Diabetic retinopathy is a complication of diabetes that occurs in the eye. One of the early signs of the disease is the appearance of an exudates wound that occurs because there is lipid or leaky fat in abnormal blood vessels and can cause blindness when its happen near of the macula. Early detection of the emergence of exudates is expected to reduce the risk of blindness to diabetic retinopathy patients.

Several previous studies have attempted to segment the exudates on the fundus image. However, because the size of the exudates is quite small compared to the overall image, the segmentation process is less to give maximum results. In this study proposes to detect regions containing exudates before segmentation process. Thus, the segmentation process will focus on specific areas only and not segment the overall image. There are three main stages in this research, namely optic disc removal, location detection of exudates, and exudates segmentation. Optic disc removal is intend to avoid false detection in the segmentation stages due to the color and contrast of optic disk that almost the same as the exudates. Optic disc removal is done by using midpoint circle algorithm. At the next stage, the detection of the location of exudates, the image will be divided into several parts of a smaller image then detected patch that contain exudates called exudates patch and patch that does not contain exudates called exudates-free patch. The area or patch is detected by using intensity thresholding because areas that containing exudates will be brighter than others. The last step is to segment the exudates patch using the saliency method then determine a threshold value to acquire the final segmentation. Saliency method could capture the unique colors in the image. So that, the exudates that have a high pixel value can be highlight in the salient image.

The result of exudates segmentation using the proposed method can be seen in figure 1. Figure 1.a) is the sample of original image and 1.b) is the segmentation result using proposed method. Detected exudates are marked with white objects that shown in figure 1.b).

 

Figure 1. a) input image, b) segmentation result

 

The diaretDB1 dataset is used to evaluate the proposed method by calculating the value of accuracy, sensitivity, and specificity. Accuracy is defined as the ratio of the number of detected exudates correctly and the number of pixels that are not exudates. It means, the accuracy calculates the performance of the method thoroughly. Sensitivity is defined as the ratio of the number of pixels identified as exudates to the total number of pixel exudates found on the ground truth. Specificity is defined as the ratio of the number of pixels identified as non-exudates to the total number of non-exudates pixels found on the ground truth. In other words, sensitivity measures the accuracy of non-exudates object in the image.

 

Table 1. Exudates segmentation result

Image

Accuracy (%)

Sensitivity (%)

Specificity (%)

1

99.93

51.56

99.96

2

97.93

62.06

98.45

3

98.99

88.77

99.00

4

98.70

61.22

98.74

5

99.68

76.14

99.70

Average

99.05

67.95

99.17

 

Based on Table 1, the result obtained that the average accuracy value of proposed method reached 99.05%. Sensitivity and specificity value of the proposed method respectively 67.95% and 99.17%. The sensitivity value obtained is quite low because there are still remains of optic disc that is not erased perfectly. In addition, in the detection process location exudates, there are still areas that contain exudates but categorized as exudates-free patch. It affects the segmentation process that causes the sensitivity level is not optimal.