Managing Diabetic Retinopathy with Deep Learning: A Data Centric Overview

Shramana Dey; Zahir Khan; T. A. PramodKumar; B. Uma Shankar; Ashis K. Dhara; Ramachandran Rajalakshmi; Rajiv Raman; Sushmita Mitra

eess.IVApr 2, 2026

Managing Diabetic Retinopathy with Deep Learning: A Data Centric Overview

Authors: Shramana Dey, Zahir Khan, T. A. PramodKumar, B. Uma Shankar, Ashis K. Dhara, Ramachandran Rajalakshmi, Rajiv Raman, Sushmita Mitra

Organizations: Indian Statistical Institute, Kolkata, West Bengal, India · Dr. Mohan’s Diabetes Specialties Centre, Madras Diabetes Research Foundation, Chennai, Tamil Nadu, India · Department of Electrical Engineering, National Institute of Technology, Durgapur, West Bengal, India · Department of Vitreo-Retina, Vision Research Foundation, Sankara Nethralaya, Chennai, Tamil Nadu, India

Abstract

Diabetic Retinopathy (DR) is a serious microvascular complication of diabetes, and one of the leading causes of vision loss worldwide. Although automated detection and grading, with Deep Learning (DL), can reduce the burden on ophthalmologists, it is constrained by the limited availability of high-quality datasets. Existing repositories often remain geographically narrow, contain limited samples, and exhibit inconsistent annotations or variable image quality; thereby, restricting their clinical reliability. This paper presents a comprehensive review and comparative analysis of fundus image datasets used in the management of DR. The study evaluates their usability across key tasks, including binary classification, severity grading, lesion localization, and multi-disease screening. It also categorizes the datasets by size, accessibility, and annotation type (such as image-level, lesion-level, and multi-disease). Finally, a recently published dataset is presented as a case study to illustrate broader challenges in dataset curation and usage. The review consolidates current knowledge while highlighting persistent gaps such as the lack of standardized lesion-level annotations and longitudinal data. It also outlines recommendations for future dataset development to support clinically reliable and explainable solutions in DR screening.

Figures & tables

Figure 1: Sample fundus images depicting the eye pathologies in lesion formation for DR, like (a) microaneurysms (MA), hemorrhages (HE), hard exudates (HX), soft exudates (SX), intraretinal microvascular abnormalities (IRMA), among others, and (b) neovascularization (NV).

Figure 2: Representative fundus images, illustrating the severity grades of DR. (a) Healthy case, and the (b)–(e) DR-affected cases.

Figure 3: Key challenges in developing high-quality, robust and standardized datasets for DR screening.

Characteristic	Suitability for DL	Clinical Implication
Large and Diverse	- Enhances generalization across populations and imaging devices.	- Enables robust performance across diverse populations and settings.
Well-Annotated	- Enables effective supervised training.	- Supports accurate diagnosis and lesion-level decision-making.
Standardized Protocols	- Provides consistent ground truth and fair benchmarking.	- Improves clinical reliability.
Balanced Representation	- Reduces bias toward majority classes; - Improves learning of rare lesions.	- Improves sensitivity for rare but severe cases.
Rich Metadata	- Supports multi-modal models; - Disentangles DR-specific effects from other factors.	- Adds context for prognosis and risk assessment.
Longitudinal Data	- Enables progression modeling and predictive analytics.	- Allows monitoring of disease evolution; - Supports prediction of treatment response and disease control.

Table 1: Key characteristics of DR datasets and their impact.

Figure 4: Chronological overview of major DR fundus image datasets (from 2003 to 2025), showing year of release, dataset type (image-level vs. lesion annotation), and availability.

Dataset	Origin	# Samples	# Patients	Resolution	DR Grade	Lesion Types; Annotation (Coarse/ Fine)	#Annotators	Comment
Moorfields 005 [ 43 ]	UK	5,537,798	45,824	Unknown	5	NA	-	Longitudinal data, Available meta data, (Reported as of July 2024)
DIARETDB0 [ 44 ]	Finland	130	-	1500 $\times$ 1152	5	MA, HE, HX, SX; Coarse	4	Lesions are marked with circle, ellipse, oval shapes or rarely polygon
DIARETDB1 [ 45 ]	Finland	89	-	1500 $\times$ 1152	5	MA, HE, HX, SX; Coarse	4	Lesions are marked with circle, ellipse, oval shapes or rarely polygon
Messidor [ 23 ]	France	1,200	-	1440 $\times$ 960, 2241 $\times$ 1488, 2304 $\times$ 1536	4	NA	-	Contains grade risk of DME (0, 1, 2)
Retinopathy Online Challenge (ROC) [ 46 ]	USA	100	-	768 $\times$ 576, 1058 $\times$ 1061, 1389 $\times$ 1383	-	MA	4	Annotations include MA only, The center location of MA are marked, Lesion structurally similar to MA also marked
E-ophtha [ 42 ]	France	EX: 82 MA: 351	-	1440 $\times$ 960 to 2544 $\times$ 1696	NA	MA, HX, SX; Fine	-	E-ophtha is divided into two set: E-ophtha-MA, E-ophtha-EX, E-ophtha-MA contains MA annotation, E-ophtha-EX contains exudate (EX) annotation, Multiple images per patient is available, Overlap of samples between E-ophtha-MA and E-ophtha-EX is unknown

Table 2: Details of DR datasets developed during 2003 - 2014.

