DATA SET RD

CATEGORIES NAME OF THE ARTICLE YEAR PUBLICATION N AUTHORS MAGAZINE COUNTRY OR PLACE TYPE OF STUDY # POPULATION (SEX) # ARTICLES KEY RESULTS AND FINDINGS TYPE OF AI (Deep learning, machine learning, networks AI APPLICATION IN THE DIAGNOSIS OF DR ADVANTAGES OF AI APPLIED LIMITATIONS OF APPLIED AI FUTURE OPPORTUNITIES AND RECOMMENDATIONS ETHICAL CONSIDERATIONS REPORTED # Artificial Intelligence Based Screening System for Diabetic Retinopathy in Primary Care 2024 Marc BagetBernaldiz, Benilde Fontoba-Poveda, Pedro RomeroAroca, Raul NavarroGil, Adriana HernandoComerma, Angel Bautista-Perez, Diagnostic s Spain Study observational retrospective AIRS: 15,297 7389 retinographies diabetic patients + 1200 images of the Messidor-2 base AIRS (image reading): - In local database: accuracy 98.6%, sensitivity 96.7%, specificity 99.8%, PPV 99.0%, NPV 98.0%, AUC 0.958. - In Messidor-2: accuracy 96.8%, sensitivity 94.6%, specificity 99.1%, PPV 90.5%, NPV 99.5%, AUC 0.918. - Classification by levels: AIRS: neural network convolutional deeply trained with images of EyePACS and base local AIRS: Diagnosis and automatic classification of Diabetic retinopathy in fundus images (no, mild, moderate, severe/proliferative). High sensitivity and specificity, even superior to others approved algorithms. Lower accuracy when applied to external databases such as Messidor-2 (difference of cameras and resolution). AIRS validation in different populations and chambers to increase applicability. Compliance with the Helsinki Declaration. Informed consent obtained from all the patients. 11 It allows reducing the load of specialists to automate screening. Need for validation in more diverse populations ethnically. Routine use in care primary to improve coverage and accessibility to screening. DRPA: 40,129 patients with type 2 diabetes DRPA: algorithm of random forests DRPA: Risk Prediction to develop any An automated unsupervised deep learning-based approach for diabetic retinopathy detection 2022 Huma Naz, Rahul Nijhawan and Neelu Jyothi Ahuja. Medical & Biological Engineering & Computing India Study experimental DIARETDB1: 89 retinal images. APTOS-2019: 3662 images. Liverpool: 134,638 images. The proposed system, called FCCNN (combination of Modified Fuzzy CMeans and CNN), obtained an accuracy of 98.66% in the DR classification. Deep Learning combined with learning not supervised (FCCNN = CNN + Modified Fuzzy CMeans) automatic detection of microaneurysms and severity classification DR in images of retina. High accuracy (98.66%) and better performance than kmeans, DEC and FCM. It was not tested in clinical settings. real, only in datasets public. Extend the model to other diseases such as Alzheimer's. Greater robustness against noise in pictures. Reduced sensitivity in injuries in foveal region and very microaneurysms small ones. To be used in systems of recommendation for early diagnosis in areas remote. Comparison with other methods: k-means: 94.43–96.63% FCM: 93.56% DEC: 90.4% Shorter time processing in comparison with others methods. Potential integration into robotics and medical computer vision. FCCNN consistently outperformed all. No data required labeled (learning unsupervised) It demonstrated robustness against noise in images: with 20% noise, it achieved 95.23% accuracy, and with 80% noise it still obtained 80.69% (Table 5, p. 14). In sensitivity and specificity: tolerance of 87.69% and specificity of 92.44% (p. 14). In terms of processing time, it was longer more efficient than other clustering methods (16–17 s vs. more than 20 s) in DEC/FCM, p. 14). 12 It detected microaneurysms (initial lesion) of the DR) with high repeatability and Diabetic retinopathy detection using Bilayered Neural Network classification model with resubstitution validation 2024 Herman Khalid Omer MethodsX Iraq Study experimental with public dataset IDRiD Dataset (Indian Diabetic Retinopathy Image Dataset, IEEE Dataport). 3,662 images (1,805 normal, 370 mild, 999 moderate, 193 severe, 295 proliferative). No were patients with previous diagnosis, images Methodology: (1) Preprocessing: histogram equalization, noise reduction with median filter, resize to 200×200 px. (2) Extraction of (3) Classification with vSLAM features (bag of visual words, landmarks: macula, optic disc, hemorrhages, microaneurysms). (4) Classification with Bilayered Neural Network (BNN) under resubstitution validation. (5) Comparison with other classifiers (SVM, KNN, Ensemble, DT). Machine learning hybrid (extraction) vSLAM + Bilayered Neural Network Both: • De novo diagnosis: distinguishes “non-DR” vs “with DR”. • Classification: gradation into 5 levels (mild, moderate, severe, proliferative). Preprocessing of images (histogram equalization, noise reduction with median filter, - High accuracy (98.5%). - Excellent sensitivity even in stadiums initials. - Preprocessing improves contrast and reduces noise. - Multi-class classification (5 levels). - Pipeline applicable to screening automated.O100 - Validated only on public dataset (IDRiD), not real clinical. - Validation with resubstitution (risk of overfitting). - High computational cost in vSLAM. - No real-time validation.O108 - Validate in real clinical cohorts. - Use cross-validation to avoid overfitting. - Compare with modern end-toend networks (EfficientNet, Transformers). - Simplify pipeline (without vSLAM). Public IDRiD dataset, without direct, actual patients. Compliant with data redistribution policies. No approval required. Additional ethics. Author declares no conflicts of interest. 13 Translated from Spanish to English - www.onlinedoctranslator.com Validation of artificial intelligence algorithm LuxIA for screening of diabetic retinopathy from a single 45° retinal color fundus Images: The Cards study 2025 Rodrigo AbreuGonzalez, Gabriela Susanna-González, Joseph PM Blair, Romina M Lasagni Vitar, Carlos Ciller, Stefanos Apostolopoulos, Sandro De Zanet, Joseph Nathan Rodríguez Martín, Carlos Bermúdez, Alfonso Luis Calle Pascual, Elena Rigo, BMJ Open Ophthalmology logy Spain Study multicentric, cross, observational Total: 945 patients with diabetes mellitus (type 1 or 2) The LuxIA algorithm showed a sensitivity of 97.1% and a specificity of 94.8% in the mtmDR detection with images Topcon. Algorithm of artificial intelligence based on learning deep (deep) learning) implemented in the Retinal platform (LuxIA) Automated screening for more-than-mild diabetic retinopathy (mtmDR) from of a single 45° color fundus image, without pupillary dilation High performance diagnosis (AUC 0.96) comparable to that of specialists. Cross-sectional design, without longitudinal evaluation. Evaluate the impact of the algorithm in real clinical practice and long-term results. Reduced sensitivity in cameras ZEISS by sample size reduced. Average age: 64.6 ± 13.5 years Implement image quality controls in the practice to optimize precision. AUC=0.96, overall accuracy of 95.2%. Confidence intervals confirmed robust estimates (sensitivity 96.8–98.2%, specificity 94.3–95.1%). Quick, non-invasive method and applicable in care primary. 14 Sex: 508 men (55%), 416 women (45%) Possibility of quality bias image variable. It increases accessibility and feasibility of screening. Validate in other countries and populations to strengthen the generalizability. Risk of human error in the manual classification of mild DR as a reference standard. In validation with ZEISS images, the Accuracy was 90.6%, specificity 92.3% and lower sensitivity (83.3%). Compatible with various devices (Topcon and Include more studies Evaluation of AIenhanced nonmydriatic fundus photography for diabetic retinopathy screening 2024 Chen-Liang Hu, YuChan Wang, WenFang Wu and Yu Xi Photodiagn osis and Photodyna mic Therapy China Study observational transversal of validation diagnostic Total: 120 patients hospitalized with diabetes (240 eyes) Comparison with FFA (gold standard): AutoEye Platform (Shanggong Medical Technology), based on deep learning (Retinal Net) Automated analysis of non-mydriatic images of 45° of the fundus for detect and classify Diabetic retinopathy at five levels: no DR, NPDR mild, moderate NPDR, Severe NPDR and PDR High accuracy and diagnostic consistency with FFA Lower accuracy in cases of severe NPDR (85.4%) and PDR (72.2%) due to limitations of the 45° field of view Expand your studies with more sample size and in communitiesAI reading group (AutoEye): sensitivity 97.5%, specificity 97.3%, PPV 98.7%, NPV 94.7%, Youden index 0.948, kappa 0.877, diagnostic accuracy global 90.6%. Sex: 59 men (118 eyes), 61 women (122 eyes) It does not require mydriasis (more comfort and speed) Incorporate wide-field techniques to improve detection of peripheral injuries and neovessels 15 Possible errors in eyes with small pupils or conditions like waterfallsIt allows mass screening in 3–5 minutesAge: 33 to 76 years Manual reading (ophthalmologist): sensitivity 96.3%, specificity 95.9%, PPV 98.1%, NPV 92.2%, index of Need for further validation in more diverse populations Confirm clinical value in routine practice beforeReduce dependence on Implementation of Artificial Intelligence-Based Diabetic Retinopathy Screening in a Tertiary Care Hospital in Quebec: Prospective Validation Study 2024 Fares Antaki, Imane Hammana, MarieCatherine Tessier, Andrée Boucher, Maud Laurence David Jetté, Catherine Beauchemin, Karim Hammamji, Ariel Yuhan Ong, MarcAndré Rhéaume, Danny Gauthier, Mona HarissiDagher, Pearse A Keane and Alfons Pomp JMIR Diabetes Canada Study prospective of clinical validation in a real environment 133 patients recruited, 115 included (230 eyes) Patient level (detection of referable disease): sensitivity 87.5 % (95% CI 71.9–95.0), specificity 66.2% (95% CI 54.3–76.3). No cases of life-threatening diabetic retinopathy The vision was overlooked. machine learning traditional with extractors manuals characteristics (FACE: Computer) Assisted Retinal Analysis). De novo diagnosis of diabetic retinopathy and edema detection diabetic macular degeneration, classifying patients in “referable” or “not referable” High sensitivity for detect disease referable and DME. Lower specificity compared to more recent deep learning models. Validate newer deep learning models in contexts similar. Informed consent of the patients. Compliance with Helsinki Declaration.Sex: 57.4% men, 42.6% women Potential to reduce workload in ophthalmology. Partial dependence on human supervision (system semi-automated). Implement audits routines to ensure performance. Use of data unidentified.Eye level (detection of any DR): sensitivity 88.2% (95% CI 76.6–94.5), specificity 71.4% (95% CI) 63.7–78.1). Average age: 55.4 years (20–90) Cost savings significant projected. Use of dilated eye exam as a reference, without multimodal images. Possible expansion of automated screening in Canada with strict policies of governance. Departmental approval and ethical exemption due to the non-interventional nature.Type 1 diabetes: 20%; Type 2: 36.5%; Unspecified: 43.5% Eye level (detection of diabetic macular edema): sensitivity 100% (95% CI 64.6– 100), specificity 81.9% (95% CI % 75.6–86.8). 13% of results inconclusive. 