Automated Optic Disc and Cup Segmentation for

Glaucoma Detection from Fundus Images Using the

Detectron2’s Mask R-CNN
2022 International Symposium on Multidisciplinary Studies and Innovative Technologies (ISMSIT) | 978-1-6654-7013-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/ISMSIT56059.2022.9932660

Fengze Wu Marion Chiariglione Xiaoyi Raymond Gao

Department of Biomedical Informatics Department of Ophathalmology and Visual Department of Ophathalmology and Visual
The Ohio State University Sciences Sciences, Department of Biomedical
Columbus, USA The Ohio State University Informatics, Division of Human Genetics
[email protected] Columbus, USA The Ohio State University
[email protected] Columbus, USA
[email protected]

Abstract—Glaucoma is a chronic, degenerative optic they notice any extensive visual field loss in one or both eyes.
neuropathy and the leading cause of irreversible blindness About half of individuals with glaucoma are not aware that they
worldwide. Individuals with glaucoma do not show typical have the disease even in developed countries [2]. Therefore,
symptoms for years and can become advanced before patients early detection and treatment are crucial for preventing vision
notice an extensive visual field loss. Therefore, early detection and loss from glaucoma.
treatment are crucial to prevent vision loss from this blinding
disease. Vertical cup-to-disc ratio (VCDR), the ratio of the vertical Vertical cup-to-disc ratio (VCDR), the ratio of the vertical
diameter of the cup over the vertical diameter of the disc in the diameter of the cup over the vertical diameter of the disc in the
optic nerve head region, is an important structural indicator for optic nerve head region, is an important structural indicator for
glaucoma. Estimation of VCDR requires accurate segmentation of glaucoma. VCDR greater than 0.7 is a typical criterion for
optic disc (OD) and optic cup (OC) on fundus images. However, glaucoma [3]. VCDR can be estimated by eye specialists during
manual annotation of the disc and cup area is time-consuming and eye exams or from fundus images. However, manual estimation
is subjective to personal experience and opinion. In this study, we and annotation of the optic disc (OD) and optic cup (OC) are
proposed an automated deep learning approach for OD and OC time-consuming and are subjective to personal experience and
segmentation and VCDR derivation from fundus images using opinion. It is imperative to develop an accurate and automatic
Detectron2, a state-of-the-art object detection platform. We algorithm for deriving VCDR from fundus images.
trained Mask R-CNN models for OD and OC segmentation and
VCDR evaluation. We assessed the performance of our method Various segmentation and classification algorithms have
using the Retinal Fundus Glaucoma Challenge (REFUGE) dataset been used to locate the boundary of the OD and OC from retinal
in terms of the Dice similarity coefficient (DSC) for OD and OC, fundus images. The existing methods include grayscale image
and the mean absolute error (MAE) for VCDR. Our method thresholding [4], contour and color clustering [5], and various
achieved highly accurate results with a DSC of 0.9622 for OD, a deep-learning architectures [6]. Deep learning techniques have
DSC of 0.8870 for OC, and an MAE of 0.0376 for VCDR on the been shown to give superior results to traditional image
hold-out testing images. This implementation surpassed all top- processing techniques in VCDR estimation [7]. With new deep
performing methods in the REFUGE challenge by improving OD learning architectures and pre-training models available, it is
and OC DSC by 0.2% and 0.4%, respectively, and reducing the
possible to further improve the VCDR estimation accuracy.
VCDR MAE by 9%. Our method provided an accurate and
automated solution for OD and OC segmentation and VCDR In this study, we developed automatic OD and OC
estimation. segmentation deep learning models and derived the
corresponding VCDR from fundus images based on
Keywords—glaucoma, fundus image, object detection, instance Detectron2’s [8] implementation of Mask R-CNN [9] and
segmentation, Detectron2, deep learning showed that our models outperform previous state-of-the-art
I. INTRODUCTION (SOTA) methods in both OD and OC segmentation and VCDR
estimation. Our method not only reduces the manual burden on
Glaucoma is a chronic, degenerative optic neuropathy and eye specialists but also provides an objective and quantitative
the leading cause of irreversible blindness worldwide. If it is left determination of VCDR from fundus images.
untreated, glaucoma can lead to permanent vision loss and
blindness. Glaucoma presents a significant disease burden. It II. MATERIALS AND METHODS
affects 70-90 million people worldwide and is responsible for
blindness in approximately 4.5 million people. The U.S. A. Dataset
economic cost of glaucoma care is estimated at $2.86 billion We used the Retinal Fundus Glaucoma Challenge
annually [1]. Individuals with glaucoma typically do not show (REFUGE) dataset [10], which consists of 1200 color fundus
symptoms for years and their case may become advanced before images with ground truth OD and OC segmentations and

