Exudate detection in fundus images using deeply-learnable features. (January 2019)

• Record Type: Journal Article
• Authors: Khojasteh, Parham; Passos Júnior, Leandro Aparecido; Carvalho, Tiago; Rezende, Edmar; Aliahmad, Behzad; Papa, João Paulo; Kumar, Dinesh Kant
• Is Part Of: Computers in biology and medicine. Volume 104(2019)
• Page Start: 62
• Page End: 69
• Publication Date: 2019-01
• Subjects: Exudate detection -- Deep learning -- Convolutional neural networks -- Deep residual networks -- Discriminative restricted Boltzmann machines -- Diabetic retinopathy; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2018.10.031
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: Presence of exudates on a retina is an early sign of diabetic retinopathy, and automatic detection of these can improve the diagnosis of the disease. Convolutional Neural Networks (CNNs) have been used for automatic exudate detection, but with poor performance. This study has investigated different deep learning techniques to maximize the sensitivity and specificity. We have compared multiple deep learning methods, and both supervised and unsupervised classifiers for improving the performance of automatic exudate detection, i.e., CNNs, pre-trained Residual Networks (ResNet-50) and Discriminative Restricted Boltzmann Machines. The experiments were conducted on two publicly available databases: (i) DIARETDB1 and (ii) e-Ophtha. The results show that ResNet-50 with Support Vector Machines outperformed other networks with an accuracy and sensitivity of 98 % and 0.99, respectively. This shows that ResNet-50 can be used for the analysis of the fundus images to detect exudates. Graphical abstract: Highlights: Automatic detection of exudates in fundus images is highly essential for assessment of diabetic retinopathy. Multiple deep learning techniques, supervised and unsupervised classifiers were investigated. Convolutional Neural Networks, pre-trained Networks and Discriminative Restricted Boltzmann Machines were used. The pre-trained network (ResNet-50) with Support Vector Machine outperforms other networks. The proposed method achieved performances with accuracy and sensitivity of 97.6% and 0.99, respectively.