Brain-inspired computer vision with applications to pattern recognition and computer-aided
diagnosis of glaucoma
Guo, Jiapan
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date: 2017
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
Guo, J. (2017). Brain-inspired computer vision with applications to pattern recognition and computer-aided diagnosis of glaucoma. University of Groningen.
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum.
Based on:
Jiapan Guo, Chenyu Shi, George Azzopardi, Nomdo M. Jansonius and Nicolai Petkov, ”Automatic analysis of retinal fundus images for glaucoma screening based on vertical cup-to-disc ratio”, submitted1.
Chapter 4
Automatic analysis of retinal fundus images
for glaucoma screening based on vertical
cup-to-disc ratio
Abstract
Glaucoma is a chronic progressive optic neuropathy that causes visual impairment or blindness, if left untreated. It is crucial to diagnose it at an early stage in order to enable timely treatment. Fundus photography is a viable option for population-based screening. A fundus photograph enables the observation of the excavation of optic disc - the hallmark of glaucoma. The excavation is quantified as vertical cup-to-disc ratio (VCDR). The manual assessment of retinal fundus images is, however, time-consuming and costly. Thus, an automated system is necessary to assist human observers. We propose a computer aided diagnosis system, which consists of localization and boundary delineation of the optic disc, segmentation of the cup, and computation of the VCDR. We also provide a reliability score that indicates the confidence of the system’s VCDR estimation. We evaluated the performance of our approach on eight publicly available data sets which have in total 1712 retinal fundus images. Our method indicated the obtained VCDR values as reliable in 1558 images; i.e. 91% of the cases. We compared the obtained VCDR values with those provided by an experienced ophthalmologist and achieve a weighted VCDR mean difference of 0.17. The system provided very reliable delineation of the optic disc (MCC=0.90), from which we obtain the height of the optic disc. The segmentation of the cup, and thus the measurement of its height, turned out to be the most problematic part of the system (MCC=0.47). Bland-Altman analysis showed that the system achieves better agreement with respect to the manual annotations for large VCDRs, which indicate pathology. Although it is not typical to rely on the classification of image alone to infer whether a case is glaucomatous or not, we computed the classification performance to obtain an indication. For a VCDR threshold of 0.7, we achieved an AUC of 0.74. The proposed system could be deployed as part of a
population-1The work presented in this paper is split into two parts. The first part is on optic disc detection
which is included in Section 3.3 in Chapter 3 while the second part is about cup segmentation and VCDR computation which are included in this chapter.
4.1. Introduction 77
4.1
Introduction
Glaucoma is a chronically progressive neuropathy that affects irreversibly the optic nerve, the neural fiber bundle that relays visual information from the eye to the brain. The worldwide number of people (aged 40-80 years) affected by glaucoma was estimated to be 64 million in the year 2013. This number is expected to increase to 76 million by 2020 and to 120 million by 2040 (Quigley and Broman, 2006; Tham et al., 2014). Glaucoma affects 1 - 2% of the population and is now the second leading cause of blindness (Quigley and Broman, 2006).
As glaucoma is initially asymptomatic and the damage is irreversible, it is vi-tal to diagnose it as early as possible in order to halt or slow down progression by adequate treatment - thus avoiding visual impairment or blindness. The diagnosis and treatment of glaucoma requires specialized physicians and sophisticated pro-cedures, such as tonometry (assessment of intraocular pressure), ophthalmoscopy (assessment of the optic nerve), and perimetry (assessment of visual function). The initial detection of glaucoma, however, does not necessarily require all of these mea-surements. Population-based glaucoma screening commonly relies on assessing the state of the optic nerve head (ONH). In addition to ophthalmoscopy by a skilled professional, fundus photography and optical coherence tomography (OCT) can be used for this purpose. With these techniques, the changes of the ONH and the retinal nerve fibre layer can be observed (Lin et al., 2007). Each technique has advantages and disadvantages. In this study we focus on fundus photography.
Fig. 4.1 shows a schematic diagram that illustrates the anatomy of the human eye. The visual pathways start with the photoreceptors in the retina, which trans-duce light into neural signals. The photoreceptors relay information to bipolar cells and these cells are connected to the retinal ganglion cells. The axons of the retinal ganglion cells leave the eye through the ONH. The ONH is also the entry point of the major blood vessels. Fig. 4.2a shows a retinal fundus image taken with a fundus camera. The optic disc is the region surrounded by the dashed white boundary and is the two-dimensional view of the ONH. The optic disc appears as a bright reddish area which usually has a vertically elliptic shape (Jonas et al., 1999). Usually, three areas can be distinguished within the optic disc: a neuroretinal rim, a cup, and blood vessels. The cup is the pale, oval region in the middle of the optic disc, marked with the black dashed boundary. It is paler than the surrounding rim because - in contrast to the rim - it is void of optic nerve fibres. It is usually slightly decentered towards the fovea. The size of the cup relative to the size of the optic disc gives an indication of the state of the optic nerve. The vertical cup-to-disc ratio (VCDR), defined as the ratio between the height of the cup and that of the optic disc, is a commonly used measure to assess the state of the optic nerve and the risk of glaucoma (Weinreb and
Pupil Iris Lens Cornea Retina Optic nerve head Optic nerve Retinal blood vessels
Figure 4.1: A schematic diagram of the human eye2.
