D-DAGNet: AN IMPROVED HYBRID DEEP NETWORK FOR AUTOMATED CLASSIFICATION OF GLAUCOMA FROM OCT IMAGES

Abstract

The introduction of Optical Coherence Tomography (OCT) in ophthalmology has resulted in significant progress in the early detection of glaucoma. Traditional approaches to identifying retinal diseases comprise an analysis of medical history and manual assessment of retinal images. Manual diagnosis is time-consuming and requires considerable human expertise, without which, errors could be costly to human sight. The use of artificial intelligence such as machine learning techniques in image analysis has been gaining ground in recent years for accurate, fast and cost-effective diagnosis from retinal images. This work proposes a Directed Acyclic Graph (DAG) network that combines Depthwise Convolution (DC) to decisively recognize early-stage retinal glaucoma from OCT images. The proposed method leverages the benefits of both depthwise convolution and DAG. The Convolutional Neural Network (CNN) information obtained in the proposed architecture is processed as per the partial order over the nodes. The Grad-CAM method is adopted to quantify and visualize normal and glaucomatous OCT heatmaps to improve diagnostic interpretability. The experiments were performed on LFH_Glaucoma dataset composed of 1105 glaucoma and 1049 healthy OCT scans. The proposed faster hybrid Depthwise-Directed Acyclic Graph Network (D-DAGNet) achieved an accuracy of 0.9995, precision of 0.9989, recall of 1.0, F1-score of 0.9994 and AUC of 0.9995 with only 0.0047 M learnable parameters. Hybrid D-DAGNet enhances network training efficacy and significantly reduces learnable parameters required for identification of the features of interest. The proposed network overcomes the problems of overfitting and performance degradation due to accretion of layers in the deep network, and is thus useful for real-time identification of glaucoma features from retinal OCT images.