Source-linked AI summary
Two-stage framework for optic disc localization and glaucoma classification in retinal fundus images using deep learning
Muhammad Naseer Bajwa, Muhammad Imran Malik, Shoaib Ahmed Siddiqui, Andreas Dengel, Faisal Shafait, Wolfgang Neumeier, Sheraz Ahmed
TL;DR
Existing optic-disc localization methods were dataset-dependent and did not generalize reliably across retinal fundus datasets. The paper uses a two-stage faster RCNN and CNN framework, achieving state-of-the-art localization across six datasets while classification on ORIGA remained difficult despite higher AUC results.
Problem
Existing disc-localization methods were heuristic, dataset-dependent, and unable to generalize reliably across different retinal fundus datasets.
Method
The paper combines fully automated faster RCNN optic-disc localization with CNN classification of the extracted disc as healthy or glaucomatous.
Results
The localization system achieved new state-of-the-art results in six of seven public datasets, while ORIGA experiments achieved higher AUC than several prior results.
Takeaways & Limitations
Fully automated deep-learning localization can eliminate dataset-specific heuristic methods across diverse datasets, supporting the proposed two-stage glaucoma-screening framework.
Takeaways & Limitations
Detailed performance measures indicate that the classifier has difficulty learning discriminative features for glaucomatous images in the ORIGA dataset.
Abstract
from arXiv · showhide
With the advancement of powerful image processing and machine learning techniques, CAD has become ever more prevalent in all fields of medicine including ophthalmology. Since optic disc is the most important part of retinal fundus image for glaucoma detection, this paper proposes a two-stage framework that first detects and localizes optic disc and then classifies it into healthy or glaucomatous. The first stage is based on RCNN and is responsible for localizing and extracting optic disc from a retinal fundus image while the second stage uses Deep CNN to classify the extracted disc into healthy or glaucomatous. In addition to the proposed solution, we also developed a rule-based semi-automatic ground truth generation method that provides necessary annotations for training RCNN based model for automated disc localization. The proposed method is evaluated on seven publicly available datasets for disc localization and on ORIGA dataset, which is the largest publicly available dataset for glaucoma classification. The results of automatic localization mark new state-of-the-art on six datasets with accuracy reaching 100% on four of them. For glaucoma classification we achieved AUC equal to 0.874 which is 2.7% relative improvement over the state-of-the-art results previously obtained for classification on ORIGA. Once trained on carefully annotated data, Deep Learning based methods for optic disc detection and localization are not only robust, accurate and fully automated but also eliminates the need for dataset-dependent heuristic algorithms. Our empirical evaluation of glaucoma classification on ORIGA reveals that reporting only AUC, for datasets with class imbalance and without pre-defined train and test splits, does not portray true picture of the classifier's performance and calls for additional performance metrics to substantiate the results.
Background
Glaucoma damages the optic nerve and may remain unnoticed until irreversible vision loss, making the optic disc a central region for automated detection. The paper addresses dataset-dependent localization and classification challenges with a two-stage deep-learning framework and supplementary ground-truth generation.
- Motivation: Glaucoma can progress to irreversible vision loss before symptoms appear, while optic-nerve damage becomes noticeable after about 40% of axons are lost.Early diagnosis can slow impairment caused by glaucoma.
- Motivation: Optic-disc structure, including the cup-to-disc ratio and related features, provides key image cues for glaucoma detection.Relevant indications are concentrated in and around the optic disc.
- Motivation: Deep learning is used because disease patterns in retinal fundus images are complex and discriminative features may be difficult to discern visually.The framework therefore focuses classification on the optic-disc region rather than the whole fundus image.
- Proposed framework: The proposed faster RCNN localizes the optic disc automatically across datasets, addressing the limited generalization of existing heuristic methods.The method evaluates localization with IOU greater than 50%.
- Proposed framework: A DCNN classifies optic discs extracted during localization into healthy and glaucoma-affected images.The classification module operates on the localized optic-disc regions.
- Data preparation: A rule-based semi-automated method generates bounding-box annotations because the publicly available datasets lacked localization ground truth.The generated annotations support training the localization model.
