Sage Journals: Discover world-class research

Abstract

Retinal image sharpness assessment is one of the critical requirements of automatic quality evaluation in telemedicine screening for diabetic retinopathy. In this paper, a new sharpness metric measuring the spread of edges is presented to quantify fundus image clarity. After edge detection on the region of interest of retinal image, the width of each edge is calculated and the histogram of region of interest generated. Based on the histogram, a distance-based factor is introduced to gain the weighted edge width, which is defined as the sharpness metric for the fundus image. The method was tested on Messidor dataset and a proprietary dataset. The results show that the proposed metric performs well over different image distortion levels and resolutions and is of low computational complexity. The weighted edge width value of gradable retinal image, which is irrelevant to resolution, is always within the range of 3–7 pixels.

Keywords

Diabetic retinopathy screening fundus image sharpness weighted edge width histogram

Introduction

According to the latest report from the International Diabetes Federation, there are about 415 million adults with diabetes worldwide.¹ Over half of these individuals will develop diabetic retinopathy (DR), which is one of the most common, serious complications of diabetes.² During the early phases of disease, there may be no obvious obstacle on the vision, so patients always realize too late to get timely examination or treatment for preventing blindness. As the number of patients with diabetes increasing, DR has become a leading cause of blindness in adults globally.¹

Studies have showed that arteries, veins, and capillaries of human body can be directly observed through fundus noninvasively.³ Hence, taking DR screening regularly is necessary to achieve early diagnosis and get the effective curing time.⁴ Digital fundus photography is used, which enables a noninvasive examination and allows image preservation and transmission for later use.⁵

However, due to the imbalance of the professional medical resources, the patients’ compliance of the screening is not satisfied. For example, no more than 50% patients with diabetes participate in screening in time in the developed countries,⁶ while as high as 82.71% of patients in China lack awareness about regular fundus examination.⁶ In order to provide quality care to patients at any location, many domestic and foreign experts are committed to telemedicine programs addressing DR.^7,8 The available achievements have proved that telemedicine is more effective to identify and determine DR severity than conventional clinic-based ophthalmoscopy does and will help preventing severe vision loss.⁷

Unfortunately, it is a common phenomenon that images with insufficient quality are collected. Jian-jun et al.⁹ pointed out that in the case of nonmydriatic retinal images, 3–30% were unsuitable for retinopathy grading. In other studies, the proportion of single field nonmydriatic (no pupil dilation) images with unsatisfied quality has been reported at 11¹⁰ and 20.8%,¹¹ respectively. In telemedicine, receiving an ungradable fundus image may lead to escaping diagnosis or being misdiagnosed. It is too much trouble for patients to return to the clinic for retaking images.¹² Besides, every examination and image is critically priceless record of that individual in longitudinal studies. Losing information for a given period may reduce the statistical power of the results. Hence, assessing fundus image quality automatically becomes the first key step before reliable lesion detection, especially in the telemedicine systems of delivering eye care.¹⁰

According to Fleming et al.,¹¹ retinal image quality should include two aspects: clarity and field definition. In this paper, we focus on the former problem. Lesions may be disguised in blurred images, therefore a diseased eye could not be timely point out. Defocus blurring of fundus image may be caused by poor camera focus, cataract, or macular edema. And poor patient fixation or saccadic eye motion during acquisition may lead to motion blurring. Some retinal images of different qualities are shown in Figure 1.

Figure 1.

(a) Qualified fundus image, (b) motion-blurred fundus image, and (c) defocus-blurred fundus image.

The clarity of the image can be seen as a function of edge strength.¹³ Strength of edges in blurred or low-contrast images is always weak. Edge information can be used to evaluate fundus image sharpness. Thus, in this paper, we develop a sharpness metric for automatic quality evaluation of retinal image based on the edge width.

The rest parts of paper are structured as follows: In the next section, an overview of previous work on sharpness evaluation for fundus images is presented. “Proposed sharpness metric for fundus images” section describes the proposed sharpness metric based on histogram of edge width. Experiment results are shown in the subsequent section, and conclusion of our work comes in the final section.

Previous work on fundus image sharpness assessment

The automatic assessment of retinal image sharpness has gained more and more attention.¹⁰ There exist two groups of methods for solving the task¹²: (i) approaches based on structure segmentation and (ii) sharpness metric.