16 Incomplete outputs: 13% of patients, mainly associated with older age and difficulties in quality of the images. Economic impact: estimated savings of CAD $245,635 per year (≈ USD $177,643) for a screening program of 5000 patients, with a reduction average CAD $49 per patient. EffNet-SVM: A Hybrid Model for Diabetic Retinopathy Classification Using Retinal Fundus Images 2025 KV Naveen, BN Anoop, KS Siju, Mithun Kumar Kar, Vipin Venugopal IEEE Access India Study experimental with public dataset APTOS 2019 Dataset (Kaggle): 3,662 images (originally 5) classes: no DR, mild, moderate, severe, proliferative). For the model regrouped into 2 classes: without DR (1805) and with DR (1857). Additional validation in HRF dataset (30 Images: 15 normal, 15 with DR). Methodology: (1) Preprocessing: resize 224×224, augmentations (rotations, flips, zoom, contrast). (2) Feature extraction with EfficientNetV2-Small ( transfer learning, GAP, BN, dropout). (3) Classification withSVM RBF ((4) Comparison against CNNs (MobileNetV2, ResNet50, DenseNet121) and other hybrids. (5) Cross-validation and testing on HRF. Results:EffNet-SVM obtained97.26% accuracy,Accuracy 96.9%, recall 97.9%, F1-score 97.4%. In HRF cross-validation: 76.7% (better than ResNet50, DenseNet121, MobileNetV2). Conclusions:The hybrid improves accuracy and generalizability, reducing false positives and negatives, with a low inference time (0.098 s), making it suitable for mass screening. Hybrid: EfficientNetV2Small (extraction) of features) + SVM RBF (classification) De novo diagnosis: It directly separates "no DR" from "with DR" based on retinographies. It does not grade severity (original classes were regrouped into a binary system). Advantages: - High accuracy (97.26%). - Better than conventional CNNs. - Low inference time (0.098 s). - Generalizes better (HRF cross-validation). - Partial explainability with Grad-CAM. Limitations: - Binary classification only (no severity levels). - Dependent on image quality (lighting, blur). - Validated only on public datasets (not real clinical setting). - SVM reduces end-to-end interpretability. - Extend to multi-category classification. - Validate in multicenter clinical cohorts. - Integrate explainable care mechanisms. - Explore multimodality (fundus + OCT/OCTA). Kaggle public dataset APTOS and HRF (no ethical approval required). Code available on GitHub. Authors declare no conflicts of interest. 17 Diabetic Retinopathy Grading by Deep Graph Correlation Network on Retinal Images Without Manual Annotations 2022 Guanghua Zhang, Bin Sun, Zhixian Chen, Yuxi Gao, Zhaoxia Zhang, Keran Li, Weihua Yang Frontiers in Medicine China Study experimental with public datasets DatasetEyePACS-1: 10,286 images of 5,158 patients; Messidor-2:1,748 images of 874 patients (France). Tags of severity: none, mild, moderate, severe, proliferative. Methodology: (1) Feature extraction withResNet-50. (2) Construction of a KNN graph between images. (3) Application ofGraph Convolutional Network (GCN)with 3 loss functions: graph-center loss, pseudocontrastive loss, transform-invariant loss. (4) Automatic generation of pseudoannotations. (5) Final classification based on graphical correlations. Results:EyePACS-1 → accuracy 89.9%, sensitivity 88.2%, Specificity 91.3%, AUC 0.953. Messidor-2 → Accuracy 91.8%, Sensitivity 90.2%, Specificity 93.0%, AUC 0.974. Performance close to that of specialists and superior to that of trained graders. Conclusions:The DGCN model can reduce the need for costly annotations and achieve clinical accuracy for automated screening. Deep learning hybrid (CNN + GCN, learning unsupervised with pseudo-labels) Automatic DR severity grading (5 levels) without the need for manual notes - Validate in multicenter clinical cohorts. - Integrate with semisupervised approaches. - Expand to other eye diseases. - Optimize training time and efficiency. Use of public data (EyePACS, Messidor-2). Ethical exemptions obtained (Quorum IRB) for EyePACS; Brest Hospital for Messidor). Funded by Shanxi and Nanjing funds. Authors declare no conflicts of interest. - It does not require expert annotations for training. - High accuracy and sensitivity (≥90%). - Performance close to that of specialists. - Scalable to large unannotated datasets. - Invariant to transformations of image. - Even lower than fully supervised models. - Complex and timeconsuming training. - Requires large datasets for robustness. - Validation only in public settings (not real clinical practice). 18 Deep neural network model for diagnosing diabetic retinopathy detection: An efficient mechanism for diabetic management 2025 Dharmalingam Muthusamy, Parimala Palani Biomedical l Signal Processing and Control India Study experimental with public dataset Public dataset of Kaggle (“Diabetic Retinopathy Resized/Arranged”) with35,126 imagesin 5 classes (No DR, mild, moderate, severe, Methodology: (1) Preprocessing with algorithm ofluminosity normalizationto reduce noise and homogenize contrast. (2) Feature extraction withSymmetric Deep Convolute Tubular Neighborhood Classifier:includesTubular Neighborhood Window,average Deep learning (hybrid model with preprocessing + tubular convolution + regression multinomial) Multilevel classification of DR (No DR, mild, moderate, severe, proliferative) in fundus images - Validation in multicenter clinical settings. - Comparison with modern end-to-end models (EfficientNet, Vision Transformers). - Integration into community screening systems. Public dataset, without Direct clinical population (no additional ethical approval required). Authors declare no conflicts of interest. - High sensitivity and specificity (>90%). - Fast processing (<30 ms). - Better performance than reference methods (DenseNet, VGG-NiN, - Validated only on public Kaggle dataset, not on real clinical cohorts. - Algorithmic complexity (pipeline with several stages). - Not tested in actual screening flow. 19 Diabetic Retinopathy Detection from Fundus Images of the Eye Using Hybrid Deep Learning Features 2022 Muhammad Mohsin Butt; DNF Awang Iskandar; Sherif E. Abdelhamid; Ghazanfar Latif; Runna Alghazo Diagnostic s International(Ma lasia; States United; Arabia Saudi; Canada) Study experimental / investigation original 3662 images of fundus examination The hybrid model (GoogleNet + ResNet18) with SVM classifier achieved an accuracy of 97.8% in binary classification (No RD vs RD) and 89.29% in multiclass classification (No RD, mildmoderate, severe-proliferative), surpassing recent methods. Deep learning (Convolutional Neural Networks, Transfer Learning) combined with classifiers of Machine Learning (SVM, RF, RBF, NB). They were extracted characteristics of fundus images through models pre-trained GoogleNet and ResNet-18 (transfer learning); they merged into a hybrid vector of 2000 characteristics that fed classifiers (mainly SVM) for binary detection and multiclass of RD. High precision; lower processing time; error reduction humans; assistance to ophthalmologists for early diagnosis; ability to process large volumes of images of shape continued. Class imbalance in the images; dependence on a single dataset (APTOS); not evaluated in realworld clinical settings; no custom CNN designed, which could limit the learning of similarities interclass. Explore the design of customized CNNs to improve capturing similarities between classes; applying techniques of data augmentation and preprocessing for remove artifacts; expand to other eye conditions such as macular degeneration and glaucoma. No approval was required ethics committee or informed consent, since data were used from public domain. The The authors declared they had no conflicts of interest. 20 Automated Microaneurysms Detection in Retinal Images Using Radon Transform and Supervised Learning: Application to Mass Diabetic Screening Retinopathy 2021 Tavakoli, M, Mehdizadeh, A, Aghayan, A, Shahri, RP, Ellis, T, Dehmeshki, J IEEE Access International(It's two united, Iran, London) Primary school, development and validation of algorithm 749 images of three databases public and a base crazy Sensitivity for detecting rd at the level of Image: 100% – Specificity: 93% on average – Detection of MAs: sensitivity 95.7% with an average of 7 false positives per image – Superior performance compared to previously published methods. Learning combined: – Method no supervised (Radon Transform to preprocessing and detection candidates) - Learning supervised with Support Vector Machine (SVM). Preprocessing for normalize color and lightning. High sensitivity and specificity. – Low false positive rate. – Processing fully automatic, fast and reproducible. – Robustness to noise and variability of the images. – Enables mass screening, reduces workload specialists. Variability in color and contrast between images can affect the performance. – Need for standardization and careful preprocessing. – It was not evaluated in real clinical settings or in populations various in real time. – Implement in population screening programs. – Integrate into health systems for early detection of DR. – Validate in larger databases and environments real clinical cases. They are not mentioned ethical considerations in The abstract; availability is appreciated. public of the bases of data. Detection of vessels and optic nerve head with Radon Transform and overlapping windows. 21 Identification of candidates to microaneurysms with unsupervised method. Final ranking of true MAs through SVM, generating DR diagnosis early. Intelligent Prediction Approach for Diabetic Retinopathy Using Deep Learning Based Convolutional Neural Networks Algorithm by Means of Retina Photographs 2020 Thomas GAS; Robinson YH; Julie EG; Shanmuganathan V; Rho S; Nam Y Computers Materials & Continues - International(Ind ia, korea) Study experimental / development and validation of model Retinal images: 45,000 for training and approximately 8 500 for testing Accuracy 98.45% in training and 92.15% in testing. Deep learning – Convolutional Neural Network (CNN) CNN trained with retinal photographs (dataset Kaggle Retinopathy Detection). High diagnostic accuracy. It requires a large number of high-quality images. Integrate the model into systems clinical screening. No reports ethical considerations explicit in the article.Time and cost reduction compared to other methods manuals. Cross-validation 94.33% and test final 88.32%. Possible variability due to differences in equipment capture. Validation in environments multicenters and populations various. 22 The model outperformed KNN, SVM, DREAM and GD-CNN in precision Scalability for large image volumes. It was not evaluated in scenarios real-time clinical data. Device Optimization laptops or mobiles. RandomizationDriven Hybrid Deep Learning for Diabetics Retinopathy Detection 2025 AM Mutawa; GR Hemalakshmi; NB Prakash; M. Murugappan IEEE Access Study experimental / development and validation of model 90 of STARE, 30 of HRF and 70 from the FFA (total 190 images). Accuracy: 96.10% Sensitivity: 95.35% Specificity: 97.06% Overall accuracy: 96.10% The model outperformed traditional diagnostic methods in all aspects metrics evaluated. Hybrid Deep Learning: Convolutional Neural Network (CNN, VGG16) combined with Radial Basis Function (RBF) and extraction of MSfeatures DRLBP, with techniques of randomization. Preprocessing of Images: Extraction of MS-DRLBP features and CNN (VGG16). Final classification using a hybrid CNN-RBF classifier with random selection of subsets High precision, sensitivity and specificity. Image sample size relatively small. Apply the approach to other medical imaging modalities and other pathologies. The study used databases only publicly accessible; the authors declare that no ethical approval was required nor consent and there is no conflicts of interest. International (Germany) mania, india) Improved segmentation of vessels and reduction of noise. Validation limited to public databases, without testing in real clinical settings. Explore deeper integrations of randomization in the CNN-RBF architecture. 