Automated method for the detection and segmentation of drusen in colour fundus image for the diagnosis of age‐related macular degeneration
2018; Institution of Engineering and Technology; Volume: 12; Issue: 6 Linguagem: Inglês
10.1049/iet-ipr.2017.0685
ISSN1751-9667
AutoresSultan Mohammad Mohaimin, Sajib Saha, Alve Mahamud Khan, Abu Shamim Mohammad Arif, Yogesan Kanagasingam,
Tópico(s)Retinal Diseases and Treatments
ResumoIET Image ProcessingVolume 12, Issue 6 p. 919-927 Research ArticleFree Access Automated method for the detection and segmentation of drusen in colour fundus image for the diagnosis of age-related macular degeneration Sultan Mohammad Mohaimin, Sultan Mohammad Mohaimin Computer Science and Engineering Discipline, Khulna University, Khulna, BangladeshSearch for more papers by this authorSajib Kumar Saha, Corresponding Author Sajib Kumar Saha Sajib.Saha@csiro.au E Health Research Centre, Commonwealth Scientific and Industrial Research Organisation, Perth, AustraliaSearch for more papers by this authorAlve Mahamud Khan, Alve Mahamud Khan Computer Science and Engineering Discipline, Khulna University, Khulna, BangladeshSearch for more papers by this authorAbu Shamim Mohammad Arif, Abu Shamim Mohammad Arif Computer Science and Engineering Discipline, Khulna University, Khulna, BangladeshSearch for more papers by this authorYogesan Kanagasingam, Yogesan Kanagasingam E Health Research Centre, Commonwealth Scientific and Industrial Research Organisation, Perth, AustraliaSearch for more papers by this author Sultan Mohammad Mohaimin, Sultan Mohammad Mohaimin Computer Science and Engineering Discipline, Khulna University, Khulna, BangladeshSearch for more papers by this authorSajib Kumar Saha, Corresponding Author Sajib Kumar Saha Sajib.Saha@csiro.au E Health Research Centre, Commonwealth Scientific and Industrial Research Organisation, Perth, AustraliaSearch for more papers by this authorAlve Mahamud Khan, Alve Mahamud Khan Computer Science and Engineering Discipline, Khulna University, Khulna, BangladeshSearch for more papers by this authorAbu Shamim Mohammad Arif, Abu Shamim Mohammad Arif Computer Science and Engineering Discipline, Khulna University, Khulna, BangladeshSearch for more papers by this authorYogesan Kanagasingam, Yogesan Kanagasingam E Health Research Centre, Commonwealth Scientific and Industrial Research Organisation, Perth, AustraliaSearch for more papers by this author First published: 01 June 2018 https://doi.org/10.1049/iet-ipr.2017.0685Citations: 5AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract Age-related macular degeneration (AMD) is one of the main reasons for visual impairment worldwide. The assessment of risk for the development of AMD requires reliable detection and quantitative mapping of retinal abnormalities that are considered as precursors of the disease. Typical signs of the latter are the so-called drusen that appear as yellowish spots in the retina. Automated detection and segmentation of drusen provide vital information about the severity of the disease. The authors propose a novel method for the detection and segmentation of drusen in colour fundus images. The method combines colour information of the object with its boundary information for the accurate detection and segmentation of drusen. To perform non-uniform illumination correction and to minimise inter-subject variability a novel colour normalisation method has been proposed. Experiments are conducted on publicly available STARE and ARIA datasets. The method achieves an overall accuracy of 96.62% which is about 4% higher than the state-of-the-art method. The sensitivity and specificity of the proposed method are 95.96 and 97.64%, respectively. 1 Introduction Age-related macular degeneration (AMD) is a retinal complication, causing abnormalities in the retina and is a leading cause of visual deficiency and irreversible blindness in the developed world [1]. Usually, aged people who are older than 50 years get affected by AMD [1]. The early stage of AMD is asymptomatic, but small bright yellowish or white lesion, called drusen [2], can be revealed through an examination of the retina. Fig. 1 shows an exemplary fundus image of a normal eye and an AMD (drusen) affected eye. Fig. 1Open in figure viewerPowerPoint Colour fundus image (a) Normal eye, (b) Drusen affected eye An increase in the size or number of drusen is a sign of the progression of the disease, leading eventually to the presence of hemorrhages (wet AMD) or to the development of geographic atrophy (late dry AMD) [3]. Presently, there is no