Fundus Image Analysis Research Articles

Multi-Label Classification of Fundus Images With EfficientNet

Convolutional neural network (CNN) has achieved remarkable success in the field of fundus images due to its powerful feature learning ability. Computer-aided diagnosis can obtain information with reference value for doctors in clinical diagnosis or screening through proper processing and analysis of fundus images. However, most of the previous studies have focused on the detection of a certain fundus disease, and the simultaneous diagnosis of multiple fundus diseases still faces great challenges. We propose a multi-label classification ensemble model of fundus images based on CNN to directly detect one or more fundus diseases in the retinal fundus images. Every single model consists of two parts. The first part is a feature extraction network based on EfficientNet, and the second part is a custom classification neural network for multi-label classification problems. Finally, the output probabilities of different models are fused as the final recognition result. And it was trained and tested on the data set provided by ODIR 2019 (Peking University International Competition on Ocular Disease Intelligent Recognition). The experimental results show that our model can be trained on fewer data sets and get good results.

Segmentation of Diabetic Retinopathy Lesions: The Common Fallacy and Evaluation of Real Segmenters

In the context of automated analysis of eye fundus images, it is an important common fallacy that prior works achieve very high scores in segmentation of lesions, and that fallacy is fueled by some reviews reporting very high scores, and perhaps some confusion with terms. A simple analysis of the detail of the few prior works that really do segmentation reveals scores between 7% and 70% in sensitivity for 1 FPI. That is clearly sub-par with medical doctors trained to detect signs of Diabetic Retinopathy, since they can distinguish well the contours of lesions in Eye Fundus Images (EFI). Still, a full segmentation of lesions could be an important step for both visualization and further automated analysis using rigorous quantification or areas and numbers of lesions to better diagnose. I discuss what prior work really does, using evidence-based analysis, and confront with segmentation networks, comparing on the terms used by prior work to show that the best performing segmentation network outperforms those prior works. I also compare architectures to understand how the network architecture influences the results. I conclude that, with the correct architecture and tuning, the semantic segmentation network improves up to 20 percentage points over prior work in the real task of segmentation of lesions. I also conclude that the network architecture and optimizations are important factors and that there are still important limitations in current work.

Dual Machine-Learning System to Aid Glaucoma Diagnosis Using Disc and Cup Feature Extraction

Glaucoma is a degenerative disease that affects vision, causing damage to the optic nerve that ends in vision loss. The classic techniques to detect it have undergone a great change since the intrusion of machine learning techniques into the processing of eye fundus images. Several works focus on training a convolutional neural network (CNN) by brute force, while others use segmentation and feature extraction techniques to detect glaucoma. In this work, a diagnostic aid tool to detect glaucoma using eye fundus images is developed, trained and tested. It consists of two subsystems that are independently trained and tested, combining their results to improve glaucoma detection. The first subsystem applies machine learning and segmentation techniques to detect optic disc and cup independently, combine them and extract their physical and positional features. The second one applies transfer learning techniques to a pre-trained CNN to detect glaucoma through the analysis of the complete eye fundus images. The results of both systems are combined to discriminate positive cases of glaucoma and improve final detection. The results show that this system achieves a higher classification rate than previous works. The system also provides information on the basis for the proposed diagnosis suggestion that can help the ophthalmologist to accept or modify it.

Automatic Detection of Diabetic Retinopathy: A Review on Datasets, Methods and Evaluation Metrics

Diabetic retinopathy (DR) is a fast-spreading disease across the globe, which is caused by diabetes. The DR may lead the diabetic patients to complete vision loss. In this scenario, early identification of DR is more essential to recover the eyesight and provide help for timely treatment. The detection of DR can be manually performed by ophthalmologists and can also be done by an automated system. In the manual system, analysis and explanation of retinal fundus images need ophthalmologists, which is a time-consuming and very expensive task, but in the automated system, artificial intelligence is used to perform an imperative role in the area of ophthalmology and specifically in the early detection of diabetic retinopathy over the traditional detection approaches. Recently, numerous advanced studies related to the identification of DR have been reported. This paper presents a detailed review of the detection of DR with three major aspects; retinal datasets, DR detection methods, and performance evaluation metrics. Furthermore, this study also covers the author’s observations and provides future directions in the field of diabetic retinopathy to overcome the research challenges for the research community.

