Optic disc and optic cup segmentation based on anatomy guided cascade network

doi:10.1016/j.cmpb.2020.105717

Computer Methods and Programs in Biomedicine

Volume 197, December 2020, 105717

https://doi.org/10.1016/j.cmpb.2020.105717 Get rights and content

Highlights

•
We propose introducing the anatomical knowledge of the optic disc and optic cup into a novel end-to-end cascade network architecture to segment small targets, via an attention mechanism, which trains rapidly from the start of training to convergence.
•
To refine the prediction, instead of post-processing, we propose segmenting the optic disc and optic cup with generative adversarial learning, which uses fully labels generated by the elliptical fitting provided in the training dataset and voids the non-elliptical fitting error in real fundus images.
•
The network framework proposed in this paper obtained state-of-the-art results in the MICCAI 2018 Retinal Fundus Glaucoma Challenge.

Abstract

Background and Objective

Glaucoma, a worldwide eye disease, may cause irreversible vision damage. If not treated properly at an early stage, glaucoma eventually deteriorates into blindness. Various glaucoma screening methods, e.g. Ultrasound Biomicroscopy (UBM), Optical Coherence Tomography (OCT), and Heidelberg Retinal Scanner (HRT), are available. However, retinal fundus image photography examination, because of its low cost, is one of the most common solutions used to diagnose glaucoma. Clinically, the cup-to-disk ratio is an important indicator in glaucoma diagnosis. Therefore, precise fundus image segmentation to calculate the cup-to-disk ratio is the basis for screening glaucoma.

Methods

In this paper, we propose a deep neural network that uses anatomical knowledge to guide the segmentation of fundus images, which accurately segments the optic cup and the optic disc in a fundus image to accurately calculate the cup-to-disk ratio. Optic disc and optic cup segmentation are typical small target segmentation problems in biomedical images. We propose to use an attention-based cascade network to effectively accelerate the convergence of small target segmentation during training and accurately reserve detailed contours of small targets.

Results

Our method, which was validated in the MICCAI REFUGE fundus image segmentation competition, achieves 93.31% dice score in optic disc segmentation and 88.04% dice score in optic cup segmentation. Moreover, we win a high CDR evaluation score, which is useful for glaucoma screening.

Conclusions

The proposed method successfully introduce anatomical knowledge into segmentation task, and achieve state-of-the-art performance in fundus image segmentation. It also can be used for both automatic segmentation and semiautomatic segmentation with human interaction.

Introduction

GLAUCOMA, one of the leading causes of irreversible blindness in the world, is usually caused by elevated Intraocular Pressure (IOP), which leads to mechanical tension and torsion of the optic nerve and loss of retinal nerve fibers [1]. For screening eye diseases, there are many approaches available, such as UBM [2], OCT [3], HRT [4], fundus photography [5] and intra-ocular pressure measurement [6]. Generally, fundus photography, as illustrated in Fig. 1(a), is still the major tool chosen by doctors and patients to evaluate glaucoma at a low cost. As shown in Fig. 1, the optic disc is a slightly vertical oval area in the retinal fundus image of a normal eye; the optic cup is the cup-shaped area in the center of the optic disc. CDR, the ratio of cup to disc diameter, represents the state of the disease [7]. The definition of CDR is illustrated in Fig. 1(b). Previous studies have shown that large CDRs are closely related to the progression of glaucoma and important in the clinical application and evaluation of glaucoma [8], [9]. Generally, a larger CDR suggests glaucoma or other diseases, such as neurological diseases. Therefore, precise segmentation of the optic disc is the premise for accurate calculation of CDRs and a guarantee for correctly diagnosing glaucoma diseases. Clinically, doctors have to draw the contour of the optic disc and optic cup manually, which is labor intensive and time-consuming. Consequently, an automatic segmentation system without human intervention or manual drawing is highly desired.