978-1-6654-7013-1/22/$31.00 ©2022 IEEE

567
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on February 27,2024 at 08:37:27 UTC from IEEE Xplore. Restrictions apply.
glaucoma labels. All images are in JPG format except ground
truth masks, which are in BMP format. The images are divided
into three sets: 400 training, 400 validation, and 400 test images.
The training images are of 2124 × 2056 pixels and the validation
and testing images are of 1634 × 1634 pixels. OD/OC polygon
annotations are included in the dataset based on the majority
voting results from seven ophthalmologists and are considered
ground truth. Manual annotations of the OD and OC by
glaucoma specialists are in BMP format where the color values
128, 255 and 0 are assigned to OD, OC, and the rest areas of the
image, respectively. We extracted the coordinates of OD/OC
bounding boxes in the format of , , , where ,
represents the left-top corner and , represents the right-
bottom corner of the bounding box. We extracted the pixel-level
coordinates of OD/OC polygons in the format of
, ,…, , where 3 and , represents the
coordinate of each pixel in the clockwise order. Table 1 shows
the summary of the REFUGE dataset we used in this study. Fig.
1 shows examples of REFUGE images and ground-truth masks.

TABLE I. THE REFUGE DATASET

Fig. 2. Overview of the proposed framework. FPN: Feature Pyramid Network,
Number RPN: Region Proposal Network, ROI: Region of Interest.
Dataset Resolution Camera
of images
1) The input image is fed to the pre-trained ResNeXt101-
Train 400 2124 × 2056 Zeiss Visucam 500
FPN model to obtain feature maps.
Valid 400 1634 × 1634 Cannon CR-2 2) The Region Proposal Network (RPN) scans the given
feature map and outputs multiple regions of interest
Test 400 1634 × 1634 Cannon CR-2
(ROI).
3) The feature map and the ROI are sent to the RoIAlign
block, for each ROI to have a fixed-size feature map.
4) Finally, the feature map is sent to fully connected
layers (for object detection and classification) and a
fully convolutional network (for pixel segmentation).
C. Training, validation, and testing
With 800 training and validation images available in the
REFUGE dataset, we used 10-fold cross-validations to train
and validate our OD/OC segmentation models. The final
models were evaluated in 400 REFUGE test images.
We trained two Mask R-CNN models: one was used to
detect OD from fundus images and the other one was used to
detect OC from segmented OD area. All the 800 color fundus
Fig. 1. Example of the REFUGE dataset. (a) Prime Images, (b) Ground-truth.
images and disc area segments were used to train the final
Optic disc and optic cup are denoted in gray and black, respectively. models. The bounding boxes of OD were cropped from ROI.
We applied image augmentation to both OD/OC training
B. Detectron2 Mask R-CNN including horizontal flip and resizing.
We used Detectron2, a Facebook AI Research’s next To accelerate the training, validation, and testing
generation library that provides SOTA object detection and procedures, we used Dual NVIDIA Tesla V100 GPU cards with
segmentation algorithms based on PyTorch, to implement our 32GB memory through the Ohio Supercomputer Center.
Mask R-CNN model, an instance segmentation algorithm. Mask Python v3.7 and PyTorch v1.8 were used to implement the
R-CNN was initially proposed by He et al. [9] with accuracy proposed method.
further improved in its Detectron2 implementation. We used the
COCO instance segmentation baselines with Mask R-CNN and D. Evaluation metrics
ResNeXt101-FPN backbone in our study. Fig. 2 shows an We used the Dice similarity coefficient (DSC) and the mean
overview of our proposed OD/OC detection and segmentation absolute error (MAE) for evaluating OD/OC instance
pipeline. segmentations and VCDR estimations, respectively.
Our proposed framework can be divided into the following DSC [10, 11] is defined as the following:
four steps: | ∩ |
2
| | | |
This work was in part funded by the Ohio Lions Eye Research Foundation.

