Automatic Retinal Image Analysis Research Articles

Automatic retinal image analysis methods using colour fundus images for screening glaucomatous optic neuropathy

ObjectivesTrain an automatic retinal image analysis (ARIA) method to screen glaucomatous optic neuropathy (GON) on non-mydriatic retinal images labelled with the additional results of optical coherence tomography (OCT) and assess...

Fundus Image-Label Pairs Synthesis and Retinopathy Screening via GANs With Class-Imbalanced Semi-Supervised Learning.

Retinopathy is the primary cause of irreversible yet preventable blindness. Numerous deep-learning algorithms have been developed for automatic retinal fundus image analysis. However, existing methods are usually data-driven, which rarely consider the costs associated with fundus image collection and annotation, along with the class-imbalanced distribution that arises from the relative scarcity of disease-positive individuals in the population. Semi-supervised learning on class-imbalanced data, despite a realistic problem, has been relatively little studied. To fill the existing research gap, we explore generative adversarial networks (GANs) as a potential answer to that problem. Specifically, we present a novel framework, named CISSL-GANs, for class-imbalanced semi-supervised learning (CISSL) by leveraging a dynamic class-rebalancing (DCR) sampler, which exploits the property that the classifier trained on class-imbalanced data produces high-precision pseudo-labels on minority classes to leverage the bias inherent in pseudo-labels. Also, given the well-known difficulty of training GANs on complex data, we investigate three practical techniques to improve the training dynamics without altering the global equilibrium. Experimental results demonstrate that our CISSL-GANs are capable of simultaneously improving fundus image class-conditional generation and classification performance under a typical label insufficient and imbalanced scenario. Our code is available at: https://github.com/Xyporz/CISSL-GANs.

An efficient approach to estimate the risk of coronary artery disease for people living with HIV using machine-learning-based retinal image analysis.

People living with HIV (PLWH) have increased risks of non-communicable diseases, especially cardiovascular diseases. Current HIV clinical management guidelines recommend regular cardiovascular risk screening, but the risk equation models are not specific for PLWH. Better tools are needed to assess cardiovascular risk among PLWH accurately. We performed a prospective study to determine the performance of automatic retinal image analysis in assessing coronary artery disease (CAD) in PLWH. We enrolled PLWH with ≥1 cardiovascular risk factor. All participants had computerized tomography (CT) coronary angiogram and digital fundus photographs. The primary outcome was coronary atherosclerosis; secondary outcomes included obstructive CAD. In addition, we compared the performances of three models (traditional cardiovascular risk factors alone; retinal characteristics alone; and both traditional and retinal characteristics) by comparing the area under the curve (AUC) of receiver operating characteristic curves. Among the 115 participants included in the analyses, with a mean age of 54 years, 89% were male, 95% had undetectable HIV RNA, 45% had hypertension, 40% had diabetes, 45% had dyslipidemia, and 55% had obesity, 71 (61.7%) had coronary atherosclerosis, and 23 (20.0%) had obstructive CAD. The machine-learning models, including retinal characteristics with and without traditional cardiovascular risk factors, had AUC of 0.987 and 0.979, respectively and had significantly better performance than the model including traditional cardiovascular risk factors alone (AUC 0.746) in assessing coronary artery disease atherosclerosis. The sensitivity and specificity for risk of coronary atherosclerosis in the combined model were 93.0% and 93.2%, respectively. For the assessment of obstructive CAD, models using retinal characteristics alone (AUC 0.986) or in combination with traditional risk factors (AUC 0.991) performed significantly better than traditional risk factors alone (AUC 0.777). The sensitivity and specificity for risk of obstructive CAD in the combined model were 95.7% and 97.8%, respectively. In this cohort of Asian PLWH at risk of cardiovascular diseases, retinal characteristics, either alone or combined with traditional risk factors, had superior performance in assessing coronary atherosclerosis and obstructive CAD. People living with HIV in an Asian cohort with risk factors for cardiovascular disease had a high prevalence of coronary artery disease (CAD). A machine-learning-based retinal image analysis could increase the accuracy in assessing the risk of coronary atherosclerosis and obstructive CAD.