approved treatment to recover from AMD; however, its progression can be delayed through treatments [3]. Routine retinal examinations and long-term follow-ups are essential for the proper care and management of AMD. Optical imaging apparatus known as Fundus camera is used to capture the retina of the eye, which is then used to assess the AMD. Typically fundus images are evaluated by specialised professionals to assess the drusen characteristics and other relevant pathologies. Large parts of the image analysis process are still done by hand, which contribute to increasing the workload for ophthalmologists [4]. Therefore, there is need to develop automatic processes that could help minimise the workload. On that perspective in this study, we propose an automated method for the detection and segmentation of drusen from colour fundus image. The method combines the colour information of the object with its boundary information and ensures accurate detection and segmentation of drusen. A novel colour correction method is also proposed to ensure robust computation of colour information in the presence of non-uniform illumination and demographic diversity of colours in the fundus images. The method has been tested on publicly available STARE [5] and ARIA [6] datasets and compared against Mittal and Kumari's method [2]. Mittal and Kumari's method [2] was found to outperform other methods and thus is considered as the state-of-the-art method for comparison here. The organisation of this paper is as follows: Section 2 describes recent methods for automated detection of drusen/AMD, the proposed method is explained in Section 3, Section 4 describes the experimental results and discussions. 2 Related works Automated detection of AMD features has been an active area of research over the last two decades and a large number of methods have been developed to date using computer vision and image processing techniques. Here we provide a brief review of some of the recent methods. Remeseiro, Beatriz et al. proposed a top down approach for early AMD detection [7]. The method used Gaussian template matching [8] for drusen detection. A green channel of the colour fundus image was selected for drusen detection. The suspected drusen area was obtained using two different templates: (1) circular template and (2) Gaussian template. A region growing segmentation was performed on the suspected drusen area to remove falsely detected drusen. Mora et al. [9] used gradient-based segmentation to detect drusen in the image. Uneven illumination correction and contrast enhancement were performed as pre-processing prior to pathology detection. Akram et al. [10] applied adaptive thresholding to extract candidate drusen area. Prior to applying adaptive thresholding a set of pre-processing was done including removal of noise using HSI space, morphological closing to suppress dark regions, contrast limited adaptive histogram equalisation to enhance contrast and Gabor filtering [11] to extract all possible bright regions. The low adaptive threshold was then applied to obtain the binary candidate drusen region. Support vector machine (SVM) [12] was applied at the end to find out true drusen. Raza et al. [13] considered Naïve Bayes classifier along with SVM to classify drusen and non-drusen area. Gabor filtering was used to eliminate confusing drusen area. Zheng et al. [14] performed illumination and colour correction to achieve more robust detection of AMD pathologies. They incorporated optimal colour descriptors and robust multiscale local image descriptors in the drusen detection process. Wong et al. [15] applied K-means [16] clustering along with SVM to detect drusen. Bhuiyan et al. [17] relied on region growing based segmentation and objects edge information to detect drusen. In [18, 19], Mookiah et al. [20] learnt wavelet features [21], performed entropy measurements and finally applied SVM to classify drusen. Prasath et al. [22] used grey-level co-occurrence matrix [23] based textural features for the detection of drusen. In [24], Van Grinsven et al. explicitly applied machine learning based technique for the detection of drusen. Kumari et al. used Otsu's thresholding and morphological operation in [25] to detect drusen in order to detect AMD. Sundaresan et al. [26] proposed a super candidate based approach for the detection and discrimination of soft and hard drusen. Retinal image was first pre-processed with sigmoid function and histogram equalisation. The median filter was applied subsequently to obtain foreground and background. Finally, adaptive