23 Use of randomization techniques that increase robustness and capacity of generalization. Image dependency of high quality. Develop theoretical models that better explain the interaction between randomization and learning deep. An Effective Diabetic Retinopathy Detection System using Deep Belief Nets and Adaptive Learning in Cloud Environment 2023 Praveen Modi; Yugal Kumar SCALABLE COMPUTIN GPRACTICE And EXPERIENCE AND India Study experimental / development and validation of model 3,200 fundus images (224 with retinopathy diabetic and the rest without the disease Accuracy: 91.28% Sensitivity: 93.46% Specificity: 94.84% F1-Score: 94.14% Evaluated with 10-fold cross validation and compared to KNN, SVM, ANN, InceptionV3, VGG16 and VGG19, showing better performance than all they. Deep Learning – Deep Belief Nets (DBN) with learning adaptive. Preprocessing (color enhancement, kmeans segmentation) and classification with DBN in a cloud environment. High precision and robustness facing problems of lighting and oversaturation. Magazine and exact year not specified. Explore techniques of segmentation based on metaheuristics. No requirements are mentioned ethical guidelines were followed; publicly available data was used (IEEE) Data Port).Validation only on public databases, no testing real clinic. Extract new features using derivatives of primer and second order. Adaptive learning that avoids over-adjustment and improves the training. 24 It depends on good quality images. quality Extend to other types of medical images.Accessible in the environment of cloud for clinical use remote. Barriers and Enablers Influencing the Implementation of Artificial Intelligence for Diabetic Retinopathy Screening in Clinical Practice: A Scoping Review 2025 Jenny Tran, Jose J. Estevez, Natasha J. Howard, Saravana Kumar Clinical & Experiment to the Ophthalmology logy Australia Scoping review 18 studies among 2018 and 2023 18 studies from 10 countries. Barriers and enablers grouped into four domains: health system, health professional, health user and Information technology. Barriers: limited clinical efficiency, imaging protocols, initial costs, mistrust, lack of training, Infrastructure problems and variability in image quality. Facilitators: immediate diagnosis, reduction of workload, training of non-medical staff, patient acceptance, offline or easily integrated systems, and accuracy AI diagnostics. review studies that they used mostly Deep Learning and analysis systems automatic images (e.g., Inception v3, EyeArt, VGGNet, ResNet, etc.). It focuses on the use of learning algorithms in-depth for analysis automatic fundus photography in environments clinical trials for screening and detection of diabetic retinopathy in different settings health contexts. High diagnostic accuracy, speed in delivering results, reduction of the specialist workload, possibility mass screening and in remote areas, improvement of access to care. Implementation costs, variability in image quality, confidentiality concerns, need for training, lack of transparency in some algorithms, infrastructure deficient technology. Develop deliberate implementation strategies, Strengthen training, improve technological infrastructure, adapt models to diverse populations and establish regulatory frameworks clear. Concerns about data confidentiality, medical liability, and perception of motives commercial; the need for transparency is highlighted in algorithms and data protection. 25 Impact of Gold Standard Label Errors on Evaluating Performance of Deep Learning Models in Diabetic Retinopathy Screening: Nationwide RealWorld Validation Study 2024 Yueye Wang, Xiaotong Han, Cong Li, Lixia Luo, Qiuxia Yin, Jian Zhang, Guankai Peng, Danli Shi, Mingguang He Journal of Medical Internet Research China observational, retrospective 736,083 images of fundus examination 237,824 participants The deep learning algorithm showed an initial sensitivity of 79.6% and Specificity was 91.6%. After correcting errors in human labels, sensitivity increased to 92.1% and specificity to 92.1%. The error rate in human labels was estimated at 1.2%, which is sufficient. to significantly affect the performance evaluation of algorithm. Deep Learning Deep learning algorithm trained to detect diabetic retinopathy referable from fundus photographs High sensitivity and specificity after label correction; screening possibility massive at low cost; improved efficiency and reduced workload for specialists. Lower initial performance in realworld environments compared to laboratory validation; dependence on the quality of human labels; the study limited to the Chinese population. Implement control of post-label quality (post-hoc) in programs of screening; validate in other populations and diseases; Consider correcting labels to improve the algorithm's sensitivity without retraining. complete. Committee Approval of Ethics of Zhongshan Ophthalmic Centre (2023KYPJ108). Fully anonymized images. Informed consent written work obtained for the Lifeline Express program. 26 Enhanced detection of diabetic retinopathy using machine learning feature-based selection and ensemble classifiers 2025 Hossein Rabbani, Shahriar Shahidi, Mohammadreza Akhlaghi, Mehdi Akhlaghi Journal of Medical Imaging (SPIE) Iran Investigation original (development and validation of AI algorithm) 749 images of three databases public and a base local. Average sensitivity for detection DR: 100% Average specificity: 93% Microaneurysm detection sensitivity: 95.7% with average of 7 false positives per image. The method showed comparable or better results than other approaches. previous. Radon transform for extraction of features and sorter supervised Support Vector Machine (SVM). The algorithm first identifies blood vessels and optical disc (masking) and then it detects microaneurysms in the preprocessed images, High sensitivity and specificity, robustness to noise, precise detection even of fine vessels, screening capacity massive and reduced workload for specialists. Variability in imaging conditions; requires normalization color; limited number of images in databases; not validated in clinical settings in real time. Expand testing in environments real clinical trials, evaluate against larger databases and diverse, optimize for real-time processing. They were used only public databases and a local base with prior approval; it is not They report ethical risks additional. 27 Effective Fundus Image Decomposition for the Detection of Red Lesions and Hard Exudates to Aid in the Diagnosis of Diabetic Retinopathy 2020 Daniela Echegaray, Pablo Morales, Carmen HernándezMatas, Ana María Mendonça, Aurélio Campilho Sensors Spain Investigation original (development and validation of AI algorithm) -High sensitivity and specificity in the detection of microaneurysms, hemorrhages and hard exudates. machine learning The system separates the images into different layers (components of background and detail) to enhance the injuries and detect them automatically Early detection of key injuries (microaneurysms, hemorrhages, exudates). It requires high-quality images enough. integrate the method into systems clinical screening. They were used exclusively public and anonymous databases; no conflicts of interest are reported no interest nor was additional consent required. It was not tested in clinical settings. in real time. Validate in more populations wide and in settings real. Decomposition of the image into components that allow separation structural details and facilitate the identification of injuries. Reduction of time and workload of the specialists. 28 It depends on the availability of public databases for validation. Explore its combination with deep neural networks for greater accuracy.Improved accuracy compared to other methods traditional preprocessing. High precision and robustness in the face of variations of lightning. A New Early Stage Diabetic Retinopathy Diagnostic Model Using Deep Convolutional Neural Networks and Main Component Analysis 2020 Mali Mohammedhasan, Harun Uğuz Treatment du Signal Türkiye Study experimental / proposal model 2063 images of retina (1468 without symptoms, 595 with symptoms). Accuracy: 91.49% Sensitivity: 94.45% Specificity: 98.93% Accuracy: 98.44% It surpassed benchmark models (AlexNet, VGG, ResNet, GoogleNet) The RUnet-PCA model showed great robustness compared to other methods. Deep Learning Neural Network Convolutional with connections residuals + reduction of dimensionality) Preprocessing of images Edge enhancement Patch generation such as data augmentation. RUnet-PCA Architecture to extract features. PCA to reduce dimensionality and optimize classification. Binary classification High precision and sensitivity in diagnosis early. Limited dataset (curated, does not represent all variability) (real clinical practice). alidar in clinical settings real and more datasets large/diverse. No reports ethical considerations explicit. Less over-adjustment thanks PCA and data augmentation. Evaluated only in images of Kaggle (needs validation) external). Expand to classification multi-level stagesPatches generated: 15,000–17,000 sub-images (48x48 pixels). 29 Combine with others biomarkers or modalities of image. Ability to detect initial symptoms (exudates, microhemorrhages). It does not include images. multimodal (e.g., OCT, angiography). Explore the interpretability and explainability of decisions. It surpasses architectures standard in performance. Diagnostic Accuracy of Artificial Intelligence-Based Automated Diabetic Retinopathy Screening in RealWorld Settings: A Systematic Review and Meta-Analysis 2024 Saint Joseph, Jerome Selvaraj, Iswarya Mani, Thandavarayan Kumaragurupari, Xianwen Shang, Poonam Mudgil, Thulasiraj Ravilla, Mingguang He American Journal of Ophthalmology logy International(Au Stralia, India) Revision systematic and meta-analysis 34 studies included with a total of 134,238 participants; gender reported in 28 studies (49% men, 51% women). Overall accuracy: 81% Sensitivity: 94% (95% CI: 92–96) Specificity: 89% (95% CI: 85–92) DOR: 127 (95% CI: 81–197) Best performance with non-mydriatic images and desktop cameras or smartphones. The AI showed comparable results to the human evaluators. Mostly Deep Learning (93%), some Machine Learning (7%). Use of neural networks deep applied to Images of the fundus (back of the eye). Algorithms classify the presence of any retinal detachment (RD), referable RD, or vision-threatening RD. compared against human evaluators. Comparable performance or superior to experts humans. Reduces workload of ophthalmologists. It allows mass screening in primary environments or rural Compatible with cameras desktop and smartphones. High heterogeneity between studies (camera type, image quality, mydriasis, type of AI). Selection bias and lack of inclusion of unpleasant images in some studies. Publications only in English. Limited applicability in contexts with poor quality of image. Integrate software quality assessment of image. Validate in more countries and diverse populations. Improve accuracy in environments low-income. Strengthen confidence in patients in the use of AI as screening tool. Adherence to the Declaration from Helsinki. AI should be considered a screening tool, not Specialist substitute. Need to inform the patients and guarantee compliance in the derivations. 