PaperlayerView exact revision on arXiv

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

arXiv:2005.14284v1cs.CVcs.LGeess.IV

TL;DR

Dataset-specific heuristic optic-disc localization methods generalized poorly, motivating a fully automated approach for glaucoma screening. The paper uses faster R-CNN for disc localization and CNN-based classification, and reports state-of-the-art localization on six of seven public datasets at IOU greater than 50%.

  • Problem

    Existing heuristic disc-localization methods were dataset-dependent and generalized poorly, while classification required focusing on the optic disc.

    Source: DRIONS-DB · p. 7

  • Method

    The paper pairs fully automated faster R-CNN optic-disc localization with CNN-based classification of diseased and healthy images.

    Source: Automated disc localization · p. 10, Conclusion · p. 14

  • Results

    Six of seven public datasets saw state-of-the-art optic-disc localization, with IOU greater than 50%.

    Source: Conclusion · p. 14

  • Takeaways & Limitations

    The authors conclude that automated localization removes the need for dataset-specific empirical or heuristic methods across a broad range of datasets.

    Source: Conclusion · p. 14

    On ORIGA, class imbalance and the lack of standard train/test splits mean AUC alone does not portray classifier performance, motivating additional metrics.

    Source: Background · p. 3

Abstract

from arXiv · show

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 can cause irreversible vision loss before symptoms are apparent, motivating optic-disc-focused automated screening. The paper presents a two-stage deep-learning framework for optic-disc localization and glaucoma classification.

Background

  • Glaucoma may remain symptomless until advanced, irreversible damage; optic-disc cues include a depressed cup, disc dilation, pallor, and cup-to-disc ratio.The source notes that optic-nerve-fibre damage and visual-field reduction become noticeable after about 40% of axons are lost.Source: Background · p. 2
  • Disc segmentation errors can significantly affect diagnosis, whereas localization-based ROI extraction retains surrounding context and is described as more resilient to localization errors.Glaucoma indicators such as disc size, cup-to-disc ratio, ISNT quadrants, and peripapillary atrophy are concentrated around the optic disc.Source: Background · p. 2
  • The proposed faster RCNN method set a new state of the art on six of seven localization datasets at IOU > 50, while DCNN classification improved AUC by 2.7% relative to the previous best.The localization method was designed to be fully automated and dataset-independent.Source: Background · p. 2
  • Because public datasets lacked bounding-box annotations, the authors generated training ground truth with a rule-based semi-automated method and reported precision, recall, and F-scores alongside AUC for ORIGA.They argue that AUC alone can misrepresent classifier performance on imbalanced datasets without standard train/test splits.Source: Background · p. 3

Related work

  • Early diagnosis depends on accurately determining optic-disc position, centre, and size, which can support subsequent automated image analysis.The paper introduces prior image-processing and machine-learning approaches that use diagnostic criteria for disc localization and glaucoma identification.Source: Related work · p. 3

Localization of optic disc

  • Prior optic-disc localization methods exploit intensity variation, brightness thresholds, vessel patterns, circular shape, and image features, but bright reflections can mislead heuristic methods.Reported techniques include adaptive filtering and thresholding, Hough transforms, and visual-dictionary classification.Source: Localization of optic disc · p. 4

Glaucoma classification

  • On ORIGA, prior methods reported AUC values from 0.778 to 0.851, while some earlier accuracy results came from small, handpicked image sets that may not generalize.Reported ORIGA results include 0.823 for reconstruction-based classification, 0.8384 for combined local and holistic AlexNet features, and 0.851 using vertical cup-to-disc ratio from segmented discs.Source: Glaucoma classification · p. 5

Methods … HRF image database

The methods cover optic-disc localization and classification and introduce the datasets used. Because those datasets lacked bounding-box ground truth for disc localization, the authors developed a ground-truth-generation mechanism.

