Efficient and robust optic disc detection and fovea localization using region proposal network and cascaded network
Introduction
Computer based automatic or semi-automatic retinal image analysis has attracted huge research attention in the past decade, for the reason that retinal images have a lot of manifestations of various eye diseases, such as diabetic retinopathy (DR), glaucoma, age-related macular degeneration and diabetic macula edema [1,2]. Fundus imaging is considered as the most established way of retinal imaging and fundus image analysis provides important quantitative indices of retinal morphology [3]. In a fundus image, there are several important anatomical landmarks including the blood vessels, the optic disc (OD), and the fovea (the center of the macula). The OD, being the most important structure in the retina, is the region where the vasculature originates and gathers or converges. On a fundus image, the OD appears as a bright, yellowish circle or ellipse and its appearance is in high contrast to its surrounding tissue. The OD morphology is an important anatomical feature for various retinal pathology analyses.
The OD is a reference landmark of another important retinal landmark, namely the fovea, which is the center of the macula. The macula, being a relatively darker region compared to its surrounding tissue, lies on the temporal side of the OD and is about 2.5 OD diameters (ODDs) far away from the OD center, slightly below [4]. The center of the macula, namely the fovea, is the most sensitive area of vision (being responsible for sharp central vision). Any lesions occurring near the fovea can result in vision damages or even blindness. In addition, the distance between a lesion and the fovea is an important criterion for judging and grading the retinopathy degree. As such, accurate and automatic detection and localization of OD and fovea is of great importance. It is usually a prerequisite for computer-based fundus image analysis and has a great significance in computer-aided diagnosis.
In this paper, we propose and validate a novel, efficient and robust method for OD detection and fovea localization, in the framework of deep neural networks. The proposed method is fully-automatic, with no need to manually define features. A large number of fundus images, obtained from three different datasets, are used for training and testing. Quantitative and qualitative evaluations are performed and comparisons with other representative methods are made.
The rest of this paper is organized as follows. In Section 2, relevant literature of this work is briefly reviewed. The proposed method as well as the datasets used in this paper are described in Section 3. Section 4 describes the experiments and results. Finally, in Section 5 discussion and conclusion are presented.
Section snippets
Related work
Various techniques have been employed to automatically or semi-automatically detect and locate the OD and fovea. For OD detection, existing works have mainly utilized features such as intensity, texture, contrast and shape to locate its position. For instance, some work [5] assumed that the OD is the brightest and largest pixel aggregation region. They calculated the variance of each candidate sub-window, and the one with the maximum variance is considered as the bounding box of OD. Osareh et
Materials and methods
The flowchart of the proposed pipeline is shown in Fig. 1, consisting of several stages. First, data preprocessing and augmentation are conducted. And then, OD detection is performed using the region proposal network [14], which is also a prerequisite for subsequent fovea localization. After that, a three-level cascaded neural network is adopted to achieve coarse-to-fine fovea localization. Note that the OD center can be roughly located as the center of the rectangular frame bounding the OD,
Experiments and results
In this work, all experiments were conducted on a server with a Gerforce GTX TITAN X 12GB GPU. The evaluations were performed on a local computer with a Gerforce GTX 750 Ti 2GB GPU. All neural network models were implemented in the Tensorflow framework. To quantify the algorithm performance, we used different evaluation criteria. For OD detection, we visually assessed the accuracy of OD’s bounding box. Specifically, we visually checked whether the predicted rectangular box is tangent to the OD.
Discussion
This paper presents a novel method for OD detection and fovea localization by combining deep learning and anatomical information. Promising results have been obtained. For OD detection, we did not directly estimate the coordinates of OD’s center, but adopted RPN to detect the bounding box of the entire OD. For fovea localization, we used an ROI-based three-level cascaded neural network, which not only reduced the interference of lesions but also achieved a coarse-to-fine localization. Compared
CRediT authorship contribution statement
Yijin Huang: Methodology, Investigation, Writing - original draft. Zhiquan Zhong: Methodology, Data curation, Formal analysis, Writing - original draft. Jin Yuan: Resources, Data curation. Xiaoying Tang: Conceptualization, Supervision, Writing - review & editing.