Dataset	Origin	# Samples	# Patients	Resolution	DR Grade	Lesion Types; Annotation (Coarse/ Fine)	Annotators	Comment
EyePACS [ 22 ]	USA	88,702	-	433 $\times$ 289 to 5184 $\times$ 3456	5	NA	-	Both left and right eye images of each patient provided
IDRiD [ 25 ]	India	516	-	4288 $\times$ 2848	5	MA, HE, HX, SX; Fine	-	81 images have fine lesion annotation, 516 images have classification labels with coarse annotation for fovea and optic disc, Contains DME grade (0, 1)
APTOS [ 27 ]	India	5,590	-	474 $\times$ 358 to 4288 $\times$ 2848	5	NA	-	Macula centered images, Only 3,662 samples are publicly available
DDR [ 24 ]	China	13,673	9,598	702 $\times$ 706 to 5184 $\times$ 3456	5	MA, HE, HX, SX; Coarse	7	Very high variability in image resolution
ODIR [ 48 ]	China	10,000	5,000	2240 $\times$ 1488	2	NA	-	Each image categorized to 8 ocular diseases classes: Normal, DR, Glaucoma, Age-related macular degeneration, Cataract, Hyper-tension, Pathological Myopia, Others, Meta-data about gender, age, etc., Images of left and right eye for each patient
DeepDRiD [ 49 ]	China	2,128	756	1956 $\times$ 1934, 3900 $\times$ 3072	5	NA	-	Dual-view images of each eye - optic disc- and fovea-centered, Each image has grades for quality, DR has grades from 0 to 5; 0 - Normal, 1 - Mild, 2 - Moderate, 3 - Severe, 4 - PDR, 5 - Ungradable, Patient level DR grade, Only 1,702 samples are publicly available, Other Meta data

Table 3: Details of DR datasets developed during 2015 - 2025.

Figure 5: Data distribution in SaNMoD, in terms of (a) the DR severity classes (grades), and (b) samples in the different categories.

Model	Acc. ( $\uparrow$ )	B. Acc. ( $\uparrow$ )	F1-score ( $\uparrow$ )		AUC-PR ( $\uparrow$ )
Model	Acc. ( $\uparrow$ )	B. Acc. ( $\uparrow$ )	No DR	DR	AUC-PR ( $\uparrow$ )
VGG16	0.9138	0.9117	0.9229	0.9023	0.8626
Densenet121	0.9150	0.9104	0.9255	0.9011	0.8722
InceptionNetV3	0.9138	0.9099	0.9240	0.9004	0.8682
EfficientNetB2	0.9185	0.9128	0.9292	0.9040	0.8820
ResNet50	0.9091	0.9023	0.9217	0.8917	0.8703

Table 4: Comparative performance in SaNMoD data, for binary classification of DR.

Model	Acc. ( $\uparrow$ )	B. Acc. ( $\uparrow$ )	F1-score ( $\uparrow$ )		AUC-PR ( $\uparrow$ )
Model	Acc. ( $\uparrow$ )	B. Acc. ( $\uparrow$ )	Non RDR	RDR	AUC-PR ( $\uparrow$ )
VGG16	0.7513	0.7550	0.6412	0.8097	0.8403
Densenet121	0.7487	0.7424	0.6275	0.8104	0.8319
InceptionNetV3	0.7566	0.7614	0.6489	0.8138	0.8443
EfficientNetB2	0.7646	0.7536	0.6426	0.8245	0.8380
ResNet50	0.7328	0.7446	0.6273	0.7918	0.8349

Table 5: Comparative performance in SaNMoD data, for binary classification of RDR.

Model	Acc. ( $\uparrow$ )	B. Acc. ( $\uparrow$ )	F1-score ( $\uparrow$ )		AUC-PR ( $\uparrow$ )
Model	Acc. ( $\uparrow$ )	B. Acc. ( $\uparrow$ )	No DME	DME	AUC-PR ( $\uparrow$ )
VGG16	0.8016	0.7805	0.8421	0.7331	0.6878
Densenet121	0.7751	0.7597	0.8156	0.7119	0.6464
InceptionNetV3	0.7884	0.7729	0.8268	0.7279	0.6646
EfficientNetB2	0.7804	0.7652	0.8200	0.7186	0.6535
ResNet50	0.7804	0.7615	0.8230	0.7108	0.6553

Table 6: Comparative performance in SaNMoD data, for binary classification of DME.

