A comprehensive diagnosis system for early signs and different diabetic retinopathy grades using fundus retinal images based on pathological changes detection

https://doi.org/10.1016/j.compbiomed.2020.104039Get rights and content

Highlights

  • A comprehensive multi-label computer-aided diagnosis system is proposed for differentiating normal from different DR cases using various benchmark datasets.

  • Four main pathological signs of DR are segmented, and six different significant features are extracted from the segmented pathological signs.

  • MLSVM classifier is proposed to detect different DR grades based on the problem transformation.

  • The performance is validated by using six performance metrics (ACC, DSC, AUC, SEN, SPE, PPV).

  • The experiments show promising results as compared with other state-of-the-art techniques.

Abstract

Multi-label classification (MLC) is deemed as an effective and dynamic research topic in the medical image analysis field. For ophthalmologists, MLC benefits can be utilized to detect early diabetic retinopathy (DR) signs, as well as its different grades. This paper proposes a comprehensive computer-aided diagnostic (CAD) system that exploits the MLC of DR grades using colored fundus photography. The proposed system detects and analyzes various retina pathological changes accompanying DR development. We extracted some significant features to differentiate healthy from DR cases as well as differentiate various DR grades. First, we preprocessed the retinal images to eliminate noise and enhance the image quality by using histogram equalization for brightness preservation based on dynamic stretching technique. Second, the images were segmented to extract four pathology variations, which are blood vessels, exudates, microaneurysms, and hemorrhages. Next, six various features were extracted using a gray level co-occurrence matrix, the four extracting areas, and blood-vessel bifurcation points. Finally, the features were supplied to a support vector machine (SVM) classifier to distinguish normal and different DR grades. To train and test the proposed system, we utilized four benchmark datasets (two of them are multi-label datasets) using six performance metrics. The proposed system achieved an average accuracy of 89.2%, sensitivity of 85.1%, specificity of 85.2%, positive predictive value of 92.8%, area under the curve of 85.2%, and Disc similarity coefficient (DSC) of 88.7%. The experiments show promising results as compared with other systems.

Introduction

Diabetic retinopathy (DR), considered a complication of diabetes, affects the retina. DR may not reveal any noticeable signs in the diabetes patients at the precocious stages and may lead to sudden blindness. The symptoms of the disease may appear as floaters or dark/empty areas in the patient's vision. The disease also causes blurring, fluctuations, impaired color, and reduction/loss in night vision [1]. The uncontrolled sugar elevation in the blood may lead to clogging in the retinal blood vessels (BV), thereby hampering their blood supply [2]. In such a case, the eye seeks to grow new BVs, but they are incapable of developing typically and leak out easily. The three early DR grades are categorized into mild, moderate, and severe non-proliferative DR (NPDR). The sophisticated grade of DR is called proliferative DR (PDR). A severe-grade NPDR is the starting of PDR grade. DR can be recognized by several changes in the retina, such as neovascularization (NV), vascular dilation or tortuosity, hemorrhages (HM), microaneurysms (MA), and exudates (EX) [3]. The last four signs refer to NPDR grades, which may all appear in the retina simultaneously, whereas the first change refers to severe NPDR or PDR [4]. Fig. 1 shows the DR grades in terms of EX, MA, HM, and NV, which increase the BV area in the retina.

Many imaging modalities can be used to detect and diagnose retinal disorders. Optical coherence tomography (OCT), colored fundus photography, OCT angiography (OCTA), and fluorescein angiography (FA) are examples of those modalities. Among them, colored fundus photography is the greatest widely utilized, especially in developing countries because of its low cost compared to the OCT machines. Unlike OCT-based diagnoses, fundus photography can capture the retinal BV network changes, which is effective in detecting many retinal diseases such as DR. OCT detects the alterations in the sub-retinal layers that can be useful in detecting the early signs of diseases, such as Macular Edema (ME) [5]. On the other hand, OCTA machines can capture the retinal vasculature changes, but the main problem of this imaging modality is the prohibitive cost [6].

Fundus cameras preserve the image quality in different situations, such as a degradation reduction in cataracts cases. Although fundus cameras are convenient for the patient because of the single flash exposition of the floodlight, the produced images have a minimal contrast [7,8]. For this issue, an ophthalmologist may not be able to precisely discover all the disease signs, which may lead to an inaccurate diagnosis of the exact disease grade. Moreover, physicians may need a long time to diagnose the patient's condition.

DR is a rapidly progressive disease. Thus, if it is not diagnosed and cured in the early stages, it can cause sudden blindness [9]. Early diagnosis of DR grades is crucial as it helps the ocular specialist to specify the appropriate treatment. Therefore, developing a precise, automated computer-aided diagnostic (CAD) system to detect DR grades is essential for the safety of the patient's eyesight. Multi-label classification (MLC) can be employed in the CAD system to differentiate the DR grades. The training set in an MLC dataset consists of instances with more than two labels. The main task of the MLC system is to predict the labels of the unclassified instances depending on the experience gained from training on instances with previously known label sets [[10], [11], [12]].

The main contributions of our proposed multi-label CAD (ML-CAD) system can be outlined in the following points:

  • We present a comprehensive framework for differentiating normal and DR cases as well as grading different DR cases from various datasets including imbalanced ML datasets.

  • The problem of non-symmetric lighting and stretching of the edge is resolved by eliminating the noise and enhancing the contrast.