Segmenting the optic disc and optic cup are admittedly challenging tasks for the following reasons: Initially, the appearance of fundus images varies for various eye diseases. The shape, color, and size of the optic disc and optic cup are different among different patients or different stages of the eye diseases [10], [11]. Furthermore, the target area, especially the optic cup area, occupies only a small portion of the entire fundus image, raising the difficulty of automatic segmentation. Also, the intricate blood vessels inside the optic disc and optic cup lead the contour of the optic disc and cup to be incomplete. Additionally, the boundary between the optic disc and the optic cup is vague, which makes accurate segmentation more challenging [12].

To date, many methods have been proposed for segmenting the optic disc and optic cup. The principal methods include (1) threshold [13], (2) active deformable model and level set [14], [15], [16], (3) statistics [17], [18], [19], (4) curve fitting based on edge extraction [15], [17], etc. Although these methods performed well in some specific evaluation criteria, it is difficult for them to perform perfectly in all aspects. Moreover, deep learning, especially the invention of convolutional neural networks, which has promoted the development of many computer vision tasks, has progressed rapidly in recent years [20], [21], [22], [23]. In particular, FCN [24] and U-Net [25] have completely changed the traditional image segmentation field. However, most of these methods consider the segmentation of the optic cup and the optic disc as two separate tasks and do not fully consider the explicit relationship between the optic disc and the optic cup [26], [27]. Besides, many algorithms require much time to post-process the segmentation results of the network with curve fitting or other strategies, thereby giving up the end-to-end learning [27].

To deal with these problems mentioned above, we propose a novel end-to-end segmentation framework, which achieves state-of-the-art performance in optic disc and optic cup segmentation. It is evident that the optic cup is always inside the optic disc, thus, the segmentation result of the optic disc provides additional information for the position of the optic cup. Considering the difference in appearance between the optic cup and the optic disc, in this paper, we divide the optic disc and cup segmentations in two, thereby enabling our model to examine the image twice from different points of view, just as a doctor would. We glance at the image first to find the area of the optic disc and then stare at it to focus on the optic cup. Moreover, the segmentation task of the optic disc and cup can be viewed as mask generation progress at another perspective. We achieve competitive performance by introducing generative adversarial learning to our end-to-end network framework. The main contributions of this paper are as follows:

1.
We propose introducing the anatomical knowledge of the optic disc and optic cup into a novel end-to-end cascade network architecture to segment small targets, via an attention mechanism, which trains rapidly from the start of training to convergence.
2.
To refine the prediction, instead of post-processing, we propose to segment the optic disc and optic cup with generative adversarial learning, which makes fully use of labels generated by the elliptical fitting provided in the training dataset and voids the non-elliptical fitting error in real fundus images.
3.
The network framework proposed in this paper obtained state-of-the-art results in the MICCAI 2018 Retinal Fundus Glaucoma Challenge [28], [29].

Section snippets

Related works

Segmentation of the optic disc and the optic cup is viewed as a pixel-wise classification. Every pixel in any position of a refuge image must be labeled. According to the segmentation methods, the major fundus image segmentation algorithms are divided into two categories: traditional approaches and deep-learning based methods.

Traditional method Traditional methods focus on (1) color intensity, (2) contour, (3) neighborhood, mainly including threshold, (4) morphology, (5) active contour model,

Methods

In this paper, we use a cascade network observe the fundus image twice to imitate doctors’ examinations. Additionally, instead of post-processing, generative adversarial learning is integrated into our method to refine segmentation results. Here we take U-Net [25] as the base structure of our proposed network framework. The architecture is illustrated in Fig. 2. To capture richer multi-resolution hierarchical semantic information and integrate anatomical knowledge, we use two cascaded U-shape

Dataset

We took the following data as our training data: 400 images from the official training dataset of REFUGE [28] and 519 images randomly sampled from the Origa650 [46] dataset. The remaining 131 images from Origa650 were used as validation data to select the proper threshold for binarization. The data in the official validated dataset of REFUGE were used as our testing data.