Models	Acc. $(\uparrow)$	B. Acc. $(\uparrow)$	F1-score $(\uparrow)$					AUC-PR $(\uparrow)$
Models	Acc. $(\uparrow)$	B. Acc. $(\uparrow)$	No-DR	Mild	Moderate	Severe	PDR	AUC-PR $(\uparrow)$
VGG16	0.7769	0.5983	0.9253	0.4493	0.6772	0.6122	0.3377	0.4459
DenseNet121	0.6128	0.4005	0.7939	0.2105	0.5442	0.2712	0.2041	0.3458
InceptionV3	0.7863	0.6506	0.9151	0.4234	0.7206	0.4842	0.6569	0.4856
EfficientNet B2	0.7898	0.5895	0.9160	0.3913	0.7550	0.6525	0.2857	0.4569
ResNet50	0.7898	0.6759	0.9226	0.5207	0.7021	0.4651	0.6497	0.4977

Table 7: Comparative performance in SaNMoD data, for multi-class grading of the severity of DR

Figure 6: Representative Grad-CAM visualization, highlighting lesion-specific activation for (a) DR, (b) RDR, and (c) DME binary classification tasks.

Medical Image Benchmarks

Medical image benchmarks compare AI methods on curated clinical images, tasks, and evaluation protocols to measure accuracy, robustness, and clinical usefulness.

See all 337 papers

Robust Diabetic Retinopathy Grading Using Dual-Resolution Attention-Based Deep Learning with Ordinal Regression
Apr 19, 2026Afshan HashmiDiabetic Retinopathy Grading Multi-Scale Feature Fusion
Tuwaiq Academy, Tuwaiq Research and Development Centre, Riyadh, Saudi Arabia
Diabetic Retinopathy Grading with CLIP-based Ranking-Aware Adaptation:A Comparative Study on Fundus Image
Mar 12, 2026Sungjun ChoDiabetic Retinopathy Grading Retinal Disease Classification
Hong Kong University of Science and Technology (HKUST), Hong Kong, China
Mobile-Ready Automated Triage of Diabetic Retinopathy Using Digital Fundus Images
Feb 25, 2026Aadi Joshi, Manav S. Sharma, Vijay Uttam Rathod +3Diabetic Retinopathy Grading Medical Image Classification
Vishwakarma Institute Of Technology, Computer Engineering Department, Pune , India · Vishwakarma Institute Of Technology, Computer Science and Engineering (AI & ML ) Department, Pune , India

Apr 19, 2026J/K move · Enter open · S save

Robust Diabetic Retinopathy Grading Using Dual-Resolution Attention-Based Deep Learning with Ordinal Regression

Afshan Hashmi

Tuwaiq Academy, Tuwaiq Research and Development Centre, Riyadh, Saudi Arabia

Diabetic Retinopathy Grading Multi-Scale Feature Fusion Cross-Dataset Generalization

Open paper AskPDF arXiv

Diabetic retinopathy (DR) is a leading cause of vision impairment worldwide, and automated grading systems play a crucial role in large-scale screening programs. However, deep learning models often exhibit degraded performance when deployed across datasets acquired under different imaging conditions. This study presents a robust dual-resolution deep learning framework for DR grading that integrates attention-based feature fusion with ordinal regression to improve cross-dataset generalization. The proposed method employs two parallel EfficientNet backbones operating at different spatial resolutions to capture complementary retinal features. A learnable attention mechanism adaptively fuses multi-resolution representations, while an ordinal regression formulation based on the cumulative link model (CORAL) explicitly accounts for the ordered nature of DR severity levels. To mitigate domain discrepancies between datasets, a preprocessing strategy combining circular cropping, contrast enhancement, and histogram matching is applied. The model was trained on the APTOS 2019 dataset and evaluated on both an internal validation split and an external Messidor-2 test set. Experimental results demonstrate strong grading performance, achieving a quadratic weighted kappa (QWK) of 0.88 on the APTOS validation set and 0.68 on the unseen Messidor-2 dataset, indicating improved robustness for cross-dataset DR grading applications.

Abstract

Medical Image Benchmarks

Explore similar work

Robust Diabetic Retinopathy Grading Using Dual-Resolution Attention-Based Deep Learning with Ordinal Regression

Diabetic Retinopathy Grading with CLIP-based Ranking-Aware Adaptation:A Comparative Study on Fundus Image

Mobile-Ready Automated Triage of Diabetic Retinopathy Using Digital Fundus Images