30 Diagnostic Accuracy of IDX-DR for Detecting Diabetic Retinopathy: A Systematic Review and Meta-Analysis 2025 Zaid Khan, Abhay M. Gaidhane, Mahendra Singh, Subbulakshmi Ganesan, Mandeep Kaur, Girish Chandra Sharma, American Journal of Ophthalmology logy International(Ind ia, Malaysia, Iraq) Revision systematic and meta-analysis 13 studies included with a total of 13,233 participants Pooled sensitivity: 95% (95% CI: 0.82–0.99) Pooled specificity: 91% (95% CI: 0.84–0.95) Area under the curve (AUC): 0.95 High performance in various countries (USA). USA, Italy, Poland, Spain, China, Netherlands, Switzerland, Austria). Validated against international standards (ETDRS, ICDR, EURODIAB). Deep Learning – autonomous system FDA approved (IDX-DR, networks) neuronal convolutional deep). IDX-DR analyzes nonmydriatic retinal images (fundus) obtained with Topcon cameras and determines the presence of referable or vision-threatening DR, without the need for human interpretation. FDA-approved system (First of its kind). Immediate results and self-employed. High sensitivity and specificity. Scalable to environments with limited resources. Potential to reduce costs and improve access in primary care. Heterogeneity between studies and populations. Possible false positives that This can generate anxiety and overload referrals. There is little representation of countries with a high prevalence of diabetes. (e.g., India, China). High initial system cost (~13,000 USD + 25 USD/patient). Evaluate cost-effectiveness in large-scale screening programs scale. Implementation in environments rural and low-income countries income. Compare IDX-DR with other approved systems (EyeArt, AI Media, RetinaVue, AEYEDS). Investigate impact on long-term clinical outcomes (preservation of vision, adherence to treatment). Concerns about privacy and security of data Need for informed consent and transparency in the use of AI. Risk of resistance from professionals in health. Need for strategies of implementation responsible and equitable. 31 Validation of a Deep Learning Model for Diabetic Retinopathy on Patients with Young-Onset Diabetes 2025 Antonio Tan-Torres III; Pradeep A. Praveen; Divleen Jeji; Arthur Brant; Xiang Yin; Lu Yang; Preeti Singh; Tayyeba Ali; Ilana Traynis; Dushyantsinh Jadeja; Rajroshan Sawhney; Dale R. Webster; Naama Hammel; Yun Liu; Kasumi Widner; Sunny Virmani; Pradeep Venkatesh; Jonathan Krause; Nikhil Tandon Ophthalmology logy and Therapy India Study prospective transversal of cohort 321 individuals with ages between 18 and 45 years (642 eyes) For moderate diabetic retinopathy or worst (moderate+ DR): sensitivity at eye level of 95.1% [95% CI 91.0-97.8] and specificity 95.3% [95% CI 93.2-97.2] in all eyes. Deep Learning Automatic detection of diabetic retinopathy moderate or worse (moderate+ DR) and presence of macular edema (DME) to from dilated fundus photography, in young population with type 1 diabetes mostly High sensitivity and specificity for moderate+ DR in young population. Sensitivity for diabetic macular edema (DME) less of what was expected, with false negatives associated with presence Sheen in pictures. Incorporate OCT imaging to improve the accuracy of macular edema detection, especially in the presence of Sheen. Informed consent of the participants; the data were anonymized/deidentified two. High image quality (~99%), which suggests clinical utility wide. Sensitivity in cohort 18-25: 97.9% [95.9-99.3]; cohort 26-45: 92.1% [87.696.0]; difference in sensitivity not statistically significant (p = 0.418). Predominantly single-country (India) cohort of patients with type 1 diabetes → possible lack of generalizability to populations with type 2 or other ethnicities/contexts. Validate the model in more ethnically diverse populations and in other countries. Compliance with the Helsinki Declaration. Good overall performance even in cohorts young people with characteristics different anatomical features (like Sheen). Evaluate performance in patients over 45 years old for comparison.Specificity for moderate+ DR: in young cohort 18-25: 97.8% [96.099.3]; in the older younger cohort: 92.1% [87.6-96.0]; this difference was statistically significant (p = 0.008). Pupil dilation is required for the taking of images, which may not be in all the environments. 32 Possible refinements in model calibration to minimize associated errors with Sheen.It may allow screening effective in population young, where there is less previous studies. OCT was not used as a reference for edema, which could improve the accuracy of DME in future research. For diabetic macular edema (DME): overall sensitivity 78.0% [65.0-87.9]; Sensitivity of young cohort 79.0% [57.9-93.6] vs. younger older cohort 77.5 % [60.8-90.6] (p = 0.893, not significant). Overall specificity for DME 94.6% [92.3-96.6]; high specificity in young people: 97.0% [94.5-99.0] vs 92.0% [88.2-95.5] in cohort Diabetic retinopathy identification based on multi-source free domain adaptation 2024 Guang-Hua Zhang, Guang Ping Zhuo, Zhao-Xia Zhang, Bin Sun, Wei-Hua Yang and Shao-Chong Zhang International to the Journal of Ophthalmology logy China Study methodological of development and validation 9598 patients, age average age 54, 48.2% men, 51.8% women). The SMPL model (Softmax-consistency) Minimization + Pseudo-Label Generator) was tested on two tasks: Method of adaptation of free domain of multiple source (multi-SFDA) network-based neuronal convolutional with generation of pseudo-tags and minimization of consistency softmax. Referable classification vs non-referable DR (DR detection) clinically significant). It does not require access to sensitive source data, preserves privacy. High computational cost for training multi-domain models. Integrate the method into ophthalmological diagnostic systems and telemedicine. Sensitive to noisy or noisy data low quality.DR referable classification (moderate or worse vs. not referable): - Accuracy 90.4%, precision 82.6%, sensitivity 96.9%, specificity 86.0 %, F1-score 0.8917. - It outperformed other domain adaptation methods (DANN, DAN, JAN, ADDA, CDAN) across all metrics. It reduces the need for large volumes of labeled data. Apply in rural areas with limited access to ophthalmologists.Normal vs abnormal DR classification (presence or absence of any degree). Limited direct clinical validity: It was evaluated only on public datasets, not on patients. Clinical validation with real patients for confirm applicability. Improves generalization across domains (train on one dataset and predict in other). Its performance in other diseases or imaging modalities. Expand to other diseases ocular and modalities of image.Normal vs. abnormal classification (any degree of DR): Accuracy 97.9%, precision 98.2%, sensitivity 97.7%, specificity 98.2% %, F1-score 0.9795. It outperforms other methods of adaptation in sensitivity, specificity and F1. 33 It once again outperformed other algorithms reference. Confusion matrix analysis (p. 9): most images were correctly classified; the rate of The error was low. Ablation studies: showed that Clinical evaluation of AI-assisted screening for diabetic retinopathy in rural areas of midwest China 2022 Shaofeng Hao, Changyan Liu, Na Li, Yanrong Wu, Dongdong Li, Qingyue Gao, Ziyou Yuan, Guanyan Li, Huilin Li, Jianzhou Yang, Shengfu Fan PLOS ONE China Study prospective 3796 patients type 1 and 2 diabetics (1295 men, 2501 women; age 19–87 years) Prevalence of DR: 22.7% by ophthalmologists and 22.5% by IA. Overall agreement: 81.6% (κ=0.752, good agreement). Sensitivity of the AI: 81.2% (95% CI: 80.3–82.1%); specificity: 94.3% (95% CI: 93.7–94.8%). At the patient level (less severe eye): sensitivity 84.6%, specificity 95.0%. When considering a random eye: sensitivity 85.3%, specificity 93.7% (κ=0.777). Differences between counties: Licheng showed better performance (sensitivity 93.2%, specificity 98.0%, κ=0.907) compared to Lucheng (sensitivity 75.5%, specificity 92.0%, κ=0.673). Image quality (cataracts, media opacities) influenced false positives negative and undiagnosable eyes (7.76%). In cases of moderate or more severe NPDR (requiring intervention), the The agreement was satisfactory. (κ=0.621). Deep Learning Automatic detection and classification of DR into five levels (no DR, mild NPDR, Moderate NPDR, NPDR severe, PDR) from fundus photographs. High specificity (>93%), comparable to ophthalmologists. Lower than expected sensitivity compared to previous studies. Improve AI accuracy with larger images quality. It allows mass screening in areas with a shortage of specialists. You cannot classify cystoid macular edema using only fundus images. Expand its use in other rural settings and in screening large-scale population. Reduces time diagnosis and facilitates rapid referrals. Difficulty processing low-quality images (small pupil, cataract). Incorporate improvements to detect complications such as macular edema It did not include patients not registered in the public system health. 34 Conclusion: AI-assisted screening Artificial intelligenceenabled screening for diabetic retinopathy: A real-world, multicenter and prospective study 2020 Yifei Zhang, Juan Shi, Ying Peng, Zhiyun Zhao, Qidong Zheng, Zilong Wang, Kun Liu, Shengyin Jiao, Kexin Qiu, Ziheng Zhou, Li Yan, Dong Zhao, Hongwei Jiang, Yuancheng Dai, Benli Su, Pei Gu, Heng Su, Qin Wan, Yongde Peng, Jianjun Liu, Ling Hu, Tingyu Ke, Lei Chen, Fengmei Xu, Qijuan Dong, Demetri Terzopoulos, Guang Ning, Xun Xu, Xiaowei Ding and Weiqing Wang BMJ Open Diabetes Research & Care China (155 centers diabetes in 26 provinces) Study prospective, multicenter 47,269 patients adults with diabetes (average age 54.3) years; 57.4% men) 94,199 images were acquired from Fundus photographs (one per eye). The deep learning algorithm classified the images into 5 stages of severity. from the Dominican Republic and also detected edema Diabetic macular disease (DME). Deep Learning (six-piece ensemble) neural networks based on architecture Inception-ResNet v2) Automatic image classification in 5 stages of DR and DME detection High sensitivity and specificity, comparable with specialists. Use of a single non-mydriatic image per eye, which may underestimate the true prevalence. Extend the use of the system to more health centers and primary care. Scalability in environments real clinical cases. Differences between the estimate prevalence by IA and by specialists (IA tended to overestimate in some cases). Improve image acquisition quality and train technicians. For validation, 31,498 were used Images of 15,805 patients Reviewed by specialists: Sensitivity for detecting referable DR: 83.3% (95% CI: 81.9–84.6). Specificity: 92.5% (95% CI: 92.1–92.9). Negative predictive value: 97.4%. Agreement with specialists: 83%. comparable to the variability interobserver (84.3%). Reduction of time and resources needed to screening programs. Evaluate the implementation for routine screening in national clinical practice. Variability in the quality of image. Ability to function with multiple camera models and in non-centers specialized. 35 Estimated prevalence detected by the system in 40,665 participants with Pleasant images: Any RD: 28.8% (95% CI: 28.4–29.3). Referable RD (moderate or worse NPDR): 24.4% (95% CI: 24.0–24.8). Vision-threatening DR (severe or worse NPDR and/or clinically MSD) Artificial intelligence using deep learning to screen for referable and visionthreatening diabetic retinopathy in Africa: a clinical validation study 2019 Valentina Bellemo, Zhan W. Lim, Gilbert Lim, Quang D. Nguyen, Yuchen Xie, Michelle YT Yip, Haslina Hamzah, Jinyi Ho, Xin Q. Lee, Wynne Hsu, Mong L. Lee, Lillian Musonda, Manju Chandran, Grace Chipalo-Mutati, Mulenga Muma, Gavin S. W. Tan, Sobha Sivaprasad, Geeta Menon, Tien Y. Wong and Daniel SW Ting The Lancet Digital Health Zambia Study prospective of clinical validation in program of screening population 1,574 patients with diabetes. 3,093 eyes examined. 4,504 images of fundus. 