Several senior glaucoma specialists created all these annotations. The size of the images from the REFUGE data set are

Discussion

•
use GAN instead of post-process.
Most contours of the optic disc and optic cup are approximately elliptical, but the shapes usually change in different stages of different eye diseases [11]. It may be not proper to apply elliptical fitting to the many samples that have strange shapes. A typical example from the validation dataset of REFUGE [28] is shown in Fig. 9. The given references for the targets are nearly elliptical, as indicated by the red circle. However, the contour of this sample is

Conclusions

We proposed a novel network architecture, which provides an end-to-end style to segment optic discs and cups automatically without any post-processing or ROI detection in advance. Our network architecture combines U-Net and inception block structures to promote the efficiency of network training and parameter optimization. With the cGAN structure, segmentation performance is improved, and the over-fitting problem of the model at low learning rates is alleviated. Moreover, we introduce an

Conflict of Interest

We declare that we have no financial and personal relationships with other people or organizations that can inappropriately influence our work, there is no professional or other personal interest of any nature or kind in any product, service and/or company that could be construed as influencing the position presented in, or the review of, the manuscript entitled, Optic Disc and Optic Cup Segmentation Based on Anatomy Guided Cascade Network

Declaration of Competing Interest

The authors bear all responsibility and have no conflicts of interest to declare.

Acknowledgement

This work is supported by National Natural Science Foundation of China (No. U1605254). The sponsoring organizations were not involved in the study design, data analysis, or interpretation.

References (48)

Mantravadi
Glaucoma
Prim. Care
(2015)
Thakur
Survey on segmentation and classification approaches of optic cup and optic disc for diagnosis of glaucoma
Biomed. Signal Process. Control
(2018)
Jonas
Ophthalmoscopic evaluation of the optic nerve head.
Surv. Ophthalmol.
(1999)
Xu
Efficient optic cup detection from intra-image learning with retinal structure priors
(2012)
J.I. Orlando et al.
Refuge challenge: a unified framework for evaluating automated methods for glaucoma assessment from fundus photographs
Med. Image Anal.
(2020)
Z.U. Rehman et al.
Multi-parametric optic disc segmentation using superpixel based feature classification
Expert Syst. Appl.
(2019)
Yang
Automatic liver segmentation using an adversarial image-to-image network
International Conference on Medical Image Computing and Computer-Assisted Intervention
(2017)
Mannino
A review of the role of ultrasound biomicroscopy in glaucoma associated with rare diseases of the anterior segment
Clin. Ophthalmol.
(2016)
Bussel
Oct for glaucoma diagnosis, screening and detection of glaucoma progression
Br. J. Ophthalmol.
(2014)
Zangwill
Heidelberg retina tomograph measurements of the optic disc and parapapillary retina for detecting glaucoma analyzed by machine learning classifiers.
Invest. Ophthalmol. Vis. Sci.
(2004)

Tuulonen

Screening for glaucoma with a non–mydriatic fundus camera

Acta Ophthalmol.

(2010)

Pronin

Measurement of intraocular pressure by patients with glaucoma.

Jama Ophthalmol.

(2017)

M. Roslin et al.

Glaucoma screening by the detection of blood vessels and optic cup to disc ratio

(2016)

Xu, et al., Optic cup segmentation for glaucoma detection using low-rank superpixel representation,...

Spaeth

The disc damage likelihood scale: reproducibility of a new method of estimating the amount of optic nerve damage caused by glaucoma.

Trans. Am. Ophthalmol.Soc.

(2002)

Aquino

Detecting the optic disc boundary in digital fundus images using morphological, edge detection, and feature extraction techniques

IEEE Trans. Med. Imaging

(2010)

Issac

An adaptive threshold based algorithm for optic disc and cup segmentation in fundus images

(2015)

Lowell

Optic nerve head segmentation

Med. Imaging IEEE Trans.