In methods falling in category (i), image clarity evaluation requires the identification of anatomical landmarks, such as detecting vessels and calculating their total length or destiny. In 2003, the first method for clarity assessment based on eye structure criteria was proposed by Usher et al.¹⁴ The area of segmented vessels over the entire image was used. Then, like Lowell et al.,¹⁵ Fleming et al.¹¹ and Hunter et al.¹⁶ judged image sharpness based on the structure of the blood vessels just within an area around the macula. More recently, based on the segmented vessel map, Yin et al.¹⁷ put forward high level image quality measures including blood vessel density and the maximum edge length or edge spread, while in Welikala et al.,¹⁸ authors calculated area, fragmentation, and complexity as the global features.

In those cases where segmentation can be reliably done, the eye structure markers will be very helpful for the later use in DR screening. However, it is always complex and computational to obtain the major anatomical structures. Especially, the failure of the segmentation is not trivial when images are of poor quality.¹² Therefore, these methods are not so suitable for telemedicine programs addressing DR.

On the other hand, (ii) sharpness metrics can provide a continuous measure in an unsupervised manner for the clarity prediction. In 1999, edge intensity histograms were first utilized by Lee and Wang¹⁹ to describe image sharpness. But as Lalonde et al.²⁰ states, “the link between intensity histogram similarity and image quality is not strong,” the results did not satisfied. In 2009, Bartling et al.²¹ calculated the degree of sharpness from the spatial frequencies of the image. In the following time, Davis et al.²² provided a real-time retinal image quality assessment procedure, where reblurred methods were introduced as a no-reference image sharpness metric. The same metric was employed in study of Pires Dias et al.¹² and of Yin et al.¹⁷

Commonly, blurring will lead to edge diffusion, so evaluation of retinal image sharpness can be made by the spread of the edge. Average edge width has been used to measure image blurriness by Marziliano et al.²³ Based on it, the cumulative probability of blur detection (CPBD),²⁴ a new local sharpness metric came. Then, CPBD value was calculated on only the green channel to characterize the strength of edges for retinal images in Yu et al.²⁵

Using sharpness metrics for fundus image quality assessment is simplicity.¹² Reduced computational complexity makes them amenable to use in DR screening. However, metrics mentioned above may be too sensitive to determine whether the retinal image is gradable in DR screening. Retinal images with slight blur are always acceptable for manual grade. For example, there must be a large range between CPBD values of blurred versions of a retinal image and the value of the normal one. But all of them may be acceptable during screening. Therefore, those sharpness metrics are not so suitable for sharpness assessment in DR screening.

As has been said, for more practical, we will pose a fundus image sharpness metric by getting weighted edge width (WEW) based on the histogram.

Proposed sharpness metric for fundus images

For evaluating blur of fundus images, we attempt to measure the spread of the edges. The computation of the proposed sharpness metric is summarized in Figure 2.

Figure 2.

Flow chart of the proposed sharpness metric.

Preprocessing

In retinal images, major edges arise from vasculature, optic disk, and lesions.¹³ According to statistics, vasculature and red lesions are always invisible in red planes of retinal photographs while there are little useful information but noise can be obtained from the blue planes.²⁶ Since each part of fundus and types of lesions can be distinguished in green planes well, only the green plane is used in our study.

Like images in Figure 1, the field including all fundus information we concerned is a circle in the fundus image. The black area beyond the circle does not contain any information. Besides, boundary of the field of view will have an adverse effect upon the following analysis. Hence, the inscribed quadrilateral of the fundus field containing major information is select as the region of interest (ROI). The next steps are conducted on the ROI of retinal image.

Edge width histogram computation

Firstly, Sobel operator is applied to achieve both horizontal and vertical edge detection. By applying a threshold to the gradient image, noises are removed.

Based on the edge map, width of each edge location is calculated. According to Lowell et al.,¹⁵ the number of pixels with increasing grayscale values in one direction as well as the number of pixels with decreasing values in the other side should be counted. The width around the edge is obtained by computing the sum of both counts except the edge point itself.

Figure 3 showing a row of the image illustrates an example of width measurement. For a detected vertical edge pixel P, local maximum P_s is defined as the start position, while local minimum P_e corresponds to the end location. Edge width of P is |P_e − P_s | = 5. When it comes to width calculation of horizontal edges, the process can easily be extended by scanning each column.

Figure 3.

One row of a fundus image.

Edges may have different widths. The probability $p (w_{i})$ of edges with width w_i can be calculated as

p (w_{i}) = \frac{n_{i}}{n}

(1)

where n_i is the total number of edges with width w_i, and n represents numbers of all edges. Histogram of edge width for ROI is then established.