56.2% women and 43.8% men. Average age: 55 years. Prevalence detected: Referable RD: 22.5% of eyes. Vision-threatening RD: 5.5%. Diabetic macular edema: 8.1%. Deep Learning (ensemble of two) CNN: VGGNet adapted and ResNet) Automatic classification of retinal images to identify referable RD, vision-threatening RD and diabetic macular edema High sensitivity for detect critical cases (almost no relevant positives were lost). Costs and requirements of infrastructure (telecommunications and computing power) can limit its adoption. Expand validation in African populations with varying degrees of severity and on different devices. Model accuracy (AUC, sensitivity, specificity): AUC for referable DR: 0.973 (95% CI: 0.969–0.978). Sensitivity: 92.25%. Specificity: 89.04%. Sensitivity to threatening DR: 99.42%. Sensitivity to macular edema diabetic: 97.19%. It worked with images obtained by non-staff doctor. Adapting referral thresholds to contexts of limited resources. A large proportion of patients do not I knew his type of diabetes (limitation of clinical data). Generalizable among different ethnicities and teams of image. Evaluate cost-effectiveness and long-term results of the implementation of AI in national programs of screening Over-referral of patients stable could overload services in contexts with few specialists. 36 Risk factors identified (by both AI and human evaluators): longer duration of diabetes, higher HbA1c levels and blood pressure elevated systolic blood pressure. Heat map visualizations: demonstrated the regions of the retina used by the model, often highlighting details that could even go unnoticed by specialists in low quality images. Collaborative Deep Learning for Privacy Preserving Diabetic Retinopathy Detection. 2022 Mahmut Karakaya, Ramazan S. Aygun, Ahmed B. Sallam 44th Annual International to the Conference e of the IEEE Engineering g in Medicine & Biology Society (EMBC) USA Article by conference (experimental, multicentric, validation with databases) 5 public databases: EyePACS (1685 images), Messidor (560), IDRiD (323), SUITABLE (703), UoA (141) Methodology:AlexNet was used with transfer learning.Each dataset trained the model separately. Then, in the In a collaborative approach, the model trained on the lowest-performing dataset was moved to the next dataset. in ascending order of performance, repeating the process until reaching the highest performing one (EMAI chain). Results:Traditional crosstraining experiments They showed variability (60–80% accuracy). Traditional dataset fusion achieved 84%. The collaborative method progressively improved the precision after each stage, reaching 93.5% accuracy (EyePACS combination → Messidor → APTOS → IDRiD). Conclusions:The collaborative approach It surpasses both individual training and traditional dataset fusion, and allows for model improvement without need to share images sensitive. Deep learning (CNN - AlexNet with transfer learning) Classification of retinal images into two classes:No DR vs. DR (they grouped together labels moderate/severe/proliferative (VAT as DR) - Extend the model to multilevel classification (mild, moderate, severe, proliferative). - Try using larger datasets and different types of datasets populations. - Apply the collaborative strategy to other architectures (ResNet, VGG, GoogleNet). - Explore implementation in real clinical settings such as federated network. No reports ethical considerations specific to the article; only the importance of the privacy in the use of medical data. - Preserves patient privacy (no share images, just trained models). - Progressive performance improvement by combining datasets. - It outperforms individual models and the fused dataset approach. - Reduces dataset biassmall or of low quality. - Dependence on the order of the datasets (EyePACS must be used first). - Reduction of classes to only binary (NoDR vs DR), losing granularity in classification. - Results vary depending on the quality of the dataset. - Lack of prospective clinical validation in patients real. 37 Automated Diabetic Retinopathy Diagnosis for Improved Clinical Decision Support. 2025 Justin Boyle, Janardhan Vignarajan, Sajib Saha MEDINFO 2023 — The Future Is Accessible Australia Study of validation of AI model in environment of telemedicine (evaluation of performance in clinical cohort real 328 unique patients (after purging of multiple records). 254 patients evaluated for DR after exclusion for poor quality of image Image quality detection: 72% accuracy at image level, 85% at level of patient. - DR detection: 85% accuracy at the image level, 87% at the patient level. - Sensitivity: up to 100% at the level of patient. - Specificity: 84–86%. - Greater importance is placed on sensitivity to avoid false positives negative. Deep Learning AI analyzes automatically Images of the fundus of the eye uploaded to the platform of Telemedicine, first evaluating quality and then broadcasting DR diagnosis, compared against ophthalmologists such as gold standard Integrates assessment automatic in a service of teleophthalmology in remote communities; high sensitivity; support for ophthalmologists; reduces risk of loss diagnosis in populations with access limited. Lower specificity (false positives); depends on quality image; algorithms trained with other bases of data, not adapted specifically to the indigenous population; relatively number limited number of patients. Extend validation to more communities; improve Image quality algorithms; optimize sensitivityspecificity balance; integrate into clinical workflows to reduce specialist workload. They are not reported in a explicit, although the context is population indigenous people in remote areas of Australia (which implies ethical sensitivity and cultural). 51 Use of Artificial Intelligence in Diabetic Retinopathy Screening: Experience in a Health Service in Santiago, Chile 2024 María C. IbáñezBruron; Andrea Cruzat; Gonzalo Orders-Caviere; Marcelo Coria Magazine Medical Chili Chili Study observational cross Total population: 366 participants 366 patients were recruited, of whom 88 (24%) had some degree of RD according to ophthalmologists, and 33 (9%) had severe or worse NPDR in at least an eye. EyeArt Software (Eyenuk, Inc.), platform artificial intelligence in the cloud Screening for detect retinopathy severe diabetic not proliferative or worse, using retinal images mydriatics, compared with clinical evaluation of fundus examination by ophthalmologist Perfect sensitivity for detect severe or worse NPDR in this sample (100 %) Participants who signed consent but withdrew before the ophthalmological examination reference → reduction of wearable sample Evaluate everyday use of algorithm and its costeffectiveness in a field study of the Health Service Metropolitan South East (SSMO) Average age: 61 ± 14 years old No training required additional staff health in units primary care ophthalmological (UAPOs) to use EyeArt Sex: 61% women; 39% men EyeArt detected 85 positive cases (23 %), with definition of positivity = severe or worse NPDR or non-diagnostic imaging evaluable. Performance depends on the threshold cut-off point: when using a less stringent cut-off point (moderate NPDR) or worse) sensitivity and specificity are reduced Develop a pilot of screening program articulated between different levels of care that answers questions about implementation and monitoring and timely referral Sensitivity for severe or worse NPDR: 100% (95% CI: 89-100). Specificity: 84% (95% CI: 80-88). Positive predictive value (PPV): 39%. Negative predictive value (NPV): 100%. The possibility of carrying out the Screening without pupillary dilation (mydriasis) reduces costs, time, risks and improves adhesion of patients False positives: many cases classified by EyeArt as positive results mild/moderate according to ophthalmologist 52 With a moderate NPDR cutoff point or worse: sensitivity 94% (95% CI: 85-98) and specificity 74% (95% CI: 6878). There were 52 false positives, of which more than half corresponded to mild or moderate DR according to ophthalmologists. The algorithm had problems with patients who dropped out before Accuracy of Autonomous Artificial Intelligence-Based Diabetic Retinopathy Screening in Real-Life Clinical Practice 2024 Eleonora Riotto, Stefan Gasser, Jelena Potic, Mohamed Sherif, Theodor Stappler, Reinier Schlingemann, Thomas Wolfensberger and Lazarus Konstantinidis Journal of Clinical Medicine Swiss Study observational retrospective in clinical practice real Total: 1350 patients (2700 images) initials) Image quality: 418 images (209 patients) were excluded for “insufficient quality” according to IDx-DR. Autonomous system IDx diagnosticDR (Digital Diagnostics, Coralville, IA, USA), based on deep learning and approved by the FDA De novo diagnosis and automatic classification of diabetic retinopathy at five levels according to the ICDR scale (no DR, mild, moderate, severe, proliferative). High sensitivity and NPV (100%) to rule out DR. Tendency to overestimate the degrees of DR (many false positives) positives). Improve specificity through algorithms advanced and more datasets various.It allows for safe screening in primary care without need to ophthalmologists. Included after excluding Unsuitable images: 1141 patients (2282 images) Patients classified without RD: 1590 Images (795 patients) → confirmed by specialists as absence of DR (100% concordance). Very low PPV for moderate (1.4%) and severe (0%). Validate in populations with higher prevalence of DR moderate and severe. 53 Few patients had moderate DR and none had severe DR confirmed in the sample. Time and cost savings in detection programsPositive AI classifications: 692 images (346 patients) classified such as mild, moderate or severe → limited agreement with specialists. Integrate techniques of explainable AI and learning in assembled to reduce false positives. Exclusion of low-quality images limits applicability in Artificial intelligencebased screening for diabetic retinopathy at community hospital. 2019 Jie He, Tingyi Cao, Feiping Xu, Shasha Wang, Haiqi Tao, Tao Wu, Liyan Sun, Jili Chen Eye (London) China Study prospective, transversal in around community 889 patients diabetics (418 men, 471 women), age The average age was 68.5 years. A total of 3,556 data points were obtained. retinal images (2 fields per eye). Methodology:Fundus non-mydriatic photographs with Topcon TRC-NW400 (maculaand disk-centered). AI based on Inception-v4,Pre-trained in ImageNet and adjusted to retinographies, installed in the community hospital. Comparison against independent diagnosis of 2 ophthalmologists (consensus if discrepancy). Results:Prevalence of detected DR: 16.1% (ophthalmologists) vs 16.3% (IA). Forany DR: Sensitivity 90.8%, specificity 98.5%, AUC 0.946. For referable DR (≥ moderate NPDR and/ or macular edema): Sensitivity 91.2%, specificity 98.8%, AUC 0.950. Conclusions:AI achieved performance comparable to specialists, with high precision and feasibility for implementation in community hospitals in China. Deep learning (CNN Inception-v4, TensorFlow) Automatic classification of 5 levels of DR according to ICDR (No DR, Mild NPDR, Moderate NPDR, NPDR) severe, PDR) + referable DR - Multicenter validation on a larger scale and in different regions. - Include detection of clinical macular edema and other pathologies. - Improve algorithms for low-quality images. - Analyze the cost-effectiveness of the implementation community. Shibei Hospital Ethical Approval (ChiCTR1800016785). Informed consent obtained. Funded by projects of the Shanghai Health Commission and local districts. Authors declare no conflicts of interest. - High sensitivity and specificity (>90%). - Implementationin situ in a community hospital. - Immediate results and reference report. - Reduces the workload of specialists and facilitates mass screening. - Meets standards international screening. - Relatively small sample size (889 patients). - Difficulty in imaging with small pupils or cataracts (nongradable). - It does not detect macular edema with certainty, only in the fundus. - Evaluation limited to a community hospital in Shanghai. 