thresholding was performed on the foreground image. Very recently Mittal and Kumari have proposed a potential method in [2] for the detection and segmentation of drusen. The method does not require any complex machine learning approach rather relies on simple gradient magnitude thresholding and hysteresis tracking; however, outperforms other state-of-the-art methods. A summary of the sensitivity and specificity obtained by different methods along with their drawbacks are provided in Table 1. Table 1. Summary of the recently developed automated methods for the detection of drusen/AMD Publications Performance measures, % Drawbacks Remeseiro et al. [7] Se = 83 Sp = 87 Method was applied only on four images. Mora et al. [9] Se = 68 Sp = 96 Kappa = 59 Sensitivity achievement was too low. Akram et al. [10] Se = 95 Sp = 98.4 Ac = 97 Performed on drusen only images. Grading was not performed. Raza et al. [13] Se = 97 Sp = 99 Ac = 98 Performed on drusen only images. Grading was not performed. Zheng et al. [14] Se = 82 Sp = 67 Ac = 80 Grading was not performed. Performed on drusen only images. Wong et al. [15] Ac = 95.46 Performed on drusen only images. Bhuiyan et al. [17] Se = 74.94 Sp = 81.17 Ac = 82.14 Only 12 images were used to quantify drusen area, achieved sensitivity was low. Van Grinsven et al. [24] Se = 85 Sp = 96 ICC = 91 & 96 No pre-processing was done which may lead to a wrong and inconclusive result. Missed some drusen area. Mookiah et al. [20] Ac = 90.19 Se = 88.89 Sp = 91.48 Extracted only greyscale features, the test was performed on drusen only images. Mookiah et al. [27] PPV = 91.48 Se = 91.11 Sp = 96.30 Ac = 93.70 PPV = 96.09 Quantification of drusen was not performed. The test was performed on drusen only image. Prasath et al. [22] Ac = 98.05 Drusen was not quantified. Performed on drusen only images. Kumari et al. [25] Se = 95 Sp = 97 Ac = 96.32 MCC = 95.56 PPV = 97.23 Misses some drusen. Performed on drusen only images. Sundaresan et al. [26] Se = 80 Sp = 90 Ac = 92.57 Test performed on drusen only images. Mittal and Kumari [2] Se = 95.36 Sp = 99.56 Ac = 96.68 Detects some other retinal pathological object as drusen. Se = sensitivity; Sp = specificity; Ac = accuracy. One major drawback of these works is that they are tested on drusen only images. Our experiments on Mittal and Kumari's method reveal that when other pathologies are also present along with drusen in the image they are likely to be misclassified as drusen and this is principally due to not incorporating colour information properly into the system. 3 Proposed method We aim to detect and segment drusen using colour and boundary information of the object. The proposed method has been inspired by the success of Mittal and Kumari's method [2]. Instead of being simple, the Mittal and Kumaris method outperforms other state-of-the-art methods. The method achieves an overall accuracy of 96.7%; however, this is for drusen only images. Our experiments reveal that when other pathologies are also present in the image along with drusen, Mittal and Kumari's method is very likely to falsely detect/segment them as drusen, even though their colour appearances are different from drusen. The main reason behind this false detection/segmentation is that Mittal and Kumari's method does not incorporate colour information into the system. Typically colour is considered as an important visual clue to detect and differentiate drusen and such pathologies by human graders. While colour provides useful information, the use of colour for automated analysis is always challenging [18] because the same colour may look different under different lighting conditions, at the same time there is also variance in pathology colours depending on demography [28]. On that perspective, for robust computation of colour values of the pathologies, we propose a two-step approach of colour normalisation. Following colour normalisation, drusen segmentation in the image is performed by explicitly using colour values. A boundary detection of the pathologies is also performed using Mittal and Kumari's method [2], which is then combined with the colour based segmentation results. A block diagram of the proposed system is shown in Fig. 2. Fig. 2Open in figure viewerPowerPoint Block diagram of the proposed system 3.1 Colour normalisation We propose a two-step approach of colour normalisation to eliminate non-uniform illumination [28] and inter patient colour variability [29]. In the first step, non-uniform and poor illumination is corrected and in the second step unified colour transformation is performed. 