Retinal Blood Vessel Segmentation using Gaussian Filter

Retinal diseases are a visual disorder that can cause human vision problems, such as blindness and low vision. Early detection of this disease is important to prevent visual impairment. In ophthalmology, segmentation of blood vessels is a major component of color retinal fundus images analysis for monitoring and diagnosing retinal diseases, like diabetic retinopathy, hypertensive retinopathy, age-related macular degeneration and arteriosclerotic. This process is effective to identify retinal diseases more accurately. This paper using Gaussian Filter to segmenting blood vessel on retinal images because this method is effective to remove noise and make blood vessels on images more clearly. These methods are applied to color fundus images from DRIVE database. These databases contain color fundus image are used to evaluate this method. The performance is evaluated by accuracy, sensitivity and specificity. The proposed method has the average of accuracy, sensitivity and specificity at 95.72%, 96.90% and 82.10%.

Interaction analysis of systolic blood pressure and glycosylated hemoglobin in diabetic retinopathy: A Chinese sample

Background: The objective of this study was to evaluate (1) the association of blood pressure (BP) profiles and glucose profiles with diabetic retinopathy (DR) and (2) the interaction between BP profiles and glucose profiles in DR in a Chinese sample. Methods: This was a cross-sectional study of 984 DR patients. Demographic data, BP profiles, glucose profiles, renal function parameters, and medical histories were recorded. All the participants underwent testing for DR with digital nonmydriatic fundus photography and image analysis, using a standardized protocol. A multiple variable logistic regression (MLR) was employed to detect the associations and interactions of BP profiles and glucose profiles with DR. Results: Systolic blood pressure (SBP) and glycated hemoglobin (HbA1c) were significantly correlated with DR ([Formula: see text] for SBP, [Formula: see text]; [Formula: see text] for HbA1c, [Formula: see text]). An interaction between SBP and HbA1c was detected in a multiple logistic regression (MLR) model after adjustment for relevant potential confounders ([Formula: see text], [Formula: see text]). Conclusion: The present study suggested that SBP and HbA1c were independently and significantly associated with DR and that SBP interacted with HbA1c to impact on DR.

The diagnostic accuracy of an intelligent and automated fundus disease image assessment system with lesion quantitative function (SmartEye) in diabetic patients

BackgroundWith the diabetes mellitus (DM) prevalence increasing annually, the human grading of retinal images to evaluate DR has posed a substantial burden worldwide. SmartEye is a recently developed fundus image processing and analysis system with lesion quantification function for DR screening. It is sensitive to the lesion area and can automatically identify the lesion position and size. We reported the diabetic retinopathy (DR) grading results of SmartEye versus ophthalmologists in analyzing images captured with non-mydriatic fundus cameras in community healthcare centers, as well as DR lesion quantitative analysis results on different disease stages.MethodsThis is a cross-sectional study. All the fundus images were collected from the Shanghai Diabetic Eye Study in Diabetics (SDES) program from Apr 2016 to Aug 2017. 19,904 fundus images were acquired from 6013 diabetic patients. The grading results of ophthalmologists and SmartEye are compared. Lesion quantification of several images at different DR stages is also presented.ResultsThe sensitivity for diagnosing no DR, mild NPDR (non-proliferative diabetic retinopathy), moderate NPDR, severe NPDR, PDR (proliferative diabetic retinopathy) are 86.19, 83.18, 88.64, 89.59, and 85.02%. The specificity are 63.07, 70.96, 64.16, 70.38, and 74.79%, respectively. The AUC are PDR, 0.80 (0.79, 0.81); severe NPDR, 0.80 (0.79, 0.80); moderate NPDR, 0.77 (0.76, 0.77); and mild NPDR, 0.78 (0.77, 0.79). Lesion quantification results showed that the total hemorrhage area, maximum hemorrhage area, total exudation area, and maximum exudation area increase with DR severity.ConclusionsSmartEye has a high diagnostic accuracy in DR screening program using non-mydriatic fundus cameras. SmartEye quantitative analysis may be an innovative and promising method of DR diagnosis and grading.

A saliency and Gaussian net model for retinal vessel segmentation

Retinal vessel segmentation is a significant problem in the analysis of fundus images. A novel deep learning structure called the Gaussian net (GNET) model combined with a saliency model is proposed for retinal vessel segmentation. A saliency image is used as the input of the GNET model replacing the original image. The GNET model adopts a bilaterally symmetrical structure. In the left structure, the first layer is upsampling and the other layers are max-pooling. In the right structure, the final layer is max-pooling and the other layers are upsampling. The proposed approach is evaluated using the DRIVE database. Experimental results indicate that the GNET model can obtain more precise features and subtle details than the UNET models. The proposed algorithm performs well in extracting vessel networks, and is more accurate than other deep learning methods. Retinal vessel segmentation can help extract vessel change characteristics and provide a basis for screening the cerebrovascular diseases.

