One of the major complications of diabetes is diabetic retinopathy. As manual analysis and diagnosis of large amount of images are time consuming, automatic detection and grading of diabetic retinopathy are desired. In this paper, we use fundus fluorescein angiography and color fundus images simultaneously, extract 6 features employing curvelet transform, and feed them to support vector machine in order to determine diabetic retinopathy severity stages. These features are area of blood vessels, area, regularity of foveal avascular zone, and the number of micro-aneurisms therein, total number of micro-aneurisms, and area of exudates. In order to extract exudates and vessels, we respectively modify curvelet coefficients of color fundus images and angiograms. The end points of extracted vessels in predefined region of interest based on optic disk are connected together to segment foveal avascular zone region. To extract micro-aneurisms from angiogram, first extracted vessels are subtracted from original image, and after removing detected background by morphological operators and enhancing bright small pixels, micro-aneurisms are detected. 70 patients were involved in this study to classify diabetic retinopathy into 3 groups, that is, (1) no diabetic retinopathy, (2) mild/moderate nonproliferative diabetic retinopathy, (3) severe nonproliferative/proliferative diabetic retinopathy, and our simulations show that the proposed system has sensitivity and specificity of 100% for grading.
Diabetic Retinopathy (DR) is a leading cause of vision loss in the working class in the world [
International clinical DR severity scale [
Severity level | Findings observable |
---|---|
No DR | No abnormalities |
| |
Mild NPDR | This is the earliest stage of retinopathy and vision is usually normal except in some cases. Small swellings known as micro-aneurysms or flame-shaped hemorrhages start to develop in the fundus quadrants |
| |
Moderate NPDR | There will be micro-aneurysms or hemorrhages of greater severity in one to three quadrants and leakage might occur, resulting cotton wool spots and exudates and so fourth to be present in the retina |
| |
Severe NPDR | Any of the following: |
| |
PDR | One or more of the following: |
There are various techniques for automatic grading of DR. These systems use one or more features such as blood vessels, exudates (EXs), micro-Aneurisms (MAs), and texture [
Yun et al. [
This work describes a new method for automatic grading of 3 main stages of DR. The proposed algorithm uses fundus fluorescein angiography (FFA) and color fundus images simultaneously. MAs appeared like white small dots in FFA and they are more distinguishable than in color fundus images [
The main setup of the proposed algorithm in this paper is as follows. In Section
We have attempted to work on database that has both FFA image and color fundus image in this DR grading system. Because bright lesions (EXs) appeared better in color image (Figure
(a) EXs in color fundus image. (b) MAs in FFA image of same fundus.
The block diagram of DR grading system. F1 to F6 indicate the features (as defined in Figure
In our previous work [
First of all we use contrast limited adaptive histogram equalization (CLAHE) algorithm [
(a) The original FFA image, (b) FFA image after CLAHE, (c) the original color image, (d) the color image after CLAHE.
We use the presented method in [
DCUT is a digital version of curvelet transform. The main motivation for using curvelet transform is dealing with 2D singularities such as edges in images instead of point singularities in 1D signals. In fact one of the desire properties of an appropriate 2D transform for image processing is directional selectivity (DS). DS is not a required property for 1D signals and so wavelet transform is nearly an optimum choice for 1D signal processing. However for high dimensional data DS plays a key role and on this base ridgelet transform [
Here we use the proposed DCUT-based OD detection method in [
Finally the OD is extracted based on this fact that OD is partly covered with vessels. Figure
(a) The original FFA image, (b) the image after amplifications of bright objects by DCUT, (c) localization of OD after applying Canny edge detector and morphological operators [
EXs are detected by performing DCUT on color fundus images. The main steps of EXs detection are presented here. Enhancement of bright lesions by applying DCUT on enhanced image and modifying its coefficients. Extracting of candidate regions for EXs by thresholding (Figure Removing of OD (Figure
(a) The original image, (b) reconstructed image after modifying DCUT coefficients, (c) the produced image after thresholding, (d) OD localization, (e) removing OD and extracting EXs.