Figure 4 shows an example of the edge width histogram computation of ROI.

Figure 4.

Edge width histogram computation. (a) Retinal image green plane, (b) ROI, (c) edge detection and (d) edge width histogram. ROI: region of interest.

Sharpness metric WEW

The average of the edge widths $\bar{w}$ for ROI is obtained by equation (2)

\bar{w} = \sum_{1}^{w_{max}} p (w_{i}) w_{i}

(2)

where

w_{max}

representing the max width of edge.

Histogram of edge width with different blurriness is shown in Figure 5. It demonstrates that, with the increasing blur, the histogram becomes diffusional, and the peak moves right.

Figure 5.

Histograms of edge width for retinal images with different blurriness. (a) Original image, (b) image filtered by a circular averaging filter with radius = 3, (c) image filtered by a circular averaging filter with radius = 9, (d) image filtered by a circular averaging filter with radius = 15, (e)–(h) are edge width histograms of (a)–(d).

Considering that blurring of invalid edges or noise may lead to too large or too small edge width, a distance-based weight is introduced in order to modify equation (2). The weight $d (w_{i})$ is defined as equation (3)

d (w_{i}) = {(\frac{w_{i}}{w_{p}})^{2}, w_{i} < w_{p} 1, w_{i} = w_{p} (\frac{w_{m} - w_{i}}{w_{m} - w_{p}})^{2}, w_{i} > w_{p}

(3)

where w_p is the location that the peak of histogram appears.

Therefore, weights from equation (3) are introduced into equation (2) resulting in a WEW, defined as our sharpness metric

WEW = \sum_{w_{i} = 1}^{w_{max}} d (w_{i}) p (w_{i}) w_{i}

(4)

The sharper the fundus image is, the lower the metric value takes.

Experiments and results

In this section, the performance of proposed sharpness metric for fundus images is presented and discussed. A MATLAB algorithm for WEW calculation was implemented on a laptop with 2 GHZ Intel Core I7 and 4 GB of RAM. To evaluate the proposed method, both motion- and defocus-blurred retinal images were used. The test sets taken from publicly available databases and proprietary dataset are first described. Then, experimental results are shown and analyzed.

Test sets

In order to develop and test the retinal image sharpness metric, publicly available retinal image dataset Messidor²⁷ was used. It takes 1200 gradable retinal images, all of which have a 45° FOV and resolutions of 2304 by 1536 pixels, 2240 by 1488 pixels, or 1444 by 960 pixels.

Besides, a proprietary dataset consisting of 30 retinal images from the first affiliated Hospital of Fujian Medical University were also adopted. Images from the dataset have a resolution of 3072 by 2048 pixels. With the help of three ophthalmologists from the first affiliated Hospital of Fujian Medical University, we defined 26 retinal images as satisfied images while the other four are unqualified.

Experiments on Messidor database

We chose some fundus images of size $2240 \times 1488$ from Messidor dataset. For each image, blur was induced in two ways:

The image was filtered with a motion filter with five different lengths $len \in {5, 10, 15, 20, 25}$ in various directions, yielding another five motion-blurred images for test.

The image was filtered with a circular averaging filter with different radius $r \in {3, 6, 9, 12, 15, 18}$ , yielding another six defocus-blurred images for test.

Sharpness metric was calculated for each test image. It is found that a strong linear relationship is consistent for all the test images. Figure 6 shows the behavior of our sharpness metric $WEW$ across distortion levels of one retinal image. As the growth of blurring degree, sharpness metric value increases as expected.

Figure 6.

Behavior of sharpness metric W EW for a fundus image (a) sharpness measurement versus length of the motion filter and (b) sharpness measurement versus radius of the circle averaging filter. WEW: weighted edge width.

We also calculated CPBD metric on each test image. The CPBD values of test images used in Figure 6(a) are presented in Figure 7. For a more intuitive observation, CPBD is used instead of 1-CPBD in the graph.

Figure 7.

Behavior of CPBD metric of blurred versions with different lengths of the motion filter. CPBD: cumulative probability of blur detection.

It is obvious that CPBD value changes sharply when a retinal image is disturbed. However, all of these motion-blurred images are acceptable in DR screening. They are qualified enough for ophthalmologists to grade. Hence, $WEW$ metric seems to be more suitable for sharpness assessment in DR screening.

