INTRODUCTION:
Diabetic retinopathy is a microvascular complication of diabetes that transpires in human beings and is also a common cause which damages the retina of the diabetic patient’s eye2. The prevalence of retinopathy varies with the age of diabetes and the duration of the disease. For the detection of diabetic retinopathy, colour fundus photographs of the retina is captured and processed. If the symptoms are identified in earlier stage, then a proper treatment can be provided. The effective treatment of diabetic retinopathy can impede the progression of the diseases where many patients are not aware of this disease. It is to be pointed out that at least 90% of the new cases of diabetic retinopathy could be reduced by giving proper treatment and regular monitoring of the eye.
Diabetic retinopathy can be diagnosed by the defects that occur in the retina. The defects may include microaneurysms, haemorrhages and exudates. Microaneurysms are the primary abnormality occurring in the eye because of diabetes. Ophthalmologist recognizes diabetic retinopathy by exploring various symptoms such as tiny, dark red spots or haemorrhages that may occur alone or in clusters. Haemorrhages are round in shape, which are found in deep layer of the retina. Exudates are of two types: hard exudates and soft exudates. Hard exudates are identified as fat and protein leaking out from the blood vessel, which prevents light from reaching the retina and causes visual impairment. The spots termed as soft exudates are often identified as the severe stages of diabetic retinopathy which is also as called cotton wool spots. These are caused by nerve fibre layer blockage and further the local nerve fibre axons gets blown up.
In this paper, a hybrid method that diagnose diabetic retinopathy through blood vessel extraction and exudates Identification from 2D fundus image is presented. The main aim is to take advantage of few effective existing algorithms and subsequently manipulate it to diagnose diabetic retinopathy. The pre-processing of the image consists of firstly extracting the green channel of the image followed by conversion of the image to greyscale. Successively, the converted grayscale image is enhanced by applying adaptive histogram equalization where it increases the contrast of the respective image. Once the pre-processing is performed, then the features of the image is extracted through maximally stable extremal regions (MSER) 2 where it detects the blobs present in the image. The main purpose of MSER is to find the correspondences between the image elements of two similar images with different viewpoints.
From the six region detectors3 (Harris-affine, Hessian-affine, MSER, edge-based regions, intensity extrema, and salient regions) that were studied and analysed, the performance of MSER is more superior to the other methods. The said comparison is done based on various parameters such as region density, region size, viewpoint change, scale change, blurriness and light change. The effectiveness of MSER is tabulated in Table 1.
Table 1. MSER effectiveness compared to other feature detection methods.
|
Parameters |
Effectiveness |
|
Region Density |
Detects 2600 for a blurred scene and 230 or light changed scene |
|
Region Size |
Detects smaller regions |
|
Viewpoint change |
Best in original images |
|
Scale change |
Second after Hessian-affine detector |
|
Blur |
Most sensitive |
|
Light change |
Best |
For detecting and extracting the blood vessels inside the eye, Kirsch’s template is well utilized so that the edges of the blood vessels are obtained. The developed hybrid method will aid the ophthalmologist to identify and diagnose diabetic retinopathy in an effective and efficient way.
LITERATURE SURVEY:
Automatic detection of hard and soft exudates using histogram thresholding4 is described. The fundus image is pre-processed using CIELab color space and then used mathematical morphology for the detection of hard and soft exudates. Detection of diabetic retinopathy through lesions5 is proposed. Here, the proposed method uses two features such as color and shape of the lesion to detect the respective lesions. The sensitivity of the proposed system is claimed to be 79.62%. An automated optic disk detection in retinal images with diabetic retinopathy6 is proposed by the authors. The detection procedure is divided in two independent methodologies. One is location method that consists of maximum difference method, maximum variance method and a low-pass filter method which works on the green channel providing the best contrast. The other method is a boundary segmentation methodology that estimates a circular approximation of the optical disk which applies mathematical morphology, edge detection and the Circular Hough Transform.
An artificial–neural network-based method7 to classify diabetic retinopathy is developed and presented. Here, an artificial–neural network-based method to classify diabetic retinopathy is implemented. For training, 6 input nodes, 6 hidden nodes, and 4 output nodes have been used. The four output nodes in the ANN corresponds to normal, diabetic Retinopathy, pre-proliferative diabetic retinopathy and proliferative diabetic retinopathy. A technique for classifying fundus images3 is proposed by the authors. Here, the fundus images are converted into a feature vector based on histograms range images at different resolutions. Then fundus images were classified by learning from feature vectors of a large number of normal and abnormal training fundus images. Segmentation of exudates with a new unsupervised approach based on ant colony optimization11 is studied and presented. A novel method for the automatic detection of exudates12 in digital fundus images is proposed. The approach is divided into three stages such as candidate extraction, contour segmentation and labelling the candidates. A state-of-art image processing13, methodology to automatically detect the presence of hard exudates in fundus images is presented. Here, the features extracted are fed in to the Echo State Neural Network as input to discriminate the fundus images.
