PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning based Computer-Aided Diagnosis Systems for Diabetic Retinopathy: A Survey

Norah Asiri, Muhammad Hussain, Fadwa Al Adel, Nazih Alzaidi

arXiv:1811.01238v2cs.CV

TL;DR

Diabetic retinopathy requires early detection, but diagnosis is difficult at scale and conventional hand-engineered approaches generalize poorly. This survey reviews deep-learning CAD methods across DR detection tasks, compares them with traditional methods, and identifies challenges and future directions. Across the reviewed problems, CNN-based techniques are most common and outperform other deep architectures, while dataset variability, annotation costs and imaging differences remain important barriers.

  • Problem

    Early DR diagnosis matters because advanced disease becomes incurable, yet large patient numbers and limited experienced technicians make screening difficult.

  • Method

    The paper surveys deep-learning CAD techniques for retinal vessel, optic-disc and lesion analysis, compares selected deep-learning and hand-engineered methods, and reviews research gaps.

  • Results

    Across reviewed DR diagnosis problems, CNN-based techniques are most common and outperform other deep architectures.

  • Takeaways & Limitations

    Deep learning provides a broad framework for detecting and diagnosing DR complications across vessels, optic discs, lesions and referable DR.

  • Takeaways & Limitations

    Clinical applicability is constrained by inconsistent benchmark grading and screening protocols, which can limit whether methods tested on those datasets work in clinical settings.

Abstract

from arXiv · show

Diabetic retinopathy (DR) results in vision loss if not treated early. A computer-aided diagnosis (CAD) system based on retinal fundus images is an efficient and effective method for early DR diagnosis and assisting experts. A computer-aided diagnosis (CAD) system involves various stages like detection, segmentation and classification of lesions in fundus images. Many traditional machine-learning (ML) techniques based on hand-engineered features have been introduced. The recent emergence of deep learning (DL) and its decisive victory over traditional ML methods for various applications motivated the researchers to employ it for DR diagnosis, and many deep-learning-based methods have been introduced. In this paper, we review these methods, highlighting their pros and cons. In addition, we point out the challenges to be addressed in designing and learning about efficient, effective and robust deep-learning algorithms for various problems in DR diagnosis and draw attention to directions for future research.

1 Introduction

Diabetic retinopathy is a major cause of blindness whose advanced stages become incurable, making early diagnosis important but difficult at health-system scale. This motivates automated CAD systems and the shift from hand-engineered machine learning toward deep-learning methods.

  • DR is a major cause of blindness among working-age people, and advanced disease becomes incurable, making early diagnosis important.
  • Large patient numbers and few experienced technicians make early DR diagnosis difficult in healthcare systems.
  • Traditional DR diagnosis commonly uses hand-engineered features with machine-learning methods and has shown promising efficiency in recognizing retinal DR regions.
  • Hand-engineered approaches require expert knowledge, intensive feature investigation and tedious parameter tuning, while generalizing poorly.
  • Deep learning has been encouraged by large datasets and GPU computing, and has outperformed traditional hand-engineered methods across computer-vision tasks.
  • The paper reviews deep-learning algorithms for DR detection, summarizes their contributions and challenges, and identifies future research directions.

2 Automatic Diabetic Retinopathy Detection

Automatic DR detection supports rapid retinal evaluation by analyzing lesions, retinal structures and disease stages from fundus images. Detection systems include lesion-level localization and classification alongside image-level screening, within broader preprocessing, feature-learning and assessment workflows.

  • Automatic image-based DR systems aim to perform rapid retinal evaluations and detect whether DR complications are present early.
  • Types of lesions: Microaneurysms are early DR signs; hemorrhages result from broken capillary walls, while exudates arise from greater leakage of damaged capillaries.
  • Optic disc: The optic disc serves as a landmark for diagnosing serious eye pathologies, checking disc neovascularization and locating structures such as the fovea.
  • DR stages: DR progresses from non-proliferative to proliferative disease, with NPDR involving small-vessel damage, fluid or blood discharge, retinal swelling and blurry vision.
  • DR grading: Grading requires trained examiners, suitable imaging conditions and adequate computer screens, with protocols used to classify pathology and make referral decisions.
  • Detection tasks: Lesion-level detection localizes and classifies lesions, whereas image-level detection assesses whether fundus images show signs of DR for screening.
  • General framework: The general detection framework includes preprocessing, feature extraction or selection, classifier choice and result assessment, using supervised or unsupervised learning.

3 Datasets and Performance Metrics

The survey introduces benchmark datasets and performance metrics commonly used to evaluate DR research methods.

  • The section overviews benchmark datasets and performance metrics commonly used for diabetic-retinopathy research.