Greve, 2004).
Outside the optic disc sometimes there is a pale ring named scleral ring, which is indicated by the white arrow in Fig. 4.3. Then outside the scleral ring there is a yellow-gray region called parapapillary atrophy (PPA), as shown in Fig. 4.3(a-b). PPA is a glaucoma-related pathology that is due to the thinning of the layers of the retina and the retinal pigment epithelium around the optic disc. PPA and the scleral ring frustrate the correct determination of the optic disc border, making the accurate estimate of the VCDR difficult. For instance, the VCDR value of the retinal fundus image in Fig. 4.3(a-b) is 0.56. It can be underestimated if the observer erroneously determines the boundary of the optic disc as that of the PPA.
The manual analysis of retinal fundus images in glaucoma population screen-ing would be a highly tedious procedure for medical experts, because of the large number of images of which only about five percent contain signs of glaucoma3. A
computer-aided diagnosis system for glaucoma screening that measures the VCDR can be used to speed up the analysis of retinal images. Our decision to compute
2This figure is taken from: http://tinyurl.com/mtkglzh.
3It is recommended for people above 40 years old to have regular eye examination. Among Caucasian
people above this age the occurrence of glaucoma is approximately 2%, and it increases rapidly with age (Wolfs et al., 2000). The occurrence is even higher in case of African descent.
4.1. Introduction 79
(a)
(b)
Figure 4.2: Example of a retinal fundus image and the optic disc and the cup. (a) A retinal fundus image (of size 564×584 pixels) captured at a field of view of 45◦from a right eye. This image is taken from the DRIVE data set (Staal et al., 2004). The white and black boundaries indicate the optic disc and the cup, respectively. The ring-shaped region between the black and the white boundaries is called neuroretinal rim. The VCDR value is 0.45. (b) A close-up view of the optic disc region.
the VCDR of retinal images as opposed to other methods that rely on image clas-sification such as convolutional neural networks (Chen et al., 2015) is motivated by two main reasons. First, the VCDR is part of the current definition of glaucoma for epidemiological studies (Foster et al., 2002) and still the preferred measure of glau-comatous optic neuropathy in longitudinal studies (Springelkamp, Iglesias, Mishra, H ¨ohn, Wojciechowski, Khawaja, Nag, Wang, Wang and MacGregor, 2017). Second, the VCDR (and optic disc area) plays an important role in the unraveling of the genetics of glaucoma (Springelkamp et al., 2014), and, especially in these so called genome-wide association studies, a very large number (1,000-100,000) of fundus im-ages have to be assessed. With the proposed work we contribute to the development of a feasible computer assisted screening program that would allow ophthalmolo-gists to focus on the images that are flagged as suspicious by the system rather than having to manually process all images.
The automatic detection of glaucoma has attracted the interest of many re-searchers. Most of the studies address parts of the problem. For instance, some studies focus on the localization of the optic disc (Sinthanayothin et al., 1999; Walter
(a)
(b)
Fovea
Scleral ring
Figure 4.3: Example of a retina with pathology. (a) A retinal fundus image with pathologies. This image is taken from the DRIVE data set (Staal et al., 2004). The region between the white and the outside gray dashed lines indicates the parapapillary atrophy. The VCDR value is 0.56. (b) A close-up view of the optic disc region.
and Klein, 2001; Li and Chutatape, 2001; Chr´astek et al., 2002; Hoover and Gold-baum, 2003; Foracchia et al., 2004; Mendonc¸a et al., 2013), and some also attempt to delineate the boundary of the optic disc (Lalonde et al., 2001; Osareh et al., 2002). Others focus on the segmentation of the cup (Wong, Liu, Tan, Fengshou, Cheung, Baskaran, Aung and Wong, 2012; Wong, Liu, Tan, Yin, Lee, Tham, lui Cheung and Wong, 2012) or some other features for the detection of glaucoma (Bock et al., 2010; Akram et al., 2015).
For an overview of the literatures on optic disc detection, we refer the interested reader to the introduction section in Chapter 3.