Acknowledgements
This study was supported by the National Key R&D Program of China (2017YFC0112404) and the National Natural Science Foundation of China (NSFC 81501546). We would like to thank Junyan Lyu from Southern University of Science and Technology (Shenzhen, Guangdong, China) for his valuable comments on the manuscript. We would also like to acknowledge Honghui Xia from Zhaoqing Gaoyao People’s Hospital (Zhaoqing, Guangdong, China) for sharing the local dataset with us.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (30)
- et al.
Proposed international clinical diabetic retinopathy and diabetic macular edema disease severity scales
Ophthalmology
(2003) - et al.
Macular edema. A complication of diabetic retinopathy
Surv. Ophthalmol.
(1984) - et al.
The use of radial symmetry to localize retinal landmarks
Comput. Med. Imaging Graph.
(2013) - et al.
Fast detection of the optic disc and fovea in color fundus photographs
Med. Image Anal.
(2009) - et al.
Fast and effective optic disc localization based on convolutional neural network
Neurocomputing
(2018) - et al.
Obtaining optic disc center and pixel region by automatic thresholding methods on morphologically processed fundus images
Comput. Methods Programs Biomed.
(2015) - et al.
Optic disc localization algorithm based on improved corner detection
Procedia Comput. Sci.
(2018) - et al.
Retinal imaging and image analysis
IEEE Rev. Biomed. Eng.
(2010) - et al.
A novel method of macula fovea and optic disc automatic detection for retinal images
J. Electron. Inf. Technol.
(2014) - et al.
Automated localisation of the optic disc, fovea, and retinal blood vessels from digital colour fundus images
Br. J. Ophthalmol.
(1999)
Comparison of colour spaces for optic disc localisation in retinal images
Detection of the optic disc in images of the retinal using the hough transform
2008 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society
Accurate and efficient optic disc detection and segmentation by a circular transformation
IEEE Trans. Med. Imaging
Detection of macula fovea in a retinal image
J. Image Graph.
Multiscale sequential convolutional neural networks for simultaneous detection of fovea and optic disc
Biomed. Signal Process. Control
Cited by (29)
A comprehensive computer-aided system for an early-stage diagnosis and classification of diabetic macular edema
2023, Journal of King Saud University - Computer and Information SciencesECLNet: Center localization of eye structures based on Adaptive Gaussian Ellipse Heatmap
2023, Computers in Biology and MedicineCitation Excerpt :Al-Bander et al. [19] regressed the location of the OD and the fovea through a two-stage convolutional neural network (CNN), which first detected and cropped out the regions of interest (ROIs) of the OD and fovea, and then refined the location of the center on the cropped ROIs image. Huang et al. [20] generated the bounding box of the OD using a region proposal network, and then obtained the location of the fovea based on the geometric relationship between OD and fovea. Babu et al. [21] used a relational network to extract the bounding boxes of OD and macular region and regressed the coordinates of the optic disc and fovea center based on the predicted bounding boxes by a two-layer CNN.
Optic disc detection and segmentation using saliency mask in retinal fundus images
2022, Computers in Biology and MedicineCitation Excerpt :Finally, they were decomposed the matrix of features was decomposed to become a sparse matrix using the combination of the structured matrix decomposition model and the fully convolution neural network. A region proposal network was used by Huang et al. [69] in order to produce multiple regions proposal of the optic disc, and the region of interest with the highest probability was selected as the OD region. This method was detected the bounding box of the optic disc without the need of a priori knowledge on the optic disc property.
A comprehensive and systematic review on classical and deep learning based region proposal algorithms
2022, Expert Systems with ApplicationsCitation Excerpt :The OD detection based on region proposal has attracted considerable research attention. An efficient method based on region proposal and deep learning is proposed to detect OD and fovea localization (Huang, Zhong, Yuan, & Tang, 2020). After data augmentation and preprocessing, OD regions are detected using region proposal network.
Fovea localization in retinal images using spatial color histograms
2024, Multimedia Tools and ApplicationsChallenges for ocular disease identification in the era of artificial intelligence
2023, Neural Computing and Applications
- 1
These authors contributed equally.