Deep Learning-Based Diabetic Retinopathy Detection: A Survey

Introduction

Artificial Intelligence (AI) and associate technologies such as Machine Learning(ML) and Deep Learning (DL) are frequently used in e-commerce, Chabot, Logistics and Supply chain, and gained momentum in the healthcare industry also, Davenport, T., & Kalakota, R. (2019). They have the capacity to extract substantial resemblance with a dataset and it can be used in the early diagnosis of chronic diseases such as cancer detection, chronic respiratory diseases, Alzheimer’s and diabetes.

Utilizations of deep learning in medicine ranges from malignant growth screening, infection observing to customized treatment recommendations, Razzak et al. (2017). Chang, V. (2018a) proposed improved Map Reduced framework to inspect malignant tumors to identify its progress, weak spots identification, and validated a cost-effective, useful way of developing analytics and visualization to influence bioinformatics. Deep learning algorithms are used for early detection and these algorithms outperforms doctors. Recently AI based algorithm called Deep Learning based Automatic Detection (DLAD) is developed for the detection of anomalous cell growth in chest radiographs, Greenfield, D. (2020). Google AI healthcare has developed a histology analysis learning algorithm LYNA Lymph Node Assistant to identify metastatic breast cancer tumors from lymph node biopsies. Application of artificial intelligence (AI) to Alzheimer’s disease may help in early detection, Mishra SG et al. (2017). With recent advances in the medical imaging, Chang, V. (2018b) exhibited that visual inspection of fusion of the medical imaging and simulation techniques, it’s possible to identify genes that trigger cancers.

Figure 1.

Process flow in Machine Learning and Deep learning methods, (Xenon Stack, 2020)

The principle contrast between ML and DL is, ML technique of classification involves processes like image preprocessing, feature extraction and classification. DL algorithms take care of raw images, automated feature extraction and classification, Schmidt-Erfurth et al. (2018). The block diagram representation of ML and DL is as shown in Figure1, Xenon Stack (2020). Different sorts of deep learning algorithms are being used in research like Convolutional Neural Network (CNN), Deep Autoencoder (DA), Recurrent Neural Network (RNN) and its variant like LSTM and so forth, Shen et al. (2017), one such deep neural network i.e. typical CNN architecture is as shown in Figure 2.

Figure 2.

Convolutional Neural Network Architecture (Sumit Saha , 2020)

Currently, DR detection is based on clinical examination or evaluation of digital color fundus photographs of the retina. This eye screening is currently labor-intensive, time-consuming, as well as subjective and can be performed by a retina specialist either directly or by first using a fundus camera to capture the retinal image, followed by a screening process. The inspiration of this paper is to provide DL applications for Diabetic Retinopathy identification. We present the structure of the automatic detection system and underlying algorithms with detailed presentations of the deep network architecture at layer level.

The paper is organized into brief introduction to Diabetic Retinopathy, deep learning based detection of DR in detail, current work in DR detection at image and lesion level. Discussion and deep learning challenges section discusses various networks performance’s in terms of sensitivity(SEN), specificity(SPE), area under the curve(AUC), and accuracy (ACC) of the algorithms along with the dataset and it also discusses challenges in deep learning like explainable machine learning, and at last conclusion section wraps up with future directions of deep learning on DR detection.