Development and Validation of a Deep Learning System to Detect Glaucomatous Optic Neuropathy Using Fundus Photographs

Deployment of Artificial Intelligence in Real-World Practice: Opportunity and Challenge

Artificial intelligence has rapidly evolved from the experimental phase to the implementation phase in many image-driven clinical disciplines, including ophthalmology. A combination of the increasing availability of large datasets and computing power with revolutionary progress in deep learning has created unprecedented opportunities for major breakthrough improvements in the performance and accuracy of automated diagnoses that primarily focus on image recognition and feature detection. Such an automated disease classification would significantly improve the accessibility, efficiency, and cost-effectiveness of eye care systems where it is less dependent on human input, potentially enabling diagnosis to be cheaper, quicker, and more consistent. Although this technology will have a profound impact on clinical flow and practice patterns sooner or later, translating such a technology into clinical practice is challenging and requires similar levels of accountability and effectiveness as any new medication or medical device due to the potential problems of bias, and ethical, medical, and legal issues that might arise. The objective of this review is to summarize the opportunities and challenges of this transition and to facilitate the integration of artificial intelligence (AI) into routine clinical practice based on our best understanding and experience in this area.

A Retrospective Comparison of Deep Learning to Manual Annotations for Optic Disc and Optic Cup Segmentation in Fundus Photographs

Purpose

Optic disc (OD) and optic cup (OC) segmentation are fundamental for fundus image analysis. Manual annotation is time consuming, expensive, and highly subjective, whereas an automated system is invaluable to the medical community. The aim of this study is to develop a deep learning system to segment OD and OC in fundus photographs, and evaluate how the algorithm compares against manual annotations.

Methods

A total of 1200 fundus photographs with 120 glaucoma cases were collected. The OD and OC annotations were labeled by seven licensed ophthalmologists, and glaucoma diagnoses were based on comprehensive evaluations of the subject medical records. A deep learning system for OD and OC segmentation was developed. The performances of segmentation and glaucoma discriminating based on the cup-to-disc ratio (CDR) of automated model were compared against the manual annotations.

Results

The algorithm achieved an OD dice of 0.938 (95% confidence interval [CI] = 0.934–0.941), OC dice of 0.801 (95% CI = 0.793–0.809), and CDR mean absolute error (MAE) of 0.077 (95% CI = 0.073 mean absolute error (MAE)0.082). For glaucoma discriminating based on CDR calculations, the algorithm obtained an area under receiver operator characteristic curve (AUC) of 0.948 (95% CI = 0.920 mean absolute error (MAE)0.973), with a sensitivity of 0.850 (95% CI = 0.794–0.923) and specificity of 0.853 (95% CI = 0.798–0.918).

Conclusions

We demonstrated the potential of the deep learning system to assist ophthalmologists in analyzing OD and OC segmentation and discriminating glaucoma from nonglaucoma subjects based on CDR calculations.

Translational Relevance

We investigate the segmentation of OD and OC by deep learning system compared against the manual annotations.

Deep learning assisted detection of glaucomatous optic neuropathy and potential designs for a generalizable model

Purpose

To evaluate ways to improve the generalizability of a deep learning algorithm for identifying glaucomatous optic neuropathy (GON) using a limited number of fundus photographs, as well as the key features being used for classification.

Methods

A total of 944 fundus images from Taipei Veterans General Hospital (TVGH) were retrospectively collected. Clinical and demographic characteristics, including structural and functional measurements of the images with GON, were recorded. Transfer learning based on VGGNet was used to construct a convolutional neural network (CNN) to identify GON. To avoid missing cases with advanced GON, an ensemble model was adopted in which a support vector machine classifier would make final classification based on cup-to-disc ratio if the CNN classifier had low-confidence score. The CNN classifier was first established using TVGH dataset, and then fine-tuned by combining the training images of TVGH and Drishti-GS datasets. Class activation map (CAM) was used to identify key features used for CNN classification. Performance of each classifier was determined through area under receiver operating characteristic curve (AUC) and compared with the ensemble model by diagnostic accuracy.

Results

In 187 TVGH test images, the accuracy, sensitivity, and specificity of the CNN classifier were 95.0%, 95.7%, and 94.2%, respectively, and the AUC was 0.992 compared to the 92.8% accuracy rate of the ensemble model. For the Drishti-GS test images, the accuracy of the CNN, the fine-tuned CNN and ensemble model was 33.3%, 80.3%, and 80.3%, respectively. The CNN classifier did not misclassify images with moderate to severe diseases. Class-discriminative regions revealed by CAM co-localized with known characteristics of GON.

Conclusions

The ensemble model or a fine-tuned CNN classifier may be potential designs to build a generalizable deep learning model for glaucoma detection when large image databases are not available.