  • We compared seven famous BV segmentation techniques and made a combination between two BV segmentation methods to extract the vessels network from various datasets accurately.

  • The proposed ML-CAD system accurately segments the four main pathological signs of DR: BV, EX, MA, and HM while other systems concentrated only on segmenting one or two signs.

  • The ML-CAD system extracts six different significant features from the segmented pathological features.

  • A multi-label (ML) support vector machine (MLSVM) classifier is utilized to diagnose the different DR grades based on the problem transformation to allow an extension if needed.

  • We validated the resulting results by using different performance metrics. Besides, we compared our proposed system with some state-of-the-art classifiers to validate the proposed system.

For reader convenience, the used abbreviations in this paper are listed in Table 1. The remainder of the manuscript is organized into six sections. Section 2 discusses the related work, current limitations, and how we overcame these limitations in the proposed system. Section 3 introduces the framework of the proposed CAD system in more detail with a brief explanation for each technique used. Section 5 describes the conducted experiments. A discussion on the experimental results is presented in Section 6. Finally, Section 7 concludes our work and presents a summary of our future work directions.

Section snippets

Related work

DR diagnosis and screening are considered as active research topics in medical image analysis. Many researchers have worked on detecting and diagnosing DR using retinal fundus images. For example, Fan et al. [13] segmented the OD from the retinal image by using CHT. They obsoleted the hypothesis of unsupervised techniques and relied on the edge structure features. Their aim was to derive benefits from the structured and deep learning in their future work of training the edge detector. Tan et

Proposed CAD system

This work is an extension of our research in [23]. The main objective of the developed CAD system is the detection of DR from colored fundus photography that contains multiple DR lesions and classifies the various disease grades. The DR grade detection assists the ophthalmologists in making the right treatment choices for the patients. The proposed CAD system utilizes six features for all the resulting segmenting preprocessed images (normal and DR signs). It selects the most correlated features

Materials

We divide this section into two subsections. The first subsection describes the utilized four benchmark datasets, which are DRIVE, STARE, IDRiD, and MESSIDOR. The second subsection discusses the specifications of the used hardware/software in our experiments.

Experimental results

This section is divided into three subsections, which are evaluation metrics, results, and discussion. In the results subsection, we present the results of the segmentation, classification, and the overall proposed CAD system on four different benchmark datasets. The segmentation results include detecting the BV, making comparisons among seven methods, and detecting the other three lesions (EX, MA, and HM). After that, we present the overall CAD system preprocessing and segmentation (BV, HM,

Discussions

The effectiveness of CAD systems for DR screening has been reported by several researchers. For example, fan et al. [13] allocated the OD from the MESSIDOR dataset and achieved 97.7% for ACC and 91.8% for SEN. This phase cannot help the ophthalmologists, but it could be one preprocessing step for the developer. The features of OD in most cases are similar to these of the EX sign. Therefore, detecting and removing OD can help in detecting EX accurately. But, it is not a DR detection method. Tan

Conclusion

We developed a new ML-CAD system that can be applied on different datasets to diagnose healthy and diabetic retinopathy grades. We used different four public datasets, DRIVE, STARE, IDRiD, and MESSIDOR. The proposed system first applies filtering and contrast enhancement. Then, the preprocessed images are segmented into vascular tree, exudates, microaneurysms, and hemorrhages. Next, the system extracts six features of BV using GLCM and bifurcation point's count in addition to the four ROI area

Acknowledgment

This work was validated and approved by Dr. Hatem Abdelkawy. He is an assistant professor at the ophthalmology department, Faculty of Medicine, Al-Azhar University, Egypt. We thank him for his great effort and time. Besides, we appreciate the time and efforts made by the editor while reviewing this paper. Finally, we thanks the reviewers for their valuable comments and recommendations to improve the quality of this work.

References (46)

  • T. Hassan et al.

    Review of Oct and Fundus Images for Detection of Macular Edema

  • A.C. Tan et al.

    An overview of the clinical applications of optical coherence tomography angiography

    Eye

    (2018)
  • M. Yung et al.

    Clinical applications of fundus autofluorescence in retinal disease

    Int j of retina and vitreous

    (2016)
  • R. Lee et al.

    Epidemiology of diabetic retinopathy, diabetic macular edema and related vision loss

    Eye and Vision

    (2015)
  • J.I. Orlando et al.

    A discriminatively trained fully connected conditional random field model for blood vessel segmentation in fundus images

    IEEE Trans. Biomed. Eng.

    (2016)
  • Z. Fan et al.

    Optic disk detection in fundus image based on structured learning

    IEEE J of biomed and health info

    (2017)
  • S. Morales et al.

    Retinal disease screening through local binary patterns

    IEEE j. biomed and health info

    (2015)
  • N. Gharaibeh et al.

    An effective image processing method for detection of diabetic retinopathy diseases from retinal fundus images

    Int. J. Signal Imag. Syst. Eng.

    (2018)
  • C. Agurto et al.

    Multiscale am-fm methods for diabetic retinopathy lesion detection

    IEEE Trans. Med. Imag.

    (2010)
  • Ramakrishnan Sundaram et al.

    Extraction of blood vessels in fundus images of retina through hybrid segmentation approach

    Mathesis

    (2019)
  • A. Imran et al.

    Comparative analysis of vessel segmentation techniques in retinal images

    IEEE Access

    (2019)
  • V. Gulshan et al.

    Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs

    Jama

    (2016)
  • Cited by (0)

    View full text