(2004)

Wong

Level-set based automatic cup-to-disc ratio determination using retinal fundus images in argali

(2008)

Joshi

Optic disk and cup segmentation from monocular color retinal images for glaucoma assessment

IEEE Trans. Med. Imaging

(2011)

Optic cup segmentation for glaucoma detection using low-rank superpixel representation

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

(2014)

J. Cheng et al.

Superpixel classification based optic disc and optic cup segmentation for glaucoma screening

IEEE Trans. Med. Imaging

(2013)

Simonyan, et al., Very deep convolutional networks for large-scale image recognition, arXiv:1409.1556,...

Deep residual learning for image recognition

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

(2016)

Cited by (31)

Automated segmentation of optic disc and cup depicted on color fundus images using a distance-guided deep learning strategy
2023, Biomedical Signal Processing and Control
Optic disc (OD) and cup (OC) regions depicted on color fundus images are important landmarks for assessing glaucoma. In this study, we developed a novel and general distance-guided deep learning strategy (DGLS) to simultaneously segment the OD and OC from color fundus images based on an available network (i.e., U-Net). The developed method used two different types of annotation regions to characterize each target object and then converted the regions into location information leveraging a distance transform in a coarse-to-fine segmentation framework. We validated the developed algorithm by applying it to simultaneously segment OD and OC from color fundus images. Experiments on four public datasets (i.e., the REFUGE, BinRushed, DirshtiGS, and Magrabia datasets) suggested that the developed DGLS achieved, on average, a Dice Score (DS) of 0.9047, a Jaccard index (JI) of 0.8387, and a Hausdorff distance (HD, in pixel) of 2.6211 for the OD and OC in the coarse segmentation stage, and 0.9065, 0.8416, and 2.5930 in the fine segmentation stage. This algorithm demonstrated a superior performance compared to the U-Net and its several variants (i.e., Attention U-Net, BiO-Net and asymmetric network) trained on a single annotation and the traditional training strategy.
Dilated residual FPN-based segmentation for mouse retinal images
2023, Heliyon
Diabetes can induce diabetic retinopathy (DR), and the blindness caused by this disease is irreversible. The early analysis of mouse retinal images, including the layer and cell segmentation properties of these images, can help to effectively diagnose this disease.
In the study, we design a dilated residual method based on a feature pyramid network (FPN), in which the FPN is adopted as the base network for solving the multiscale segmentation problem concerning mouse retinal images. In the bottom-up encoding pathway, we construct our backbone feature extraction network via the combination of dilated convolution and a residual block, further increasing the range of the receptive field to obtain more context information. At the same time, we integrate a squeeze-and-excitation (SE) attention module into the backbone network to obtain more small object details. In the top-down decoding pathway, we replace the traditional nearest-neighbor upsampling method with the transposed convolution method and add a segmentation head to obtain semantic segmentation results.
The effectiveness of our network model is verified in two segmentation tasks: ganglion cell segmentation and mouse retinal cell and layer segmentation. The outcomes demonstrate that, compared to other supervised segmentation methods based on deep learning, our model attains the utmost precision in both binary segmentation and multiclass semantic segmentation tasks.
The dilated residual FPN is a robust method for mouse retinal image segmentation and it can effectively assist DR diagnosis.
Towards an extended EfficientNet-based U-Net framework for joint optic disc and cup segmentation in the fundus image
2023, Biomedical Signal Processing and Control
Automatic and accurate segmentation of the optic disc (OD) and optic cup (OC) is crucial for clinical glaucoma screening. However, the segmentation accuracy cannot meet the requirements of clinical diagnosis yet. This paper aims to propose a novel joint segmentation framework to improve accuracy.
We suggested an extended EfficientNet-based U-Net, named EE-UNet. First, we use the EfficientNet to extract distinguishable features under different scales of receptive fields. Second, we employ the Conditional Random Field as a Recurrent Neural Network (CRF-RNN) to extend the U-Net framework, maintaining an end-to-end mode. Third, the Ranger optimizer is suggested for faster and more stable convergence with minimal computational cost. Finally, we designed a multi-label loss function to balance the foreground and background pixels.
Extensive experiments were trained and verified on the datasets REFUGE and GAMMA and tested on Drishti-GS1 and RIM-ONE-v3. The proposed EE-UNet achieved a high DICE score in OD/OC segmentation (0.9624/0.9228, respectively), outperforming the state-of-the-art (SOTA) methods.
Quantitative and qualitative results verified the superior performance of the proposed method in OC and OD segmentation, thus proving that it can be a promising tool in the early screening of large-scale glaucoma.
PY-Net: Rethinking segmentation frameworks with dense pyramidal operations for optic disc and cup segmentation from retinal fundus images
2023, Biomedical Signal Processing and Control
