Automatic optic nerve head localization and cup-to-disc ratio detection using state-of-the-art deep-learning architectures

Abstract

Computer vision has greatly advanced recently. Since AlexNet was first introduced, many modified deep learning architectures have been developed and they are still evolving. However, there are few studies comparing these architectures in the field of ophthalmology. This study compared the performance of various state-of-the-art deep-learning architectures for detecting the optic nerve head and vertical cup-to-disc ratio in fundus images. Three different architectures were compared: YOLO V3, ResNet, and DenseNet. We compared various aspects of performance, which were not confined to the accuracy of detection but included, as well, the processing time, diagnostic performance, effect of the graphic processing unit (GPU), and image resolution. In general, as the input image resolution increased, the classification accuracy, localization error, and diagnostic performance all improved, but the optimal architecture differed depending on the resolution. The processing time was significantly accelerated with GPU assistance; even at the high resolution of 832 × 832, it was approximately 170 ms, which was at least 26 times slower without GPU. The choice of architecture may depend on the researcher's purpose when balancing between speed and accuracy. This study provides a guideline to determine deep learning architecture, optimal image resolution, and the appropriate hardware.Recently, there have been considerable advances in computer technology and breakthroughs in artificial-intelligence algorithms for "deep learning". In the field of ophthalmology, automated image analysis using deep learning is attracting increasing attention, and researchers have reported that it can outperform a human doctor 1,2 . One of the significant innovations in image-classification deep-learning algorithms was AlexNet 3 , which was developed by Alex Krizhevsky. AlexNet competed in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2012 and was overwhelmingly victorious. This algorithm opened a new horizon of computer vision using deep convolutional neural networks (CNNs).Since AlexNet was developed, there have been numerous modifications to CNN architectures. One of these modifications-and another significant leap-was ResNet 4 , which is famous for its "residual connection. " Before ResNet, deep CNN architectures had a "degradation problem": with the increase of the depth of a neural network, the accuracy becomes saturated and decreases rapidly 5,6 . ResNet solved this problem via residual connection, which adds the output of the lower layer to the higher layer. This direct addition of the lower-layer output delivers undegraded information to the higher layer and makes deep neural network training more effective 4 . ResNet won 1 st place in the 2015 ILSVRC.Another state-of-the-art approach for solving the degradation problem is DenseNet, which was introduced by Huang et al. 7 in 2016. This architecture feed-forwards each layer to every other layer. ResNet also forwards the output of the lower layer directly to the higher layer; ho...

The use of deep learning technology for the detection of optic neuropathy

“…In the field of DL, there are two types of research related to the optic disc: one is related to the detection of the optic disc in the whole fundus image (19) and the other is related to the classification of the optic disc in the cropped image (20). While there has been much research on the optic disc classification algorithm, the target detection of the optic disc has not been thoroughly studied (21).…”

“…Park et al (21) studied and compared the performance of the most advanced DL architecture in detecting optic disc and VCDR in fundus images. The training data set was composed of 1959 eyes with normal fundus, glaucoma, and other optic neuropathy (in which VCDR >0.4 accounted for 94.3%) which were randomly divided into a training data set and a verification data set at a ratio of 9:1.…”

Enhancing the Performance of Vehicle Passenger Detection under Adverse Weather Conditions Using Augmented Reality-Based Machine Learning Approach

“…Each bounding box label contains the following information: classification (c), box center coordinates (x, y), box width (w), and box height (h) (15). Approximately 90% (8,778) of the images were used for training purposes, and the remaining 10% (975) of the images were adopted as the testing set in the holdout method (17)(18)(19). Ivasˇic´-Kos et al (17) compared the results of the different ratios of training and testing sets (90:10 vs. 80:20) in machine learning study and found no significant difference of performance between the two ratio settings.…”

“…Ivasˇic´-Kos et al (17) compared the results of the different ratios of training and testing sets (90:10 vs. 80:20) in machine learning study and found no significant difference of performance between the two ratio settings. Tao et al (18) and Park et al (19) used 90% of the total data set for training and 10% for testing the machine learning algorithms in their study and stated the ratio of 90:10 (training: testing) would be more effective for a traffic data set of less than 10,000 items. Furthermore, the cross-validation process was executed to include up to 30% of the total data set in the test data set.…”

Automatic Detection of Coronary Metallic Stent Struts Based on YOLOv3 and R-FCN

“…Wu et al [ 11 ] developed a deep learning model (BMSNet) with the YOLOv3 architecture for assisting haematologists in the interpretation of bone marrow smears for faster diagnosis and disease monitoring. Park et al [ 12 ] compared the performance of various state-of-the-art deep-learning architectures, including YOLOv3, for detecting the optic nerve head and vertical cup-to-disc ratio in fundus images. Safdar et al [ 13 ] highlighted the most suitable Data Augmentation technique for medical imaging by using YOLOv3.…”