The studies by Wong, Liu, Tan, Fengshou, Cheung, Baskaran, Aung and Wong (2012); Wong, Liu, Tan, Yin, Lee, Tham, lui Cheung and Wong (2012) and Bock et al. (2010) focused on the cup segmentation and automatic glaucoma detection. Wong, Liu, Tan, Fengshou, Cheung, Baskaran, Aung and Wong (2012) proposed a spatial heuristic ensembling approach which fuses different methods to segment the cup. Later, in another work, Wong, Liu, Tan, Yin, Lee, Tham, lui Cheung and Wong (2012) proposed a method to identify the boundary of the cup by determining the vessel kinking points, which indicate the cup excavation. Bock et al. (2010) proposed an automatic system to estimate the glaucoma risk index which indicates the
probabil-4.2. Proposed method 81
ity of a retina being glaucomatous. Their system extracts features, such as FFT and B-spline coefficients, from the spatial and the frequency domains followed by a sup-port vector machine classification. Most of these methods were, however, tested on proprietary data sets or public data sets with ground truth data that is not publicly available.
Some other approaches (Abr`amoff et al., 2007; Xu et al., 2008) detect the optic nerve head in stereo retinal fundus images which provide depth information. Sys-tems that rely on sophisticated equipments are, however, not considered suitable for population screening as they are too time consuming and require expensive re-sources.
In this work, we propose a computer-aided diagnosis system that uses novel computer vision algorithms to estimate the VCDR in retinal fundus images. Our approach is divided into four main steps. First, our system localizes the optic disc and approximates its boundary with an ellipse. Then, it segments the optic disc re-gion into cup and neuroretinal rim. Finally, we compute the VCDR and provide a reliability indicator. We evaluate the proposed approach on eight public data sets and compare the obtained results with the manual annotation provided by a glau-coma expert from the University Medical Center Groningen.
The rest of the chapter is organized in the following way. In Section 4.2, we present our proposed method that segments the optic disc and the cup and eval-uates the VCDR. In Section 4.3, we describe the data sets and the corresponding manual annotation followed by the experiments along with the experimental re-sults. Finally, we discuss some aspects of the proposed method in Section 4.4 and draw conclusions in Section 4.5.
4.2
Proposed method
4.2.1
Overview
We propose a novel approach for assisting population-based glaucoma screening and we provide annotation by an experienced ophthalmologist4 for eight public
data sets. In our approach, we first localize the optic disc by using two types of trainable COSFIRE (Combination of Shifted Filter Responses) filters (Azzopardi and Petkov, 2013b): one type configured to be selective for the divergent points of vessel trees and the other type configured to be selective for circular bright regions. We then fit an ellipse that approximates the boundary of the detected optic disc. Next, we eliminate the blood vessels inside the optic disc and apply K-means clustering
I
vI
dI
eI
cReliability
scor
e
(a)
(b)
(c)
(d)
(e)
(f)
(g)
VCDR
Figur e 4.4: Schematic overview of the pr oposed appr oach. For a given (a) retinal fundus image we first apply (b) a set of vasculatur e-selective COSFIRE filters to detect the diver gent point of the major vessels. Then, (c) we cr op a lar ge region ar ound the maximum response and apply (d) a set of disc-selective COSFIRE filters to detect bright disc patterns. (e) W e cr op a small region ar ound the maximum response of the disc-selective COSFIRE filters and fit an ellipse to appr oximate the boundary of the detected optic disc. The black boundary in (f) indicates the delineated disc boundary . W e then employ K -means clustering to segment the disc region into the neur or etinal rim and the cup. (g) The blue boundary inside the disc indicates the segmented cup. Finally , we compute the VCDR accor ding to the segmented cup and disc and pr ovide a reliability scor e.4.2. Proposed method 83
to segment the delineated optic disc into two regions: the cup and the neuroretinal rim. Finally, we compute the VCDR value and provide a reliability score of the measurement that is a function of the indicators provided in each step. Fig. 4.4 illustrates a schematic overview of the proposed procedure.
4.2.2
Localization and Delineation of the Optic Disc
For the localization and delineation of the optic disc boundary, we apply the COS-FIRE approach introduced in Section 3.3 in Chapter 3.
4.2.3
Segmentation of the Cup by K-means Clustering
Our experience shows that the edge-based method we used above for the delin-eation of the optic disc is not sufficiently robust for the segmentation of the cup. This is due to the fact that the boundaries of the cups are typically diffuse and they are often occluded by blood vessels. The approach that we use to determine the cup consists of two steps. In the first step, we apply edge-preserving smoothing (details are provided in 4.3.2) and detect the thickest blood vessel, which divides the delin-eated optic disc into two main regions. Of these two regions, we consider the one with the highest average intensity and remove all blood vessels within it, Fig. 4.5(b-d). In the second step, we use the hue, saturation and luminance (HSL) of the pixels within the selected region to cluster them by the K-means algorithm (K=2). We consider the resulting cluster of which the pixels have a smaller average distance to the center of the optic disc to be the cup, while the other cluster to represent the neuroretinal rim. The two centroids of the K-means algorithm are initialized as fol-lows. One centroid is the 3D HSL vector of the pixel with the highest luminance, and the other centroid is the 3D HSL vector of the pixel whose hue value is ranked in the 20th percentile5along the boundary of the selected region. The width of this
boundary is a fraction 0.025 of the maximum diameter of the delineated optic disc6,
Fig. 4.5e. The black and the red stars in Fig. 4.5f indicate the locations of the respec-tive centroids. Fig. 4.6a shows the result of the K-means clustering. We consider the white region as the cup since the average distance of the pixels is closer to the disc center while the gray region is the neuroretinal rim.