To corroborate our metrics, $WEW$ metric was calculated for all the gradable retinal images in Messidor.

Figure 8 illustrates the result of $WEW$ metric for several subsets with different resolutions in Messidor. We found that WEW metric values of most retinal images in Messidor dataset seem to be irrelevant to resolution of the image. And the metric for a gradable retinal image is in the range of 3–7 pixels.

Figure 8.

(a) WEW metric for the 2240 × 1488 fundus images in Base11 subset, (b) WEW metric for the 1440 × 960 fundus images in Base22 subset, and (c) WEW metric for the 2304 × 1536 fundus images in Base33 subset. WEW: weighted edge width.

Performance on proprietary dataset

Here we extended our experiments to the proprietary dataset. WEW metric was calculated for each fundus image which was marked by manual in advance. The results are shown in Figure 9. Metric values of most images are around 5 pixels. However, one blurred image takes low WEW metric value while the value of another acceptable image is beyond 7 pixels. The total AUC is 93.3% when images with value range in 3–7 were marked as “qualified.”

Figure 9.

WEW metric for the 3072 × 2048 fundus images from proprietary dataset. WEW: weighted edge width.

As shown in the Figures 8 and 9, there are still some gradable or blurred images with unexpected values. The main cause of error is mainly due to the problem of reliable determination of edge. When blur is increasing, the total number of edges extracted by Sobel operator goes down, which reduces the probability of edge with various widths. In that case, a serious blurred image with very low WEW value will be mistaken for the clear picture. With the help of an advanced edge operator, the performance will be improved.

Besides, some pictures should have been acceptable are defined useless. Low contrast, uneven illumination or artifacts affect the metric. A proper preprocessing would certainly make it more robust.

Although the method has several disadvantages as mentioned above, low complexity was one of our prime objectives in applying this metric.

Computation time comparison

Comparison between the computation time of CPBD and WEW is given. Retinal images were analyzed after the same preprocessing. The average running time for each resolution is listed in Table 1. It is obvious that the proposed WEW metric has a better real-time performance than CPBD metric does. The advantage becomes more evident as image resolution increases.

Table 1.

Comparison of average computation time in different scales.

Resolution	CPBD (s)	WEW (s)
1440 × 960	3.05	1.55
2240 × 1488	8.10	2.42
2304 × 1536	9.25	2.63
3072 × 2048	22.09	7.43

CPBD: cumulative probability of blur detection; WEW: weighted edge width.

Conclusion

In this paper, we develop a retinal image sharpness metric based on edge width histogram. The WEW metric is of low computational complexity and in good agreement with ophthalmologists’ sharpness assessment. The metric for a gradable retinal image is in the range of 3–7 pixels. According to the satisfied AUC and low complexity, it has the potential to be an efficient feature in quality evaluation for DR screening application.

Future research will focus on the measurement of field definition of the fundus image and the image quality evaluation process for DR screening.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Natural Science Foundation of China (No. 61671152, No. 60827002), by the Natural Science Foundation of Fujian Province, China (No.2016J0129), by the Educational Commission of Fujian Province of China (No. JAT160070).

References

Aguiree

Brown

Cho

et al.

IDF diabetes atlas: seventh edition, Brussesl: International Diabetes Federation, 2015.

Jiang

et al.

Increasing prevalence of type 2 diabetes in Chinese adults in shanghai. Diabetes Care 2012; 35: 1028–1030.

Goto

Katsuki

Ikui

et al.

Pathological studies on the intracerebral and retinal arteries in cerebrovascular and noncerebrovascular diseases. Stroke 1975; 6: 263–269.

Yen

Leong

. A sorting system for hierarchical grading of diabetic fundus images: a preliminary study. IEEE Trans Inf Technol Biomed 2008; 12: 118–130.

Xia

Wang

et al.

Clinical analysis of the application of non-mydriatic fundus photography for diabetic retinopathy screening. Int Eye Sci 2013; 13: 2267–2269.

Xun

Zou

. Community screening and prevention of diabetic retinopathy. Chin J Ophthalmol Otorhinolaryngol 2008; 8: 276–279.

Guariguata

Whiting

Hambleton

et al.

Global estimates of diabetes prevalence for 2013 and projections for 2035. Diabetes Res Clin Pract 2014; 103: 137–149.

Das

Raman

Ramasamy

et al.

Telemedicine in diabetic retinopathy: current status and future directions. Middle East Afr J Ophthalmol 2014; 22: 174.