Lifestyle Changes Reduced Estimated White Matter Hyperintensities Based on Retinal Image Analysis.

This study evaluates if there is an association between lifestyle changes and the risk of small vessel disease (SVD) as measured by cerebral white matter hyperintensities (WMH) estimated by the automatic retinal image analysis (ARIA) method. We recruited 274 individuals into a community cohort study. Subjects were assessed at baseline and annually with the Health-Promoting Lifestyle Profile II Questionnaire (HPLP-II) and underwent a simple physical assessment. Retinal images were taken using a non-mydriatic digital fundus camera to evaluate the level of WMH estimated by ARIA (ARIA-WMH) to measure the risk of small vessel disease. We calculated the changes from baseline to one year for the six domains of HPLP-II and analysed the relationship with the ARIA-WMH change. A total of 193 (70%) participants completed both the HPLP-II and ARIA-WMH assessments. The mean age was 59.1 ± 9.4 years, and 76.2% (147) were women. HPLP-II was moderate (Baseline, 138.96 ± 20.93; One-year, 141.97 ± 21.85). We observed a significant difference in ARIA-WMH change between diabetes and non-diabetes subjects (0.03 vs. -0.008, respectively, p = 0.03). A multivariate analysis model showed a significant interaction between the health responsibility (HR) domain and diabetes (p = 0.005). For non-diabetes subgroups, those with improvement in the HR domain had significantly decreased in ARIA-WMH than those without HR improvement (-0.04 vs. 0.02, respectively, p = 0.003). The physical activity domain was negatively related to the change in ARIA-WMH (p = 0.02). In conclusion, this study confirms that there is a significant association between lifestyle changes and ARIA-WMH. Furthermore, increasing health responsibility for non-diabetes subjects reduces the risk of having severe white matter hyperintensities.

Comprehensive automatic processing and analysis of adaptive optics flood illumination retinal images on healthy subjects.

This work presents a novel fully automated method for retinal analysis in images acquired with a flood illuminated adaptive optics retinal camera (AO-FIO). The proposed processing pipeline consists of several steps: First, we register single AO-FIO images in a montage image capturing a larger retinal area. The registration is performed by combination of phase correlation and the scale-invariant feature transform method. A set of 200 AO-FIO images from 10 healthy subjects (10 images from left eye and 10 images from right eye) is processed into 20 montage images and mutually aligned according to the automatically detected fovea center. As a second step, the photoreceptors in the montage images are detected using a method based on regional maxima localization, where the detector parameters were determined with Bayesian optimization according to manually labeled photoreceptors by three evaluators. The detection assessment, based on Dice coefficient, ranges from 0.72 to 0.8. In the next step, the corresponding density maps are generated for each of the montage images. As a final step, representative averaged photoreceptor density maps are created for the left and right eye and thus enabling comprehensive analysis across the montage images and a straightforward comparison with available histological data and other published studies. Our proposed method and software thus enable us to generate AO-based photoreceptor density maps for all measured locations fully automatically, and thus it is suitable for large studies, as those are in pressing need for automated approaches. In addition, the application MATADOR (MATlab ADaptive Optics Retinal Image Analysis) that implements the described pipeline and the dataset with photoreceptor labels are made publicly available.