Enhanced Visualization of Retinal Microvasculature in Optical Coherence Tomography Angiography Imaging via Deep Learning

Background: To investigate the effects of deep learning denoising on quantitative vascular measurements and the quality of optical coherence tomography angiography (OCTA) images. Methods: U-Net-based deep learning denoising with an averaged OCTA data set as teacher data was used in this study. One hundred and thirteen patients with various retinal diseases were examined. An OCT HS-100 (Canon inc., Tokyo, Japan) performed a 3 × 3 mm² superficial capillary plexus layer slab scan centered on the fovea 10 times. A single-shot image was defined as the original image and the 10-frame averaged image and denoised image generated from the original image using deep learning denoising for the analyses were obtained. The main parameters measured were the OCTA image acquisition time, contrast-to-noise ratio (CNR), peak signal-to-noise ratio (PSNR), vessel density (VD), vessel length density (VLD), vessel diameter index (VDI), and fractal dimension (FD) of the original, averaged, and denoised images. Results: One hundred and twelve eyes of 108 patients were studied. Deep learning denoising removed the background noise and smoothed the rough vessel surface. The image acquisition times for the original, averaged, and denoised images were 16.6 ± 2.4, 285 ± 38, and 22.1 ± 2.4 s, respectively (P < 0.0001). The CNR and PSNR of the denoised image were significantly higher than those of the original image (P < 0.0001). There were significant differences in the VLD, VDI, and FD (P < 0.0001) after deep learning denoising. Conclusions: The deep learning denoising method achieved high speed and high quality OCTA imaging. This method may be a viable alternative to the multiple image averaging technique.

Artificial Intelligence to Detect Papilledema from Ocular Fundus Photographs

BACKGROUND

Nonophthalmologist physicians do not confidently perform direct ophthalmoscopy. The use of artificial intelligence to detect papilledema and other optic-disk abnormalities from fundus photographs has not been well studied.

METHODS

We trained, validated, and externally tested a deep-learning system to classify optic disks as being normal or having papilledema or other abnormalities from 15,846 retrospectively collected ocular fundus photographs that had been obtained with pharmacologic pupillary dilation and various digital cameras in persons from multiple ethnic populations. Of these photographs, 14,341 from 19 sites in 11 countries were used for training and validation, and 1505 photographs from 5 other sites were used for external testing. Performance at classifying the optic-disk appearance was evaluated by calculating the area under the receiver-operating-characteristic curve (AUC), sensitivity, and specificity, as compared with a reference standard of clinical diagnoses by neuro-ophthalmologists.

RESULTS

The training and validation data sets from 6779 patients included 14,341 photographs: 9156 of normal disks, 2148 of disks with papilledema, and 3037 of disks with other abnormalities. The percentage classified as being normal ranged across sites from 9.8 to 100%; the percentage classified as having papilledema ranged across sites from zero to 59.5%. In the validation set, the system discriminated disks with papilledema from normal disks and disks with nonpapilledema abnormalities with an AUC of 0.99 (95% confidence interval [CI], 0.98 to 0.99) and normal from abnormal disks with an AUC of 0.99 (95% CI, 0.99 to 0.99). In the external-testing data set of 1505 photographs, the system had an AUC for the detection of papilledema of 0.96 (95% CI, 0.95 to 0.97), a sensitivity of 96.4% (95% CI, 93.9 to 98.3), and a specificity of 84.7% (95% CI, 82.3 to 87.1).

CONCLUSIONS

A deep-learning system using fundus photographs with pharmacologically dilated pupils differentiated among optic disks with papilledema, normal disks, and disks with nonpapilledema abnormalities. (Funded by the Singapore National Medical Research Council and the SingHealth Duke-NUS Ophthalmology and Visual Sciences Academic Clinical Program.).

Effects of Study Population, Labeling and Training on Glaucoma Detection Using Deep Learning Algorithms

Purpose

To compare performance of independently developed deep learning algorithms for detecting glaucoma from fundus photographs and to evaluate strategies for incorporating new data into models.

Methods

Two fundus photograph datasets from the Diagnostic Innovations in Glaucoma Study/African Descent and Glaucoma Evaluation Study and Matsue Red Cross Hospital were used to independently develop deep learning algorithms for detection of glaucoma at the University of California, San Diego, and the University of Tokyo. We compared three versions of the University of California, San Diego, and University of Tokyo models: original (no retraining), sequential (retraining only on new data), and combined (training on combined data). Independent datasets were used to test the algorithms.

Results

The original University of California, San Diego and University of Tokyo models performed similarly (area under the receiver operating characteristic curve = 0.96 and 0.97, respectively) for detection of glaucoma in the Matsue Red Cross Hospital dataset, but not the Diagnostic Innovations in Glaucoma Study/African Descent and Glaucoma Evaluation Study data (0.79 and 0.92; P < .001), respectively. Model performance was higher when classifying moderate-to-severe compared with mild disease (area under the receiver operating characteristic curve = 0.98 and 0.91; P < .001), respectively. Models trained with the combined strategy generally had better performance across all datasets than the original strategy.

Conclusions

Deep learning glaucoma detection can achieve high accuracy across diverse datasets with appropriate training strategies. Because model performance was influenced by the severity of disease, labeling, training strategies, and population characteristics, reporting accuracy stratified by relevant covariates is important for cross study comparisons.

Translational Relevance

High sensitivity and specificity of deep learning algorithms for moderate-to-severe glaucoma across diverse populations suggest a role for artificial intelligence in the detection of glaucoma in primary care.