Glaucoma is one of the leading causes of irreversible blindness across the globe. Early diagnosis is therefore essential for the provision of timely treatments and reducing the loss of vision. From the retinal fundus images, the calculation of the cup to disc ratio (CDR) is an efficient indicator to distinguish between glaucomatous and non-glaucomatous cases. Therefore, precise segmentation of optic disc and cup from the retinal images are important steps for improved diagnosis of the disease. This paper presents a robust segmentation pipeline for optic disc and cup segmentation utilizing the U-Net architecture. In the upsampling half of the model, a spatial pyramid based decoder cascaded with an intermediate decoder is introduced. To incorporate channel-wise attention throughout the framework squeeze and excite blocks are incorporated. To enhance the relevant spatial features in the feature representations of the fundus images, spatial attention module is utilized. Furthermore to achieve multi-scale context extraction modified receptive field blocks (MRFB) are added to the encoding layers of the network. In addition, the MRFBs are also introduced in the auxiliary decoder for fine-tuning of the feature representation. The performance of the proposed model has been evaluated on three publicly available retinal fundus databases, DRISHTI-GS, RIM-ONE, and REFUGE and the model has achieved a dice score of 97.10, 96.12, 96.48 respectively on the optic disc and 93.38, 92.92, 93.85 on optic cup segmentation, outperforming the state-of-the-art methods.
Presumptive discerning of the severity level of glaucoma through clinical fundus images using hybrid PolyNet
2023, Biomedical Signal Processing and Control
Citation Excerpt :
The pixels of each part are grouped and are named as superpixels. In the entire surface of the image, watershed algorithm predicts the watershed lines and the catchment basins [22,23]. There are different algorithms to identify the region by observing the dark and light pixels that are low and high respectively.
Glaucoma is the major leading cause of non-curable visual defects and blindness across the world. Glaucoma is one of the eye diseases which cause an optic nerve, due to increased pressure called intraocular pressure. If the damage gets worse, it causes permanent vision loss within a few years. Timely diagnosis, early and accurate prediction are the most important factors for the prevention of visual blindness. For experts, it is uncompromising task for detecting glaucoma manually. In most of the retinal images, different investigations are carried out for identifying an automated detection tool for diagnosing glaucoma.
The main aim of this paper is to build a machine learning model for predicting glaucoma and its severity level. To predict the severity level of glaucoma, firstly, the segmentation of the blood vessel is performed and classification is accomplished using proposed U-Net model. Secondly, the segmentation of the Optic Disc (OD) and the Optic Cup (OC) is performed to determine the Cup to Disk Ratio (CDR). The proposed hybrid PolyNet classifies the severity of the glaucoma disease. The earlier detection of glaucoma disease is important for preventing the severity of the eye disease.
The segmentation of blood vessels is implemented using the U-Net model, the sensitivity and the specificity are measured as 97% and 88% respectively. The accuracy and the F1 score have been measured as 96% and 98% respectively. Hence, the proposed deep learning architecture with hybrid PolyNet models is implemented for OD and OC separately to produce an appropriate result. By increasing the number of layers, more features are identified. The current work deals with the retinal fundus images and achieves the highest accuracy of 96.21% in the ACRIMA dataset with sensitivity and specificity of about 97.32% and 94.26% respectively.
Blood vessel segmentation is achieved with U-Net and the segmentation of the optic disc and the optic curve have been achieved and the severity level of glaucoma has been identified using hybrid PolyNet with high accuracy.
Medical image segmentation using deep semantic-based methods: A review of techniques, applications and emerging trends
2023, Information Fusion
Citation Excerpt :
In contrast to NENet-seg, the newly designed model accurately determined OD and OC boundaries. An advanced model is suggested by Bian et al. [120]. It works by segmenting optic disc (OD) and optic cup (OC) without relying on post-processing or ROI spotting, as shown in Fig. 29.
Semantic-based segmentation (Semseg) methods play an essential part in medical imaging analysis to improve the diagnostic process. In Semseg technique, every pixel of an image is classified into an instance, where each class is corresponded by an instance. In particular, the semantic segmentation can be used by many medical experts in the domain of radiology, ophthalmologists, dermatologist, and image-guided radiotherapy. The authors present perspectives on the development of an architectural, and operational mechanism of each machine learning-based semantic segmentation approach with merits and demerits. In this regard, researchers have proposed different Semseg methods and examined their performance in a variety of applications such as medical image analysis (e.g., medical image classification and segmentation). A review of recent advances in Semseg techniques are presented in this paper by applying computational image processing and machine learning methods. This article is further presented a comprehensive investigation on how different architectures are helpful for medical image segmentation. Finally, advantages, open challenges, and possible future directions are elaborated in the discussion part, beneficial to the research community to understand the significance of the available medical imaging segmentation technology based on Semseg and thus deliver robust segmentation solutions.