Assessment of image quality on color fundus retinal images using the automatic retinal image analysis

Image quality assessment is essential for retinopathy detection on color fundus retinal image. However, most studies focused on the classification of good and poor quality without considering the different types of poor quality. This study developed an automatic retinal image analysis (ARIA) method, incorporating transfer net ResNet50 deep network with the automatic features generation approach to automatically assess image quality, and distinguish eye-abnormality-associated-poor-quality from artefact-associated-poor-quality on color fundus retinal images. A total of 2434 retinal images, including 1439 good quality and 995 poor quality (483 eye-abnormality-associated-poor-quality and 512 artefact-associated-poor-quality), were used for training, testing, and 10-ford cross-validation. We also analyzed the external validation with the clinical diagnosis of eye abnormality as the reference standard to evaluate the performance of the method. The sensitivity, specificity, and accuracy for testing good quality against poor quality were 98.0%, 99.1%, and 98.6%, and for differentiating between eye-abnormality-associated-poor-quality and artefact-associated-poor-quality were 92.2%, 93.8%, and 93.0%, respectively. In external validation, our method achieved an area under the ROC curve of 0.997 for the overall quality classification and 0.915 for the classification of two types of poor quality. The proposed approach, ARIA, showed good performance in testing, 10-fold cross validation and external validation. This study provides a novel angle for image quality screening based on the different poor quality types and corresponding dealing methods. It suggested that the ARIA can be used as a screening tool in the preliminary stage of retinopathy grading by telemedicine or artificial intelligence analysis.

Machine-learning method for localization of cerebral white matter hyperintensities in healthy adults based on retinal images.

Retinal vessels are known to be associated with various cardiovascular and cerebrovascular disease outcomes. Recent research has shown significant correlations between retinal characteristics and the presence of cerebral small vessel disease as measured by white matter hyperintensities from cerebral magnetic resonance imaging. Early detection of age-related white matter changes using retinal images is potentially helpful for population screening and allow early behavioural and lifestyle intervention. This study investigates the ability of the machine-learning method for the localization of brain white matter hyperintensities. All subjects were age 65 or above without any history of stroke and dementia and recruited from local community centres and community networks. Subjects with known retinal disease or disease influencing vessel structure in colour retina images were excluded. All subjects received MRI on the brain, and age-related white matter changes grading was determined from MRI as the primary endpoint. The presence of age-related white matter changes on each of the six brain regions was also studied. Retinal images were captured using a fundus camera, and the analysis was done based on a machine-learning approach. A total of 240 subjects are included in the study. The analysis of various brain regions included the left and right sides of frontal lobes, parietal–occipital lobes and basal ganglia. Our results suggested that data from both eyes are essential for detecting age-related white matter changes in the brain regions, but the retinal parameters useful for estimation of the probability of age-related white matter changes in each of the brain regions may differ for different locations. Using a classification and regression tree approach, we also found that at least three significant heterogeneous subgroups of subjects were identified to be essential for the localization of age-related white matter changes. Namely those with age-related white matter changes in the right frontal lobe, those without age-related white matter changes in the right frontal lobe but with age-related white matter changes in the left parietal–occipital lobe, and the rest of the subjects. Outcomes such as risks of severe grading of age-related white matter changes and the proportion of hypertension were significantly related to these subgroups. Our study showed that automatic retinal image analysis is a convenient and non-invasive screening tool for detecting age-related white matter changes and cerebral small vessel disease with good overall performance. The localization analysis for various brain regions shows that the classification models on each of the six brain regions can be done, and it opens up potential future clinical application.

Facilitating diabetic retinopathy screening using automated retinal image analysis in underresourced settings

AbstractAimTo evaluate an automated retinal image analysis (ARIA) of indigenous retinal fundus images against a human grading comparator for the classification of diabetic retinopathy (DR) status.MethodsIndigenous Australian adults with type 2 diabetes (n = 410) from three remote and very remote primary‐care services in the Northern Territory, Australia, underwent teleretinal DR screening. A single, central retinal fundus photograph (opportunistic mydriasis) for each eye was later regraded using a single ARIA and a UK human grader and national DR classification system. The sensitivity and specificity of ARIA were assessed relative to the comparator. Proportionate agreement and a Kappa statistic were also computed.ResultsRetinal images from 391 and 393 participants were gradable for ‘Any DR’ by the human grader and ARIA grader, respectively. ‘Any DR’ was detected by the human grader in 185 (47.3%) participants and by ARIA in 202 (48.6%) participants (agreement =88.0%, Kappa = 0.76,), whereas proliferative DR was detected in 31 (7.9%) and 37 (9.4%) participants (agreement = 98.2%, Kappa = 0.89,), respectively. The ARIA software had 91.4 (95% CI, 86.3–95.0) sensitivity and 85.0 (95% CI, 79.3–89.5) specificity for detecting ‘Any DR’ and 96.8 (95% CI, 83.3–99.9) sensitivity and 98.3 (95% CI, 96.4–99.4) specificity for detecting proliferative DR.ConclusionsThis ARIA software has high sensitivity for detecting ‘Any DR’, hence could be used as a triage tool for human graders. High sensitivity was also found for detection of proliferative DR by ARIA. Future versions of this ARIA should include maculopathy and referable DR (CSME and/or PDR). Such ARIA software may benefit diabetes care in less‐resourced regions.