Jian-jun

Xiao-yan

. Some issues of diagnosis based on ocular fundus image reading in teleophthalmology. Ophthalmol China 2014; 23: 217–220.

10.

Sim

Keane

Tufail

et al.

Automated retinal image analysis for diabetic retinopathy in telemedicine. Curr Diabetes Rep 2015; 15: 1–9.

11.

Fleming

Philip

Goatman

et al.

Automated assessment of diabetic retinal image quality based on clarity and field definition. Invest Ophthalmol Vis Sci 2006; 47: 1120–1125.

12.

Pires Dias

Miguel

Oliveira

OCM

et al.

Retinal image quality assessment using generic image quality indicators. Inform Fusion 2014; 19: 73–90.

13.

Imani

Pourreza

Banaee

. Retinal image quality assessment using shearlet transform, Berlin: Springer, 2015.

14.

Usher D, Himaga M, Dumskyj M, et al. Automated assessment of digital fundus image quality using detected vessel area. In: Proceedings of medical image understanding and analysis, Citeseer, Sheffield, 10–11 July 2003, pp. 81–84. Durham: The British Machine Vision Association and Society for Pattern Recognition.

15.

Lowell J, Hunter A, Habib M, et al. Automated quantification of fundus image quality. In: Proceedings of the 3rd European medical and biological engineering conference, Prague, 20–21 November 2005, p.1618. Berlin: Springer.

16.

Hunter A, Lowell JA, Habib M, et al. An automated retinal image quality grading algorithm. In: 2011 33th annual international conference of the IEEE Engineering in Medicine and Biology Society, (EMBC), Boston, 30 August–3 September 2011, pp.5955–5958. New York: IEEE Press.

17.

Yin F, Wong DWK, Yow AP, et al. Automatic retinal interest evaluation system (ARIES). In: 2014 36th annual international conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Chicago, 26–30 August 2014, pp.162–165. New York: IEEE Press.

18.

Welikala

Fraz

Foster

et al.

Automated retinal image quality assessment on the UK Biobank dataset for epidemiological studies. Comput Biol Med 2016; 71: 67–76.

19.

Lee SC and Wang Y. Automatic retinal image quality assessment and enhancement. In: Medical Imaging 1999, San Diego, 21–23 February 1999, pp.1581–1590. Bellingham: SPIE.

20.

Lalonde M, Gagnon L and Boucher MC. Automatic visual quality assessment in optical fundus images. In: Proceedings of vision interface, Ottawa, 7–9 June 2001, pp. 259–264. Toronto: CIPPRS.

21.

Bartling

Wanger

Martin

. Automated quality evaluation of digital fundus photographs. Acta Ophthalmol 2009; 87: 643–647.

22.

Davis H, Russell S, Barriga E, et al. Vision-based, real-time retinal image quality assessment. In: 22nd IEEE international symposium on computer-based medical systems, Albuquerque, 3–4 August 2009, pp.1–6. New York: IEEE Press.

23.

Marziliano

Dufaux

Winkler

et al.

Perceptual blur and ringing metrics: application to jpeg2000. Signal Process Image Commun 2004; 19: 163–172.

24.

Narvekar

Karam

. A no-reference image blur metric based on the cumulative probability of blur detection (CPBD). IEEE Trans Image Process 2011; 20: 2678–2683.

25.

Yu H, Agurto C, Barriga S, et al. Automated image quality evaluation of retinal fundus photographs in diabetic retinopathy screening. In: 2012 IEEE southwest symposium on image analysis and interpretation, Santa Fe, 22–24 April 2012, pp.125–128. New York: IEEE Press.

26.

Wang

Jin

et al.

Human visual system-based fundus image quality assessment of portable fundus camera photographs. IEEE Trans Med Imaging 2016; 35: 1046–1055.

27.

Decencière

Zhang

Cazuguel

et al.

Feedback on a publicly distributed image database: the Messidor database. Image Anal Stereol 2014; 33: 231–234.

A retinal image sharpness metric based on histogram of edge width

Abstract

Keywords

Introduction

Previous work on fundus image sharpness assessment

Proposed sharpness metric for fundus images

Preprocessing

Edge width histogram computation

Sharpness metric WEW

Experiments and results

Test sets

Experiments on Messidor database

Performance on proprietary dataset

Computation time comparison

Conclusion

Footnotes

Declaration of conflicting interests

Funding

References