In order to improve the contrast of EXs, intensity of gray levels in green channel
Again we use DCUT for segmenting vessels [ Inverting FFA image. Curvelet-based contrast enhancement. Taking DCUT of the match filtered response of enhanced retinal image. Removing low frequency component, and amplifying all other coefficients. Applying inverse DCUT. Thresholding using the mean of the pixel values of the image. Applying length filtering and removing misclassified pixels. Actually the cross-section of retinal vessels has a Gaussian shaped intensity profile [
Figure
(a) Inverse of FFA, (b) matched filter response of enhanced image, (c) high frequency component of image, (d) the produced image after thresholding, (e) the extracted vessels by applying length filtering.
As completely discussed in [ Vessel extraction based on DCUT. OD extraction based on DCUT. ROI definition based on this fact that macula locates on 2.5 OD diameters away from center of OD. Finding end-points of vessels in ROI region. For this reason the following steps are proposed. A 3 The center of these end-points is obtained by averaging all end-points’ coordinator and then the average distance of all end-points to this center is calculated. The final end-points are selected by comparing each end-point’s distance against this mean value and discarding end-points that their distances are greater than mean value. Connecting selected end points to each other.
(a) Vessel mapped image, (b) showing end-points in predefined ROI, (c) removing end-points which are far from center of FAZ, (d) connecting selected end-points, (e) extracted FAZ from DCUT method, (f) subtracting vessels from original image, (g) applying morphological closing, (h) extracted FAZ by thresholding (g), (i) applying logical and between extracted FAZ in (e) and (h), (j) showing the final FAZ region.
MAs are the earliest clinical sign of DR. In this paper, a new method for detecting MAs is presented. In order to detect MAs, segmented vessels, are subtracted from original enhanced FFA image (Figure
(a) Original image after removing vessels, (b) applying morphological dilation on (a), (c) background image by applying morphological erosion on (b), (d) mapping zero pixels of (c) on (b), (e) removing background, (f) produced image after thresholding, (g) detected MAs, (h) MAs on original image.
In order to classify different stages of DR, we must extract appropriate and significant features. The feature set should be selected such that the between-class discrimination is maximized while the within-class discrimination is minimized. In this section, we explain about the selected features for DR grading. Area of detected FAZ: In [ Circularity of detected FAZ: FAZ region is an oval shape in normal retinal images. So stage of DR could have great influence on shape of this region. Analyzing variance of distance between points around FAZ and center of FAZ could be a good feature for DR grading. Total number of MAs and number of MAs in FAZ boundary: As shown in Table Total area of EXs: The higher stage of the DR would have more EXs due to damages or leakages of the blood vessels. Area of blood vessels: In some higher stages, main blood vessels are damaged (especially around macula) and new vessels are created (Neovascularization). These new vessels are thin. So, higher stages have fewer blood vessels because of damage.
In the last step SVM is used to classify DR severity. This grading is based on extracted features from:
SVM isasetofrelated supervisedlearningmethodsusedfor grading [
In this paper, we apply SVM twice. First one is for separating grade1 (normal) from grade 2 (mild and moderate NPDR) and grade 3 (severe NPDR and PDR). Second one is separating grade 2 and grade 3.
As we have shown in block diagram of Figure
Average of extracted features in different stages.
No DR | Mild/Moderate |
Severe NPDR/PDR | |
---|---|---|---|
No. data | 30 | 25 | 15 |
Area of EXs | 15 | 450 | 2987 |
No. of MAs in FAZ | 0 | 13.8 | 15.04 |
Total no. of MAs | 0.94 | 110 | 73.76 |
Area of vessels | 52567.14 | 45654 | 29521 |
Area of FAZ | 3498.25 | 5773 | 13624 |
Variance | 12.93 | 27.07 | 110.84 |
Extracted features from color and FFA images.
In this paper, a curvelet-based algorithm for DR grading was introduced. In this algorithm it is necessary to detect OD, vessels, FAZ, EXs, and MAs that all of them are detected by employing curvelet transform. In the next step 6 features were obtained from extracted FAZ and detected lesions and used as an input vector for SVM classifier. This algorithm was able to completely distinguish between “normal Stage”, “mild/moderate NPD”, and “severe NPDR/PDR”.
As an extension of this study, it is suggested to extract more features to increase the ability of algorithm for grading all stages of DR. This stage needs collecting more data (including both FFA and color fundus images) for each DR grade.
In this paper, only curvelet transform is used as an oriented transform that is able to separate the image to several subimages with specific time-frequency and orientation components. Other directional transforms such as dual tree complex wavelet transform [