3.1.1 Non-uniform/poor illumination correction Illumination correction is performed in the luminance channel only. First, the RGB colour fundus image is transformed into the LAB colour space [4]. It is worth mentioning, LAB colour space is more correlated with human vision than other colour spaces such as RGB and CYMK [30], and is frequently used for object detection and classification of the natural scene [31]. Following transformation in LAB space, we implement the background subtraction method [19] for illumination correction. Let, IL is the luminance channel, and IC1 and IC2 are the chrominance channels of the image after RGB to LAB conversion and before illumination correction. Let IL-cor be the luminance corrected image, which is computed as below (1)where is the background image generated after applying a Gaussian blur with a window size of w × w on . No processing is done at this stage on IC1 and IC2. Following illumination correction on the luminance channel, the image is transformed back to the RGB space. Fig. 3 shows a typical non-uniform illuminated image and its corresponding illumination corrected image. Fig. 3Open in figure viewerPowerPoint Sample image (a) Original, (b) After illumination correction 3.1.2 Unified colour correction Weighted Von Kries model [32] is applied to perform unified colour correction to minimise the inter subject variability of colours between fundus images. We propose here a novel approach to compute the weights of the model in the context of retinal imaging. In [32], the area of the segmented memory colour objects and the colour difference between the elliptic model of the memory colour object and the segment representative is used to compute the weights. In this work, we rely on mean RGB values of the optic disk (OD) and blood vessels to compute the weights. Let and be the mean RGB values of the OD and blood vessels, respectively. Then the colour correction for an input pixel i of the illumination corrected image I, having intensity [R, G, B] is corrected to [Rc, Gc, Bc] using the following formula: (2) (3) (4)Here and are the average of the mean RGB values of the OD and blood vessels computed over 1000 images from the Kaggle dataset [33]. Fig. 4 shows the unified colour corrected image of Fig. 3b. Fig. 4Open in figure viewerPowerPoint Colour normalised photograph of Fig. 3a 3.1.3 OD segmentation Detection of the OD is a general step in the automatic extraction of the anatomical structures of the retina. In an image of a healthy retina, the OD appears to be approximately circular, roughly one-sixth the width of the image in diameter, brighter than the surrounding area and is the convergent area of the blood vessels network [34]. All of the mentioned properties are typically considered for the detection of OD [7]. In this work, we have detected the OD using the Hough transform based method proposed in [7], which considered applying Sobel or Canny mask for the detection of edges and finally Hough transform for the detection of circles. In our implementation, we have used the Sobel mask for the detection of edges, as it has been verified in [7] that the Sobel mask performs better than the Canny mask. 3.1.4 Blood vessels segmentation Retinal blood vessels segmentation is an important step for many retinal image analysis tasks [35] and is an important pre-processing step here to perform unified colour correction. In this work for the segmentation of blood vessels, we implement the method proposed in [36]. Here we provide a brief review of the method. At each pixel position, a window of size W × W pixels is considered and the average grey level is computed as . Twelve lines of length W pixels oriented at 12 different directions (angular resolution of 15°) passing through the centred pixels are considered and the average of grey levels of pixels along each line is computed. The line with the maximum value is called the 'winning line' and its value is defined as . The line response at a pixel is then computed as below (5)The generalised line detector is computed as below (6)In the combination process, the same weight is assigned for each scale and the final segmentation is the linear combination of line responses of different scales. The response at each image pixel is defined as below (7)where is the number of scales used, is the response of the line detector at scale L, and is the value of the inverted green channel at the corresponding pixel. The basic line detector works on the inverted green channel of a retinal image where the vessels appear brighter than the