Abstract
Digital fundus photography plays a major role in the diagnosis of different retinal pathologies like hypertension, diabetic retinopathy and Glaucoma. To identify abnormal components on the retina, retinal features should be detected accurately. Retinal vessel structure is one of the important landmarks of the retina. So precise detection of retinal vessel structure is imperative. This paper presents a simple, robust retinal vessel extraction approach based on the line detectors and morphological operations. As vessel detection is basically a problem of a line detection, the green channel retinal image is applied to morphological opening using a line as structuring element. The resultant image is again applied with the line detectors and thresholded using Otsu’s thresholding. The proposed algorithm overcomes the fundamental issues of scale and orientation avoiding the need of multiple thresholds with improved values of performance measure as compared to the state of the art techniques. The proposed algorithm is applied on 3 standard databases-HRF (healthy and Diabetic), DIARETDB1 and DRIVE. Area under the ROC curve (AUC) of 97% was achieved with 91% Sensitivity and 97% Specificity for DRIVE dataset. The proposed algorithm achieved an Accuracy of 97%, Sensitivity of 85 % and Specificity of 97% for HRF database. On DIARETDB1 database too observed very good results.
Introduction
Retinopathies related to diseases such as diabetes and cardiovascular diseases have an ever increasing importance as a cause of blindness and visual loss [1]. Current statistics show, 39 million people are blind and 246 million have low vision worldwide. Timely diagnosis of eye diseases can prevent blindness [1, 2]. The global Vision 2020 initiative has an effect to prevent blindness, particularly from ocularinfections. Research indicates that at least 90% of these new cases could be avoided if there were proper and automatic screening and monitoring of the eyes. Abnormal features of the retina can be identified in-time with the aid of fundus photography through innovative image processing algorithms. Advanced medical therapies save a person from blindness [3].
Automated examination of retinal images plays a significant role in the diagnosis and analysis of different retinal pathologies like hypertensive, diabetic pathology and Glaucoma. Currently, a timely eye check-up appears to be the best method with nearly worldwide coverage of the people at risk [1]. Early detection of abnormal features, which are not often directly visible by clinical investigation, has the potential to reduce the global burden of diabetes and cardiovascular disease. Engineering tools such as digital image processing combined with innovative machine learning to allow identification and automated classification of features, lesions, and vascular changes in fundus images of the retina [2, 3].
A significant number of lesions have been found on the retina due to diabetics (progressive disease). The evolution of disease takes place from mild Non Proliferative Diabetic Retinopathy (NPDR) to Severe Proliferative Diabetic Retinopathy (PDR) [3]. Usually, first and the most common sign is microaneurysms or small haemorrhages – tiny capillary dilations. As the disease progresses, retinal capillaries start leaking fluid forming exudates – lipid deposits that appear as bright yellow/white lesions in the photograph. If the leakage is located around most acute vision, i.e. macula, it leads to sight threatening macular edema [1, 2]. Proliferative diabetic retinopathy (PDR) develops from obstructed capillaries which cause micro infarcts called soft exudates. This PDR stage quickly turns into Severe PDR where extensive lack of oxygen causes formation of small and fragile vessels either near the optic disc or in the retinal periphery. These newly developed vessels are disposed to bleed, resulting in neovascularisation, pre-retinal haemorrhages, fibrosis and ultimately detachment of the retina resulting in blindness [1].
In order to recognise the abnormalities like Microaneurysms, haemorrhages and exudates, the normal retinal landmarks namely blood vessels, optic disc and macula need to be detected accurately. Optic disc and macula detections are achieved with a greater accuracy in the earlier mentioned literature. Retinal blood vessels are combinations of arteries and veins, originating from the optic nerve towards the retina. These vessels normally possess maximum light reflectance on the centreline that seems more significant on arteries than veins and the oxygenated blood supply to arteries makes them brighter than veins [3]. In general, the methods of the literature for vessel detection try to extract the features like position, grey level profiles of vessels, linearity and steady variation in intensity lengthwise.
