Towards Practical Unsupervised Anomaly Detection on Retinal Images

Retinopathy of Prematurity (ROP) is a blinding disease, which primarily occurs on premature infants whose birth weights is less than 1250 grams or gestation is less than 31 weeks. ROP has become the leading cause of preventable childhood blindness throughout the world. Nowadays, more and more researchers start attempting to develop auto or semi-auto methods based on digital image analysis to diagnose ROP. However, factors like high measurement errors or redundant analysis phrases make traditional analysis methods difficult to assist diagnose ROP perfectly. In this paper, we develop an automated system to analyse premature infants’ retinal images using deep neural networks. We primarily try to solve two problems. (1) the existence of ROP, normal or ROP; (2) the severity of ROP, mild-ROP or severe-ROP. Deep neural networks take the advantages of strong representation ability and enable to nonlinear mapping, attain high accuracies and great generalization performances on retinal fundus image datasets.

Accurate retinal vessel segmentation plays a vital role in the early detection and diagnosis of ophthalmologic dis- eases such as diabetic retinopathy, glaucoma, and hypertension. Traditional image processing and machine learning techniques often struggle to capture the fine vessel structures and complex variations present in retinal fundus images. To overcome these limitations, this study proposes a hybrid deep learning framework that combines Convolutional Neural Networks (CNNs) and Gen- erative Adversarial Networks (GANs) for precise retinal vessel segmentation. In the proposed method, the CNN component serves as the feature extractor, learning multi-scale spatial and contextual features from retinal images, while the GAN framework enhances segmentation accuracy by employing a discriminator network that distinguishes between real and generated vessel maps. This adversarial learning setup encourages the generator to produce vessel segmentations that closely resemble ground truth images. Additionally, preprocessing techniques such as contrast enhancement and noise reduction are applied to improve the visibility of retinal structures, and post-processing steps refine the vessel boundaries for higher accuracy. Experimental results on publicly available retinal image datasets, such as DRIVE and STARE, demonstrate that the pro- posed GAN-CNN model outperforms conventional CNN-based segmentation methods in terms of accuracy, sensitivity, specificity, and Dice coefficient. The system achieves robust performance even under varying illumination and noise conditions. These findings indicate that integrating adversarial learning with convolutional feature extraction can significantly enhance retinal vessel segmentation, paving the way for more efficient and automated retinal disease screening systems

Using Deep Learning Models to Characterize Major Retinal Features on Color Fundus Photographs

Vision impairment is becoming a major health concern, especially in elderly people. While in the medical field, manually detecting ocular pathology has significant difficulty. Therefore, deep learning diagnostic techniques are widely used for identifying eye diseases and play a crucial role in diagnosing vision-related problems. Examination of fundoscopy allows for analyzing eyes for diagnosis of eye diseases, including Diabetic retinopathy (DR), Cataracts, Glaucoma, Age‑related macular degeneration, Pathologic Myopia, and more. In this paper, we propose a concise review of introducing most of the DL, hybrid, and ensemble models utilized for the purpose of identification and classification of eye diseases. Various datasets, feature extraction techniques, and metrics for performance evaluation are discussed. The chosen papers come from conferences and scholarly publications published from 2019 to 2024. We evaluate the performance of chosen researches using different datasets, the most common ones include ocular disease intelligent recognition, Indian DR image dataset, EyePACS, methods to evaluate segmentation and indexing techniques in the field of retinal ophthalmology, methods to evaluate segmentation and indexing techniques in the field of retinal ophthalmology-version 2, DIARETDB, Structured analysis of the retina, high-resolution fundus, digital retinal images for vessel extraction, online retinal fundus image dataset for Gl analysis and research, retinal fundus multi-disease image dataset and Kaggle datasets. The detection studies that have been reviewed show that the accuracy of these approaches varies between 73% and 99%, the sensitivity ranges from 69% to 99% and precision is between 89% and 99%. The results show that great accuracy is consistently achieved with DL algorithms compared to traditional Machine learning approaches. However, there are still some challenges and limitations remaining including excessive resource consumption and over-fitting due to dataset size and diversity issues. This review offers useful insight for researchers and healthcare professionals to comprehend AI technologies properly for the detection, classification, and diagnosis of retinal diseases. We succinctly summarize the methodologies of all the chosen studies and focus on the elements that define the aim of the studies.