3.1 Retinal Fundus Image Datasets

DR research uses diverse public retinal-image datasets spanning lesion detection, grading, vessel analysis and optic-disc or glaucoma-related tasks. These datasets differ in image conditions, annotations, populations and clinical scope.

  • Public benchmark datasets cover DR, retinal vessels, lesions, optic discs and glaucoma-related analysis across multiple detection tasks.
  • MESSIDOR: MESSIDOR provides expert image-level DR grades and macular-edema risk levels from 45° non-mydriatic fundus images.
  • MESSIDOR: MESSIDOR grades DR using microaneurysm, hemorrhage and new-vessel criteria, while macular-edema risk depends on hard-exudate distance from the macula.
  • e-ophtha: The e-ophtha datasets provide ophthalmologist-verified locations of microaneurysms and exudates, while healthy images may contain reflections and optical artifacts that mislead detection.
  • EyePACS: EyePACS contains high-resolution images from different conditions, expert image-level annotations and DR grades ranging from 0 to 4.
  • Vessel datasets: DRIVE contains 40 images split into 20-image training and test sets, with 33 images showing no DR and 7 showing mild early DR.
  • Vessel datasets: STARE includes images of 13 eye diseases with pixel-level annotations for blood vessels and the optic nerve but no grading.
  • Dataset diversity: Other datasets vary in population, imaging conditions and annotation goals, including pediatric images, lesion maps, vessel widths, optic-disc boundaries and glaucoma measures.

3.2 Performance Metrics

The survey describes classification and segmentation metrics used to assess DR algorithms, covering instance-level correctness, ranking, probabilistic prediction, overlap, and boundary localization.

  • Classification metrics: Accuracy measures correctly classified instances among all instances, with TP, TN, FP, and FN defining classification outcomes.The evaluated instance may be a fundus image, patch, or pixel, depending on the task.
  • Classification metrics: F-score combines precision and recall into a single performance metric.
  • Classification metrics: ROC curves plot true-positive rate against false-positive rate, while AUC summarizes this relationship on a 0-to-1 scale.Higher AUC values indicate better performance.
  • Classification metrics: Log loss evaluates probabilistic classification by penalizing false classifications using class-assignment probabilities.
  • Segmentation metrics: Segmentation assessment includes overlapping error, intersection over union, DSC, region precision recall, and boundary-based evaluation.Boundary-based evaluation measures pointwise distance between predicted and ground-truth closed curves, ideally approaching zero.

4 Overview of Deep Learning

The survey introduces CNNs, autoencoders, RNNs, and DBNs as deep-learning architectures for DR analysis, emphasizing their differing feature-learning and training strategies.

  • Convolutional Neural Networks: CNNs use convolutional, pooling, and fully connected layers to encode local and global information in end-to-end feature learning.Models discussed include AlexNet, VGGNet, GoogLeNet, Inception, and ResNet.
  • Autoencoders: Autoencoders reconstruct their input and support stacked architectures trained through unsupervised layer-wise pre-training followed by supervised fine-tuning.
  • Autoencoders: Sparse and denoising autoencoders promote sparse representations or recover corrupted inputs, respectively, and denoising variants have been used in DR detection.
  • Recurrent Neural Networks: RNNs combine current inputs with outputs from previous iterations, allowing shared parameters to model context across input positions or time.
  • Deep Belief Networks: DBNs cascade restricted Boltzmann machines and use probabilistic similarity with greedy unsupervised layer-wise pre-training.The probabilistic formulation is described as preventing degenerate solutions.

5 Literature Survey

The literature survey organizes deep-learning research for DR diagnosis around four clinically important tasks: vessel segmentation, optic-disk analysis, lesion analysis, and image-level referral diagnosis.

  • The survey categorizes DR detection tasks into retinal blood-vessel segmentation, optic-disk localization and segmentation, lesion detection and classification, and image-level diagnosis for referral.

5.1 Retinal Blood Vessel Segmentation

Deep-learning vessel-segmentation methods span CNNs, autoencoders, ensembles, domain adaptation, transfer learning, and CNN–CRF systems, with preprocessing, augmentation, and adaptation shaping reported performance.

  • Problem: Retinal vessel segmentation is difficult because vessels have low contrast, variable morphology, noisy backgrounds, and confounding pathologies such as microaneurysms and hemorrhages.
  • CNN-based methods: CNN methods include pixel-wise classification, image-to-image regression, structured prediction, boundary detection, auxiliary classifiers, and CNN–CRF formulations.
  • CNN-based methods: Liskowski and Krawiec’s CNN performed best among CNN methods for accuracy, AUC, and sensitivity, using preprocessing and an augmented training dataset.
  • Autoencoder-based methods: Autoencoder approaches include cross-modality transformation, stacked denoising autoencoders, and two-level ensembles for vessel segmentation.
  • Autoencoder-based methods: Domain adaptation improved SAE-DNN vessel segmentation by adapting weights with unlabeled target samples and fine-tuning with a small labeled target set.
  • Comparative analysis: Cross-modality transformation and two-level SAE ensembles led SAE methods in accuracy, with no significant difference between those two approaches.
  • Comparative analysis: CNN and SAE methods showed no noticeable accuracy difference; CNNs’ larger parameter counts increase overfitting risk, while larger labeled datasets or augmentation may help.