background. The original green channel is included in the combination since it provides additional information to discriminate the proximity between the blood vessels and other structures such as pathologies and the OD. 3.2 Generation of drusen probable map In this stage, segmentation of drusen is performed using the colour information and a drusen probable binary map is generated. The given image is the first colour normalised and then is transformed into the LAB space. For each pixel of the LAB image a colorometric distance from the reference chromaticity point Lavg, Aavg, Bavg is measured as below (8)If the measured distance is less than a predefined threshold T, the pixel is considered as drusen pixel, otherwise not. More specifically, we perform the following test on each pixel of the LAB image to generate the drusen probable binary map (9)The values of Lavg, Aavg, Bavg, and T were determined experimentally. A set of N = 20 drusen only images {I1, I2, …, IN} from STARE and ARIA datasets [5] were used for the experiment. An experienced grader outlined the pathology of these images. For each image of the set, the mean LAB values of the drusen were computed. Lavg, Aavg, Bavg were computed by taking the average of these values over the 20 images. Let Di was the set of drusen pixels for image Ii, then for each pixel p(x, y) Di, the following colorometric distance was measured (10)We computed the maximum colorometric distance per image, which was then used to compute the threshold T as below (11) 3.3 Boundary extraction of drusen Drusen boundary is extracted using Mittal and Kumari's method [2]. A list of pre-processing including homomorphic filtering for removing artefact, green channel selection to ensure better contrast and Gaussian smoothing for noise removal are performed prior to boundary extraction. Let gGF be the image obtained after these pre-processings. To extract the boundary gGF is filtered with a Sobel mask in the horizontal and vertical directions to get the first derivative of image intensity in these directions. These two images are then used to compute the magnitude, gf, and direction, of the edge gradient. Mathematically these operations are defined as below (12) (13) (14)After obtaining gradient edge magnitude and direction, non-maxima suppression is performed to remove unwanted pixels which may not constitute the edge. Non-maxima suppression still gives some wrong edge pixels due to noise and colour variation across the image, and gradient magnitude thresholding is performed to filter out those pixels. Gradient magnitude threshold operation can be mathematically expressed as below (15)where is the high threshold and is the low threshold. The edge obtained after the thresholding can be represented as below (16)where T is a predefined threshold. Following gradient magnitude threshold, still we may have some weak edge pixels that are falsely detected, and edge tracking by hysteresis is performed to suppress those unwanted weak edge pixels. Let E is the image generated after edge tracking. Iterative edge thinning, end point recovery, end point labelling and boundary detection by edge linking operations are performed on E subsequently, to generate the drusen boundary. It is worth mentioning while the above operations are performed to get the boundary of drusen only, our experiments reveal that the boundary of some other pathology can also be unwillingly detected if they are present in the image. 3.4 Drusen segmentation using colour and border information In this step, the results obtained in Sections 3.2 and 3.3 are combined. First, we fill the inside region of the object boundary obtained in Section 3.3 by applying region filling technique [37]. Let IR is the binary image obtained after region filling, and IDC is the image obtained in Section 3.2; then the drusen segmented image ID is generated based on the following operation: (17)where is the logical AND operator, which performs pixel by pixel logical AND operation between images. 