Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Retinal Pixels to Patients: Evolution of Deep Learning Research in Diabetic Retinopathy Screening

Muskaan Chopra, Lorenz Sparrenberg, Armin Berger, Sarthak Khanna, Jan H. Terheyden, Rafet Sifa

TL;DR

This survey analyzes deep learning for diabetic retinopathy screening from 2016 to 2025, tracing the field's shift from private-data CNN success to privacy-preserving, generalizable, and clinically trusted AI pipelines. It synthesizes foundations, reproducibility challenges, domain shifts, data-efficiency methods, architectures, federated learning, and evaluation protocols, highlighting open gaps in multi-center validation and calibration. The authors offer a practical agenda emphasizing open datasets, standardized benchmarks, clinically meaningful five-class grading, and transparent reporting to accelerate real-world deployment. By linking methodological advances such as self-supervised learning, domain generalization, interpretable models, and privacy-preserving training to translational goals, the work underscores implications for broader medical imaging beyond DR.

Abstract

Diabetic Retinopathy (DR) remains a leading cause of preventable blindness, with early detection critical for reducing vision loss worldwide. Over the past decade, deep learning has transformed DR screening, progressing from early convolutional neural networks trained on private datasets to advanced pipelines addressing class imbalance, label scarcity, domain shift, and interpretability. This survey provides the first systematic synthesis of DR research spanning 2016-2025, consolidating results from 50+ studies and over 20 datasets. We critically examine methodological advances, including self- and semi-supervised learning, domain generalization, federated training, and hybrid neuro-symbolic models, alongside evaluation protocols, reporting standards, and reproducibility challenges. Benchmark tables contextualize performance across datasets, while discussion highlights open gaps in multi-center validation and clinical trust. By linking technical progress with translational barriers, this work outlines a practical agenda for reproducible, privacy-preserving, and clinically deployable DR AI. Beyond DR, many of the surveyed innovations extend broadly to medical imaging at scale.

From Retinal Pixels to Patients: Evolution of Deep Learning Research in Diabetic Retinopathy Screening

TL;DR

This survey analyzes deep learning for diabetic retinopathy screening from 2016 to 2025, tracing the field's shift from private-data CNN success to privacy-preserving, generalizable, and clinically trusted AI pipelines. It synthesizes foundations, reproducibility challenges, domain shifts, data-efficiency methods, architectures, federated learning, and evaluation protocols, highlighting open gaps in multi-center validation and calibration. The authors offer a practical agenda emphasizing open datasets, standardized benchmarks, clinically meaningful five-class grading, and transparent reporting to accelerate real-world deployment. By linking methodological advances such as self-supervised learning, domain generalization, interpretable models, and privacy-preserving training to translational goals, the work underscores implications for broader medical imaging beyond DR.

Abstract

Diabetic Retinopathy (DR) remains a leading cause of preventable blindness, with early detection critical for reducing vision loss worldwide. Over the past decade, deep learning has transformed DR screening, progressing from early convolutional neural networks trained on private datasets to advanced pipelines addressing class imbalance, label scarcity, domain shift, and interpretability. This survey provides the first systematic synthesis of DR research spanning 2016-2025, consolidating results from 50+ studies and over 20 datasets. We critically examine methodological advances, including self- and semi-supervised learning, domain generalization, federated training, and hybrid neuro-symbolic models, alongside evaluation protocols, reporting standards, and reproducibility challenges. Benchmark tables contextualize performance across datasets, while discussion highlights open gaps in multi-center validation and clinical trust. By linking technical progress with translational barriers, this work outlines a practical agenda for reproducible, privacy-preserving, and clinically deployable DR AI. Beyond DR, many of the surveyed innovations extend broadly to medical imaging at scale.

Paper Structure

This paper contains 44 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Clinical context and grading
  3. Promise of deep learning
  4. Challenges in early research
  5. Objectives and contributions of this survey
  6. Organization of the paper
  7. Foundational Breakthroughs and the Reproducibility Gap
  8. Generalization and Domain Shift
  9. Methodological Advances (2019 - 2024)
  10. Handling Class Imbalance
  11. Learning with Limited Labels
  12. Domain Generalization and Adaptation
  13. Building Clinical Trust
  14. Architectures and Interpretability
  15. From CNN Backbones to Ensembles
  16. ...and 29 more sections

Figures (2)

  • Figure 1: Taxonomy and structure of this survey. The survey begins with foundational breakthroughs and the reproducibility gap (Section \ref{['sec:foundations']}), addresses challenges of generalization and domain shift (Section \ref{['sec:domain_shift']}), and methodological advances from 2019 - 2024 (Section \ref{['sec:methods']}), then examines architectures and interpretability (Section \ref{['sec:arch_xai']}), federated learning and privacy (Section \ref{['sec:federated']}), datasets (Section \ref{['sec:datasets']}), performance benchmarking (Section \ref{['sec:bench']}), evaluation protocols (Section \ref{['sec:evaluation']}), and finally discussion and future directions (Section \ref{['sec:discussion']}).
  • Figure 2: Example fundus photographs Kaggle2015EyePACS illustrating the spectrum of diabetic retinopathy.