Ground truth free retinal vessel segmentation by learning from simple pixels

Retinal vessel segmentation is fundamental for the automatic retinal image analysis and ocular disease screening. This paper aims to learn a ground truth free feature aggregation strategy for the vessel segmentation. Five vesselness maps modelling the vessels'profile, appearance, and shape are first generated. Together, the histogram of the local binary pattern and the green colour are extracted. In each vesselness map, the pixels with large vesselness values are regarded as simple positive samples. The pixels with small vesselness values are regarded as simple negative samples, and the pixels with mediocre values are treated as difficult pixels. The simple positive samples and simple negative samples near the difficult pixels consist of the training dataset while the rest vesselness maps together with the local binary pattern histogram, and green colour channel are used as the features to learn a strong classifier. Then, without leveraging any ground truth, multiple kernel boosting is used to combine four support vector machine kernels to learn a strong vessel model for each image. Applying this learnt model to the pixels with mediocre values in the single vesselness map, their label will be determined. Totally, five strong vessel models are learnt. Finally, pixels with the majority supports from the strong vessel models are labelled as vessel pixels. The proposed method achieves accuracy of 94.83%, sensitivity of 72.59%, and specificity of 98.11% on DRIVE dataset, and accuracy of 95.51%, sensitivity of 78.09%, and specificity of 97.56% on STARE. It outperforms the state-of-the-art unsupervised methods and achieves comparable performances to the supervised methods.

A machine learning approach for retinal images analysis as an objective screening method for children with autism spectrum disorder.

BackgroundAutism spectrum disorder (ASD) is characterised by many of features including problem in social interactions, different ways of learning, some children showing a keen interest in specific subjects, inclination to routines, challenges in typical communication, and particular ways of processing sensory information. Early intervention and suitable supports for these children may make a significant contribution to their development. However, considerable difficulties have been encountered in the screening and diagnosis of ASD. The literature has indicated that certain retinal features are significantly associated with ASD. In this study, we investigated the use of machine learning approaches on retinal images to further enhance the classification accuracy.MethodsForty-six ASD participants were recruited from three special needs schools and 24 normal control were recruited from the community. Among them, 23 age-gender matched ASD and normal control participant-pairs were constructed for the primary analysis. All retinal images were captured using a nonmydriatic fundus camera. Automatic retinal image analysis (ARIA) methodology applying machine-learning technology was used to optimise the information of the retina to develop a classification model for ASD. The model's validity was then assessed using a 10-fold cross-validation approach to assess its validity.FindingsThe sensitivity and specificity were 95.7% (95% CI 76.0%, 99.8%) and 91.3% (95% CI 70.5%, 98.5%) respectively. The area under the ROC curve was 0.974 (95% CI 0.934, 1.000); however, it was noted that the specificity for female participants might not be as high as that for male participants.InterpretationBecause ARIA is a fully automatic cloud-based algorithm and relies only on retinal images, it can be used as a risk assessment tool for ASD screening. Further diagnosis and confirmation can then be made by professionals, and potential treatment may be provided at a relatively early stage.