The present codes implemented are either computationally heavy or not accurate to such an extent that the ophthalmologist can analyse the image further to get more information. Hard and soft exudates detection4 described here shows less accuracy due to false detection. However, the accuracy is low due to artifacts, additive noise and fainted exudates. The detection of diabetic retinopathy through lesions has few problems. Their algorithm depends on the detection of optic disk and blood vessels and subsequently makes the results dependent on the detection of optic disk and blood vessels. The limitation of the method is that the authors did not compared to existing method and classification feature were not good. The authors have to be able to analyse low quality images, but the images of several megabytes in size would not be acceptable because, it needs large storage requirements. The accuracy and robustness of locating the optical disk has to be increased. A survey of the classical and the methods for classifying and diagnosing7 the type of retinal disease and detecting its features after diagnosis at an earlier stage of the disease is presented. Although a lot of work has been done, automatic diagnosis of retinal diseases at an earlier stage still remains an open problem. The classification of the system accuracy requires more range images and a larger number of neighborhood windows. One solution is to perform a feature selection procedure during the training stage to identify the most distinctive histogram features. Then the range images with less distinctive histogram features need not to be calculated.
MATERIALS:
Image from standard diabetic retinopathy database such as DIARETDB1 and HEI-MED are utilized for this empirical studies. For the experimentation, 50 fundus images of 256 X 256 size from each database is obtained. Experiments were carried out with all the three color components such as red, green and blue in the RGB color model for each and every images obtained from the database. Among these three color components, the green component is more sensitive and has the highest contrast which helps us to manipulate and obtain the desired results in a better way.
Fig.1: Architecture of the proposed system
METHODS:
The proposed method comprises of three different phases such as the pre-processing phase, blood vessel extraction and exudates identification. The overall architecture of the proposed system is shown below in Figure 1.
The proposed system utilizes and manipulates 2d fundus images as input. The images consists of anomalies like exudates and haemorrhages, which are the major signs of diabetic retinopathy. The input fundus images are of colour (RGB), from which green channel is extracted. The green channel is used for the extraction of exudates as well as blood vessels. The green channel image is further converted to greyscale image. This output is again processed using adaptive histogram method to increase the contrast thereby easing out the process of extracting the exudates and blood vessels. In Contrast Limited Adaptive Histogram Enhancement (CLAHE), the contrast limiting procedure is applied to each and every neighborhood from which a transformation function is derived. CLAHE10 helps to prevent the over amplification of noise that adaptive histogram equalization can give rise to. Morphological dilation and erosion operations is performed on the image followed by morphological closing. This pre-processed image is taken as the input for vessel extraction which can be used for identifying the area where blood leakage has happened. On the other hand, the output pre-processed image is sent through multiple MSER feature extraction functions under various conditions which identifies the required features. This process is performed to maximize the amount of exudates identified in the input image thereby allowing a thorough analysis.
Blood Vessel Extraction:
In the blood vessel extraction process, few pre-processing steps has to be carried out for effectively extracting the blood vessels. Initially, the colored fundus image is converted to green plane since it is more sensitive and has the maximum contrast when compared with other two color planes. Further, the green component image is enhanced using contrast limited adaptive histogram enhancement which subsequently leads to better vessel extraction. Once the enhancement process is performed, utilizing the Kirsch’s template the respective blood vessels are extracted.
By considering a single mask and rotating it to 8 different orientations such as North, North-West, West, South-West, South, South-East, East and North-East the edges are obtained. The maximum value that is obtained by the convolution of each mask with the respective image gives the edge magnitude. The maximum magnitude that is produced defines the edge direction.
The Kirsch masks are defined by the following mask representations.
By applying the different mask, the respective edge directions are obtained. Subsequently, by applying the appropriate threshold to the edge detected image, the desired blood vessels are extracted.
Exudates Identification:
For identifying the exudates from the fundus images, the pre-processed fundus image is sent through multiple MSER feature extraction process under various conditions that identifies the required features. This process is performed to maximize the amount of exudates identified in the input image thereby allowing a thorough analysis. MSER implements a sweep threshold of intensity from black to white, performing a simple luminance thresholding of the image followed by extraction of the connected components. It finds out a threshold when an extremal region is “Maximally Stable”. Because of the discrete nature of the image, the region below or above may be coincident with the actual region. The next step is to approximate a region with an ellipse which is optional. These region descriptors identified are kept as features. The MSER8,9 algorithm has been adapted to color images, by replacing thresholding of the intensity function with agglomerative clustering, based on color gradients. By detecting MSERs in multiple resolutions9, robustness to blur, and scale change can be improved. MSER is used extensively in this work to detect exudates which in turn allows the system to identify if the patient is diabetic retinopathic or not.
Image Im is a mapping Im:
. External regions are well defined on
images if:
1.
is totally ordered, i.e. reflexive,
antisymmetric and transitive binary relation 
exists.
Here in this paper, only 
is considered, but external regions can be
defined on real-valued images (
).
2.
An adjacency relation 
is defined.
Here in this paper, 4-neighbourhoods are used, i.e. 
are adjacent (
iff 
Region 
is a contiguous subset of 
i.e. for each 
there is a sequence 
and 
.
Region Boundary (Outer) 
, ie, the boundary 
of is the set of pixels being adjacent to at
least one pixel of but not belonging to .
Extremal Region 
is a region such that for all 
(maximum intensity region) or 
(minimum intensity region).
Maximally stable Extremal Region (MSER). Let 
be a sequence of nested extremal regions,
i.e. 
. Extremal region 
is maximally stable iff 
has a local minimum at 
(
denotes cardinality). 
is a parameter of the method.
RESULTS:
The work fulfils its purpose by identifying blood vessels and detecting the exudates if present. The input fundus image is shown in Fig.1 and the green channel of the input image is shown in Fig. 2. The green channel is considered since it contains the most amount of features compared to red channel and blue channels.