- Evaluation considerations: AUC-only reporting can misrepresent classifier performance on imbalanced datasets without standardized train-test splits, motivating additional metrics.The paper reports precision, recall, and F-scores alongside AUC.
Methods
The methodology uses publicly available retinal datasets and addresses their lack of optic-disc localization annotations through a semi-automatic ground-truth generation workflow. ORIGA supplies clinical labels and disc-related annotations, while HRF contributes a small set of categorized fundus images with vessel-segmentation references.
- Datasets: The study uses multiple publicly available retinal fundus datasets, none of which provides bounding-box ground truth for optic-disc localization.This absence motivates the development of a ground-truth generation mechanism.
- Datasets: ORIGA contains 650 retinal fundus images for benchmarking computer-aided segmentation and classification.It provides manual optic-disc and optic-cup segmentation, cup-to-disc ratios, and healthy or glaucomatous labels.
- Ground-truth generation: The semi-automatic ground-truth workflow proposes optic-disc locations heuristically and then has an expert manually verify and correct them.This process generated localization ground truth for all seven datasets used in the study.
- Datasets: The HRF database contains 15 healthy, 15 glaucomatous, and 15 diabetic-retinopathy images.Each image has binary gold-standard vessel segmentation from experts and clinicians.
OCT & CFI
The OCT & CFI dataset contains paired OCT and colour fundus images from 50 healthy people, but lacks optic-disc and vessel annotations.
- The dataset contains OCT and colour fundus images of both eyes from 50 healthy persons.
- The images were collected at Feiz Hospital’s Ophthalmology Department in Isfahan, Iran.
- The dataset provides no ground truth for optic-disc segmentation, blood-vessel segmentation, or optic-disc localization.
DRIVE
DRIVE is a 40-image retinal vessel-segmentation dataset with predefined train and test splits, while the paper identifies optic-disc localization as a broader automation challenge.
- DRIVE contains 40 retinal fundus images collected in the Netherlands through a diabetic-retinopathy screening programme.
- The dataset is split into 20 training images and 20 test images.
- Training images have one manual vessel-segmentation mask, whereas test images have two manual masks.
- DRIVE provides no annotation for optic-disc localization.
- Whole-image CNN input does not adequately focus classification on the optic disc, motivating region-of-interest extraction and robust localization.
Semi-automated GT generation for OD localization
The paper generates optic-disc localization ground truth by combining heuristic proposals with expert verification, using multi-dataset validation to assess robustness.
- A heuristic algorithm proposes optic-disc locations that are manually verified and corrected by an expert.
- This process generated localization ground truth for all seven datasets used in the study.
- 780 images were divided into 525 training, 48 validation, and 207 testing images for heuristic evaluation.
- Empirical parameters were manually selected using validation accuracy, including retinal-rim crop margin and maximum valid disc radius.
- The mixture of three datasets introduced inter-dataset variation to validate heuristic accuracy and robustness.
Heuristic algorithm for OD localization
The heuristic localizes the optic disc through thresholding, morphology, and geometric fitting, but must address misleading bright regions such as rim fringes and reflections.
- The algorithm begins by identifying the optic disc as typically the brightest region in a retinal fundus image.
- Misleading bright spots can arise from disease or image-capture conditions, including retinal-rim fringes and macular reflections.
- Otsu thresholding estimates the retinal rim centre and radius, enabling removal of the bright fringe.
- Adaptive binarization uses the mean of the top 1% brightest pixels, followed by erosion and dilation to suppress noise and connect regions.
- The method estimates disc centre and radius, draws an enlarged circle for localization, and then obtains expert verification and corrections.
Results of heuristic localization
The heuristic localization approach achieved strong overlap with manually annotated optic discs but remained dataset-specific, motivating a fully automated alternative.
- Results: Only 3 of 573 train-and-validation images and 1 of 207 test images caused localization failures.These failures occurred across three datasets.
- Results: More than 96% of optic discs were localized with over 50% of the actual disc present in the prediction.The average overlap between predicted disc area and ground truth was around 70% for test images.
- Limitation: The heuristic approach was dataset-specific and might not work well on diverse real-world fundus images.The authors therefore corrected its annotations and used them as ground truth for automated localization.