- This paper presents a comparative analysis between a traditional convolutional neural network (CNN) model and a hybrid ResNet50 + AdaBoost architecture for non-invasive cardiovascular risk prediction using retinal fundus images. The baseline model, trained on a small dataset for binary classification, achieves reasonable accuracy but lacks scalability and granularity. In contrast, the proposed model leverages deep residual learning for feature extraction and ensemble-based classification to predict five distinct cardiovascular risk levels. Evaluated on a class-balanced dataset of over 11,500 images, the hybrid model achieved a testing accuracy of 91.48% and a macro F1-score of approximately 0.93. The system also integrates data visualization and database logging to support clinical explainability. The findings demonstrate the proposed method’s superiority in risk stratification, robustness, and deployment readiness for real-world applications. Key Words: Retinal Imaging, Cardiovascular Risk Prediction, ResNet50, AdaBoost, Deep Learning, Ensemble Learning, Fundus Images, Heart Attack Classification

Retinal image quality classification makes a great difference in automated diabetic retinopathy screening systems. With the increase of application of portable fundus cameras, we can get a large number of retinal images, but there are quite a number of images in poor quality because of uneven illumination, occlusion and patients movements. Using the dataset with poor quality training networks for DR screening system will lead to the decrease of accuracy. In this paper, we first explore four CNN architectures (AlexNet, GoogLeNet, VGG-16, and ResNet-50) from ImageNet image classification task to our Retinal fundus images quality classification, then we pick top two networks out and jointly fine-tune the two networks. The total loss of the network we proposed is equal to the sum of the losses of all channels. We demonstrate the super performance of our proposed algorithm on a large retinal fundus image dataset and achieve an optimal accuracy of 97.12%, outperforming the current methods in this area.

Early detection of diabetic retinopathy is a serious disease for diabetics to minimize their sightlessness risks. The different approaches take a much longer time for a very large training dataset. In classifying to better the accuracy of diabetic retinopathy, a novel technique called MAP Concordance Regressive Camargo’s Index-Based Deep Multilayer Perceptive Learning Classification (MAPCRCI-DMPLC) has been introduced with minimum time consumption. The novel model of MAPCRCI-DMPLC comprises the input layer, hidden layers, and output layer for detecting diabetic retinopathy at an early stage through high accuracy and less moment consumption. The proposed MAPCRCI-DMPLC model collected the retinal fundus images from the dataset as input. After that, we carried out image preprocessing using the MAP-estimated local region filtering-based preprocessing technique in the first hidden layer. In the second hidden layer, Camargo’s index-based ROI extraction is performed to identify the infected region. Then, Concordance Correlative Regression is applied for texture feature extraction. Then the color feature is extracted, beginning the image. The features extracted to the output layer to classify the different levels of DR using the swish activation function through higher accuracy. An investigational assessment using a retinal image dataset on factors such as peak signal-to-noise ratio (PSNR), disease detection accuracy (DDA), false-positive rate (FPR), and disease detection time (DDT), regarding the quantity of retinal fundus images and image dimension. The quantitative and qualitatively analyzed outcome shows a better presentation of our proposed MAPCRCI-DMPLC technique when compared through the five state-of-the-art approaches.

Glaucoma detection from retinal fundus images using graph convolution based multi-task model

The structure of the optic disc is impacted by the devastating eye disease glaucoma. Most doctors rely on viewing and screening the retinal fundus images to detect it. Identifying the optic disc is a crucial stage for automated diagnosis of glaucoma. This study presents a transfer learning-based approach to locate the optic disc in the retinal fundus image. The fundus images are categorized into four distinct classes according to the positions of optic disc to learn the in-depth specific features.Based on the YOLOv3 model, the optic disc detection method is then fine-tuned using the specialised image dataset. The model involved is trained and evaluated on 2178 retinal fundus images obtained from database at a nearby hospital and publically accessible datasets. In 2103 of the retinal fundus images, the suggested approach successfully detected the optical disc, yielding an average accuracy of 96.56%.

Non-invasive biopsy diagnosis of diabetic kidney disease via deep learning applied to retinal images: a population-based study.

Background Diabetic Retinopathy (DR) remains a leading cause of preventable blindness, with rising prevalence particularly in low-resource settings. Traditional DR screening relies on specialist ophthalmologic assessment, posing barriers to early detection. Recent advances in deep learning have enabled the development of automated tools that may enhance the scalability and accessibility of DR screening. Objectives To compare the performance of two convolutional neural network (CNN) architectures-EfficientNet-B5 and ResNet-50-for automated DR detection using publicly available retinal fundus image datasets. Methods We utilized a transfer learning approach on two large datasets: EyePACS 2015 (n = 35,126) for training/validation and APTOS 2019 (n = 3,662) for external testing. Images were preprocessed with resizing, cropping, and contrast enhancement. The EfficientNet-B5 and ResNet-50 models were independently trained and evaluated using key metrics: accuracy, sensitivity, specificity, area under the ROC curve (AUC), and Quadratic Weighted Kappa (QWK), with classification across DR severity levels. Results EfficientNet-B5 achieved a QWK of 92%, sensitivity of 98%, specificity of 93%, and AUC of 96% on the test dataset. ResNet-50 demonstrated similarly high performance with an accuracy of 96%, QWK of 86%, sensitivity of 94%, specificity of 97%, and AUC of 96%. Both models showed robust generalization across severity grades, indicating strong discriminative capabilities. Conclusions The results highlighted that our models are viable architectures for AI-driven DR screening, with EfficientNet-B5 showing slightly higher agreement scores and ResNet-50 offering stronger specificity. These models hold promise for deployment in DR screening programs, especially in resource-limited settings where ophthalmologic expertise is scarce. Future work should focus on clinical validation and integration into public health systems to improve early DR detection and reduce vision-related morbidity. Key messages • Integrating artificial intelligence (AI) into diabetic retinopathy screening can enhance early detection and support equitable access to eye care, particularly in resource-limited settings. • AI-based screening tools offer a scalable and effective approach to strengthen public health strategies aimed at preventing vision loss from diabetic complications.