4 Experiments and results A total of 72 images comprising 60 images from STARE dataset [5] and 12 images from ARIA dataset [6] have been used for the experiment. Out of 60 images from STARE dataset, 25 are drusen only images and rest have other pathologies as well along with drusen. Two of the 12 images from ARIA have other pathologies along with drusen. An experienced grader outlines the pathology of these images. The proposed method and Mittal and Kumari's method [2] are used for the automated segmentation of drusen; which is then compared against the manual segmentation. For the drusen only images, we have not observed any difference between the proposed and Mittal and Kumari's method that would be relevant in practice (Table 2). Table 2. Performance comparison of the proposed drusen detection method with Mittal and Kumari's method for drusen only images Drusen category Number of images Total number of drusen in ground truth Number of correctly detected drusen Accuracy Proposed method Mittal and Kumari's method Proposed method Mittal and Kumari's method Small ( 125 µm) 18 1245 1091 1091 87.63 87.63 Few images have more than one category of drusen. Fig. 5 shows an example segmentation of drusen by these two methods for drusen only case. Fig. 5Open in figure viewerPowerPoint Segmentation of drusen (a) Proposed method, (b) Mittal and Kumari's method [2] For the images which had other pathologies as well with drusen, the proposed method performs better than the Mittal and Kumari's method. Fig. 6 shows the boundary of detected drusen by these two methods on an exemplary image that have other pathology as well with drusen. The red circles in Fig. 6b show some of the wrongly detected pathologies as drusen by Mittal and Kumari's method, which are not detected by the proposed method. Fig. 7 shows another example of segmented drusen by automated methods, in comparison with human grander. Fig. 6Open in figure viewerPowerPoint Detected boundary of drusen (a) Proposed method, (b) Mittal and Kumari's method [2] Fig. 7Open in figure viewerPowerPoint Drusen boundary by (a) Mittal and Kumari's method, (b) Proposed method, (c) Human grader For the quantitative comparison, sensitivity, specificity, accuracy, Matthew's correlation coefficient (MCC) and positive predictive value (PPV) are computed by comparing the detection result with ophthalmologists' hand drawn ground truth. In line with [2], sensitivity, specificity, accuracy, MCC, and PPV are computed as below (18) (19) (20) (21) (22)Here, is the number of correct predictions that a drusen pixel is positive, is the number of incorrect predictions that a drusen pixel is negative, is the number of incorrect predictions that a drusen pixel is positive and is the number of correct predictions that a drusen pixel is negative. Table 3 shows the overall sensitivity, specificity, accuracy, MCC and PPV obtained by these two methods. Tables 4 and 5, respectively, show the average sensitivity, specificity, accuracy, MCC, and PPV obtained by these two methods in STARE and ARIA datasets. Figs. 8 and 9 and Table 6 detail the obtained quality measure per image by these two methods. It is worth mentioning since we have obtained identical results by these two methods for drusen only case, hence the results obtained for drusen with other pathology images are only reported in the tables and figures. Table 3. Overall sensitivity, specificity, accuracy, MCC and PPV obtained by the proposed and Mittal and Kumari's methods Mittal and Kumari Proposed method Mean Std. Mean Std. Sensitivity, % 95.50 4.79 95.96 4.17 Specificity, % 95.22 1.36 97.64 0.93 Accuracy, % 92.84 1.25 96.62 0.94 MCC, % 95.41 2.56 96.84 1.91 PPV, % 92.95 1.25 96.68 0.84 Table 4. Average sensitivity, specificity, accuracy, MCC and PPV obtained by the proposed and Mittal and Kumari's methods in STARE dataset Mittal and Kumari Proposed method Mean Std. Mean Std. Sensitivity, % 96.06 5.02 96.44 4.37 Specificity, % 95.18 1.42 97.64 0.83 Accuracy, % 92.78 1.31 96.71 0.97 MCC, % 95.62 2.68 97.04 1.95 PPV, % 92.90 1.32 96.81 0.87 Table 5. Average sensitivity, specificity, accuracy, MCC and PPV obtained by the proposed and Mittal and Kumari's methods in ARIA dataset Mittal and Kumari Proposed Method Mean Std. Mean Std. Sensitivity, % 85.64 0.78 87.54 0.58 Specificity, % 95.96 0.29 97.63 0.26 Accuracy, % 93.83 0.21 94.99 0.07 MCC, % 91.80 0.46 93.25 0.66 PPV, % 93.96 0.11 94.47 0.59 Table 6. Obtained sensitivity, specificity, accuracy, MCC and PPV (in percentage) per image by the proposed and Mittal and Kumari's methods in ARIA dataset Mittal and Kumari Proposed method Image 1 Image 2 Image 1 Image 2 Sensitivity 85.09 86.19 87.95 87.13 Specificity 96.16 95.76 97.44 97.81 Accuracy 93.68 93.98 95.03 94.95 MCC 91.47 92.12 92.78 93.722 PPV 93.88 94.03 94.89 94.05 Fig. 8Open in figure viewerPowerPoint Sensitivity, specificity, and accuracy obtained per image by different methods in STARE dataset. Each dot represents an image Fig. 9Open in figure viewerPowerPoint MCC and PPV obtained per image by different methods in STARE dataset. Each dot represents an image 5 Discussions and conclusion We propose a new