5Empirically, the region that is close to the periphery of the optic disc is most likely to be the
neuroreti-nal rim which has the minimum intensity value in the hue channel after removing vessels. We choose the point which has the intensity value ranked the 20th percentile instead of the minimum in order to avoid selecting the residual vessel pixels in this region.
6We use a boundary of more than one-pixel thickness in order to be less sensitive to the noisy pixels
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.5: The initialization of the K-means clustering. (a) The optic disc region image (of size 209 × 153 pixels) that is obtained from the previous steps. (b) Results of edge preserving smoothing, and (c) vessel elimination. (d) The considered part which has the higher average intensity. (e) Region in which the pixels are considered candidates for the centroid of neu-roretinal rim. (f) The markers indicate the initialized centroids of the K-Means clustering.
Next, we use morphological closing followed by an opening to connect the iso-lated regions and remove small islands. The empirical sizes of the disc-shaped struc-turing elements for these morphological operations are fractions of 0.1 and 0.05 of the maximum axis of the determined optic disc, respectively. Fig. 4.6b shows the cup and rim segmentation after the morphological operations. We fit an ellipse to the white region with its center being the center of mass and its vertical and hori-zontal axes being the height and width of the concerned component, respectively. The blue ellipse in Fig. 4.6c illustrates the segmented cup.
4.2. Proposed method 85
(a)
(b)
(c)
Figure 4.6: Cup segmentation by the K-means clustering. (a) Result of the K-Means cluster-ing. (b) Segmentation of the cup and the neuroretinal rim after the morphological operations. (c) The final determined cup which is indicated by the inner ellipse in blue.
Hd Hc
Figure 4.7: Computation of the VCDR. The black dot indicates the center of the cup. The VCDR here is equal to 0.37.
4.2.4
Vertical Cup-to-disc Ratio (VCDR)
We compute the VCDR as the ratio Hc
Hd of the height of the ellipse representing the
cup and the height of the delineated disc with respect to the center of the cup. For the considered example in Fig. 4.7, the VCDR value is 0.37.
4.3
Implementation and experimental results
4.3.1
Data sets and the manual annotation from the
ophthalmolo-gist
We experiment on the same eight data sets of retinal fundus images, which we used for the optic disc detection. Most of the data sets were originally created for the purpose of vessel segmentation or diabetic retinopathy screening. Among these data sets, only 50 out of 101 images from the DRISHTI-GS1 data set (Sivaswamy et al., 2014) have annotations of glaucoma-related features, namely boundaries of the optic discs and the cups. The DRIONS (Carmona et al., 2008) and the ONHSD (Lowell et al., 2004) data sets provide annotations of the optic disc boundaries and the HRF data set (Kohler et al., 2013) gives the centers of the optic discs. None of the other data sets provide any annotations of glaucoma-related features.
One of our contributions in this work is the annotation7of the optic disc and the
cup boundaries for the first eight data sets made by an experienced ophthalmologist from the University Medical Center Groningen (UMCG). He used an annotation tool to mark the leftmost, the rightmost, the topmost and the bottommost boundary points of the optic disc as well as those of the cup in each image. Fig. 4.8 shows some examples of the manual annotations. We then fit two ellipses to these eight points to represent the annotated boundaries of the optic disc and the cup in each image.
4.3.2
Edge Preserving Smoothing
Papari et al. (2007) proposed a smoothing algorithm which was used to achieve artistic effects on photographic images. The algorithm smooths the texture details while preserving the edges. It has two parameters, which are the standard devia-tion of the Gaussian funcdevia-tion for smoothing and the sharpness of the edges. In the implementation before the delineation of the disc boundary, we set the size of the brush stroke to 3 pixels while for the one in the cup segmentation, we set it to a fraction (=0.1) of the maximum diameter of the determined optic disc. We set the sharpness to 15 in both cases. For further detail on the edge smoothing algorithm, we refer to (Papari et al., 2007) and the online implementation8.
7These annotations (specifications of the optic disc and cup boundaries) can be downloaded from
http://matlabserver.cs.rug.nl.
4.3. Implementation and experimental results 87
Figure 4.8: Examples of manual annotations of optic discs and cups in retinal fundus images (taken from the DRISHTI-GS1, DIARETDB1 and the DRIVE data sets, respectively) by the ophthalmologist. The black stars indicate the boundary points of the optic discs while the green stars are the boundary points of the cups. The bottom row shows the close-up view of these examples.