The vessels are connected in the retina, forming a binary tree like structure. Blood vessel detection suffers from many issues such as methods based on matched filters, suffer from the problems of scale and orientation [5, 15]. Accurate detection and extraction of blood vessel structure is still a challenging task as most of the methods mentioned in the literature fail to detect vessels in the presence of lesions producing false positives and reduced value of sensitivity, which lowers overall accuracy [7]. Thin vessel detection is another major problem faced by many of state of the art methods. In general, methods using edge detectors fail to recognize correct and false edges. Branching, bifurcation, and crossover in vascular structure can further complicate the profile model. Central Vessel reflex, variations in intensity profiles and low contrast images along with background noise increases the problems in most of the blood vessel detection algorithms [4].
This paper proposes an innovative, simple and robust method for retinal vessel extraction. The use of morphological opening and line detectors helps in enhancing all line paths of vessels equally avoiding the need of multiple thresholds and helping to extract thin vessels. It also overcomes the fundamental issues of scaling and orientation as it is working on multiple origins and size of images equally. The vessel extraction achieved by this method produces precise results even in the presence of lesions at an initial stage of Diabetic Retinopathy.
Related work
Blood vessel detection techniques can be broadly classified into four categories – matched filters, edge detectors, morphological operators and pattern recognition techniques. Chaudhari et al. [5] addressed the problem of detecting retinal blood vessels by approximating Grey level profile of a cross section of retinal blood vessel by a Gaussian shaped curve and applying the concept of matched filter detection to detect piecewise linear segment of blood vessels giving an accuracy of 0.8733. Joes Staal et al. [6] presented a ridge based vessel segmentation methodology from colour images of the retina on DRIVE and STARE database given an accuracy of 0.9442, 0.9516. An innovative spatially weighted fuzzyc-means (SWFCM) clustering algorithm for vessel detection with accuracy of 0.8911 on DRIVE presented by Giri Babu Kande et al. [7].
A Radius based Clustering Algorithm (RACAL) using distance based principle to map the distributions of the image pixels, is proposed by Salem et al. [8]. Yang et al. [9] presented an automatic hybrid method including the combination of mathematical morphology and a fuzzy clustering algorithm. Zhang et al. [10] generalized the classical matched filter and extended with a first-order derivative of the Gaussian (MF-FDOG) on DRIVE and STARE database and given 0.7120, 0.9724 and 0.9382 as sensitivity, specificity and accuracy. Diego Marín et al. [11] invented a method using a neural network (NN) scheme for pixel classification. A new system is proposed by Keith A. Goatman et al. [12] for detection of New Vessels on the Optic Disc.
Delibasis et al. [13] presented an automatic parametric model-based tracking algorithm for vessel segmentation and diameter estimation. Akram et al. [14] proposed retinal vessel extraction using multi-layered thresholding based technique producing an Accuracy of 0.9469. Fraza et al. [15] reported an automated method using a unique combination of techniques for vessel centre line detection and morphological bit plane slicing, methodology is evaluated on DRIVE and STARE databases, producing an average accuracy of 0.9430,sensitivity 0.7152 and specificity 0.9768, but method is dependent upon vessel centre line reflex and scale and orientation problem.
Montoro et al. [16] focused on studying the appearance of the retinal vascular network in different colour spaces like RGB and HSV to extract the most distinct vessel features and classify the retinal vascular network as arteries and veins. Chakraborti et al. [17] proposed a novel self-adaptive matched filter for retinal blood vessel detection, a synergistic mixture of the vesselness filter with high sensitivity and the matched filter with high specificity is obtained using orientation histogram. Shuangling Wang et al. [18] presented a retinal blood vessel segmentation algorithm based on feature and ensemble learning given sensitivity 0.8173, specificity 0.9733 and accuracy of 0.9767. Temitope Mapayi et al. [19] proposed a local adaptive thresholding technique depending on grey level co-occurrence matrix.