5.2 Optic Disc Feature

Optic-disc analysis in DR CAD uses CNNs, SAEs, and hybrid pipelines for localization and segmentation, with methods differing in preprocessing, architecture, and evaluation databases. Comparisons report strong localization results but caution that segmentation rankings are difficult across heterogeneous datasets.

  • OD detection: OD detection supports DR analysis because the optic disc can be confused with bright lesions such as exudates.Detection includes localizing the OD and, in some methods, segmenting its boundaries.
  • CNN-based methods: CNN-based OD and OC segmentation pipelines combine localization, enhancement, pixel classification, probability maps, and boundary prediction.Examples include nine-layer CNNs, patch-based CNNs, modified U-Nets, and multi-scale CNNs with boosting.
  • CNN-based methods: Ensemble CNNs with entropy sampling extend boosting-based learning by selecting informative points to learn convolutional filters.The extended method uses ensemble CNNs and entropy sampling for OD and OC segmentation.
  • CNN-based methods: Residual and fully convolutional approaches jointly segment the optic disc and cup, including ResU-Net with GAN-based mapping and FCNs validated on multiple datasets.These methods learn mappings from retinal images to OD and OC segmentation maps.
  • Comparative findings: Faster RCNN localization followed by vessel removal and a shape-constrained level set achieved the best localization results for MESSIDOR in the reviewed comparison.M-Net instead segments OD and OC jointly and estimates the cup-to-disc ratio, reporting state-of-the-art results on ORIGA and SCES.
  • Comparative findings: OD segmentation performance is difficult to rank because methods were evaluated on different databases with different metrics.Only one reviewed SAE method segmented the OD, using pixel features to distinguish OD from parapapillary atrophy.

5.3 Lesion Detection and Classification

The survey reviews deep-learning methods for lesion detection and classification across DME, exudates, microaneurysms, and hemorrhages. CNN-based approaches generally show stronger reported performance, while joint image- and lesion-level systems and efficient pretrained models address screening needs.

  • Scope: Deep-learning methods address DME, exudate, microaneurysm, and hemorrhage detection and classification.DME methods are mainly categorized as CNN-based or autoencoder-based approaches.
  • DME detection: DME pipelines combine end-to-end CNN classification, exudate localization and segmentation, preprocessing, augmentation, and multi-stage residual classification.Methods use datasets including MESSIDOR, HEI-MED, and e-ophtha.
  • DME detection: CNN-based methods outperform DBN-based methods for DME detection, while DBNs use significantly fewer learnable parameters.The survey notes that DBNs had not been used end-to-end in the compared methods.
  • Exudate detection: Exudate detection is challenging because lesion size, shape, and contrast vary substantially; reviewed methods use CNNs for localization, patch classification, or visualization-based detection.Some systems combine exudate maps with optic-disc, vessel, and bright-border information.
  • Exudate detection: Pre-trained ResNet-50 with SVM outperformed other compared deep-learning patch methods for exudate detection on DIARETDB1 and e-ophtha.The survey also describes the method as computationally efficient with comparable results.
  • Microaneurysm and hemorrhage detection: For microaneurysm detection, Haloi’s CNN appeared to outperform other methods in sensitivity, specificity, AUC, and accuracy, while CNN methods jointly detected referable DR and red lesions efficiently.The survey attributes the apparent advantage to pixel augmentation rather than image-based augmentation.

5.4 Classification of Fundus Images for Referral

Image-level referral systems grade fundus images for referable DR, primarily using CNNs trained on large datasets and sometimes adding lesion localization or pathological-region visualization. Reported performance is strong, but cross-study comparison is limited by differing datasets and grading protocols.