4.3.3
Implementation of the proposed approach
Disc segmentation and estimation of its reliability
We implement the method proposed in Section 3.3 for the segmentation of the op-tic disc. We first apply the configured vasculature COSFIRE filters to detect the divergence of the major vessel tree. We denote by Iv a reliability indicator of this
detection; it is equal to 1 when there is a filter that responds to the image, otherwise it is 0. Next, in the same way, we apply the disc-selective COSFIRE filters to detect the bright disc in order to precisely localize the optic disc. We then denote by Id
the reliability indicator of the bright disc, which is 1 when one of the disc-selective filters responds, otherwise 0. Finally, we fit an ellipse to the disc to approximate its boundary. We use three different ranges of values as pre-estimated disc disc sizes. For each range of values, we result in an ellipse that best fits the optic disc. We de-note by Iea reliability indicator that concerns the delineation of the disc boundary.
It can have one of four values (0, 1/3, 2/3, or 1) that represents the proportion of the number of similar ellipses.
Cup segmentation and VCDR computation
In the cup segmentation step, we apply the K-means clustering to segment the neu-roretinal rim and the cup region. In order to measure the effectiveness of the clus-tering of the two regions, we compute the difference in the HSL color space between the centroids of the resulting two regions. We denote this difference by Icwhich we
use as an indicator for cup-rim separability. In the last step, we compute the vertical cup-to-disc ratio in the way as described in Section 4.2.4.
Reliability estimation
Finally, for each retinal fundus image we compute a reliability score (RS) which is a function of the indicators obtained from each step:
RS = Iv∧ Id∧ (Ie> 0) ∧ (Ic> 0.03)
The resulting value of RS is binary; a value of 1 indicates that the obtained result is reliable, otherwise it is not. A unreliable score (RS = 0) can be due to the uncertain presence of the vessel divergent points or the bright disc region, the presence of PPA and the lack of obvious cup excavation. In such cases, the images can be referred to ophthalmologists for manual examination.
4.3.4
Experimental Results
We evaluate the performance of the proposed approach on the following aspects: disc localization, disc segmentation, cup segmentation and VCDR values.
Number of Images with Reliable Results
We evaluate on the images with reliable results (RS = 1) which amount to a total of 1558 images. We list the number of reliable images for each data set in the first column in Table 4.1.
Performance of Optic Disc Localization
We compare the automatically obtained results with the ones provided by an expe-rienced ophthalmologist. We report the results of the localization accuracy and the average localization error in Table 4.1. The relative error of the optic disc localization is computed as the distance between the automatically determined disc center and the one from the manual annotation divided by the height of the disc in the manual annotation.
4.3. Implementation and experimental results 89
Table 4.1: Localization accuracy and average localization error. The second column represents the number of images with reliable results in each data set. The values in brackets represent the total number of images in the concerned data sets.
Data set No. of images withRS = 1
(total No.) Accuracy (%)
Average localization error CHASEDB1 25 (28) 100 0.09 DIARETDB1 82 (89) 100 0.05 DRISHTI-GS1 86 (101) 100 0.05 DRIONS 101 (110) 100 0.05 DRIVE 36 (40) 100 0.04 HRF 41 (45) 100 0.08 MESSIDOR 1100 (1200) 99.52 0.06 ONHSD 87 (99) 98.78 0.07 Overall 1558 (1712) 99.79 0.06
Table 4.2: Performance measurements of the disc segmentation on the test data sets. We mark in bold the best result for each measurement.
Data set Se Sp MCC RHE CHASEDB1 0.92 0.99 0.89 0.09 DIARETDB1 0.97 0.99 0.91 0.07 DRISHTI-GS1 0.98 0.99 0.93 0.07 DRIONS 0.97 0.99 0.89 0.09 DRIVE 0.96 0.99 0.89 0.08 HRF 0.94 0.99 0.90 0.08 MESSIDOR 0.95 0.99 0.90 0.07 ONHSD 0.92 0.99 0.86 0.08 Weighted mean 0.95 0.99 0.90 0.07
Performance of Optic Disc Segmentation
We then evaluate the performance of the optic disc segmentation in the same way as introduced in Chapter 3. We list the mean values of the sensitivity, specificity, Mathew Correlation Coefficient (MCC) as well as the relative disc height error (RHE) for each data set in Table 4.2.
Performance of Cup Segmentation
We also measure the quality of the cup segmentation using the same performance indicators mentioned above, and report the mean results per data set in Table 4.3.
Performance of VCDR Estimation
Next, we compute the errors of the automatically obtained VCDR values with re-spect to those from the manual annotations by the ophthalmologist. In Fig. 4.9, we
Table 4.3: Performance measurements of the cup segmentation on the test data sets. We mark in bold the best result for each measurement.