568
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on February 27,2024 at 08:37:27 UTC from IEEE Xplore. Restrictions apply.
where and are the ground truth and predicted
segmentations of the OD/OC regions, respectfully.
VCDR [12] is calculated as the following:
/
where is the vertical diameter of OD/OC.
VCDR MAE [10, 11, 13] is defined as:
1
!" Σ%& | '( ) *( |

where '( and *( are VCDR values calculated based

on predicted and ground truth OD/OC polygon masks.
III. RESULTS
We implemented the proposed OD/OC segmentation
pipeline using the Detectron2 platform in PyTorch. We used a
grid search for hyperparameter tuning. For each group of
hyperparameters, we applied 10-fold cross-validation and
selected the set of hyperparameters that yield the best average
DSC for OD/OC segmentation. We found that hyperparameters
such as: learning rate = 0.01, maximum iterations = 13500, and
batch size = 64 gave the best DSC for OD segmentation and Fig. 3. Examples of detection and segmentation results from our proposed
model. (a) Input images, (b) Ground-truth bounding boxes, (c) Predicted
learning rate = 0.001, maximum iterations = 6750, and batch bounding boxes, (d) Grround-truth masks, and (e) Predicted masks.
size = 64 gave the best DSC for OC segmentation. Masks for optic disc and optic cup are dennoted in gray and black,
respectively.
In order to study the performance of our proposed method,
we compared it with the results in the REFUGE Challenge The best results of the three top-performing teams in
using the same dataset and evaluation metrics. We applied our
REFUGE were 0.9602 for OD DSC, 0.8837 for OC DSC, and
models to the 400 hold-out images in the REFUGE testing
0.0414 for VCDR MAE. Our method achieved an OD DSC of
dataset and extracted OD/OC regions from all the images.
Based on our hardware configuration, the training time for both 0.9622, an OC DSC of 0.8870, and a VCDR MAE of 0.0376,
models is around three hours. The proposed method took 4.42s which outperformed the top-performing REFUGE teams by
on average for model loading, 0.12s on average per fundus improving OD and OC DSCs by 0.2% and 0.4%, respectively,
image for OD detection, and 0.08s on average per ROI for OC and reducing the VCDR MAE by 9.0%. Our results outperform
detection. The proposed method is much faster than manual all REFUGE teams in average OD/OC DSC score and VCDR
annotations. This shows the time-saving advantage of this MAE.
method in the application. Fig. 3 shows examples of the OD and
OC detection and segmentation results from our proposed IV. CONCLUSION AND DISCUSSION
models. In this study, we proposed an automated method for detecting
Table 2 shows the performance of our models in terms of and segmenting OD and OC from fundus images and deriving
DSC for OD and OC segmentations, and MAE for VCDR the corresponding VCDR. Our method used Detectron2’s
estimation in comparison to the top-performing methods in the implementation of Mask R-CNN for object detection and
REFUGE challenge. instance segmentation. We trained two models for detecting
OD and OC separately. We evaluated the performance of our
TABLE II. OPTIC DISC/CUP SEGMENTATION RESULTS IN THE TEST SET work using DSC for OD/OC segmentation and MAE for
VCDR. Our models outperformed the top-ranking methods in
Disc Avg. Cup Avg. VCDR
Method Training Set
DSC DSC MAE the REFUGE challenge in both OD/OC segmentation and
REFUGE VCDR estimation. Since images in the REFUGE dataset are
Proposed 0.9622 0.8870 0.0376
train and valid from a single population (i.e., Chinese) future works need to
REFUGE REFUGE include more diverse datasets to improve the generalizability of
0.9602 0.8826 0.0450
rank #1 train and valid
REFUGE REFUGE train
the presented method. Moreover, other (advanced) deep
0.9464 0.8837 0.0414 learning algorithms, such as ensemble learning, may further
rank #2 and ORIGA
REFUGE
REFUGE train 0.9525 0.8728 0.0456
increase the model performance. In summary, our proposed
rank #3 method provides an accurate and automated tool for segmenting
Abbreviations: Avg.: average, DSC: Dice similarity coefficient, MAE: mean OD/OC and determining VCDR from fundus images,
absolute error, VCDR: vertical cup-to-disc ratio. addressing the critical need for early detection to reduce
glaucoma-related blindness.