Methods

  • The reviewed datasets supplied no optic-disc localization bounding-box ground truth, prompting development of a ground-truth-generation mechanism.The methodology covers optic-disc localization and classification and begins with an introduction to the publicly available datasets used.Source: Methods · p. 5

ORIGA (−light)

  • ORIGA provides manual optic-disc and optic-cup segmentations, cup-to-disc ratios, and healthy or glaucomatous labels for benchmarking segmentation and classification algorithms.The dataset uses a custom tool to generate the manual segmentations.Source: ORIGA (−light) · p. 5
  • ORIGA contains 482 healthy images and 168 glaucomatous images and has served as a standard dataset in some recent glaucoma-classification research.Source: ORIGA (−light) · p. 5

HRF image database

  • The HRF image database is provided by the Department of Ophthalmology at Friedrich-Alexander University Erlangen-Nuremberg, Germany.Source: HRF image database · pp. 5–6

Methods

  • Fig. 3 presents the workflow for the semi-automatic ground-truth-generation mechanism.Source: Figure 3 · p. 6

HRF image database

  • HRF contains 15 healthy, 15 glaucomatous, and 15 diabetic-retinopathy images, each with expert-provided binary gold-standard vessel segmentation.Source: HRF image database · p. 6

OCT & CFI · DIARETDB1

OCT & CFI and DIARETDB1 are described as datasets with different purposes and annotation coverage. Neither provides ground truth for optic-disc localization.

OCT & CFI

  • OCT & CFI contains OCT and colour fundus images of both eyes from 50 healthy people recruited at Feiz Hospital in Iran.Source: OCT & CFI · p. 6
  • OCT & CFI was collected for comparing macular OCTs between normal right and left eyes, and lacks ground truth for optic-disc or vessel segmentation and optic-disc localization.Source: OCT & CFI · p. 6

DIARETDB1

  • DIARETDB1 contains 89 colour fundus images collected at Kuopio University Hospital in Finland.Source: DIARETDB1 · p. 6
  • 84 DIARETDB1 images had at least mild non-proliferative diabetic retinopathy, while five were healthy.Source: DIARETDB1 · p. 6
  • DIARETDB1 was created primarily to benchmark automated diabetic-retinopathy detection methods.Source: DIARETDB1 · p. 6
  • Four independent medical experts marked microaneurysms, haemorrhages, and hard and soft exudates in DIARETDB1.Source: DIARETDB1 · p. 6
  • DIARETDB1 provides neither ICDR severity grades nor ground truth for optic-disc localization.Source: DIARETDB1 · p. 6

DRIVE · DRIONS-DB

The authors argue that classification requires an optic-disc region of interest, while dataset-specific heuristic localization methods did not generalize; supervised localization also required ground truth absent from the seven datasets, motivating a minimally assisted generation method.

DRIVE

  • DRIVE contains 40 images split equally into train and test sets, with one vessel-segmentation mask per training image and two per test image.It provides no optic-disc localization annotations.Source: DRIVE · p. 6

DRIONS-DB

  • DRIONS-DB contains 110 retinal images collected at Miguel Servet Hospital in Saragossa, Spain, for optic nerve head segmentation benchmarking.Two independent experts traced ONH contours using an annotation tool.Source: DRIONS-DB · p. 6
  • Messidor provides 1,200 high-resolution colour fundus images, each with diabetic-retinopathy grade and macular-edema risk annotations.Its images were collected from three ophthalmologic departments under a French Ministry of Research and Defense-funded project.Source: DRIONS-DB · p. 6
  • The authors found that whole-fundus CNN input did not focus classification on the optic disc, while existing heuristic localization methods did not generalize across datasets.They therefore sought a robust, fully automated, dataset-independent localization method.Source: DRIONS-DB · p. 7
  • Because none of the seven datasets used in the work supplied localization ground truth, the authors devised a generation method requiring minimal human involvement.The generated annotations were intended to train and benchmark automated disc localization.Source: DRIONS-DB · p. 7
  • The paper proceeds from ground-truth generation to optic-disc extraction and then classification.The stated classification task was to distinguish healthy from glaucomatous retinal fundus images.Source: DRIONS-DB · pp. 6–7

Semi-automated GT generation for OD localization

A heuristic proposes optic-disc locations that experts manually verify and correct as needed, producing localization ground truth for seven datasets. The heuristic was evaluated on three datasets using a held-out test set and images rescaled to a common size.