method for the automated detection and segmentation of drusen from the colour fundus image in the context of automated analysis of AMD. The method focuses on the robust computation of colour information for more accurate segmentation of pathology which has an associated colour. A novel two-step approach of colour normalisation has been proposed here that overcomes non-uniform/poor illumination and minimises the inter-subject variability of colours in retinal images. Fig. 10 shows that the colorometric difference (ΔE) between the drusen of the individual image and the reference chromaticity point of drusen, minimises and becomes more stable when a colour correction is also performed following illumination correction. Thus, the proposed colour correction creates the environment so that a unique threshold to segment the colour object can be more reliably defined and used. Fig. 10Open in figure viewerPowerPoint Colorometric difference (ΔE) between the drusen of individual image and the reference chromaticity point of drusen, prior to and after colour correction by the proposed method The colour-based segmentation results are finally combined with the extracted drusen boundary by the Mittal and Kumari's method [2] to produce more accurate results. The above operations add about 2.5 s/image extra processing time [on an Intel Core i7 (CPU 2.90GHz, RAM 8 GB) machine, on images from ARIA and STARE dataset] on Mittal and Kumari's method. Of this added time 80% is spent on performing vessel segmentation, 10% is spent on OD segmentation and rest for other computation. Even though in retinal pathology detection accuracy is more crucial than computational time, it is worth mentioning that if required, reduction of this added time is also possible through applying the fast vessel segmentation method. The method has been tested on images from STARE and ARIA datasets and is found to have an accuracy of 96.62% with sensitivity, specificity, MCC and PPV of 95.96, 97.64, 96.84 and 96.68%, respectively. To the best of our knowledge, Mittal and Kumari's method [2] outperforms other methods in the domain and achieved an accuracy of 96.68% with sensitivity and specificity of 95.36 and 99.56%, respectively. However, this is for drusen only images, and our experiments reveal that when other pathologies are also present in the images along with drusen, Mittal and Kumari's method finds the boundary of these pathologies as well. For drusen with other pathology images, we obtained an accuracy of 92.84% with sensitivity, specificity, MCC and PPV of, respectively, 95.50, 95.22, 95.41 and 92.95% by the Mittal and Kumari's method. Our proposed method ensures robust incorporation of colour information into the system and thus ensures an overall accuracy improvement of 4% in comparison with Mittal and Kumari's method. 6 References 1https://nei.nih.gov/health/maculardegen/armd_facts, Last accessed: 20 January 2017 2Mittal D., and Kumari K: 'Automated detection and segmentation of drusen in retinal fundus images', Comput. Electr. Eng., 2015, 47, pp. 82– 95 3Phan T.V., Seoud L., and Chakor H. et al: 'Automatic screening and grading of age-related macular degeneration from texture analysis of fundus images', J. Ophthalmology, 2016, 2016, pp. 1– 11 doi:10.1155/2016/5893601 4Mookiah M.R.K., Acharya U.R., and Chua K. et al: 'Computer-aided diagnosis of diabetic retinopathy: a review', Comput. Biol. Med., 2013, 43, (12), pp. 2136– 2155 5 STARE dataset.: Available at: http://cecas.clemson.edu/~ahoover/stare/, Last accessed: 20 January 2017 6 ARIA dataset.: Available at: http://www.eyecharity.com/aria_online, Last accessed: August 24 2014 7Remeseiro B., Barreira N., and Calvo D. et al: 'Automatic drusen detection from digital retinal images: AMD prevention', Comput. Aided Syst. Theory-EUROCAST, 2009, 2009, pp. 187– 194 8Russ J.C.: ' The image processing handbook' ( CRC Press, Inc., Boca Raton, 1999, 3rd edn.) 9Mora A.D., Vieira P.M., and Manivannan A. et al: 'Automated drusen detection in retinal images using analytical modelling algorithms', Biomed. Eng. Online, 2011, 10, (1), p. 59 10Akram M.U., Mujtaba S., and Tariq A.: ' Automated drusen segmentation in fundus images for diagnosing age related macular degeneration'. Int. Conf. on Electronics, Computer and Computation (ICECCO), 2013, pp. 17– 20 11Sung J., Bang S.Y., and Choi S.: 'A Bayesian network classifier and hierarchical Gabor features for handwritten numeral recognition', Pattern Recogn. Lett., 2006, 27, pp. 66– 75 12Cortes C., and Vapnik V.: 'Support-vector networks', Mach. Learn., 1995, 20, (3), pp. 273– 297 13Raza G., Rafique M., and Tariq A. et al: ' Hybrid classifier based drusen detection in colored fundus images'. IEEE Jordan Conf. on Applied Electrical Engineering and Computing Technologies (AEECT), 2013, pp. 1– 5 14Zheng Y., Vanderbeek B., and Daniel E. et al: ' An automated drusen detection system for classifying age-related macular degeneration with color fundus photographs'. 