Proposed vessel segmentation method
The proposed blood vessel segmentation method is an approach based on the combination of morphological analysis with line detectors. Since vessels and lesions are morphologically different. Morphological opening is done using line as SE which will enhance line segments and suppress lesion part of the retina. The results of it will be further enhanced using line detectors and Morphological opening giving additional advantage. Then the resultant image is thresholded using Otsu’s thresholding followed by morphological cleaning. Experimental results have shown that proposed methodology achieved a high value of sensitivity even in the presence of lesions and it is free from the problems of scale and orientation.
Pre-processing
In pre-processing section the input retinal image is first converted into green channel as blood vessels are clearly seen into green channel as shown in Fig. 2(b). The retinal fundus image from specified database is given as an input image as shown in Fig. 2(a).
Image smoothing
Morphological opening helps to smooth out bright lesions and optic disc. Morphological opening operation with SE as a ‘disk’ of radius 5 using the following Equation (1).
Here f is pre-processed color image and b € SB, Where SB is structuring element of size ‘S’. This gives smooth regions for dark lesions, but it need contrast enhancement [20].
Contrast Limited Adaptive Histogram Equalization works on small sub sections in an image called ‘tiles’. Each tile’s contrast is enhanced by specifying ‘distribution’. Here we have preferred use of flat histogram distribution.
The equation of CLAHE is given by,
Where φ
w
is sigmoid function and given by,
φ f max and φ f min are the maximum and minimum intensity values of morphologically processed smooth green channel image respectively. m w and σ w are the mean and variance of intensity values within the window [20].
Morphological opening is applied on extracted green channel image in order to smooth out bright objects like white lesions and optic disc. The resultant image is shown in Fig. 2(c). But this image is an image with poor contrast so it is applied with the contrast enhancement technique i.e. CLAHE. Output of it is shown in Fig. 2(d). where contrast between blood vessels, background, and other lesions are enhanced.
Vessel boost
The demand for improvement in appearance of a vessel pattern arises due to lack of contrast which typically causes problem for thin vessels. Directional matched [5] and Gabor filters [14] are preferred for this drive, but we have preferred the use of morphological opening to enhance the vascular pattern and thin vessels. The morphological open operation is erosion followed by dilation, using the same structuring element for both operations. We have used ‘line’ as a structuring element to enhance blood vessels since vessel segmentation is basically a problem of the line detection.
Assuming the object X and the structuring element B, X is first eroded and further dilated by B. We need to fix the lengths of the line based on the selected database. Here we have kept the length of line from 2 to 5. Image opening is made in 45°, 90°, 135°, 180° and maximum of the response is found out from all directions in order to enhance vessels and subdue other components in the image.
The image obtained after considering the maximum response of opening operation is convolved with average filter. The average filter is applied on each pixel in image to smooth an image, which takes the average around neighboring pixels. The background image with homogenized background is obtained by subtracting response of average filter from the contrast enhanced image. The resultant image is then convolved with line detectors in 4 directions, i.e. horizontal, vertical, +45° and –45°. The masks for convolution are as shown in Fig. 3 below.
By performing a convolution operation, maximum of them is obtained. An inverted and resultant of line detector image results are added together to increase the brightness. Again, applying SE as a line of size 2 to 5 in all directions, i.e. 45°, 90°, 135°, 180° image is opened in all directions and maximum from all directions is found out as shown in Fig. 4(a).
The choices of morphological opening followed by the line detectors benefits in enhancement of the vessel structure as vessels are basically lines in nature. The vessels in dark region are present with less intensity which might not get reflected after thresholding. So all vessel pixels of resultant image are further enhanced by multiplying a suitable factor with an image. Thus enhancing all vessels pixels equally and avoiding the need of multiple thresholds. The resultant image is then applied with Otsu’s thresholding. The output image is as shown in Fig. 4(c).
Blood vessel extraction
The output image is logically ANDed with mask image from database in order to remove mask as shown in Fig. 5(a). The thresholded image still consists of noise as well non-vessel pixels, they are removed by using morphological ‘clean’ operation. Thus avoiding the need of further post processing algorithms for vessel extraction. The clear blood vessel extracted image as shown in Fig. 5(b).