Semi-automated GT generation for OD localization

  • Heuristic optic-disc proposals were manually checked and corrected by an expert to generate localization ground truth for all seven datasets.The workflow is depicted in Figure 3.Source: Semi-automated GT generation for OD… · p. 7
  • Of 780 images from three high-resolution datasets, 525 were used for training, 48 for validation, and 207 for testing.Validation selected empirical parameters, which were fixed during testing; images were processed at 1500 × 1500 pixels and then restored to their original size.Source: Semi-automated GT generation for OD… · p. 7
  • Figure 5 shows correct heuristic localization examples from HRF, OCT, CFI, and ORIGA, plus the only illustrated failure, on an ORIGA image.The supplied passage identifies this as the only incorrect localization example shown.Source: Figure 5 · p. 7
  • The three-dataset mixture was intended to test heuristic accuracy and robustness across inter-dataset image-resolution variation.All images were rescaled to a fixed size for processing before being returned to their original size.Source: Semi-automated GT generation for OD… · pp. 7–8

Heuristic algorithm for OD localization

The heuristic method estimates the optic disc (OD) location through retinal-rim cropping, adaptive thresholding, and morphological cleanup, then localizes a candidate disc for expert verification.

Heuristic algorithm for OD localization

  • Bright rim fringes from misalignment and reflective cloud-like spots near the macula can mislead brightness-based localization.The rim fringe can result from incorrect eye placement during image capture, while macular-region spots can arise from light reflection in the ocular fundus.Source: Heuristic algorithm for OD localization · pp. 8–9
  • Otsu thresholding estimates retinal dimensions for circular cropping, removing rim artifacts despite imprecise fundus segmentation.The crop uses a mask with radius smaller than the estimated retinal radius.Source: Heuristic algorithm for OD localization · p. 9
  • The method thresholds each image at the mean intensity of its top 1% brightest pixels, then erodes and dilates to suppress noise and join bright regions.These operations provide a better approximation of the OD core before its centre and radius are calculated.Source: Heuristic algorithm for OD localization · p. 9
  • A larger-radius circle marks the approximate OD location, and an expert manually verifies and corrects candidate locations where needed.The method calculates the candidate centre and radius after morphological processing.Source: Heuristic algorithm for OD localization · p. 9

Results of heuristic localization

Heuristic localization achieved substantial overlap with manual disc annotations, but the approach was dataset-specific. Its annotations were corrected and used as ground truth for the subsequent automated localization method.

Results of heuristic localization

  • Visual inspection found failures on 3 of 573 test-plus-validation images and 1 of 207 test images from three datasets.The failures are shown in Fig. 5.Source: Results of heuristic localization · p. 9
  • More than 96% of optic discs were localized with over 50% of the actual disc present, while average test-image overlap was around 70%.Overlap was measured using IOU between predicted bounding boxes and manual ground truth.Source: Results of heuristic localization · p. 9
  • About 52% of predicted discs contained over 70% of the actual disc; minimum IOU was over 20%, versus some researchers’ criteria accepting IOU >0.Those criteria counted localization as correct when the predicted-center distance was less than the expected disc diameter.Source: Results of heuristic localization · p. 9
  • Despite promising results, the heuristic approach was dataset-specific and might not generalize to diverse real-world fundus images.Its annotations were corrected and supplied as ground truth to the subsequent automated localization method.Source: Results of heuristic localization · p. 9

Two-stage method for automated disc localization and classification: the proposed approach

The proposed system first localizes the optic disc and then classifies it as healthy or glaucomatous. Its method combines Faster R-CNN localization with a CNN classifier.