Two-stage method for automated disc localization and classification: the proposed approach
The proposed system uses faster RCNN to localize and extract the optic disc, followed by a CNN that classifies the extracted region as healthy or glaucomatous. The design combines object-detection modules with a focused convolutional classifier and regularization components.
- Proposed approach: The two-stage solution first performs optic-disc localization and then classifies the extracted disc as healthy or glaucomatous.The optic disc is used as the region of interest because most glaucoma-related information is concentrated there.
- Automated localization: Faster RCNN combines a Region Proposal Network, CNN classifier, and bounding-box regression to generate precise disc locations.The RPN proposes candidate rectangles, the CNN assesses object presence, and regression fits the box to the disc.
- Glaucoma classification: The CNN classifier uses four convolutional layers followed by three fully connected layers, with softmax producing class probabilities.The described architecture includes pooling, local response normalization, and dropout layers.
- Convolutional layers: Convolutional layers learn increasingly complex image features by applying multiple kernels across the input.The first convolutional layer uses 96 kernels of size 11 × 11 × 3.
- Regularization and downsampling: Pooling downsamples feature maps, while max pooling in this network uses a 3 × 3 window with stride 2.The architecture also uses local response normalization and dropout probability 0.5 to support generalization.
Results and discussion
The automated localization method generalized across unseen datasets and outperformed dataset-tailored heuristics, while classification results were evaluated with multiple metrics because ORIGA is imbalanced and lacks a standard split. The classifier achieved higher AUC than comparison studies, but its confusion matrix shows substantial difficulty identifying glaucomatous images.
- Automated localization: 84.65% average overlap was achieved on DIARETDB1, 84.13% on DRIVE, and 80.46% on DRIONS-DB.These values are reported for predicted versus actual optic-disc bounding boxes.
- Evaluation metrics: Because ORIGA is class-imbalanced, the evaluation reports precision, recall, and F-scores in addition to AUC.The authors state that accuracy alone does not portray true classifier performance on such datasets.
- Classification errors: 391 of 412 healthy images were correctly classified, whereas only 48 of 139 glaucomatous images were correctly classified.The confusion matrix records 21 healthy images misclassified as glaucomatous and 91 glaucomatous images misclassified as healthy.
- AUC comparison: AUC provides a quantitative basis for comparing binary classifiers, and Table 6 reports the proposed model as superior to comparative studies.The ROC curve varies the positive-confidence threshold and plots recall against specificity.
- Random training: The model achieved 84.87% AUC for some random 99-image training and 551-image testing splits, exceeding four comparison results.The ORIGA dataset has no standard train-test split, so the model was repeatedly trained with random splits.
Conclusion
The paper presents fully automated optic-disc localization that achieves state-of-the-art performance across diverse datasets, while glaucoma classification remains difficult despite improved AUC. The authors therefore emphasize the need for additional classifier metrics and further work on reliable glaucomatous-image recognition.
- The fully automated localization method achieves new state-of-the-art results on six of seven publicly available datasets with IOU greater than 50%.
- The two-stage system combines optic-disc localization with classification of images as healthy or glaucoma-affected.
- ORIGA classification remains difficult because fine-grained discriminative image details appear to be lost deeper in the network.
- The authors call for additional performance measures and classifiers tailored to identify glaucomatous images reliably.
Authors’ contributions
The authors describe contributions spanning methodology, experimentation, supervision, ophthalmological validation, funding, and use of publicly available databases. They report no competing interests.
- MNB designed and implemented the methodology and wrote the manuscript, while MIM and SA conceptualized the project and administered experiments.
- SAS supported method design and implementation; AD and FS supervised the project; WN provided ophthalmological insight and validated ground truth.
- Muhammad Naseer Bajwa received PhD funding support from NUST and Pakistan’s Higher Education Commission.
- All databases used in the work are publicly available.
- The authors declare that they have no competing interests.
Publisher’s Note
The publisher states that Springer Nature remains neutral regarding jurisdictional claims in published maps and institutional affiliations.
- Springer Nature remains neutral regarding jurisdictional claims in published maps and institutional affiliations.