54 Artificial Intelligence Detection of Diabetic Retinopathy 2023 Jennifer I. Lim, Carl D. Regillo, SriniVas R. Sadda, Eli Ipp, Malavika Bhaskaranand, Chaithanya Ramachandra, Kaushal Solanki (EyeArt Study Subgroup) Ophthalm ology Science USA (multicentric, 10 centers) Rehearsal prospective, pivotal, multicenter 521 patients (999 eyes) with diabetes ≥18 years, without eye treatment Previous. Recruited between 2017–2018 in 10 centers. Methodology:Two 45° NM images (disccentered and macula-centered) captured with Canon CR-2 cameras; automatic analysis withEyeArt AIin the cloud; dilation if image is not evaluable. Gold standard: centralized reading (Wisconsin Fundus Photograph Reading Center) with 4 dilated stereoscopic fields (ETDRS). AI vs dilated ophthalmoscopy was compared (general practitioners and retina specialists). Results:For mtmDR: EyeArt sensitivity 96.4%, specificity 88.4%, gradability 97.4%. General practitioners: sensitivity 20.6%, specificity 99.8%. Retinologists: sensitivity 59.5%, specificity 98.9%. EyeArt and retinologists did not miss any cases of visionthreatening DR. Conclusions:EyeArt significantly outperforms general ophthalmologists in sensitivity and also surpasses retina specialists, although with less specificity. It is a viable screening tool in primary care, reducing the workload for specialists. Deep learning (EyeArt: multiple neural networks convolutional integrated, trained with > 375,000 images) Automated screening in 2 NM fields → classification in mtmDRandvtDRat eye level - Improve specificity without compromising sensitivity. - Validation in more diverse populations and clinical workflows. - Extend to other retinal pathologies. - Evaluate impact on adherence and longterm visual outcomes. - Integrate into national screening programs. Approved by University of Illinois at Chicago IRB;informed consent obtained. ClinicalTrials.gov Registration NCT03112005. Financed byNIHandEyenuk Inc. Several authors Eyenuk employees/ shareholders. - Very high sensitivity (>96%). - Avoids loss of severe cases (vtDR). - Speed: results in <1 min. - Operation at the primary care point. - Reduces the burden on specialists. - Broad coverage with minimal training technique. - Lower specificity (~88%) → more false positives and overreference. - Dependence on NM images (may miss lesions) peripheral). - It does not evaluate other pathologies (glaucoma, cataract, refraction). - Frequent false positives in nonDR pathologies (drusen, occlusions, ERM). 55 Evaluation of an Artificial Intelligence System for the Detection of Diabetic Retinopathy in Chinese Community Healthcare Centers 2022 Xiuqing Dong, Shaolin Du, Wenkai Zheng, Chusheng Cai, Huaxiu Liu and Jiangfeng Zou Frontiers in Medicine China Study cross, multicenter 443 patients diabetics (848 eyes); 283 men (63.9%) and 160 women (36.1%); average age 52 years According to the reference standard (ophthalmologists): 582 eyes (68.6%) without RD. 33 (3.9%) mild NPDR. 121 (14.3%) moderate NPDR. 68 (8.0%) Severe NPDR. 44 (5.2%) PDR. 56 (6.6%) with diabetic macular edema (DME). Deep Learning (network convolutional developed by Shanghai EagleVision Medical Technology Co., Ltd., CARE system) Automatic detection of DR and DME in single-field retinal images without dilatation. High specificity in any RD and mtmDR. Low sensitivity for vtDR and DME. Conduct studies multicenters with greater sample size. High frame rate pleasant (99.1%). Using single-field photography may miss injuries peripherals. Improve the algorithm for advanced DR detection and DME.Feasibility for use in community hospitals. 56 Limited sample size for severe DR subgroups Add images multimodal (e.g., OCT)Speed in generating reportsCARE system performance (Table 2, page 4): Any RD: sensitivity 75.2%, specificity 94.0%, PPV 85.1%, NPV Diagnostic Accuracy of Hand-Held Fundus Camera and Artificial Intelligence in Diabetic Retinopathy Screening 2023 Martina Tomić, Romano Vrabec, Đurđica Hendelja, Vilma Kolarić, Tomislav Bulum, Dario Rahelić Biomedicine nes Croatia Study transversal of validation (IDF Diabetic Retinopathy Screening Project) 160 patients with Undiagnosed type 2 diabetes RD previous (320 eyes), 89 men/71 women, middle age 65 years, duration median diabetes 14 years old Methodology: (1) All patients: clinical examination with slit-lamp biomicroscopy. (2) Retinal photography with standard Zeiss VISUCAM camera. (3) Photography with TANG portable camera (2 fields, macular and optic disc) → evaluated by DeepDR AIand an external ophthalmologist from IDF. (4) Comparison between methods (AI, standard, clinical). Results:AI + portable camera had an AUC of 0.921 vs. clinical examination and 0.883 vs. standard camera. Sensitivity was 89.1% and specificity 100% compared to clinical examination; sensitivity was 83.2% and 100% specificity in front of the camera Deep learning (DeepDR, CNN trained to DR classification) De novo diagnosis: distinguishes between “no DR” vs “with DR” (mild NPDR to PDR). Basic classification: It is grouped into two categories: (1) Non-DR/MiDR (monitoring) and (2) ≥Moderate NPDR/PDR (moderate or severe non-proliferative retinopathy and retinopathy) proliferative (PDR) (refer). Advantages: - High sensitivity and specificity (>83–89% and 100%). - Portable, fast and simple method. - Suitable for use by nonmedical personnel (nurses). - Reduces the burden on specialists. - Comparable to the standard clinical examination. Limitations: - Single-center study, small sample. - Variability in image quality (22.5% medium level). - Differences in the detection of mild NPDR. - Without multicenter validation. - Validate in multicenter studies and population. - Improve detection of mild NPDR. - Integrate into telemedicine and national screening programs. - Expand to other eye pathologies. Approved by the Ethics Committee of Merkur University Hospital (Croatia, protocol 04/38-299). Informed consent signed. Funding: no specific support. Authors declare no conflicts of interest. 57 Deep Learning Algorithm Detects Presence of Disorganization of Retinal Inner Layers (DRIL)–An Early Imaging Biomarker in Diabetic Retinopathy 2023 Rupesh Singh, Srinidhi Singuri, Julia Batoki, Kimberly Lin, Shiming Luo, Dilara Hatipoglu, Bela Anand-Apte, Alex Yuan Translation to the Vision Science & Technology USA Study transversal / development of model 664 patients (1201 eyes, 5992 B-scans OCT) 53.1% women (638) 46.9% men (563) Average age: 69 years (±11) The best performance was obtained with modified GoogleNet (transfer Learning, weight freeze + data augmentation): Accuracy: 88.3%, Specificity: 90.0%, Sensitivity: 82.9%, AUC: 0.93 Deep Learning – Neural Networks Convolutional CNNs training with OCT B-scans manually labeled for presence or absence made of DRILL. Rapid automation of the detection of early biomarkers in DR. Retrospective study and of a center only. Extend to multicategory classification (severity of DR, DRIL and DME). Approved by the Cleveland Clinic ethics committee. Dataset limited in diversity population. Compliance with Helsinki Declaration and HIPAA. Expand the dataset to include greater ethnic and clinical diversity.High accuracy with dataset relatively small. 58 Transfer learning with adjustment of initial weights and data augmentation (rotation, reflection). Need for further validation external and multiclinic.Cohen's kappa > 0.85 in evaluation DRIL manual. Integration into screening programs and clinical trials. Anonymized images before analysis.Visual explainability with Grad-CAM maps. Possible exclusionary bias low quality images.Grad-CAM showed that the network focused correctly on the central fovea. CNN optimization for reduce time training. Binary classification (“DRIL present” / “DRILL Potential for support clinical and standardization Advancing Diabetic Retinopathy Screening: A Systematic Review of Artificial Intelligence and Optical Coherence Tomography Angiography Innovations 2025 Alireza Hayati, Mohammad Reza Abdol Homayuni, Reza Sadeghi, Hassan Asadigandomani, Mohammad Dashtkoohi, Sajad Eslami, Mohammad Soleimani Diagnostic s International (Ira n, States joined) Revision systematic 32 studies included; sizes sample Individuals vary between 76 and 2640 images; some with more than 1000 participants. Deep learning algorithms (CNNs, ViTs (hybrids) outperformed ML traditional in sensitivity, specificity and AUC. Some models achieved >99% of accuracy and AUC. The multimodal combination (OCTA + fundus + structural OCT) improved results. Especially high sensitivity for Severe DR, but limited in stages early. ViTs showed an advantage in interpretability and overall accuracy. Deep Learning (CNNs, ViTs, multibranch CNNs, CNN+ML hybrids); some studies with traditional ML and ANNs. Automated analysis of OCTA to identify microvascular changes early and classify severity. Use of CNNs, ViTs and hybrid architectures for binary detection (DR/no) DR) and classification multilevel (mild NPDR, moderate, severe, PDR). Multimodal models that OCTA + fundus + integrate clinical data High sensitivity and specificity. Load reduction in ophthalmologists. Possibility of mass screening and telemedicine. Diagnostic potential early non-invasive. Interpretive improvement with XAI and ViTs. Variability between devices OCTA and protocols of acquisition. Lack of standardization and homogeneous datasets. Generalization problems (many models trained in internal bases). DL's "black box" that reduces clinical confidence. Difficulty in detecting very DR early. Create large, multicenter, and public datasets. Validate in scenarios real clinical cases. Standards for acquisition from OCTA and report of results. Using federated learning to protect data and extend training. Cost-based evaluations effectiveness and applicability in telemedicine. Need for informed consent and Privacy protection. Concern about bias. algorithmic and lack of generalization. Call for implementation responsible and equitable. Emphasis on XAI (explainable AI) to increase clinical confidence. 59 A Classification Tree Model with Optical Coherence Tomography Angiography Variables to Screen Early-Stage Diabetic Retinopathy in Diabetic Patients 2022 Hongyan Yao, Shanjun Wu, Zongyi Zhan and Zijing Li Journal of Ophthalmology logy China Study transversal with development and validation of models predictive Total: 241 patients with T2DM Two models were compared: logistic regression and classification tree, both wearingOCTA variables (superficial capillary plexus density (SCP), deep capillary plexus density (DCP), peripapillary radial plexus density (RPCP), FAZ area, BCVA, duration of DM). Tree of classification (classification and regression tree, CART) Identification and discrimination by DR early (mild to moderate NPDR) vs absence of DR in patients with type 2 diabetes, using OCTA variables Simple and practical process for clinical decisions. General and not very predictive precise. Conduct prospective studies with larger populations. No calculations required complex real-time systems, suitable for clinical use straight. Small sample size and hospital → bias. Improve the simplification of the OCTA process to do so more accessible. Retrospective study. Validate and refine the model. 