Source-linked AI summary
Application of Deep Learning in Fundus Image Processing for Ophthalmic Diagnosis -- A Review
Sourya Sengupta, Amitojdeep Singh, Henry A. Leopold, Tanmay Gulati, Vasudevan Lakshminarayanan
TL;DR
Ophthalmic diagnosis from retinal fundus images faces heavy clinical workloads, limited publicly available data, and reliance on manual or predefined image-processing procedures. This review surveys deep-learning applications, datasets, architectures, and performance measures across retinal segmentation, lesion detection, and disease classification. It concludes that deep learning generally outperformed traditional methodologies, while standardized metrics, sufficient annotated data, and robustness to domain shift remain important boundaries.
Problem
Ophthalmic diagnosis involves diseases that may cause vision loss, while clinical data are underused and traditional systems depend on manual annotation, segmentation, and predefined image features.
Method
The review examines published deep-learning studies and datasets for fundus-image segmentation, lesion detection, disease classification, and associated performance measures.
Results
Deep-learning methods outperformed traditional methodologies in most of the reviewed cases across ophthalmic computer-aided diagnosis applications.
Takeaways & Limitations
The review provides a cross-disease account of deep-learning architectures and performance outcomes for ophthalmic diagnosis using retinal fundus images.
Takeaways & Limitations
Comparisons are constrained by scarce large annotated datasets, nonstandardized performance indicators, and domain shifts between training and real-world test images.
Abstract
from arXiv · showhide
An overview of the applications of deep learning in ophthalmic diagnosis using retinal fundus images is presented. We also review various retinal image datasets that can be used for deep learning purposes. Applications of deep learning for segmentation of optic disk, blood vessels and retinal layer as well as detection of lesions are reviewed. Recent deep learning models for classification of diseases such as age-related macular degeneration, glaucoma,diabetic macular edema and diabetic retinopathy are also reported.
1. Introduction
Retinal diseases can cause severe vision loss, while diagnosis is time-intensive and supported by limited publicly available clinical data. This review examines deep learning for fundus-image segmentation and classification across ophthalmic diseases and datasets.
- More than 40 million people in the United States suffer from acute eye-related diseases that may cause complete vision loss if untreated.
- Glaucoma, diabetic retinopathy, and age-related macular degeneration are among the common retinal diseases discussed.The passage also describes glaucoma diagnosis through optic cup-to-disk ratio and related retinal measures, and diabetic retinopathy lesions such as microaneurysms.
- Figure 1 illustrates retinal morphologies and pathological manifestations in a fundus photograph.
- Overburdened health-care systems make diagnosis and treatment error-prone and time-intensive, despite routine generation of clinical data.Much of these data are rarely used for computer-aided diagnostics and are not publicly available.
- Traditional retinal computer-aided diagnosis systems rely on predefined templates and kernels applied to manually annotated and segmented image regions.
- Deep learning learns relevant representations from data without manual feature extraction and is reviewed here for fundus-image segmentation and disease classification.The review covers retinal structures, lesions, datasets, and ophthalmic diseases including glaucoma and diabetic retinopathy.
2. Application in Retinal Image Processing Techniques
The review surveys fundus-image datasets and deep-learning applications spanning image segmentation, lesion detection, and disease classification. It covers optic-disc and nerve-head processing, diabetic-retinopathy lesions, retinal vessels, and AMD, glaucoma, and related disease studies.
- Datasets: The review catalogs fundus-image datasets supporting optic-disc segmentation, lesion detection, glaucoma detection, and retinal image analysis.Examples include MESSIDOR, ONHSD, ORIGA, RIGA, STARE, and REFUGE.
- Segmentation: Segmentation supports automatic cropping of regions of interest and helps address image distortions, noise, and nonuniform illumination during retinal-image processing.The review discusses deep-learning approaches for optic-disc, cup, nerve-head, and other retinal structures.
- Review scope: Published research on deep learning for fundus-based ophthalmic diagnosis increased significantly from 2014 through the literature reviewed up to December 2018.The authors collected papers using Google Scholar queries covering deep learning, ophthalmology, segmentation, classification, fundus photos, datasets, and retina.
- Lesion segmentation and detection: Deep-learning studies detect retinal lesions relevant to diabetic-retinopathy screening, including microaneurysms and larger hemorrhages.Reported approaches include pixel-based neural networks, stacked sparse autoencoders, CNN preprocessing layers, and CNN–codebook combinations.
- Retinal blood-vessel segmentation: Retinal blood-vessel segmentation studies include hybrid random-forest and deep-neural-network systems, CNN ensembles, and fully convolutional architectures.An early hybrid method did not outperform conventional approaches, while later work used 12 CNNs and patch-based training.
- Disease classification: The review reports deep-learning classification studies for AMD, glaucoma, and other ophthalmic diseases using architectures such as CNNs, VGG16, residual networks, and contextualized CNNs.AMD studies combined deep visual features with patient information, while glaucoma studies used CNN-based classification and optic-structure segmentation.
3. Conclusion and Future Research
The review surveys deep learning applications in ophthalmic diagnosis and identifies data scarcity, inconsistent evaluation metrics, and domain shift as major barriers to robust real-world use.
- Review scope: The review surveys state-of-the-art deep learning approaches and traditional methods for computer-aided ophthalmic diagnosis using retinal fundus images.It presents applications across retinal image analysis and compares reported deep learning and traditional methodologies.
- Reported findings: Deep learning methods outperformed traditional methodologies in most reported cases, while reducing reliance on manual feature extraction through data-driven learning.The review notes convolutional neural networks as widely used for classification, detection, and segmentation of fundus structures.
- Future research: Large datasets and manual annotations remain scarce, limiting the data required for deep learning in ophthalmic imaging.The review identifies generative models, including GANs and variational auto-encoders, as possible ways to synthesize clinically relevant annotated fundus images.
- Future research: Unstandardized performance indicators prevent straightforward comparison of deep learning architectures for the same disease state.Reported studies may emphasize different measures, such as accuracy or AUC, while generalized metrics including G-mean and MCC are also suggested.
- Future research: Camera-setting differences create domain shift because training and test images may come from different distributions in real-world diagnosis.The review reports decreased accuracy under domain shift and highlights deep domain adaptation as a direction for robust models.