A Two Layer Sparse Autoencoder for Glaucoma Identification with Fundus Images.

Glaucoma is a type of eye condition which may result in partial or consummate vision loss. Higher intraocular pressure is the leading cause for this condition. Screening for glaucoma and early detection can avert vision loss. Computer aided diagnosis (CAD) is an automated process with the potential to identify glaucoma early through quantitative analysis of digital fundus images. Preparing an effective model for CAD requires a large database. This study presents a CAD tool for the precise detection of glaucoma using a machine learning approach. An autoencoder is trained to determine effective and important features from fundus images. These features are used to develop classes of glaucoma for testing. The method achieved an F - measure value of 0.95 utilizing 1426 digital fundus images (589 control and 837 glaucoma). The efficacy of the system is evident, and is suggestive of its possible utility as an additional tool for verification of clinical decisions.

Deep Learning Fundus Image Analysis for Diabetic Retinopathy and Macular Edema Grading

Diabetes is a globally prevalent disease that can cause visible microvascular complications such as diabetic retinopathy and macular edema in the human eye retina, the images of which are today used for manual disease screening and diagnosis. This labor-intensive task could greatly benefit from automatic detection using deep learning technique. Here we present a deep learning system that identifies referable diabetic retinopathy comparably or better than presented in the previous studies, although we use only a small fraction of images (<1/4) in training but are aided with higher image resolutions. We also provide novel results for five different screening and clinical grading systems for diabetic retinopathy and macular edema classification, including state-of-the-art results for accurately classifying images according to clinical five-grade diabetic retinopathy and for the first time for the four-grade diabetic macular edema scales. These results suggest, that a deep learning system could increase the cost-effectiveness of screening and diagnosis, while attaining higher than recommended performance, and that the system could be applied in clinical examinations requiring finer grading.

Robust intensity variation and inverse surface adaptive thresholding techniques for detection of optic disc and exudates in retinal fundus images

Diabetic Retinopathy (DR) is an adverse change in retinal blood vessels leads to blindness for diabetic patients without any symptoms. Diabetes is characterized by imbalance level of glucose in the human body. The optic disc (OD) is the major retinal landmark. Localization of OD is an important step in fundus image analysis and to develop Computer Aided Diagnosis tool for DR. OD center detection is necessary to reduce false positive rate in the detection of exudates (EXs). EXs is the white lesion present in the retina which is the early symptom for the diagnosis of DR. OD is detected using intensity variation algorithm and EXs is segmented using inverse surface adaptive thresholding algorithm. This algorithm achieves better result in localizing OD and segmenting EXs when compared to literature-reviewed methods. The maximum intensity variance method is used to locate OD with average ACC of 96.54%, 98.65%, 98.12%, 99.23%, 99.81% and 98.47% in DIARETDB0, DIARETDB1, MESSIDOR, DRIVE, STARE and Bejan Singh Eye Hospital databases with less computation time of 102ms, 108ms, 120ms, 93ms, 110ms and 131ms. The inverse surface adaptive thresholding method has achieved an SE of 97.43%, 98.87%, 99.12%, 97.21%, 98.72%, and 96.63%, a SPE of 91.56%, 92.31%, 90.21%, 90.14%, 89.58%, 92.56% and an ACC of 99.34%, 99.67%, 98.34%, 98.87%, 99.13%, 98.34% for DIARETDB0, DIARETDB1, MESSIDOR, DRIVE, STARE and Bejan Singh Eye Hospital databases respectively.