Automated disc localization

  • Faster R-CNN uses proposals, CNN classification, and box regression, achieving 100% correct localization at 60% IOU and 97.11% average overlap across three datasets.Source: Automated disc localization · p. 10

Classification of glaucoma

  • The classifier processes the extracted optic-disc ROI—where most glaucoma information lies —through four convolutional and three fully connected layers, ending in softmax class probabilities.Source: Classification of glaucoma · p. 10

Convolutional layers

  • The CNN learns features through convolution; its first layer uses 96 kernels of size 11 × 11 × 3, and ReLU adds nonlinearity.Source: Convolutional layers · p. 10

Pooling layers

  • Max pooling downsamples feature maps to reduce their size and help control overfitting; the network uses 3 × 3 windows with stride 2.Source: Pooling layers · pp. 10–11

Normalization layers

  • Local Response Normalization normalizes activations around excited neurons, creating competition among kernel outputs and supporting generalization.Source: Normalization layers · p. 11

Dropout layers

  • Dropout zeros hidden-neuron outputs with probability 0.5 during training to reduce overfitting and encourage robust features.Source: Dropout layers · p. 11

Results and discussion

The automated detector outperformed dataset-specific heuristics on four datasets, while ORIGA classification results varied across random splits and accuracy alone was insufficient for imbalanced data.

Results with random training

Results of automated disc localization

  • The detector outperformed dataset-specific heuristics despite not being trained on the four datasets; mean overlaps were 84.65% for DIARETDB1, 84.13% for DRIVE, and 80.46% for DRIONS-DB.The passage also begins reporting an overlap for another dataset, but its value and dataset name are incomplete.Source: Results of automated disc localization · p. 12

Results of classification

  • To limit healthy-class bias during training, stratified sampling ensured each batch included glaucoma images; training used Adam, cross-entropy loss, and a 0.0001 learning rate.Source: Results of classification · p. 12

Results with random training

  • Across more than 1,500 random runs using 99 training images and the remainder for testing, the best split reached 79.67% overall accuracy.The dataset has no standard train/test split.Source: Results with random training · p. 12
  • Because ORIGA is class-imbalanced, accuracy alone does not portray classifier performance, so precision, recall, and F1-score are also considered.Precision, recall, and F-score capture different aspects of class-wise prediction performance.Source: Results with random training · p. 12

Conclusion

The paper concludes that fully automated Faster R-CNN optic disc localization generalizes across public datasets, while CNN glaucoma classification on ORIGA remains difficult despite higher AUC.

Conclusion

  • Faster R-CNN localization set new state-of-the-art on six of seven public datasets at IoU greater than 50%, without dataset-specific heuristic methods.The method is described as robust and accurate across a broad spectrum of datasets.Source: Conclusion · p. 14
  • Despite higher AUC on ORIGA with random training and k-fold cross-validation, the CNN struggles to learn discriminative features for glaucomatous images.The authors suggest fine-grained discriminative details are lost as the network hierarchy increases.Source: Conclusion · p. 14
  • More effort is needed to tailor classifiers that identify glaucomatous images reliably.Source: Conclusion · p. 14

Authors’ contributions

MNB led the methodology and manuscript, while the other authors contributed to project design, implementation, supervision, ophthalmological validation, and manuscript review. The authors also report PhD funding for MNB, public availability of the databases, and no competing interests.

Authors’ contributions

  • MNB designed and implemented the methodology and wrote the manuscript; all authors approved the final version.MIM and SA conceptualised the project, administered experiments, and revised the draft; SAS assisted with method design and implementation, AD and FS supervised the project, and WN provided ophthalmological insight and validated ground truth.Source: Authors’ contributions · p. 14

Funding

  • MNB received PhD funding from NUST and HEC Pakistan under the Prime Minister Program for Development of PhDs in Science and Technology.Source: Funding · p. 14

Availability of data and materials

  • The authors state that all databases used in the work are publicly available.The passage lists ORIGA, HRF Image Database, OCT & CFI, DIARETDB1, DRIVE, DRIONS-DB, and Messidor.Source: Availability of data and materials · p. 14

Competing interests

Author details

  • The listed affiliations span Technische Universität Kaiserslautern, DFKI, NCAI, NUST, and an ophthalmology clinic in Kaiserslautern.Source: Author details · p. 14

Publisher’s Note

Springer Nature states that it remains neutral regarding jurisdictional claims in published maps and institutional affiliations.

  • Springer Nature remains neutral regarding jurisdictional claims in published maps and institutional affiliations.Source: Publisher’s Note · p. 16

Related papers

Loading 2005.14284v1…