Association between obesity, common chronic diseases and health promoting lifestyle profiles in Hong Kong adults: a cross-sectional study

BackgroundRecent surveys revealed that the health status of many people from Hong Kong is far from ideal. Although non-communicable diseases are largely preventable, few relevant health promotion and disease prevention programs are available. Thus, we assessed the health indicators of Chinese adults in Hong Kong to investigate the relationship between obesity, common chronic diseases, and health-promoting lifestyle profiles to provide inspirations for decision makers in formulating targeted disease prevention and health management programs.MethodsThis is a secondary analysis of a data set of 270 community-dwelling Hong Kong adults who were within the eligible age range between 18 and 80 years without eye diseases that affect retinal photographs. The study exposure variable, health-promoting lifestyle profiles, was measured using the Health-Promoting Lifestyle Profile II (HPLP-II) questionnaire. The primary outcome variable, obesity, was defined using body mass index and waist-hip ratio. The secondary study outcome, estimated chronic diseases, including of anemia, chronic kidney disease, and cardiovascular disease, were estimated using automatic retinal image analysis from the retinal images. Data were analyzed using tests of proportion, the independent sample t-tests, Welch’s t-test, and binary logistic regression models.ResultsAll HPLP-II subscales had positive responses (≥ 2.5). Significant differences were noted between men and women in the health responsibility and nutrition subscales (Health Responsibility: p = 0.059; Nutrition: p = 0.067). Regression models revealed that nutrition (adjusted odds ratio [AOR] = 0.41; p = 0.017), physical activity (AOR = 0.50; p = 0.015), interpersonal relations (AOR = 2.14; p = 0.016), and stress management (AOR = 2.07; p 0.038) were associated with obesity; while spiritual growth (AOR = 0.24; p = 0.077) and interpersonal relations (AOR = 5.06; p 0.069) were associated with estimated chronic kidney disease.ConclusionsImproving health behaviors may control or alleviate the prevalence of obesity and chronic kidney disease. These findings could arouse concern about lifestyle behaviors and promote self-assessment of health-promoting lifestyles to the general public. The study also provided new insights into the relationship between the HPLP-II and other common chronic diseases that warrant further study.

Optic Disk Segmentation through Edge Density Filter in Retinal Images

Accurate and Efficient Detection of Optic Disk (OD) is an important task for the diagnosis of various eye related diseases such as Diabetic Retinopathy, Glaucoma, and papilledema in automatic retinal image analysis. This paper presents a novel and robust OD detection framework which starts with the detection of OD pixel, a center of OD region. Next, based on the ODP, a sub-image extraction is accomplished and it is processed for blood vessel removal through morphological technique. Finally, a new filter, called as Edge Density Filter (EDF) is accomplished to find out the extract the OD region in the Blood vessel removed, Edge detected and binarized sub-image. Two publicly available datasets, MESSIDOR and DRIVE were used to evaluate the performance of proposed method and the performance is measured through performance metrics such as Overlap Score, Sensitivity and Accuracy.

A review on automatic identification of fovea in retinal fundus images

Identification of retinal diseases is a very significant area of ophthalmology. Regular procedures are extremely specific, which rely on manual observation and highly prone to error. Hence, it is extremely fundamental to set up an automatic system for screening of vision threatening diseases like diabetic retinopathy (DR) and diabetic maculopathy (DM). Patients who are suffering from DR are at high risk to have DM which may lead to blindness, if not detected and treated appropriately at the appropriate time. Automatic analysis of retinal images requires knowledge and the properties of anatomical structures and retinal lesions. Thus, locating fovea plays a vital role in the analysis of retinal images. In recent times image processing has become a very effective tool for the detection and analysis of abnormalities in retinal images. This survey paper depicts the fundamental terminology related to automatic detection of macula and fovea. Literature review of various methods used for finding fovea in retinal fundus images is discussed. Detection issues involved in fovea are also discussed in this paper.