Retinal vessel segmentation by a divide-and-conquer funnel-structured classification framework

Accurate vessel segmentation is a fundamental and challenging task for the retinal fundus image analysis. Current approaches typically train a global discriminative model for retinal vessel classification that is difficult to fit the complex pattern of vessel structure. In this paper, we propose a novel divide-and-conquer funnel-structured classification framework for retinal vessel segmentation. More specifically, a dividing algorithm, named multiplex vessel partition (MVP), is proposed to divide retinal vessel into well constrained subsets where vessel pixels with similar geometrical property are grouped. A set of homogeneous classifiers are trained in parallel to form discriminative decision for each group. This decomposes a complex classification problem into a number of relatively simpler ones. Moreover, a funnel-structured vessel segmentation (FsVS) framework is proposed to reclassify the uncertain samples caused by imperfect grouping of pixels. This alleviates the problem in data partition at the dividing phase and further enhances the complexity and discriminative capability of the decision model. Both quantitative and qualitative experimental comparisons on three publicly available databases show that the proposed framework produces high performance for retinal vessel segmentation, achieving 95.47–96.46% vessel segmentation accuracy, 83.72–85.79% local vessel segmentation accuracy, 78.63–81.92% F1-score and 76.55–80.13% Matthew correlation coefficient respectively, better than the state-of-the-art methods.

Age-related changes of the retinal microvasculature.

PurposeBlood vessels of the retina provide an easily-accessible, representative window into the condition of microvasculature. We investigated how retinal vessel structure captured in fundus photographs changes with age, and how this may reflect features related to patient health, including blood pressure.ResultsWe used two approaches. In the first approach, we segmented the retinal vasculature from fundus photographs and then we correlated 25 parameterized aspects ("traits")—comprising 15 measures of tortuosity, 7 fractal ranges of self-similarity, and 3 measures of junction numbers—with participant age and blood pressure. In the second approach, we examined entire fundus photographs with a set of algorithmic CHARM features. We studied 2,280 Sardinians, ages 20–28, and an U.S. based population from the AREDS study in 1,178 participants, ages 59–84. Three traits (relating to tortuosity, vessel bifurcation number, and vessel endpoint number) showed significant changes with age in both cohorts, and one additional trait (relating to fractal number) showed a correlation in the Sardinian cohort only. When using second approach, we found significant correlations of particular CHARM features with age and blood pressure, which were stronger than those detected when using parameterized traits, reflecting a greater signal from the entire photographs than was captured in the segmented microvasculature.ConclusionsThese findings demonstrate that automated quantitative image analysis of fundus images can reveal general measures of patient health status.

Mathematical morphology for microaneurysm detection in fundus images.

Fundus image analysis is the basis for the better understanding of retinal diseases which are found due to diabetes. Detection of earlier markers such as microaneurysms that appear in fundus images combined with treatment proves beneficial to prevent further complications of diabetic retinopathy with an increased risk of sight loss. The proposed algorithm consists of three modules: (1) image enhancement through morphological processing; (2) the extraction and removal of red structures, such as blood vessels preceded by detection and removal of bright artefacts; (3) finally, the true microaneurysm candidate selection among other structures based on feature extraction set. The proposed strategy is successfully evaluated on two publicly available databases containing both normal and pathological images. The sensitivity of 89.22%, specificity of 91% and accuracy of 92% achieved for the detection of microaneurysms for Diaretdb1 database images. The algorithm evaluation for microaneurysm detection has a sensitivity of 83% and specificity 82% for e-ophtha database. In automated detection system, the successful detection of the number of microaneurysms correlates with the stages of the retinal diseases and its early diagnosis. The results for true microaneurysm detection indicates it as a useful tool for screening colour fundus images, which proves time saving for counting of microaneurysms to follow Diabetic Retinopathy Grading Criteria.

The Fluorescein Angiographic Signs of Aggressive Posterior Retinopathy of Prematurity

The purposeis to study the fluorescence angiographic signs of posterior aggressive retinopathy of prematurity.Patients and methods. A prospective study of 83 infants with aggressive posterior retinopathy of prematurity (APROP), who were on treatment in the neonatal center in 2012–2018 years. Gestational age was between 22 and 30 weeks. Depending on the diagnostic methods, patients were divided into two groups: 1 (main) — 51 (61.4 %) children who, in addition to standard ophthalmic examination methods, received 132 series of FA using a pediatric retinal camera RetCam3; 2 (control) — 32 (38.6 %) children without FA. After the procedure, the somatic state of the child was monitored for 24 hours for the registration of drug dye side effects.Results. Angiographic predictors of the disease and objective signs of APROP manifestation were presented in group 1 children a week earlier than in the group 2. This made laser treatment possible on 2.0 ± 0.5 weeks earlier in the group 1 (32.8 ± 1.2 weeks) compared to group 2 (34.8 ± 1.7 weeks) (p &lt; 0.05) and to increase its efficiency by 8.2 % (86.3 and 78.1 % respectively).Conclusion. Comparative analysis of angiograms and color fundus images obtained with RetCam3 in the studied groups of children demonstrates the advantages and safety FA for the detection angiographic predictors of APROP, early objective signs of disease manifestation, it provides timely and effective laser treatment of complicated categories of very premature infants.

