Efficient and robust optic disc detection and fovea localization using region proposal network and cascaded network

https://doi.org/10.1016/j.bspc.2020.101939Get rights and content

Highlights

  • Three different datasets of large sample sizes were used for validation purpose.

  • Optic disc detection accuracies were 99.59%, 100% and 100% for the three datasets.

  • Fovea localization accuracies were 97.8%, 99.03% and 99.25% for the three datasets.

  • It took less than 5 s for the entire pipeline to process an image of size 2880 × 2136.

Abstract

The optic disc (OD) and fovea are two important anatomical landmarks of the retina. Localization of OD and fovea plays an important role in retinal image analysis. This paper proposes a novel, efficient and robust method for OD detection and fovea localization. The proposed method consists of two steps. First, we use region proposal network to generate multiple OD region proposals, and then OD’s bounding box is identified as the proposal with the highest probability. Second, we create a region of interest containing the fovea based on the geometrical relationship between OD and fovea, and then employ a three-level cascaded convolutional neural network to locate the fovea, which is treated as a regression problem. The proposed method is trained and evaluated on three sets of fundus images (the first dataset contains 1955 fundus images and the other two are publicly available datasets respectively containing 516 and 1200 fundus images). According to 5-fold cross-validation experiments, our OD detection accuracy is respective 99.59%, 100%, and 100% for the three datasets and the corresponding fovea localization accuracy is respective 97.8%, 99.03%, and 99.25%. The proposed method is found to be robust and effective in OD detection and fovea localization, and its accuracy is comparable, or even superior to representative state-of-the-art algorithms.

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)

  • A. Osareh et al.

    Comparison of colour spaces for optic disc localisation in retinal images

    (2002)
  • X. Zhu et al.

    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

    (2008)
  • S. Lu

    Accurate and efficient optic disc detection and segmentation by a circular transformation

    IEEE Trans. Med. Imaging

    (2011)
  • W. Zhou et al.

    Detection of macula fovea in a retinal image

    J. Image Graph.

    (2018)
  • B. AI-Bander et al.

    Multiscale sequential convolutional neural networks for simultaneous detection of fovea and optic disc

    Biomed. Signal Process. Control

    (2018)
  • Cited by (29)

    • ECLNet: Center localization of eye structures based on Adaptive Gaussian Ellipse Heatmap

      2023, Computers in Biology and Medicine
      Citation 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 Medicine
      Citation 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 Applications
      Citation 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.

    View all citing articles on Scopus
    1

    These authors contributed equally.

    View full text