10th Int. Symp. on Biomedical Imaging (ISBI), 2013, pp. 1448– 1451 15Wong D.W.K., Liu J., and Cheng X. et al: ' THALIA-An automatic hierarchical analysis system to detect drusen lesion images for AMD assessment'. 10th Int. Symp. on Biomedical Imaging (ISBI), 2013, pp. 884– 887 16Forgey E.: Cluster analysis of multivariate data: efficiency vs. Interpretability of classification', Biometrics, 1965, 21, (3), pp. 768– 769 17Bhuiyan A., Kawasaki R., and Sasaki M. et al: 'Drusen detection and quantification for early identification of age related macular degeneration using color fundus imaging', J. Clin. Exp. Ophthalmol., 2013, 4, (305), p. 2 18Sande V.D., Koen T.G, and Snoek C.: 'Evaluating color descriptors for object and scene recognition', IEEE Trans. Pattern Anal. Mach. Intell., 2010, 32, (9), pp. 1582– 1596 19Gonzalez R.C., and Woods R.E.: ' Digital image processing' ( Pearson Prentice Hall, 2008, 3rd edn.) 20Mookiah M.R.K., Acharya U.R., and Koh J.E.W. et al: 'Automated diagnosis of age-related macular degeneration using greyscale features from digital fundus images', Comput. Biol. Med., 2014, 53, pp. 55– 64 21Chui C.K.: ' An introduction to wavelets' ( Elsevier, 2016) 22Prasath A.R., and Ramya M.M.: 'Detection of macular drusen based on texture descriptors', Res. J. Inf. Technol., 2015, 7, (1), pp. 70– 79 23Haralick R.M., and Shanmugam K.: 'Textural features for image classification', IEEE Trans. Syst. Man Cybern., 1973, 3, (6), pp. 610– 621 24Grinsven V., Mark J.J.P., and Lechanteur Y.T.E. et al: 'Automatic drusen quantification and risk assessment of age-related macular degeneration on color fundus images automatic drusen quantification on color fundus images', Invest. Ophthalmol. Vis. Sci., 2013, 54, (4), pp. 3019– 3027 25Kumari K., and Mittal D.: 'Automated drusen detection technique for age-related macular degeneration', J. Biomed. Eng. Med. Imaging, 2015, 2, (1), p. 18 26Sundaresan V., Ram K., and Selvaraj K. et al: ' Adaptive super-candidate based approach for detection and classification of drusen on retinal fundus images'. Proc. Ophthalmic Medical Image Analysis Second Int. Workshop, OMIA 2015, Munich, Germany, pp. 81– 88, available at: http://ir.uiowa.edu/omia/2015_Proceedings/2015/11 27Mookiah M.R.K., Acharya U.R, and Koh J.E.W. et al: 'Decision support system for age-related macular degeneration using discrete wavelet transform', Med. Biol. Eng. Comput., 2014, 52, (9), pp. 781– 796 28Saha S.K., Xiao D., and Kanagasingam Y.: 'A novel method for correcting non-uniform/poor illumination of color fundus photographs', J. Digit. Imaging, 2017, pp. 1– 9 29Saha S.K., Xiao D., and Fernando B. et al: 'Deep learning based decision support system for automated diagnosis of age-related macular degeneration (AMD)', Invest. Ophthalmol. Vis. Sci., 2017, 58, (8), p. 25 30Fairchild M.D.: ' Color appearance models' ( John Wiley & Sons, Chichester, 2005, 2nd edn.) 31Baldevbhai P.J., and Anand R.S.: 'Color image segmentation for medical images using L*a*b* color space', IOSR J. Electron. Commun. Eng., 2012, 1, (2), pp. 24– 45 32Moreno A., Fernando B., and Kani B. et al: ' Color correction: a novel weighted Von Kries model based on memory colors', Proc. Int. Workshop on Computational Color Imaging, Milan, Italy, April 2011, pp. 165– 175 (LNCS, 6626) 33 Kaggle dataset for diabetic retinopathy detection: Available at: https://www.kaggle.com/c/diabetic-retinopathy-detection/data, Last accessed: 20 January 2017 34Park M., Jin J.S., and Luo S.: ' Locating the optic disc in retinal images'. Int. Conf. on Computer Graphics, Imaging and Visualisation, 2006, pp. 141– 145 35Saha S.K., Xiao D., and Frost S. et al: 'A two-step approach for longitudinal registration of retinal images', J. Med. Syst., 2016, 40, (12), p. 277 36Nguyen U.T.V., Bhuiyan A., and Park L.A.F. et al: 'An effective retinal blood vessel segmentation method using multi-scale line detection', Pattern Recognit., 2013, 46, (3), pp. 703– 715 37Soille P.: ' Morphological image analysis: principles and applications' ( Springer Science & Business Media, Berlin, 2003, 2nd edn.) Citing Literature Volume12, Issue6June 2018Pages 919-927 FiguresReferencesRelatedInformation
Referência(s)