Traditionally, medical discoveries are made by observing associations, making hypotheses from them and then designing and running experiments to test the hypotheses. However, with medical images, observing and quantifying associations can often be difficult because of the wide variety of features, patterns, colours, values and shapes that are present in real data. Here, we show that deep learning can extract new knowledge from retinal fundus images. Using deep-learning models trained on data from 284,335 patients and validated on two independent datasets of 12,026 and 999 patients, we predicted cardiovascular risk factors not previously thought to be present or quantifiable in retinal images, such as age (mean absolute error within 3.26 years), gender (area under the receiver operating characteristic curve (AUC) = 0.97), smoking status (AUC = 0.71), systolic blood pressure (mean absolute error within 11.23 mmHg) and major adverse cardiac events (AUC = 0.70). We also show that the trained deep-learning models used anatomical features, such as the optic disc or blood vessels, to generate each prediction.

Heart disease represents a significant public health problem worldwide, and early discovery is required in order to improve clinical outcomes and decrease mortality. Retinal imaging can provide a noninvasive methodology for evaluating microvascular health, with strong associations to cardiovascular disease. This work introduces a simple deeplearning method to predict heart disease from retinal fundus images. The proposed system consists of preprocessing retinal fundus images, then extracting features with convolutional neural networks (CNN), and finally predicting heart disease using binary classification. The model displays strong predictive capabilities, demonstrating that retinal vascular patterns contain important cardiovascular risk biomarkers. The results suggest that retinal images could be an effective, scalable screening tool for early detection of heart disease.

Background The rising burden of cardiovascular disease (CVD) has spurred the development of innovative, non-invasive screening methods. Advances in artificial intelligence (AI) and deep learning (DL) applied to retinal imaging offer a promising avenue, as changes in the retinal vasculature can mirror systemic vascular health. Purpose To synthesize evidence from studies using artificial intelligence (AI) algorithms trained on retinal images to predict cardiovascular disease (CVD) risk. Methods A systematic literature search was conducted on MEDLINE and Embase databases up to February 2024 in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guideline using relevant search terms such as; "cardiovascular disease," "artificial intelligence," "deep learning," "retinal imaging," "colour fundus photography," etc. Inclusion criteria were the development of a DL model applied to any ophthalmic imaging modality for predicting CVD risk or outcome. Results Of 9880 studies which were screened, 13 studies were included. All studies included general population databases, while 7 (54%) studies used databases that included patients with pre-existing CVD risk factors. All studies used retinal fundus images as input for the DL models, and most models (92%) analysed characteristics of the retinal vasculature (e.g., vessel calibre, venular dilatation, arteriolar narrowing, microaneurysms) for their prediction. Overall, 18 different CVD risk factors were predicted through DL models, with age (n=13; 100%), sex (n=11; 85%) and smoking status (n=9; 69%) being the most common. Five (38%) studies analysed binary CVD outcomes including incident myocardial infarction, stroke, and coronary atherosclerotic disease; Four (31%) studies compared CVD risk prediction to traditional CVD risk scores (e.g., Framingham risk score, European Systematic Coronary Risk Evaluation, etc.). These studies were able to accurately stratify cumulative CVD events into low, moderate, and high-risk groups via fundus imaging, demonstrating stratification comparable to established CVD risk scores and cardiac imaging modalities. In total, 8 (62%) studies performed an external validation: the area under the receiver operating characteristic curve ranged between 68.2% and 85.9%. Accuracy, specificity and sensitivity were measured in 4 (31%) studies. Respectively, they ranged between 58.3%-82.0%, 40.4-66.0% and 81.0-89.1%. Only one AI model (Reti-CVD) was made publicly available for clinical use. Conclusions In conclusion, recent studies using DL applied to fundus imaging to predict CVD risk mostly examine retinal vasculature to make predictions and can stratify CVD risk comparably to other clinical risk scores. Though many report promising performance, the majority have not been used in real clinical settings. Additional research is required to enable their clinical implementation in a primary care context, or in an ophthalmological setting.

Anomaly detection via gating highway connection for retinal fundus images