-matching and 3D filtering (BM3D) algorithm, flattening the retinal layers to avoid the effects arising from different macula positions, and excluding non-retinal layers by cropping. DeepOCT identified OCT with ME with the rates of 99.20%, 100%, and 98.40% for accuracy, sensitivity, and specificity, respectively. The DeepOCT provides a standardized analysis, a lightweight architecture by reducing the number of trainable parameters, and high classification performances for both large- and small-scale datasets. It can analyze medical images at different levels with simple feature learning, whereas the literature uses complicated pre-trained feature learning architectures.