  • Task definition: Image-level referral detection grades DR severity using only image-level annotations, with some methods jointly detecting lesions without lesion-level labels.The reviewed methods use CNN models for this problem.
  • CNN-based referral methods: Inception-v3 achieved high sensitivity and specificity for referable DR using EyePACS-1 and MESSIDOR-2 images graded by multiple ophthalmologists.EyePACS-1 contains 9,963 images from 4,997 patients.
  • CNN-based referral methods: Referral systems include joint lesion localization, BoVW-style feature encoding, five-stage CNN grading, and ResNet grading with convolutional visualization.Training and evaluation span datasets such as Kaggle, MESSIDOR, EyePACS, and e-ophtha.
  • Comparative findings: Gulshan et al.’s method outperformed other reviewed methods in sensitivity, specificity, and AUC, with performance comparable to a panel of seven certified ophthalmologists.The survey also reports strong cross-dataset performance for Gargeya and Leng’s method.
  • Comparative findings: Performance comparisons are difficult because the reviewed methods used different datasets for training and testing.Benchmark grading variability and differing screening protocols further limit transfer to clinical settings.
  • Clinical interpretation: Visualization of pathological regions can aid real-time clinical validation of automated referral diagnoses.This addresses the black-box nature of CNN models noted in the survey.

6 Discussion

The survey finds that CNN-based deep-learning methods generally outperform traditional hand-engineered approaches across retinal vessel, optic-disc, and lesion tasks, but performance varies across datasets and deep models remain difficult to interpret and vulnerable to overfitting.

  • Challenges: Cross-method comparisons are difficult because studies use different datasets and metrics; the survey therefore argues that methods should be evaluated on the same datasets to estimate real performance gains.CNN models also require many parameters and annotated images, creating overfitting risk when available datasets are small.
  • Retinal vessel segmentation: Deep learning produced mixed vessel-segmentation comparisons: traditional methods achieved sensitivity 96.48% and accuracy 97.59% versus 78.11% and 95.35% for a deep-learning method on DRIVE, while deep learning performed better on STARE and CHASE.The comparison indicates that results depend on the evaluation dataset.
  • Optic-disc localization: For optic-disc localization, deep learning reached accuracy 99.9% versus 99% on MESSIDOR, matched 100% on DRIVE, but trailed a traditional method achieving 100% versus 98.88% on DIARETDB1.These results show that neither approach dominates every dataset-specific comparison.
  • Optic-disc segmentation: Deep-learning methods achieved higher optic-disc segmentation accuracy than traditional methods on MESSIDOR, 96.4% versus 86%, and DRIVE, 92.68% versus 75.56%.Figure 4 summarizes optic-disc localization and segmentation performance across traditional and deep-learning methods.
  • Lesion detection: For lesion detection, a traditional exudate method reached accuracy 98.87% versus 98.2% for deep learning, while deep learning achieved higher sensitivity at 99%; deep learning also exceeded traditional MA results across four reported metrics.On MESSIDOR, the MA comparison was sensitivity 97% versus 94%, specificity 95% versus 90%, AUC 0.982 versus 0.942, and accuracy 95.4% versus 90%.
  • Overall comparison: CNN-based methods generally outperform traditional methods for retinal vessel segmentation, optic-disc detection and segmentation, and diabetic-retinopathy lesion detection.The survey nevertheless identifies robustness, interpretability, and overfitting as unresolved challenges.

7 Gaps and Future Directions

Deep learning has advanced DR diagnosis, but robust clinical deployment remains constrained by interpretability, dataset quality and diversity, image variability, authenticity, and screening conditions.

  • Deep-learning DR models often function as black boxes and usually lack interpretations that could validate their diagnostic usefulness clinically.
  • Robust models require large, expertly annotated datasets, yet pixel- and image-level labeling is expensive and class imbalance biases training toward normal cases.Proposed directions include learning from limited data, GANs, more realistic augmentation, and models with fewer parameters.
  • Fundus-image illumination, capture angle, and other artifacts create non-uniform inputs that require preprocessing, artifact correction, or more sophisticated augmentation.Artifact-correction approaches themselves need broadly annotated data covering different image artifacts.
  • Systems should first verify retinal-image authenticity because user-friendly editing software makes image tampering easy.
  • Screening performance is bounded by population and acquisition differences, including retinal pigmentation, mydriasis, field of view, and ungradable images.The review reports that mydriasis with multiple shots is more sensitive and specific than non-mydriasis with a single shot; ungradable images generally range from 7–17%.
  • Benchmark datasets vary in grading, protocols, and composition, so methods trained and tested on one dataset may not generalize to clinical settings or across datasets.The review advocates standardized classification and agreement among at least three graders, alongside cross-dataset validation.

8 Conclusion

The conclusion presents deep learning as a basis for more robust and accurate DR diagnosis across retinal-image analysis tasks, while summarizing the survey’s coverage of tasks, datasets, metrics, and architectures.

  • Deep learning has been applied to retinal vessel segmentation, optic-disc detection and segmentation, DR-lesion detection and classification, and referable-DR detection.
  • The survey covers DR biomarkers, diagnostic tasks, task frameworks, datasets, evaluation metrics, and deep-learning architectures used in DR diagnosis.
Loading 1811.01238v2…