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.

Retinal blood vessel extraction employing effective image features and combination of supervised and unsupervised machine learning methods.

In medicine, retinal vessel analysis of fundus images is a prominent task for the screening and diagnosis of various ophthalmological and cardiovascular diseases. In this research, a method is proposed for extracting the retinal blood vessels employing a set of effective image features and combination of supervised and unsupervised machine learning techniques. Further to the common features used in extracting blood vessels, three strong features having a significant influence on the accuracy of the vessel extraction are utilized. The selected combination of the different types of individually efficient features results in a rich local information with better discrimination for vessel and non-vessel pixels. The proposed method first extracts the thick and clear vessels in an unsupervised manner, and then, it extracts the thin vessels in a supervised way. The goal of the combination of the supervised and unsupervised methods is to deal with the problem of intra-class high variance of image features calculated from various vessel pixels. The proposed method is evaluated on three publicly available databases DRIVE, STARE and CHASE_DB1. The obtained results (DRIVE: Acc = 0.9531, AUC = 0.9752; STARE: Acc = 0.9691, AUC = 0.9853; CHASE_DB1: Acc = 0.9623, AUC = 0.9789) demonstrate the better performance of the proposed method compared to the state-of-the-art methods.

VAMPIRE® fundus image analysis algorithms: Validation and diagnostic relevance in hypertensive cats.

To validate a retinal imaging software named VAMPIRE® (Vascular Assay and Measurement Platform for Images of the Retina) in feline patients and test the clinical utility in hypertensive cats. One hundred and five healthy cats were enrolled. They represented the normal dataset used in the validation (group 1). Forty-three hypertensive cats with no noticeable retinal abnormalities were enrolled for the clinical validity of the software (group 2). Eleven points (4 veins, 4 arteries, and 3 arterial bifurcations) were measured for each digital image. Repeatability and reproducibility of measurements were assessed using two independent operators. Data were statistically analyzed by the Mann-Whiney and Tukey box plot. Significance was considered when P<0.05. Two hundred and ten retinal images were analyzed for a total of 2310 measurements. Total mean was 9.1 and 6.1 pixels for veins and arteries, respectively. First, second, and third arteriolar bifurcations angles were 73.6°, 76.9°, and 85.4°, respectively. A comparison between groups 1 and 2 showed a statistically significant reduction in arteriolar diameter (mean 3.3 pixels) and branch angle (55°, 47.8° and 59.9°) associated with increasing vein diameter (mean 24.15 pixels). Current image analysis techniques used in human medicine were investigated in terms of extending their use to veterinary medicine. The VAMPIRE® algorithm proved useful for an objective diagnosis of retinal vasculature changes secondary to systemic hypertension in cats, and could be an additional diagnostic test for feline systemic hypertension.

PixelBNN: Augmenting the PixelCNN with Batch Normalization and the Presentation of a Fast Architecture for Retinal Vessel Segmentation.

Analysis of retinal fundus images is essential for eye-care physicians in the diagnosis, care and treatment of patients. Accurate fundus and/or retinal vessel maps give rise to longitudinal studies able to utilize multimedia image registration and disease/condition status measurements, as well as applications in surgery preparation and biometrics. The segmentation of retinal morphology has numerous applications in assessing ophthalmologic and cardiovascular disease pathologies. Computer-aided segmentation of the vasculature has proven to be a challenge, mainly due to inconsistencies such as noise and variations in hue and brightness that can greatly reduce the quality of fundus images. The goal of this work is to collate different key performance indicators (KPIs) and state-of-the-art methods applied to this task, frame computational efficiency–performance trade-offs under varying degrees of information loss using common datasets, and introduce PixelBNN, a highly efficient deep method for automating the segmentation of fundus morphologies. The model was trained, tested and cross tested on the DRIVE, STARE and CHASE_DB1 retinal vessel segmentation datasets. Performance was evaluated using G-mean, Mathews Correlation Coefficient and F1-score, with the main success measure being computation speed. The network was 8.5× faster than the current state-of-the-art at test time and performed comparatively well, considering a 5× to 19× reduction in information from resizing images during preprocessing.

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.