Microaneurysm detection in fundus images using a two-step convolutional neural network

Background and objectivesDiabetic retinopathy (DR) is the leading cause of blindness worldwide, and therefore its early detection is important in order to reduce disease-related eye injuries. DR is diagnosed by inspecting fundus images. Since microaneurysms (MA) are one of the main symptoms of the disease, distinguishing this complication within the fundus images facilitates early DR detection. In this paper, an automatic analysis of retinal images using convolutional neural network (CNN) is presented.MethodsOur method incorporates a novel technique utilizing a two-stage process with two online datasets which results in accurate detection while solving the imbalance data problem and decreasing training time in comparison with previous studies. We have implemented our proposed CNNs using the Keras library.ResultsIn order to evaluate our proposed method, an experiment was conducted on two standard publicly available datasets, i.e., Retinopathy Online Challenge dataset and E-Ophtha-MA dataset. Our results demonstrated a promising sensitivity value of about 0.8 for an average of >6 false positives per image, which is competitive with state of the art approaches.ConclusionOur method indicates significant improvement in MA-detection using retinal fundus images for monitoring diabetic retinopathy.

Quality and content analysis of fundus images using deep learning

Automatic retinal image analysis has remained an important topic of research in the last ten years. Various algorithms and methods have been developed for analysing retinal images. The majority of these methods use public retinal image databases for performance evaluation without first examining the retinal image quality. Therefore, the performance metrics reported by these methods are inconsistent. In this article, we propose a deep learning-based approach to assess the quality of input retinal images. The method begins with a deep learning-based classification that identifies the image quality in terms of sharpness, illumination and homogeneity, followed by an unsupervised second stage that evaluates the field definition and content in the image. Using the inter-database cross-validation technique, our proposed method achieved overall sensitivity, specificity, positive predictive value, negative predictive value and accuracy of above 90% when tested on 7007 images collected from seven different public databases, including our own developed database—the UoA-DR database. Therefore, our proposed method is generalised and robust, making it more suitable than alternative methods for adoption in clinical practice.

Correlation of vasogenic white matter lesions and retinal vascular network parameters

Objective: To investigate the correlation of vasogenic white matter lesions with retinal vascular network parameters using fully automatic retinal image analysis of fundus photographs. Methods: A total of 106 patients with cerebral small vessel disease who were hospitalized in Department of Neurology, the First Affiliated Hospital of Sun Yat-sen University during March and October 2015, and were able to undertake cerebral MRI and fundus photography in a sitting position were included. They were divided into two groups (mild or moderate-severe) according to the Fazekas scores of periventricular white matter lesions and deep white matter lesions shown by MRI. The clinical data and retinal vascular network parameters were compared between mild and moderate-severe groups. Results: According to the severity of periventricular white matter lesions, Logistic regression analysis showed that after adjusting baseline information, decreased asymmetry index of artery (OR=1.71, 95%CI 1.02-2.88, P<0.05)was associated with periventricular white matter lesions. As for deep white matter lesions, Logistic regression analysis showed that after adjusting baseline information, decreased central retinal artery equivalent(OR=5.19, 95%CI 1.06-25.44, P<0.05), decreased asymmetry index of artery (OR=2.96, 95%CI 1.42-6.17, P<0.05), decreased asymmetry index of venule (OR=2.99, 95%CI 1.48-6.02, P<0.05) and increased central retinal vein equivalent (OR=0.14, 95%CI 0.03-0.67, P<0.05) were associated with deep white matter lesions. Conclusions: White matter lesions of different places could be contributed to different pathological process. Therefore, the early diagnosis and observation of them are applicable to different retinal vascular network parameters.