Data set Se Sp MCC RHE VCDR CHASEDB1 0.55 0.99 0.53 1.04 0.18 DIARETDB1 0.45 0.99 0.41 1.78 0.19 DRISHTI-GS1 0.41 0.99 0.55 0.35 0.13 DRIONS 0.53 0.99 0.58 1.04 0.10 DRIVE 0.46 0.99 0.48 0.68 0.13 HRF 0.41 0.99 0.51 0.89 0.12 MESSIDOR 0.42 0.99 0.46 0.64 0.13 ONHSD 0.37 0.99 0.42 0.82 0.13 Weighted mean 0.43 0.99 0.47 0.75 0.13
show the box and whisker diagram of the VCDR values that characterize the distri-bution of the data. Fig. 4.10 illustrates the box and whisker diagrams of the VCDR errors for all test images with RS = 1, which show the distribution of the errors in each data set. As we see from the plots, the median values of the VCDR errors are around 0.1, which are indicated by the red horizontal lines in the middle of the boxes. In Fig. 4.11a, we illustrate the distribution of the obtained reliable VCDR values in all test data sets with respect to the manual annotation of the ophthalmol-ogist. While Fig. 4.11b shows the inter-observers between two ophthalmologists on 100randomly selected images. We also illustrate in Fig. 4.12 the Bland-Altman plot (also known as the Tukey mean-difference plot) of the automatically determined VCDR values with respect to the manual annotation from the ophthalmologist. Such a plot is used to analyze the agreement between two measurements. The mean dif-ference (bias) between the automatically obtained VCDRs and those from the oph-thalmologist was −0.06 and the mean difference ±1.96 times the standard deviation (limits of agreement) +0.40 and −0.52, respectively. For the inter-observer variabil-ity (Fig. 4.11b), these values were 0.03, 0.19, and −0.13, respectively. As indicated by the shaded region in Fig. 4.12, the proposed approach achieved a better agreement (smaller difference) on images with a large VCDR (> 0.7) according to the manual annotations. This is important, because a VCDR above 0.7 indicates pathology.
Performance of the classification of glaucomatous retinas
In clinical practice, glaucoma experts do not rely on a VCDR value alone to conclude whether a patient has glaucoma or not. However, as mentioned before, the VCDR is an important measure for research and contributes to clinical care and screen-ing. In order to get an indication of the screening performance of our method, we computed the sensitivity and specificity of our approach in the following way. We used VCDRoph = 0.7 as a threshold to label the images as healthy or
glaucoma-4.3. Implementation and experimental results 91
CHASEDB1 DIARETDB1 DRISHTI-GS1 DRIONS DRIVE HRF MESSIDOR ONHSD
VCDR
values
Figure 4.9: Box-whisker plot of the VCDR values provided manually by the ophthalmologist on all images in each data set.
tous (corresponding to the 97.5thpercentile of the general popularion; (Wolfs et al.,
2000)). We then took a threshold value VCDR*. We computed the sensitivity as the ratio of the number of images that have the automatically obtained VCDR above the VCDR* threshold and the manually provided VCDR above the VCDRoph(area
Q3 in Fig. 4.11), and the number of images that have the manually provided VCDR above the VCDRoph (area Q3+Q4 in Fig. 4.11). The specificity is the ratio of the
number of images that have the automatically obtained VCDR below the VCDR* and the manually provided VCDR below the VCDRoph (area Q1 in Fig. 4.11), and
the number of images that have the manually provided VCDR below the VCDRoph
(area Q1+Q2 in Fig. 4.11). The sensitivity and specificity are functions of the VCDR* threshold and we illustrate this in Fig. 4.13. This figure shows the receiver operating characteristic (ROC) curve of the automatic method. The closer such a curve is to the top-left corner, the better the performance of the approach is. In the same way, we obtained ROC curves for threshold VCDRophequal to 0.5 and 0.8 (presented in
the same figure). We achieved an area under the curve (AUC) of 0.51, 0.74 and 0.79 for VCDRophequal to 0.5, 0.7, and 0.8, respectively.
CHASEDB1 DIARETDB1 DRISHTI-GS1 DRIONS DRIVE HRF MESSIDOR ONHSD
VCDR
err
ors
Figure 4.10: Box-whisker plot of the VCDR errors on reliable images (RS = 1) in each data set.
4.4
Discussion
We propose a systematic approach that computes the vertical cup-to-disc ratio to as-sist ophthalmologists in population-based glaucoma screening. We experiment on eight public data sets with a total of 1712 retinal fundus images and evaluate the per-formance for each step of the approach. We compare the localization perper-formance of our method with a recently published approach proposed by Akram et al. (2015) and report comparable results in the previous chapter in Table. 3.2. The images whose optic discs are not correctly localized in the DIARETDB1, DRIVE and HRF data sets are later indicated as unreliable cases by the proposed reliability score. This is indicated by the results shown in Table 4.1, in which these three data sets have lo-calization accuracies of 100% after removing unreliable cases. In Fig. 4.14, we show the four images from the DRIVE data set that our algorithm labeled as unreliable as well as the segmentation results. The first image is marked as unreliable due to the uncertain presence of the optic disc pattern. The problems of the other three images are due to the presence of PPA outside the disc boundaries.
In the evaluation of optic disc and cup segmentation, we achieve a weighted mean MCC of 0.9024 and 0.4730, respectively. For the evaluation of the VCDR
val-4.4. Discussion 93 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Ophthalmologist VCDR Automatically determined VCDR Ophthalmologist VCDR Another glaucoma expert VCDR (a) (b) Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4
Figure 4.11: Distribution of the VCDRs on reliable images. (a) Distribution of the automati-cally obtained VCDR values on reliable images with respect to the manual annotation from the ophthalmologist. The x-axis represents the VCDR value provided by the ophthalmologist while the y-axis is the automatically obtained VCDR value. (b) Distribution of the VCDR val-ues provided by a trained (non-medical) expert with respect to the manual annotation from the ophthalmologist. The x-axis represents the VCDR value provided by the ophthalmologist while the y-axis is the VCDR value provided by the second observer. The intensity values in both matrices indicate the number of images falling in the grid regions. The vertical and hor-izontal lines represent the VCDR* threshold for the classification of the images. The ones that fall into Q1, Q2, Q3 and Q4 areas are the true negatives, false positives, true positives and false negatives, respectively. The sensitivity is computed as #Q3
#Q3+#Q4 while the specificity
is #Q1
#Q1+#Q2. # indicates the number of images falling in the region. For the chosen cut-off
value of 0.7 in (a), Q1=1311, Q2=166, Q3=18 and Q4=63.
ues, we achieve a weighted mean VCDR error of 0.17 on 1558 out of 1712 images, to which the system could provide reliable results. The means are weighted with the number of images in each data set. We do not compare with existing methods on the evaluation of disc and cup segmentation as well as VCDR values because those methods use either proprietary data sets or public data sets with manual annotation that is not publicly accessible. Here we mention two other studies that propose the automatic calculation of VCDR. These are the ones authored by Yin et al. (2012) and Mittapalli and Kande (2016) who reported a VCDR error of 0.10 on 325 images and 0.13on 59 images, respectively. The used data sets and the manual annotation are, however, not publicly available.
The results show that the step that has room for large improvements is the cup segmentation, which is in particular indicated by the large height errors (RHE) of the cups. The reason is probably due to the fact that most of the normal retinal fundus
Mean of two measurements Dif fer ence of two measur ements +1.96SD +0.3977 MEAN -0.0631 -1.96SD -0.5239
Figure 4.12: Bland-Altman plot of the automatically determined VCDR values with respect to the manual annotation from the ophthalmologist. The x-axis and the y-axis of the plot are the mean and the difference of the VCDR values obtained from the proposed approach and the ophthalmologist, respectively. Each blue circle represents an image and the blue line indicates the mean value while the red dashed lines indicate the mean value ±1.96 times the standard deviation. The images falling in the shading concern VCDR values above the 97.5th
percentile of the general population (indicating pathology)
.
images do not have an obvious cup excavation. We show an example in Fig.4.15a. In such cases it is difficult for the proposed system to determine the cup region as indicated by the ophthalmologist (Fig.4.15b). This is mainly due to the fact that the pixels (except the vessel pixels) in the disc region have similar intensity values and the K-means algorithm always seeks to cluster all the pixels into two groups, which in most of the cases overestimates the cup region (Fig.4.15c). In future, we aim to investigate other segmentation algorithms that can deal with such challenging images.
In healthy eyes, the VCDR ranges from 0 to approximately 0.7 (97.5thpercentile
(Wolfs et al., 2000)); in glaucomatous eyes from approximately 0.5 to - ultimately - 1.0 (Springelkamp, Wolfs, Ramdas, Hofman, Vingerling, Klaver and Jansonius, 2017). The agreement between our approach and the annotation of the ophthalmol-ogist was best for VCDRs above 0.7 (Fig.4.12, and - related to that - our approach was able to detect VCDRs above 0.7 and especially above 0.8 with a reasonable sen-sitivity and specificity (Fig. 4.13). For lower VCDRs, the approach cannot be used. Fortunately, the greater VCDRs are the relevant ones to detect from the point of view of preventing blindness. For screening, a high specificity is the most important issue
4.4. Discussion 95 1 - Specificity Sensitivity VCDRoph =0.5 VCDRoph =0.7 VCDRoph =0.8
Figure 4.13: ROC curves of the automatically obtained VCDR values with respect to the man-ual annotation provided by the ophthalmologist. We obtain these curves by ranging the VCDR* threshold between 0 and 1 in intervals of 0.05. The area under the curve (AUC) is equal to 0.51, 0.74 and 0.79 when we set the VCDRophto 0.5, 0.7 and 0.8, respectively.