Machine learning applied to retinal image processing for glaucoma detection: review and perspective

INTRODUCTION

This is a systematic review on the main algorithms using machine learning (ML) in retinal image processing for glaucoma diagnosis and detection. ML has proven to be a significant tool for the development of computer aided technology. Furthermore, secondary research has been widely conducted over the years for ophthalmologists. Such aspects indicate the importance of ML in the context of retinal image processing.

METHODS

The publications that were chosen to compose this review were gathered from Scopus, PubMed, IEEEXplore and Science Direct databases. Then, the papers published between 2014 and 2019 were selected . Researches that used the segmented optic disc method were excluded. Moreover, only the methods which applied the classification process were considered. The systematic analysis was performed in such studies and, thereupon, the results were summarized.

DISCUSSION

Based on architectures used for ML in retinal image processing, some studies applied feature extraction and dimensionality reduction to detect and isolate important parts of the analyzed image. Differently, other works utilized a deep convolutional network. Based on the evaluated researches, the main difference between the architectures is the number of images demanded for processing and the high computational cost required to use deep learning techniques.

CONCLUSIONS

All the analyzed publications indicated it was possible to develop an automated system for glaucoma diagnosis. The disease severity and its high occurrence rates justify the researches which have been carried out. Recent computational techniques, such as deep learning, have shown to be promising technologies in fundus imaging. Although such a technique requires an extensive database and high computational costs, the studies show that the data augmentation and transfer learning techniques have been applied as an alternative way to optimize and reduce networks training.

Usefulness of Denoising Process to Depict Myopic Choroidal Neovascularisation Using a Single Optical Coherence Tomography Angiography Image

Quality of single optical coherence tomography angiography (OCTA) images of myopic choroidal neovascularisation (mCNV) is poorer than in averaged images, although obtaining averaged images takes much time. This study evaluated the clinical usefulness of novel denoising process for depicting mCNV. This study included 20 eyes of 20 patients with mCNV. Ten en face images taken in a 3 × 3 mm macular cube were obtained from outer-retina-to-choriocapillaris layer. Three image types were prepared for analysis; single images before and after the denoising process accomplished deep learning (single and denoising groups, respectively) and up to 10 images were averaged (averaging group). Pairwise comparisons showed vessel density, vessel length density, and fractal dimension (FD) were higher; whereas, vessel density index (VDI) was lower in single group than in denoising and averaging groups. Detectable CNV indices, contrast-to-nose ratio, and CNV diagnostic scores were higher in denoising and averaging groups than in single group. No significant differences were detected in VDI, FD, or CNV diagnostic scores between denoising and averaging groups. The denoising process can utilise single OCTA images to provide results comparable to averaged OCTA images, which is clinically useful for shortening examination times with quality similar to averaging.

Health Economic and Safety Considerations for Artificial Intelligence Applications in Diabetic Retinopathy Screening

Systematic screening for diabetic retinopathy (DR) has been widely recommended for early detection in patients with diabetes to address preventable vision loss. However, substantial manpower and financial resources are required to deploy opportunistic screening and transition to systematic DR screening programs. The advent of artificial intelligence (AI) technologies may improve access and reduce the financial burden for DR screening while maintaining comparable or enhanced clinical effectiveness. To deploy an AI-based DR screening program in a real-world setting, it is imperative that health economic assessment (HEA) and patient safety analyses are conducted to guide appropriate allocation of resources and design safe, reliable systems. Few studies published to date include these considerations when integrating AI-based solutions into DR screening programs. In this article, we provide an overview of the current state-of-the-art of AI technology (focusing on deep learning systems), followed by an appraisal of existing literature on the applications of AI in ophthalmology. We also discuss practical considerations that drive the development of a successful DR screening program, such as the implications of false-positive or false-negative results and image gradeability. Finally, we examine different plausible methods for HEA and safety analyses that can be used to assess concerns regarding AI-based screening.

Artificial Intelligence in Ophthalmology: Evolutions in Asia

Artificial intelligence (AI) has been studied in ophthalmology since availability of digital information in ophthalmic care. The significant turning point was availability of commercial digital color fundus photography in the late 1990s, which caused digital screening for diabetic retinopathy (DR) to take off. Automated Retinal Disease Assessment software was then developed using machine learning to detect abnormal lesions in fundus to screen DR. The use of this version of AI had not been generalized because the specificity at 45% was not high enough, although the sensitivity reached 90%. The recent breakthrough in machine learning is the invent of deep learning, which accelerates its performance to be on par with experts. The first 2 breakthrough studies on deep learning for screening DR were conducted in Asia. The first represented collaboration of datasets between Asia and the United States for algorithms development, whereas the second represented algorithms developed in Asia but validated in different populations across the world. Both found accuracy for detecting referable DR of >95%. Diversity and variety are unique strengths of Asia for AI studies. There are many more studies of AI ongoing in Asia not only as prospective deployments in DR but in glaucoma, age-related macular degeneration, cataract, and systemic disease, such as Alzheimer's disease. Some Asian countries have laid out plans for digital health care system using AI as one of the puzzle pieces for solving blindness. More studies on AI and digital health are expected to come from Asia in this new decade.