View all citing articles on Scopus

View full text

Optic disc and optic cup segmentation based on anatomy guided cascade network

Highlights

Abstract

Background and Objective

Methods

Results

Conclusions

Introduction

Section snippets

Related works

Methods

Dataset

Discussion

Conclusions

Conflict of Interest

Declaration of Competing Interest

Acknowledgement

Prim. Care

Biomed. Signal Process. Control

Surv. Ophthalmol.

Med. Image Anal.

Expert Syst. Appl.

A review of the role of ultrasound biomicroscopy in glaucoma associated with rare diseases of the anterior segment

Clin. Ophthalmol.

Oct for glaucoma diagnosis, screening and detection of glaucoma progression

Br. J. Ophthalmol.

Heidelberg retina tomograph measurements of the optic disc and parapapillary retina for detecting glaucoma analyzed by machine learning classifiers.

Invest. Ophthalmol. Vis. Sci.

Screening for glaucoma with a non–mydriatic fundus camera

Acta Ophthalmol.

Measurement of intraocular pressure by patients with glaucoma.

Jama Ophthalmol.

Glaucoma screening by the detection of blood vessels and optic cup to disc ratio

The disc damage likelihood scale: reproducibility of a new method of estimating the amount of optic nerve damage caused by glaucoma.

Trans. Am. Ophthalmol.Soc.

Detecting the optic disc boundary in digital fundus images using morphological, edge detection, and feature extraction techniques

IEEE Trans. Med. Imaging

An adaptive threshold based algorithm for optic disc and cup segmentation in fundus images

Optic nerve head segmentation

Med. Imaging IEEE Trans.

Level-set based automatic cup-to-disc ratio determination using retinal fundus images in argali

Optic disk and cup segmentation from monocular color retinal images for glaucoma assessment

IEEE Trans. Med. Imaging

Optic cup segmentation for glaucoma detection using low-rank superpixel representation

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

Superpixel classification based optic disc and optic cup segmentation for glaucoma screening

IEEE Trans. Med. Imaging

Deep residual learning for image recognition

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition