1. Introduction

Rapid development in the field of ophthalmology has increased the demand of computer aided classifiers for various eye diseases. Papilledema is an eye disease in which the optic disc of the eye is swelled due to increase in intracranial pressure. It causes blurred vision, headaches and nausea in its initial stages. If it is left untreated, it can result in permanent loss of vision and it might lead to death in some cases. Papilledema occurs due to increase in intracranial pressure which happens due to increased blood pressure i.e. hypertension. Any patient showing papilledema from hypertension has malignant hypertension and should be considered as a medical emergency. Papilledema is not just caused by hypertension; there are several other causes too. It is well known that papilledema itself is not just a disease, but a symptom of several other diseases [1]. If papilledema is found in a patient, its underlying cause should be found as soon as possible, because higher stages of papilledema indicate that some serious disease (such as brain tumor or malignant hypertension) might be progressing in the patient. Therefore papilledema should be detected on time before it is too late to save the patient’s life [1].

Different signs of papilledema as seen on a fundus image include blurring of the optic disc boundary, appearance of a circumferential ring around the optic disc, dilation of blood vessels, appearance of hemorrhages or cotton wool spots, loss of major vessels as they leave the optic disc and blurring of the vessels on the optic disc [1]. Figure 1 shows a fundus image from STARE data set with papilledema. In the image, it can be seen that the boundary of optic disc is blurred and not quite distinguishable. The blood vessels are also dilated while some vessels are blurred.

 

Fig. 1 Fundus image with papilledema: It can be noted that the optic disc boundary is blurred and blood vessels are dilated.

Download Full Size | PDF

In literature, there have been several papilledema case studies carried out from medical perspective, where papilledema was found along with presence of some life threatening diseases, thus helping doctors in their diagnosis. It is recommended that every patient of papilledema should be treated very carefully. Generally, a computed tomography (CT) scan, magnetic resonance imaging (MRI) scan or lumbar puncture are used to detect the real cause of papilledema but these are not very viable sources. Neurosurgical operations also provides a convenient way to cure papilledema however it also fails in majority of the cases, as described in [2]. The best way to detect papilledema is through thorough physical examination and maintaining the patient checkup history. Therefore, it is of grave importance to do a complete physical examination and get a thorough history of the patient to diagnose the underlying cause of papilledema and not just undergo neurosurgical operations for treating papilledema. Papilledema can be easily diagnosed through regular checkups [3].

Some of researchers have worked on the automated detection of papilledema using fundus and OCT images. The major sign of papilledema is optic disc swelling as shown in Fig. 1 and automated detection of papilledema mainly focuses on the analysis of optic disc. L. Tang et al. used digital stereoscopic color fundus images to evaluate volume of the optic disc in [4]. By measuring optic disc volume, papilledema can be diagnosed and managed because increase in volume would indicate presence of papilledema. The authors have used a total of 29 sets of fundus photographs along with optic nerve head centered spectral domain optic coherence domain tomography scans (SD-OCT scans) obtained from fifteen patients having papilledema. In order to estimate the 3-dimensional shape of the optic nerve head, an automated multi-scale stereo correspondence algorithm is used. This algorithm automatically finds dense association between the stereo images giving a stereo derived depth map which is further interpolated to smooth all the noise present using a thin plate spine (TPS). For determining the change in volume of the optic nerve head using OCT, an automated 3-D segmentation technique was applied which segments the OCT volume. Finally the correlation between SD-OCT volume and stereo volume measurements was assessed through a reference plane which also helps in grading papilledema using Friesen scale. The volumetric measurements of retinal surface elevation found from both stereo fundus photographs and OCT scans are positively correlated (correlation coefficient r2 = 0.60; P < 0.001). Therefore, it is concluded that the elevation of the retinal surface in patients suffering from papilledema are extracted from stereoscopic fundus photographs and it correlates well with the results that are obtained from OCT scans. Stereoscopic color imaging of the optic nerve head together with some method of automated shape reconstruction can prove to be a low-cost substitute of SD-OCT scans that has potential for a more cost-effective diagnosis and management of papilledema.

Another method to determine the degree to which OCT can discriminate variation in the retinal nerve fiber layer thickness (RNFL) between healthy eyes, eyes with mild papilledema and eyes with pseudo-papilledema has been reported in [5]. The data set contained a total of 41 subjects out of which 17 were normal (healthy) people, 11 people had congenitally crowded optic nerves and 13 people had mild papilledema. All of these subjects went through fundus photography, complete neuro-ophthalmic examination and testing of visual fields. Among these patients, spinal fluid pressure measurements were calculated in 5 patients of pseudopapilledema and eleven patients of mild papilledema. The OCT scans were made on both eyes of all the subjects. These scans were circular in shape and had 3.38mm diameter surrounding the optic disc. The swelling in the fundus images (papilledema) was graded by 2 experts after analyzing the fundus photographs. Authors reported that the RNFL thickness in superior and inferior parts was greater showing more correlation between all groups of subjects. The mean RNFL thickness in normal subjects and patients with papilledema was significantly different however it showed no significant difference between patients with congenitally crowded optic nerves and those having papilledema. This study concluded that although OCT gave statistical difference in RNFL thickness between patients of papilledema, pseudo-papilledema and normal people, it did not distinguish patients with papilledema from those having congenitally crowded optic nerves.

Apart from this, an automated system to monitor different stages of papilledema has been reported in [6]. Authors extracted 24 features based on different signs of papilledema. The algorithm was applied on a local data set consisting of images with papilledema of grades 0 to 4 taken from 39 patients showing progression of papilledema for over 2 years. For classification, the decision tree methodology was used. However, no work was done to classify grade 5 papilledema cases. The algorithm was evaluated by calculating Cohen’s weighted k coefficient which tells if the result from algorithm is in agreement with ground truth. The value of k varies from −1 to 1 where −1 means the results are in complete disagreement and 1 means the results are in complete agreement. Out of 100 images, 64 images were correctly classified into their respective grades of papilledema. Another method for classifying papilledema from fundus images employs 13 different features [7]. This method gives an accuracy of 96.67%, when applied to 30 images from the STARE data set. One of the biggest limitations of this algorithm is that it does not use any vascular features or other features specifically related to blurring of optic disc boundary. Previously we have presented an automated method to detect papilledema from fundus images by extracting 10 different vascular and textural features in [8] but here, we propose a fully automated decision support system by incorporating different combinations of color, texture, vessel and disc margin based features for automated diagnosis of papilledema from fundus images. We have also validated our proposed system on more images from STARE as compared to [8], particularly the proposed system was tested on 160 fundus images in which 90 of them are from STARE data set and 70 of them are from our local data set that is acquired from AFIO. The rest of this paper is structured as follows. Section 2 gives an in-depth description of our proposed methodology. Section 3 demonstrates our results, followed by Section 4 which discusses the evaluation extracted feature set. Section 5 depicts the overall results followed by Section 6, which gives a discussion on our proposed implementation and outlines conclusions.

1.1 Data set description

There are two data sets that have been used in this research. The first one is the online STARE data set and the second one is our custom data set acquired from AFIO. STARE data set has annotations for optic nerve (ON) swelling [17]. These images are further re-annotated with the help of two ophthalmologists. These new annotations were made by the ophthalmologists without any prior information about annotations from STARE database. Apart from this, a local data set was annotated by two ophthalmologists working in AFIO. We considered a subset of 90 and 70 images from STARE and local data set respectively for which both ophthalmologists gave same annotations. The annotations have been carried out using Modified Friesen Scale [18]. Modified Friesen Scale defined 6 grades for papilledema, however, in this research we are only focusing on two i.e. normal and papilledema. The images are not further graded into sub grades of papilledema. Table 1 shows original Modified Friesen Scale [18].

Table 1. Modified Friesen Scale [18]

View Table | View all tables in this article

Table 2. Summary of data set used in this research

View Table | View all tables in this article

2. Proposed methodology

We propose a fully automated system to detect papilledema from fundus images. First of all, the input fundus images are preprocessed. After preprocessing, several features are extracted from each candidate image. Out of those features, the optimal features are selected and combined to form a feature matrix to be used for classification between normal and papilledema images. Figure 2 shows a block diagram of our proposed system.

 

Fig. 2 Block diagram of proposed automatic papilledema detection system.

Download Full Size | PDF

2.1 Preprocessing

The preprocessing is the first step in our proposed system for the localization of optic disc region and for the extraction of blood vessels from input fundus image. As we know, papilledema causes swellings within the optic disc region, also shown in Fig. 3, so rather than processing the whole fundus image, the proposed system only extracts the specific region of interest (ROI) that depicts swelled optic disc and blood vessels.

 

Fig. 3 Fundus image with papilledema.

Download Full Size | PDF

Since optic disc is a bright and circular region from which all blood vessels emerge, these properties (intensity and vessel density) can be used to determine its position [9]. First of all, vessels are enhanced for which 2-D Gabor wavelet is used due to its ability to enhance directional structures. After vessel enhancement, vessel segmentation is performed through a multilayered thresholding technique [10]. The reason for using segmentation method proposed in [10] is because it is quite robust to noise and it can even detect small vascular patterns with an average accuracy of 94.85% [10].The images are then cropped such that the resulting image has optic disc and a small region surrounding the optic disc (2-disc diameter). Apart from this, we have also extracted the optic disc boundary by removing blood vessels around the optic disc region using the method which we proposed in [11] and the boundary of OD is smoothened using ellipse fitting. The extracted disc boundary plays a vital role in extracting color features as described in section 2.2.1 below.

2.2 Feature extraction

The preprocessing step results into our required ROI which includes optic disc and a small region containing optic disc as well as vascular segmentation of this region. Figure 4 shows the ROI that can be used to extract different representative features.

 

Fig. 4 The ROIs obtained because of preprocessing step (a) original fundus image (b) ROI after optic disc localization (c) ROI after vessel segmentation.

Download Full Size | PDF

The feature extraction process is divided into four sub-steps, which include:

2.2.1 Color features

In case of normal images, the optic disc boundary is sharp, bright and clearly visible while for the images with papilledema the optic disc boundary becomes dull, less bright and blurry. Therefore, as papilledema advances to higher grades, the RGB values of the pixels on the optic disc starts to vary. The color subset in feature extraction calculates these varied values. A total of 3 features are extracted from the color images. For these 3 features, the ROI obtained after optic disc localization is directly used whereas for rest of the subsets, the image is first converted into gray-scale image. The features that have been used from this color image include maximum, minimum and mean pixel value. In order to calculate these values, 8 pixels, at an equal distance on the optic disc boundary are automatically selected. Figure 5 shows selected pixel values on a normal image and on an image with papilledema. The quantitative values for colored features against these two images are also given in Fig. 5.

 

Fig. 5 Selected pixels and their respective values on disc boundary of (a) normal image, and (b) image with papilledema.

Download Full Size | PDF

Note that the selected pixel values are all lying on the optic disc boundary. This is done to exploit the fact that the images with papilledema would have blurred disc margins. As one can see that the boundary of image with papilledema is less sharp as compared to that of a normal image without papilledema. Therefore, calculating pixel values on the boundary gives good information for detecting papilledema. To automatically select the 8 pixels, we first extracted the center of the localized optic disc region. Then with the radial distance from the center, we have calculated 8 pixels with equal interval between 0 to 2π. The radial distance is determined by finding the difference between the boundary of optic disc from its center as shown in Fig. 5. Since the optic disc is not entirely circular so that’s why some of the boundary pixels do not honor the length of the radial line. In that case, we picked those boundary pixels which lies on the same orientation as radial line and are closest to the radial lines as shown in Fig. 6.

 

Fig. 6 Automated extraction of 8 selected pixels, (a) radial lines are shown in white color which are used to extract the optic disc boundary pixels at 8 equal intervals, (b) extracted pixels are mapped onto the optic disc image.

Download Full Size | PDF

From these selected pixels, we calculated the following three features.

2.2.2 Texture features

Texture is one of the most important characteristics which can be used to identify different objects or region of interests. In any image, the texture gives a great deal of information such as contrast, regularity or uniformity, etc. present in that image. Texture of an image is associated with the spatial distribution of its intensity values. Different methodologies can be used to extract a significant number of quantitative texture features from images which can then be used to obtain further classification results [12]. In a fundus image, as the swelling of the optic disc increases due to papilledema, its peripapillary texture properties changes [6]. These textural variations are due to abnormal optic disc pathology resulted from the spread of papilledema. Based on this motivation, we have extracted some of the useful textural features that contributes positively in discriminating the diseased and healthy pathology.

The first four textural features subset have been extracted using gray-level co-occurrence matrix (GLCM) of the image. The GLCM deals with statistical properties of spatial distribution of image gray level values. In a gray-scale image, each pixel has some intensity value called the gray level. The GLCM tabulates how frequently different gray-level combinations co-occur in the image or a particular section of the image [13]. In our proposed implementation, we have extraction following texture based features:

 

Fig. 7 The selected profile paths in case of (a) healthy fundus image, and (b) image with papilledema, and their computed profiles shown respectively in (c) and (d).

Download Full Size | PDF

2.2.3 Vascular features

When papilledema occurs, it affects the vascular structure of the fundus image. As the disease progresses, the vascular properties vary substantially. Vessels become obscured, dilated and in some cases tortuosity of vessels occurs. With the increase in the grade of papilledema, the obscuration of major vessels also increases making them less defined. Keeping in mind these properties, several vessel features have been extracted from fundus images which are [6]:

 

Fig. 8 Fundus images, their vessel segmentation and respective VDI.

Download Full Size | PDF

2.2.4 Disc margin obscuration features

The optic disc can be divided into four equal poles of 90°. They are called superior, nasal, inferior and temporal poles. The nasal and temporal poles are further subdivided into three equal segments of 30° each which are called superior, central and inferior segments of nasal or temporal poles. Figure 9 shows the division of optic disc into poles and the subdivision of nasal pole. When the optic disc is swollen due to papilledema, optic disc (OD) margin gets blurry. As the swelling progresses, so does the obscuration of OD margin. According to modern Friesen scale [16], the nasal pole is first to start becoming blurred. In a higher degree of papilledema, the temporal pole also becomes obscured along with the nasal pole. In case of extreme papilledema, the obscuration progresses all around the optic disc. The superior and inferior poles have large variability in the level of obscuration of disc margin due to which they are not used. Based on these properties, the disc margin obscuration features are extracted. Although Echegaray et al. [6], have proposed disc margin obscuration features but we have defined the disc margin obscuration features in a different way. Authors in [6] measure the blurriness in different sectors of optic disc by calculating standard deviation of the radial distance from the center of the optic disc to each point on its margin (where the change in pixel intensity was greatest every 2°). In our algorithm, we have used mean intensity value of gradient magnitude in different sectors of optic disc for calculating the disc margin obscuration features which are:

 

Fig. 9 The division of OD into poles and subdivision of nasal pole.

Download Full Size | PDF

Table 3. Summary of extracted features

View Table | View all tables in this article

2.3. Feature selection

All extracted 26 features were combined to form a feature vector. However, all of these features were not used directly for classification. The optimum features were selected through Wilcoxon rank sum test so that redundant, irrelevant or noisy features responsible in reduction of performance of proposed system can be removed. The Wilcoxon rank sum test analyzes whether the normal or abnormal class median values present in the data set differ significantly or not. If the features do not have different values of median for both classes then those features are considered less effective for classification. Table 4 shows the results obtained after applying Wilcoxon rank sum test to our original set of 26 features. The features are written in descending order of their absolute score. The features become less significant as their absolute score decreases and the p-value increases. The top ten features obtained according to this test are features F15, F16, F7, F25, F12, F13, F4, F11, F2 and F18 (i.e., representative mean area of all regions in vascular map, STD of areas in vascular map regions, homogeneity, maximum nasal to temporal ratio of OD boundary obscuration, profile value, maximum area in the vascular map regions, contrast, cluster shade, minimum pixel value of the colored fundus images and VDI features).

Table 4. Performance of features calculated using Wilcoxon rank sum test

View Table | View all tables in this article

2.4. Classification

After feature selection, the selected features were combined to form a feature set which is then used for classification between normal and papilledema cases using the SVM classifier. We chose SVM as a classification algorithm in our proposed implementation because it is fast and requires less time for training as compared to other supervised classifiers. Also, the major benefit of SVM is that it requires only its support vectors to hold a decision boundary instead of storing all of its training samples, which makes it memory efficient as well. Also it can easily classify non-linearly separable samples through the use of its non-linear hyper-plane kernels. In order to perform classification, the data set was first divided into training and testing samples such that the training samples were used to train the classifier while the testing samples were used to test the output of the classifier. It is essential that the training samples must not have any information about the testing samples. If the training and testing samples are not divided properly then it is possible that same samples can be used for training as well as testing thus giving overly optimistic or rather invalid results. We used 10-folds cross validation i.e. whole data set was divided into 10 equal folds or subsets having samples of both classes. Out of 10 folds, 9 were used for training the classifier while the remaining one fold was used for testing. This process was repeated 10 times so that each time a different fold was used for testing. Thus after obtaining results for 10 times, their average was taken and used as the final classification result.

3. Results

The proposed algorithm was tested on two different data sets including a local data set and an online data set. First data set that we have used contains images which are locally collected from AFIO, Rawalpindi, Pakistan. This data set contains a total of 70 RGB fundus images including 40 normal images and 30 abnormal images (images with papilledema). All these images are of 1504 × 1000 resolution. The second data set that we used for our evaluation is a publically available data set called STARE [17]. It contains approximately 400 RGB fundus images of 700 × 605 resolution. However, out of those 400 images, only 10 images with optic disc swelling (papilledema) are present in the data set. Since we wanted to have a bigger data set to test our proposed methodology, we needed more images with papilledema. To compensate this shortage of fundus images with papilledema, we obtained another 30 papilledemic images from online and these images were annotated by multiple expert ophthalmologists working in AFIO.

3.1 Group-wise features evaluation

We divided the extracted features into 4 groups or categories namely color, texture, vascular and disc margin obscuration features. In order to evaluate these groups, we performed group-wise classification using all possible combination of different groups. Therefore, for 4 groups, total 14 possible group-wise combinations were used. The evaluation results on the combined data set are shown in Table 5.

Table 5. Group-wise features and their evaluation results

View Table | View all tables in this article

It can be observed from the results table that the group which performs well individually as compared to other groups was ‘vascular features’ giving an accuracy of 75.9%, followed by ‘color features’ with accuracy of 72.3%. One can also note that the worst classification results in case of an individual feature set were obtained for the ‘disc margin obscuration features’ with an accuracy of 59.2%. We report that the best results were obtained in case of combined color, texture and disc margin obscuration features, giving an accuracy of 87.8%, specificity of 90.6% and sensitivity 84.1%.

It can be further noted that the color features when combined with other groups increase the accuracy of that group significantly. For example, the accuracy in case of texture features alone was 65.6% which was increased to 79.8% when combined with color features. Similarly the accuracies in case of vascular features and disc margin obscuration features were increased respectively from 75.9% to 84.3% and from 59.2% to 72.9% after combining them with color features. Similar to color features, the vascular features also seem to increase the accuracy of groups. The accuracy for texture features was increased from 65.62% to 80.9% by combining them with vascular features. Similarly the disc margin obscuration had an increase in accuracy from 59.2% to 78%. The color and vascular features when combined together gave an accuracy of as high as 84.34% (which is the second best out of all group-wise combinations and the best of all 2 group combinations). It can be concluded that the most significant groups are color features and vascular features.

4. Individual features evaluation

All features were individually evaluated using Wilcoxon rank sum (Table 4) and top features were used one by one to study the effect of number of features on accuracy, sensitivity and specificity. The results were summarized in the form of a graph indicating this effect, as shown in Fig. 10.

 

Fig. 10 Effects of number of top features on percentage accuracy, sensitivity, and specificity results.

Download Full Size | PDF

The best percentage accuracy, sensitivity, and specificity results (respectively 87.4%, 88.7%, and 85.72%) were obtained using top 12 features. An abrupt change in all 3 performance measures can be noticed at top 4 features; otherwise the graph remains rather smooth from top 5 features to top 10 features.

5. Overall results

In the end, the proposed method was evaluated on all databases. Table 6 gives summary of our evaluation results.

Table 6. Proposed system performance

View Table | View all tables in this article

It can be seen that although same methodology was applied to AFIO and STARE data set and the ratio of images with and without papilledema were same for both cases, but the performance measures for STARE are more accurate than AFIO. It is so because the AFIO data set contained images that were mostly of mild papilledema cases (grade 1, 2, 3), whereas in case of STARE the images were mostly of severe papilledema cases (grade 4, 5). Apart from this, we have extended our algorithm originally proposed in [7] to automatically diagnose and grade papilledema from digital retinal images and we have also compared our proposed system with the state of the art solution proposed in [6] on STARE data set and we obtained better performance in terms of accuracy as shown in Table 7.

Table 7. Performance comparison

View Table | View all tables in this article

6. Discussion and conclusion

This paper has proposed an automated system for detecting papilledema from fundus images. The input images were first preprocessed (i.e., including optic disc localization, images cropping to focus on optic disc and a small surrounding portion, and vessel segmentation operations). Based on different properties and signs of papilledema, a total of 26 features were extracted and categorized into 4 groups; color features, texture features, vascular features and disc margin obscuration features. The group color features had 3 features and they were calculated from colored fundus images; texture features had 9 features which were extracted from grey-scale fundus images; vascular features had 6 features calculated after vessel segmentation of fundus images and the disc margin obscuration group containing 8 features and calculated from gray-scale fundus images. All the above mentioned features were evaluated using Wilcoxon rank sum test and followed by 10 folds cross validation and classification through SVM classifier. In order to evaluate the system, we calculated accuracy, sensitivity and specificity. To further evaluate the significance of different groups of features, all groups were tested individually as well as with different combinations. The effect of different number of top features on performance measures was also computed. The proposed system was applied on a local data set, acquired from AFIO that had 70 fundus images in which 30 have papilledema symptoms and 40 were healthy. AFIO data set has been annotated by multiple expert ophthalmologists who works in AFIO. Apart from this, we have applied our proposed system on a publically available STARE data set [17] that contained 90 fundus scans in which 40 had papilledema symptoms and 50 were healthy. STARE data set is an online data set that has been annotated by multiple expert ophthalmologists. These annotations act as a ground truth on which our proposed system was validated. We have also applied the proposed system on the combination of both databases. The performance measures of proposed techniques are given in Table 6. We have shown herein and conclude that our proposed methodology gives more accurate results for even higher degrees of papilledema cases. Apart from this, we have compared our proposed system with [6] on the same STARE data set and our proposed algorithm achieved better performance in terms of accuracy as shown in Table 6. In this paper, we have extended our proposed system originally proposed in [7] to incorporate efficient and early diagnosis of low grade papilledema from fundus scans. This has been done by adding a block in our original proposed system to automatically extract vascular features from the candidate retinal image. However, our proposed system is sensitive to the quality of acquired fundus scans and poor quality scan such as shown in Fig. 11 can lead to incorrect classification. Since our local AFIO data set contained such poor-quality scans so that’s why we have lesser performance ratings on our local data set as compared to the STARE data set.

 

Fig. 11 Poor quality fundus scans from our local AFIO data set.

Download Full Size | PDF

In the future, the grading of papilledema based on Friesen scale [16] can be done for a more detailed analysis of papilledema. Moreover, this work can be extended by incorporating OCT along with fundus scan to give an early and robust classification of papilledema.