Optic disc and optic cup segmentation based on anatomy guided cascade network
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)
Glaucoma
Prim. Care
(2015)Survey on segmentation and classification approaches of optic cup and optic disc for diagnosis of glaucoma
Biomed. Signal Process. Control
(2018)Ophthalmoscopic evaluation of the optic nerve head.
Surv. Ophthalmol.
(1999)Efficient optic cup detection from intra-image learning with retinal structure priors
(2012)- et al.
Refuge challenge: a unified framework for evaluating automated methods for glaucoma assessment from fundus photographs
Med. Image Anal.
(2020) - et al.
Multi-parametric optic disc segmentation using superpixel based feature classification
Expert Syst. Appl.
(2019) Automatic liver segmentation using an adversarial image-to-image network
International Conference on Medical Image Computing and Computer-Assisted Intervention
(2017)A review of the role of ultrasound biomicroscopy in glaucoma associated with rare diseases of the anterior segment
Clin. Ophthalmol.
(2016)Oct for glaucoma diagnosis, screening and detection of glaucoma progression
Br. J. Ophthalmol.
(2014)Heidelberg retina tomograph measurements of the optic disc and parapapillary retina for detecting glaucoma analyzed by machine learning classifiers.
Invest. Ophthalmol. Vis. Sci.
(2004)
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
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 ControlTowards an extended EfficientNet-based U-Net framework for joint optic disc and cup segmentation in the fundus image
2023, Biomedical Signal Processing and ControlPY-Net: Rethinking segmentation frameworks with dense pyramidal operations for optic disc and cup segmentation from retinal fundus images
2023, Biomedical Signal Processing and ControlPresumptive discerning of the severity level of glaucoma through clinical fundus images using hybrid PolyNet
2023, Biomedical Signal Processing and ControlCitation 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.
Medical image segmentation using deep semantic-based methods: A review of techniques, applications and emerging trends
2023, Information FusionCitation 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.