(de Vries et al., 2012; Vaahtoranta-Lehtonen et al., 2007). Hence, the left part of the ROC curve (Fig.4.13) is the most important part.
The vasculature-selective COSFIRE filters are effective to locate the side at which the optic disc is placed. In order to improve the localization precision, we then apply a set of disc-selective COSFIRE filters within the selected region. The proposed two-stage localization process turned out to be essential to reduce the number of false detections. Any bright lesions, such as hard exudates, would affect the performance of disc-selective filters if they had to be applied to the entire image.
From the training set with 397 retinal fundus images, we configure up to 27 vasculature-selective COSFIRE filters with which we are able to detect most of the vessel trees in all retinal fundus images from the eight data sets. In this way, we demonstrate the robustness of the configured filters, which were tested on com-pletely different data sets from the training set. We make all these 27 filters publicly available9.
In this work, we propose a complete system for glaucoma screening. The system consists of four main steps: the localization of the optic disc, delineation of the disc
Figure 4.14: (Top row) The four retinal fundus images (‘07 test.tif’, ‘26 training.tif’, ‘31 training.tif’ and ‘34 training.tif’) from the DRIVE data set that are indicated as unreliable. (Bottom row) The corresponding close-up views of the segmented optic discs and the cups. The black ellipses indicate the delineated optic discs while the blue ellipses are the cups. The reliability indicators (Iv, Id, Ie, Ic)of the four images are (1,0,0.33,0.1),(1,1,0,0.02),(1,1,0,0.07)
and (1,1,0,0.06), respectively. The presence of PPA is the main reason for the unreliable results.
(a) (b) (c)
Figure 4.15: Example of unobvious cup excavation. (a) An example of a retinal fundus image of unobvious cup excavation. (b) The manual segmentation of the optic disc and the cup obtained from the eight boundary points provided by the ophthalmologist. (c) The automated segmentation.
boundary, the segmentation of the cup and the determination of the VCDR. In the first step, we use trainable COSFIRE filters for the detection of the vessel divergent points and bright disc patterns. The approach we used for the delineation of the disc boundary is based on fitting of ellipses. For the segmentation of the cup, we used K-means clustering which is popular for its simplicity and effectiveness.
4.5. Conclusions 97
One of the contributions of this work is the manual annotation data of 1712 im-ages from eight data sets, which we make available online10. The manual annotation
has been generated by an ophthalmologist at the UMCG hospital in Groningen, who marked the locations of the topmost, the bottommost, the leftmost and the rightmost boundaries of both the optic disc and the cup of all images. A subset has been an-notated by another ophthalmologist, showing a very small bias and inter-observer variability (Fig. 4.11b).
The rise of deep learning approaches has also motivated other researchers to investigate such techniques for the diagnosis of pathologies from retinal fundus im-ages. There are already quite a few publications (Wong and Bressler, 2016; Gulshan et al., 2016; Chen et al., 2015) using deep learning methods for diabetic retinopa-thy screening as well as glaucoma detection. For instance, the work by Chen et al. (2015) used convolutional neural networks (CNNs) to classify the retinal fundus im-ages into normal and glaucomatous. Their system can be considered as a black box, which gives as output a value that indicates the probability whether an image has signs of glaucoma or not. The visual features that they use to come to such a conclu-sion cannot be interpreted by ophthalmologists. On the other hand, our approach is based on the same features - in particular the VCDR - that glaucoma experts use in practice for their diagnosis and is used widely in glaucoma research (Foster et al., 2002). Such features can then be used to monitor the progression of every individ-ual, while the uninterpretable features from black box methods do not offer this possibility.
4.5
Conclusions
We propose a novel approach for the detection of glaucoma-related features in oph-thalmology. It consists of four steps, namely optic disc localization, optic disc delin-eation, cup delindelin-eation, and computation of the vertical cup-to-disc ratio (VCDR). For a total number of 1558 images from eight data sets we achieve a mean VCDR difference of 0.17 with respect to a glaucoma expert. The system provides very re-liable delineation of the optic disc (MCC=0.90), from which we obtain the height of the optic disc. The segmentation of the cup, and thus the measurement of its height, turned out to be the most problematic part of the system (MCC=0.47). There is room for further improvements of the method, especially regarding the segmen-tation of small cups. Bland-Altman analysis showed that the system achieves better agreement with respect to the manual annotations for large VCDRs, which indicate pathology. Although it is not typical to rely on the classification of image alone to
infer whether a case is glaucomatous or not, we computed the classification per-formance to obtain an indication. For a VCDR threshold of 0.7, we achieved an AUC of 0.74. The proposed system could be deployed as part of a population-based glaucoma screening. We also made available online the manual annotations by an experienced ophthalmologist of glaucoma-related features of eight benchmark data sets in order to facilitate comparison of future methods and thus further this field.