Retinal blood vessel segmentation based on heuristic image analysis

Automatic analysis of retinal fundus images is becoming increasingly present today, and diseases such as diabetic retinopathy and age-related macular degeneration are getting a higher chance of being discovered in the early stages of their development. In order to focus on discovering those diseases, researchers commonly preprocess retinal fundus images in order to detect the retinal landmarks - blood vessels, fovea and the optic disk. A large number of methods for the automatic detection of retinal blood vessels from retinal fundus images already exists, but many of them are using unnecessarily complicated approaches. In this paper we demonstrate that a reliable retinal blood vessel segmentation can be achieved with a cascade of very simple image processing methods. The proposed method puts higher emphasis on high specificity (i.e. high probability that the segmented pixels actually belong to retinal blood vessels and are not false positive detections) rather than on high sensitivity. The proposed method is based on heuristically determined parametric edge detection and shape analysis, and is evaluated on the publicly available DRIVE and STARE datasets on which it achieved the average accuracy of 96.33% and 96.10%, respectively.

Retinal image analytics detects white matter hyperintensities in healthy adults.

ObjectiveWe investigated whether an automatic retinal image analysis (ARIA) incorporating machine learning approach can identify asymptomatic older adults harboring high burden of white matter hyperintensities (WMH) using MRI as gold standard.MethodsIn this cross‐sectional study, we evaluated 180 community‐dwelling, stroke‐, and dementia‐free healthy subjects and performed ARIA by acquiring a nonmydriatic retinal fundus image. The primary outcome was the diagnostic performance of ARIA in detecting significant WMH on MRI brain, defined as age‐related white matter changes (ARWMC) grade ≥2. We analyzed both clinical variables and retinal characteristics using logistic regression analysis. We developed a machine learning network model with ARIA to estimate WMH and its classification.ResultsAll 180 subjects completed MRI and ARIA. The mean age was 70.3 ± 4.5 years, 70 (39%) were male. Risk factor profiles were: 106 (59%) hypertension, 31 (17%) diabetes, and 47 (26%) hyperlipidemia. Severe WMH (global ARWMC grade ≥2) was found in 56 (31%) subjects. The performance for detecting severe WMH with sensitivity (SN) 0.929 (95% CI from 0.819 to 0.977) and specificity (SP) 0.984 (95% CI from 0.937 to 0.997) was excellent. There was a good correlation between WMH volume (log‐transformed) obtained from MRI versus those estimated from retinal images using ARIA with a correlation coefficient of 0.897 (95% CI from 0.864 to 0.922).InterpretationWe developed a robust algorithm to automatically evaluate retinal fundus image that can identify subjects with high WMH burden. Further community‐based prospective studies should be performed for early screening of population at risk of cerebral small vessel disease.

Multi-Cell Multi-Task Convolutional Neural Networks for Diabetic Retinopathy Grading.

Diabetic Retinopathy (DR) is a non-negligible eye disease among patients with Diabetes Mellitus, and automatic retinal image analysis algorithm for the DR screening is in high demand. Considering the resolution of retinal image is very high, where small pathological tissues can be detected only with large resolution image and large local receptive field are required to identify those late stage disease, but directly training a neural network with very deep architecture and high resolution image is both time computational expensive and difficult because of gradient vanishing/exploding problem, we propose a Multi-Cell architecture which gradually increases the depth of deep neural network and the resolution of input image, which both boosts the training time but also improves the classification accuracy. Further, considering the different stages of DR actually progress gradually, which means the labels of different stages are related. To considering the relationships of images with different stages, we propose a Multi-Task learning strategy which predicts the label with both classification and regression. Experimental results on the Kaggle dataset show that our method achieves a Kappa of 0.841 on test set which is the 4th rank of all state-of-the-arts methods. Further, our Multi-Cell Multi-Task Convolutional Neural Networks (M2CNN) solution is a general framework, which can be readily integrated with many other deep neural network architectures.

Development of an automatic retinal image analysis (ARIA) for screening cerebral small vessel disease in the community

Background: Cerebral white matter hyperintensity (WMH) on MRI of elderly subjects is associated with an increased risk for incident stroke and dementia. WMH is considered a manifestation of sporadic small vessel disease (SVD). Since it commonly harbors in the brain for years before clinical symptoms appear, this subclinical phase provides a therapeutic window for exploring preventive treatments. However, currently there is no economical, convenient and sensitive screening tool for WMH.