60 Logistic regression model: included BCVA (LogMAR) and SCP density as significant predictors (OR=60.30, 95% CI 2.40–1513.82, p=0.013; SCP density OR=0.86, 95% CI 0.78–0.96, p=0.006). OCTA is non-invasive, quantitative, fast and without pupillary dilation. Lack of validation in cohorts large multicentric. Explore OCTA combination with other methods quantitative. AUC: 0.75 (95% CI 0.66–0.85), sensitivity 63%, specificity 83% in Automatic diagnosis of diabetic retinopathy using Vision Transformer based on wide-field optical coherence tomography angiography 2024 Zenan Zhou, Huanhuan Yu, Jiaqing Zhao, Xiangning Wang, Qiang Wu and Cuixia Dai Journal of Innovative Optical Health Sciences China Study methodological of development and technical validation with images of OCTA 288 patients diabetics and 97 healthy subjects (images acquired) with SSdevice OCT). Models were trained and validated with 12x12 mm WF-OCTA images centered on the fovea. Vision Transformer applied to WF-OCTA images Automatic classification of imaging at four levels: no DR, mild NPDR, moderate/severe NPDR and PDR Greater accuracy and speed in comparison with traditional CNN. Need for more WFimages High-quality OCTA for generalization. WF-OCTA could become the diagnostic standard for DR. All participants gave their consent. informed After increasing the data (38,777 images generated), the model achieved: Accuracy for DR detection (binary): 99.55% Sensitivity: 99.49% Specificity: 99.57% Accuracy for classification by stadiums: 99.20% Ability to capture more relevant details through the mechanism of attention. Limitations of the initial dataset (small number, data increase was required). As the availability of WFimages increases OCTA, the vision transformer model and its extensions More may be applied widely used in diagnostics clinical. 61 Not yet evaluated in environments broad clinical practice.Rapid diagnosis (0.027 (s per image). Possible substitute no invasive angiography fluorescein. Performance by category (Table 2, page 7): Normal → Accuracy 98.56%, Recall Detection of Diabetic Retinopathy Using Extracted 3D Features from OCT Images. 2022 Mahmoud Elgafi, Ahmed Sharafeldeen, Ahmed Elnakib, Ahmed Elgarayhi, Norah S. Alghamdi, Mohammed Sallah, Ayman El-Baz Sensors International:Egi pto (Mansoura University), USA. U.S. (University) of Louisville), Saudi Arabia (Princess Nourah Univ.), Egypt (Higher Institute of Engineering and Technology) Article by investigation (experimental, with images clinics) 188 cases (100 normal,88 RD); OCT acquired in University of Louisville Hospital Methodology Three-dimensional OCT volumes were obtained at the University of Louisville Hospital (Zeiss-Cirrus HDOCT 5000). Each volume included 5 Bscan slices per eye, with minimal signal accepted 7/10. The 12 retinal layers were automatically segmented using a model based on random fields of Markov and shape priors. Two types of features were extracted from each layer: primer reflectivity 3D order and thickness. Neural networks were trained individual per each Neural networks artificial with backpropagation (fusion NN) Automatic segmentation of 12 retinal layers in OCT, extraction of 3D features (reflectivity and thickness), normal/RD classification High accuracy, sensitivity, and specificity; leverages 3D information, better than 2D approaches Need for more data; training with limited cases; comparison with 2D methods may not be completely fair Test with datasets larger; integrate other ML/DL algorithms; evaluate robustness and generalization clinic Approval of theIRB of the University of Louisville; compliance with Declaration of Helsinki; informed consent obtained; no conflicts of interest declared interests 62 A deep learning model for identifying diabetic retinopathy using optical coherence tomography angiography 2021 Gahyung Ryu, Kyungmin Lee, Donggeun Park, Sang Hyun Park, Min Sagong Scientific Reports South Korea Study retrospective, external validation 301 eyes in initial cohort (51 healthy, 51 DM without DR, 53 mild NPDR, 49 NPDR moderate, 48 NPDR severe, 49 PDR). 240 valid datasets after exclusions. External validation: 195 additional eyes (120 datasets finals) Methodology:OCTA 3×3 mm² and 6×6 mm² (Optovue). Ground truth based on ultra-widefield fluorescein angiography (UWF FA)reviewed by retinal specialists. CNN basedResNet101 pretrained on ImageNet,trained end-to-end. It was compared to an ML model that used UNet for Segmentation and extraction of 4 features (vascular density, skeletal density, fractal dimension, FAZ area). Results:CNN achieved AUC 0.93–0.97 for DR onset detection (sensitivity 91–98%, specificity 85– 93%) and AUC 0.94–0.98 for DR Deep learning (CNN ResNet101, end-toend). Compared with classic ML feature-based. Automatic DR classification (none, NPDR - non-proliferative, PDR - proliferative) and referable state from OCTA images (SCP, DCP, full retina, combined). - Expand dataset to a larger scale and across multiple centers. - Include cases with macular edema and low quality to test robustness. - Directly compare OCTA vs fundus in automated diagnosis. - Explore finer multicategory classification and multi-modal integration (fundus + OCTA). Approved by IRB of Yeungnam University Medical Center (No. 2020-02-003). Written consent waived because it is a retrospective study. Compliance with Declaration of Helsinki. Authors declare no conflicts of interest. - Use of OCTA, which captures microvasculature in depth. - Robust ground truth with UWF FA (more reliable than standard fundus). - End-to-end model avoids extraction bias features manual. - Good performance even with 3×3 mm² images (small macular area). - External validation consistent. - Sample size still relatively small. - Exclusion of cases with macular edema, low quality images or with artifacts. - Lack of validation in real-world multicenter clinical practice. - It was not directly compared with fundus photography in the same cohort. 63 Early Diabetic Retinopathy Detection from OCT Images Using Multifractal Analysis and Multi-Layer Perceptron Classification 2025 Ahlem Aziz, Necmi Serkan Tezel, Seydi Kaçmaz, Youcef Attallah Diagnostic s (MDPI) Türkiye Study experimental with public dataset Public dataset Retinal OCT Image Classification—C8 (Kaggle),24,000 retinal images with 8 conditions; for this study: 6000 images balanced (normal vs DR). No clinical/ demographic data s. Methodology: (1) Preprocessing: Gaussian blurring + CLAHE + binarization. (2) Extraction of9 multifractal descriptors: generalized dimensions (Db, Di, Dc), singularity spectrum (αmin, αmax, αcenter, f(α)max), spectral width, symmetric shift. (3) Classification with various algorithms (LR, SVM, DT, RF, XGBoost, LightGBM, Gradient Boosting, MLP). (4) Validation with5-fold cross-validation. Results:MLP performed better: accuracy 98.02%, precision 98.24%, recall 97.80%, specificity 98.84%, F1-score 98.01%. XGBoost and LightGBM also performed better. Machine learning hybrid (extraction) features manual multifractals + classification with MLP) Binary classification (normal vs DR) in OCT images, based on retinal structural complexity Advantages: - High accuracy and sensitivity (>98%). - Use of standard OCT, more accessible than OCTA. - Interpretable features (explainability) mathematics). - Low cost computational (without need for large GPUs). - Scalable and clinically applicable model. Limitations: - Public dataset without clinical metadata (age, sex, evolution). - Binary classification only (normal vs DR, without severity grading). - Validation restricted to public OCTs, not real clinical cohorts. - Validate in clinical and multicenter settings. - Extend to multilevel classification (mild, moderate, severe, proliferative). - Integrate with other functional biomarkers (e.g., ERG). - Merge OCT with other modalities (fundus, OCTA). Kaggle public dataset (no ethical approval required). No funding. External. Authors declare no conflicts of interest. 64 Diabetic Retinopathy Screening Using Smartphone-Based Fundus Photography and Deep-Learning Artificial Intelligence in the Yucatan Peninsula: A Field Study 2023 John J. Wroblewski, Ermilo SanchezBuenfil, Miguel Inciarte, Jay Berdia, Lewis Blake, Simon Wroblewski, Alexandria Patti, Gretchen Suter and George E. Sanborn Journal of Diabetes Science and Technology Mexico Study of field, observational, retrospective, no interventional validation of AI algorithms Total: 248 patients with diabetes 2130 images were acquired. Deep learning (neural networks) convolutional integrated into the Media and systems EyeArt) Automatic detection of any degree of diabetic retinopathy in images captured with portable cameras fundus mounted on smartphones. High sensitivity and specificity in detection from the Dominican Republic. Unconventional definition of “ground truth” (clinical examination) combined with reading masked images, no international standard). Implement mass screening programs in regions rural areas using these technologies. “Reference truth”: 129 patients 119 had some degree of DR and 119 did not They had RD. Sex: 212 women (average age 56.4) years, range 5-80), 36 men (age) average age 55.9 years, range 12-73) Media allows screening immediate offline in areas remote without internet. Basic training in operators to improve image quality. Media (offline): evaluated the 248 patients. Sensitivity: 94% (95% CI 88–97) Specificity: 94% (95% CI 88–98) Positive predictive value (PPV): 95% Negative predictive value (NPV): 93% LR+: 15.95, LR-: 0.07, statistical J=0.879 The use of untrained operators affected image quality, especially at EyeArt. 65 EyeArt offers analysis in the cloud with results automatic standardized. Validate the use in different populations with varying prevalences and distinct phenotypes.No grading by severity levels was performed (only presence/absence of RD).Possibility of implementation in screening campaigns Include severity grading in future studies. EyeArt (online): rated only 156 Limited cohort and prevalence Comparison of smartphone-based retinal imaging systems for diabetics retinopathy detection using deep learning. 2020 Mahmut Karakaya, Recept E. Hacisoftaoglu BMC Bioinforma tics (Suppl. MCBIOS '19) USA Study experimental (images) synthetic and public dataset) DatasetEyePACS Kaggle:13,624 original images retina (No DR, Mild, Moderate, Severe, Proliferative). To balance: 686 were selected No DR and 686 DR. Images of smartphones were simulatedwearing circular cutouts Methodology: (1) Initial capture of synthetic images with an eye model to evaluate field of view. (2) Simulation of the field of view of each device on images of EyePACS. (3) CNNAlexNet with transfer learning (last layers replaced by 2 classes, trained with SGD, minibatch=2, LR=1e−5, momentum=0.9). (4) Evaluation in 690 balanced images. Results:Accuracy with original images: 81.6% (AUC 0.88). With Deep learning (CNN AlexNet with transfer learning) Binary classification (No DR vs DR) from original images and simulated systems smartphone - Validate with imagesreal captured with each device. - Expand to classification by severity (mild-proliferative). - Compare with other more modern CNNs (ResNet, Inception). - Implement in clinical studies in rural areas with equipment shortages. Use of Kaggle public data (EyePACS). No. Additional ethical approval was required. Authors declare no conflicts of interest. Funded by NIH/NIGMS (INBRE P20GM103429). - It allows for the evaluation of the feasibility of portable and economical systems. - Sample simulation impact of the visual field. - AlexNet with TL achieves acceptable results with little data. - iNview close to standard camera image performance. - Smartphone images were simulated, not real images of patients. - Small balanced dataset (686 vs 686). - Binary classification, not multicategory. - No clinical validation in the field. - Results limited to EyePACS. 