Results
The Proposed algorithm is applied on standard databases such as DRIVE, HRF (Healthy and Diabetic) and DIARETDB1. The results for DRIVE and HRF databases are compared with ground truth images provided along with the dataset.
Materials
IMAGERET–Proposed algorithm’s performance is observed on publically available DIARETDB 1 database from IMAGERET project. Out of 89 colour fundus photograph’s 84 images are having signs of at least mild NPDR and all experts involved in annotation agreed that 4 images do not have any signs of DR. The images were acquired at the Kuopio University Hospital. The digital fundus camera with a FOV of 50 degrees was used to take images [22].
DRIVE – To evaluate the performance of the proposed vessel segmentation technique DRIVE database is used. The fundus images for the DRIVE database were acquired from a diabetic retinopathy screening drive in the Netherland. The database contains 400 images, people affected due to diabetics between the age group 25–90 were involved in the screening process. 40 images are arbitrarily chosen, no signs of DR are present in 33 images whereas 7 of them are showing mild DR signs. Each of the images is compressed with JPEG standard with a pixel resolution of 786×584 and 8 bits per pixel [23].
Chosen 40 images are divided into two sets – training and testing, containing 20 images in each. The test dataset is provided with two manual segmentations; one treated as ground truth and another one is used to compare results with computer generated data. A single vascular structure is provided for the images from training dataset. This dataset is specialized in vessel segmentation and makes two sets of images available, one for training and one for testing purposes [23].
HRF – The database is provided by the Pattern Recognition Lab, the Department of Ophthalmology, Germany and the Brno University of Technology, Brno. This database has been made available to support relative studies on automatic segmentation algorithms on retinal fundus images. The images are having resolution of 3504×2336 captured with Canon EOS 20D. The public database contains 15 images of, healthy retina, retina affected due to diabetic retinopathy, and of the glaucoma patients each. The segmented vascular structure is provided as a gold standard for each image of database, along with the masks with defined FOV. The group of specialists in the analysis of retinal images and expert doctors from ophthalmology clinics participated in generating the ground truths [24].
Evaluation parameters
TP is the value of True Positive pixels means vessel pixels present in both ground truth as well as segmented image. TN is the value of True Negative pixels i.e. pixels absent in both cases. FP is False Positive – pixels which are absent in ground truth, but present in segmented image and FN False Negative – the pixels present in the ground truth image but absent in segmented image. The parameter TPR should approach to unity, whereas the parameters FPR should be ideally zero. Sensitivity replicates the capability of an algorithm to detect the true vessel pixels. Specificity shows the capability to detect number of pixels which do not belong to vessel, expressed as 1–False Positive Rate.
Evaluation Parameters are as indicated in the following Table 1.
Vessel segmentation results
Table 2 summarizes vessel segmentation results of the proposed algorithm on DRIVE, HRF Diabetic and HRF Healthy databases. The results are compatible with the methods mentioned in earlier literature. The results are also observed great on 89 images from Diaretdb1 database. Results on DRIVE and HRF (both Diabetic and Healthy) database are obtained by pixel to pixel comparison of vessel segmented image and ground truth images from database document. Experimental results have shown the values of parameters like Accuracy and Specificity obtained using the proposed algorithm are similar to the other methods mentioned in literature rather it exceeds in some cases. It is worth mentioning here that the average value for sensitivity is improved greatly as compared to the state of the art methods in addition to it the value of specificity is greater in most cases than the other reported methodologies in literature.
Conclusions
The proposed work found very suitable for the extraction of blood vessel structure which is a major retinal landmark. The retinal vasculature is obtained using simple morphological operations and line detectors together avoiding the need of multiple thresholds. The advantage for the practical application is that the algorithm works on retinal images from multiple origins and can be used by different operators working with different equipment. It is important to point out that the parameters used for segmentation are invariant. The robustness with fast implementation and simplicity makes this work suitable for being implemented into a complete automated retinopathy screening process. The proposed algorithm has achieved a high value of sensitivity as compared to most of the earlier proposed methodologies.