569
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on February 27,2024 at 08:37:27 UTC from IEEE Xplore. Restrictions apply.
[1] K. Allison, D. Patel, and O. Alabi, "Epidemiology of glaucoma:
the past, present, and predictions for the future," Cureus, vol. 12,
no. 11, 2020.
[2] R. Bourne, "Worldwide glaucoma through the looking glass," vol.
90, ed: BMJ Publishing Group Ltd, 2006, pp. 253-254.
[3] J. G. Crowston, C. R. Hopley, P. R. Healey, A. Lee, P. Mitchell,
and S. Blue Mountains Eye, "The effect of optic disc diameter on
vertical cup to disc ratio percentiles in a population based cohort:
the Blue Mountains Eye Study," Br J Ophthalmol, vol. 88, no. 6,
pp. 766-70, Jun 2004, doi: 10.1136/bjo.2003.028548.
[4] A. Issac, M. Parthasarthi, and M. K. Dutta, "An adaptive threshold
based algorithm for optic disc and cup segmentation in fundus
images," in 2015 2nd international conference on signal
processing and integrated networks (SPIN), 2015: IEEE, pp. 143-
147.
[5] I. Fondón et al., "Automatic cup-to-disc ratio estimation using
active contours and color clustering in fundus images for glaucoma
diagnosis," in International Conference Image Analysis and
Recognition, 2012: Springer, pp. 390-399.
[6] K. Park, J. Kim, and J. Lee, "Automatic optic nerve head
localization and cup-to-disc ratio detection using state-of-the-art
deep-learning architectures," Scientific reports, vol. 10, no. 1, pp.
1-10, 2020.
[7] H. Veena, A. Muruganandham, and T. S. Kumaran, "A review on
the optic disc and optic cup segmentation and classification
approaches over retinal fundus images for detection of glaucoma,"
SN Applied Sciences, vol. 2, no. 9, pp. 1-15, 2020.
[8] A. K. Yuxin Wu, Francisco Massa, Wan-Yen Lo, Ross Girshick.
"Detectron2." https://github.com/facebookresearch/detectron2
(accessed 07 July, 2022).
[9] K. He, G. Gkioxari, P. Dollár, and R. Girshick, "Mask r-cnn," in
Proceedings of the IEEE international conference on computer
vision, 2017, pp. 2961-2969.
[10] J. I. Orlando et al., "Refuge challenge: A unified framework for
evaluating automated methods for glaucoma assessment from
fundus photographs," Medical image analysis, vol. 59, p. 101570,
2020.
[11] H. He, L. Lin, Z. Cai, and X. Tang, "Joined: Prior guided multi-
task learning for joint optic disc/cup segmentation and fovea
detection," arXiv preprint arXiv:2203.00461, 2022.
[12] B. J. Fan, D. Y. Wang, L. R. Pasquale, J. L. Haines, and J. L.
Wiggs, "Genetic variants associated with optic nerve vertical cup-
to-disc ratio are risk factors for primary open angle glaucoma in a
US Caucasian population," Investigative ophthalmology & visual
science, vol. 52, no. 3, pp. 1788-1792, 2011.
[13] H. Fu, J. Cheng, Y. Xu, D. W. K. Wong, J. Liu, and X. Cao, "Joint
optic disc and cup segmentation based on multi-label deep network
and polar transformation," IEEE transactions on medical imaging,
vol. 37, no. 7, pp. 1597-1605, 2018.

570
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on February 27,2024 at 08:37:27 UTC from IEEE Xplore. Restrictions apply.