66 Media – An offline, smartphone-based artificial intelligence algorithm for the diagnosis of diabetic retinopathy 2020 Bhavana Sosale, Aravind R Sosale, Hemanth Murthy, Sabyasachi Sengupta, Muralidhar Naveenam Indian Journal of Ophthalm ology India Study observational, cross 304 patients with diabetes; analysis final in297 patients (59.2% without DR, 11.7% NPDR mild, 13.8% NPDR moderate, 11.1% Severe NPDR, 4% PDR, 20.4% with macular edema). Average age 55 ± 11 years, 42% women. Methodology:Three dilated fields of vision per eye captured with camera Remidio NM FOP 10.Anonymized and graded imagestwo expert retina specialists;In case of disagreement, a consensus diagnosis was made. That agreed-upon diagnosis became the gold standard. The Medios AI algorithm, based onCNN (MobileNet for quality control + two CNN networks for injuries),It was executedOffline on the same smartphone (iPhone 6). Results: Sensitivity for DR Deep learning (CNN: MobileNet for quality + CNNs for DR) Automatic classification of fundus images into: No DR vs DR, Referible DR (≥NPDR) moderate or DME), and VTDR - Validate in multicenter studies and with non-mydriatic images. - Expand training for full severity classification. - Extend detection to other diseases (e.g. macular degeneration). - Compare performance with other algorithms (Google AI, IDx-DR, EyeArt). Ethical approval institutional. Informed consent was obtained. The following were used: anonymized images. Conflicts: Two authors are linked to the company Medios Technologies (software provider). There was no direct funding. - Worksoffline,without needing internet or servers. - High sensitivity for referable DR and VTDR. - Direct integration into smartphone camera - portable. - Fast and user-friendly workflow for technicians. - Potential of scalability in contexts - Small sample size (297 patients). - Dilated images only (not tested on non-mydriatic images). - Algorithm trained only for DR, not for other retinal pathologies. - Slightly lower specificity (some false positives). 67 Simple, Mobile-based Artificial Intelligence Algorithm in the detection of Diabetic Retinopathy (SMART) study. 2020 Bhavana Sosale, Ramachandra Aravind Sosale, Hemanth Murthy, Srikanth Narayana, Usha Sharma, Sahana GV Gowda, Muralidhar Naveenam BMJ Open Diabetes Research & Care India Study prospective, cross, clinical validation 922 individuals with diabetes mellitus recruited; analysis final in900 patients (648 without RD, 252 with RD; Severity: 51 NPDR mild, 163 moderate, 3 severe, 35 PDR). Methodology:Non-mydriatic retinal images (centered retinal disc and macula) obtained with Remidio FOP. Each patient → 4 images. Five Retinologists classified them using the international DR severity scale (gold standard = majority). The Medios AI algorithm (on iPhone 6, offline) included a quality control network (MobileNet) and two Inception-V3 CNNs to classify referable DR. Results:For any DR: sensitivity 83.3%, specificity 95.5%, AUC 0.90. For referable DR: sensitivity 93%, specificity 92.5%, AUC 0.88. For vision-threatening DR (STDR): sensitivity 95.2–98.7%. Kappa AI vs. specialists: 0.8. Conclusions:The smartphone-integrated offline algorithm is feasible, with high performance comparable to international standards, and facilitates screening in lowresource environments. Deep learning (CNN: MobileNet for quality + Inception-V3 for DR) Automatic classification of NM images in:DR present/absent and RD preferable/notreferable, with binary reference output - Expand to multilevel classification (mild, moderate, severe, proliferative). - Validation in other LMICs. - Evaluate cost-effectiveness and workflow. - Incorporate detection of macular edema and post-treatment lesions. Approved by the Ethics Committee of Diacon Hospital (NCT03572699). Informed consent obtained. Funded by Diacon Hospital. Conflicts: some authors with Media relations Technologies/Remidio. - Worksoffline ((without internet). - Integrated into a portable camera, low cost. - High sensitivity for RDR and STDR. - Comparable to FDA standards and international. - Allows screening scalable community. - Only non-mydriatic images; may fail in small pupils. - Algorithm trained for binary/referable DR (not full degrees). - Does not detect post-laser healing as active DR. - Lack of validation in other countries and multicenter contexts. 68 Diabetic Retinopathy Screening Using Artificial Intelligence and Handheld Smartphone-Based Retinal Camera 2022 Fernando Korn Malerbi, Rafael Ernane Andrade, Paulo Henrique Morales, José Augusto Stuchi, Diego Lencione, Jean Vitor de Paulo, Mayana Pereira Carvalho, Fabrícia Silva Nunes, Roseanne Montargil Rocha, Daniel A. Journal of Diabetes Science and Technology and Brazil Study transversal, realworld screening 940 patients with DM2; 824 with images obtained; 679 with images pleasant (age mean 60.8 ± 11.4 years; 64.9% women).They didn't have previous diagnosis from the Dominican Republic,were participants of a screening campaign (Itabuna Diabetes Methodology:Images obtained with a portable cameraPhelcom Eyer (2 fields/ eye, after mydriasis). Degree of DR performed by a specialist independent (gold standard) and by algorithmPhelcomNet (CNN based on Xception, trained on >10,000 Images from the same device). Comparison sensitivity/specificity for mtmDR (more than mild). Results:Sensitivity 97.8%, specificity 61.4%, AUC 0.89. False Deep learning (CNN based on Xception, “PhelcomNet”) De novo diagnosis: Detects the presence of DR in a population without a prior diagnosis. Classification practice: “no/mild DR” vs “>mild DR (mtmDR)”. Advantages: - High sensitivity (>97%). - Portable device, feasible in campaigns massive. - Integration with telemedicine. - Operation in real environment of scarce resources. - Heatmaps (Grad-CAM) They contribute to interpretability. Limitations: - Relatively low specificity (~61%). - Only one specialist as a reference reader (risk of bias). - No complete classification (only binary mtmDR vs no). - High proportion of ungradable images in older patients or those with cataracts. - Multicenter and longitudinal validation. - Optimize specificity without compromising sensitivity. - Expand to a 5-level classification. - Integrate into public screening programs. Approved by the Ethics Committee of the Federal University of São Paulo (#1260/2015). Informed consent obtained. Funding partial: FAPESP. Authors linked to Phelcom Technologies (declared conflict). 69 Diagnostic Accuracy of Community-Based Diabetic Retinopathy Screening with an Offline Artificial Intelligence System on a Smartphone 2019 Sundaram Natarajan, Astha Jain, Radhika Krishnan, Ashwini Rogye and Sobha Sivaprasad. JAMA Ophthalmology logy India Study prospective, transversal and population 213 patients included: 110 women (51.6%) and 103 men (48.4%). System sensitivity and specificity Offline for referable RD (RDR): Sensitivity: 100.0% (95% CI: 78.2–100.0). Specificity: 88.4% (95% CI: 83.2–92.5). Deep Learning (CNNs based on Inception-V3, integrated into a ensemble) Binary classification (RD) referable vs. non-RD) in retinal images captured by smartphone, with possibility of identifying any DR It works offline, no requires internet. Sample size relatively small. Validation in more cohorts large (minimum 1050) patients). Approved by the Committee Ethics of the Aditya Jyot Eye Hospital (Mumbai). Informed consent obtained. Declaration of Helsinki. High sensitivity (100%) for RD referable. It does not classify the grades. Individuals from the Dominican Republic. Improve the algorithm to discriminate between mild DR and referable.Rapid integration into community programs. Somewhat reduced specificity, especially in mild DR and lesions unrelated. For any Dominican Republic: Sensitivity: 85.2% (95% CI: 66.3–95.8). Specificity: 92.0% (95% CI: 97.1–95.4). 71 Standard for implementation in national programs of screening. It offers maps of activation that facilitates interpretability. Need for larger samples to reduce the margin of error statistical. Even when including lower quality images captured by non-specialized personnel, the sensitivity for RDR AI-Human Hybrid Workflow Enhances Teleophthalmology for the Detection of Diabetic Retinopathy 2023 Eliot R. Dow, Nergis C. Khan, Karen M. Chen, Kapil Mishra, Chandrashan Perera, Ramsudha Narala, Marina Basina, Jimmy Dang, Michael Kim, Marcie Levine, Anuradha Phadke, Marilyn Tan, Kirsti Weng, Diana V. Do, Darius M. Moshfeghi, Vinit B. Ophthalmology logy Science USA Study prospective of cohort 2,012 patients with type 1 diabetes or type 2, ≥18 years, without a prior diagnosis of RD or exam in the last year. AI alone (IDx-DR): - Sensitivity: 95.5% (95% CI) 86.7–100). - Specificity: 60.3% (95% CI) 47.7–72.9). - Accuracy: 70%. - Gradability: 62.5%. - Positive predictive value (PPV): 47.7%. Deep learning Autonomous detection of more-than-mild diabetic retinopathy (MTMDR), defined such as ETDRS ≥35 or presence of edema diabetic macular degeneration. High sensitivity. Low specificity, with many false positives. Improve gradability with dilation or better chambers. Formal individual consent exempt, it was reported verbally to the patients. It allows screening in primary care without ophthalmologists. Limited gradability in patients without dilation (36.5%) unpleasant). Conduct prospective trials and randomized. 72 Increased access to screening in populations various. Extend the hybrid strategy to other screening programs.Dependence on the quality of the image. Human teleophthalmology (consensus of specialists): - Sensitivity: 69.5% (95% CI) 50.7–88.3). Adjust AI calibration for balance sensitivity and specificity. It can reduce the burden on specialists and costs in health programs. The time between screening and exam In-person care (median 52 days) may delay care in cases Automated multidimensional deep learning platform for referable diabetic retinopathy detection: a multicentre, retrospective study 2022 Guihua Zhang, JianWei Lin, Ji Wang, Jie Ji, Ling-Ping Cen, Weiqi Chen, Peiwen Xie, Yi Zheng, Yongqun Xiong, Hanfu Wu, Dongjie Li, Tsz Kin Ng, Chi Pui Pang and Mingzhi Zhang BMJ Open China Study multicentric, retrospective 83,465 images of 39,836 eyes of 21,716 patients. Average age: 60 years Men: 44%, Women: 56%. The DL system included five classifiers: image quality, retinopathy, maculopathy gradability, maculopathy and photocoagulation. Deep Learning (neural networks) convolutional: Inception-V3, Xception and InceptionResNet-V2 in ensemble) Automatic detection of referable DR and maculopathy based on photographs of retina, generating decisions at the image level, eye and patient Accuracy and sensitivity comparable to experts. Use of non-stereoscopic imaging → risk of misdiagnosis of DME. Expand prospective studies with rating multicategory (DR 0–5). Assessment multidimensional consistent with guidelines clinics. Less frequent injuries (e.g. intraretinal microvascular abnormality, venous beading) not always detected. Train with more data on rare injuries to fine classification.In external validation, it achieved accuracy. of 91.5–98.0%, sensitivity of 91.7–97.8% and specificity of 90.7–98.1% according to the classifier. 73 Explainability using SHAPheat maps CAM. Apply stratified analysis by age and duration of diabetes and devices. Retrospective study with incomplete data (age, duration of diabetes, devices).For the detection of referable DR in image level, eye and patient: Accuracy: 91.8–96.7% Feasibility for mass screening in real-world contexts Only one classification was adopted