Comparison of Canny’s and Snakes’ Algorithm as Applied to Diabetic Retinopathy

Abstract

Aims: To ascertain the effectiveness of edge detection and contour detection algorithms to identify hard exudates and hemorrhage region in fundus images of diabetic patients. Study Design: Canny’s algorithm was selected as the edge detection algorithm, and Snakes’ algorithm was selected as the contour detection algorithm. Place and Duration of Study: A total of 212 fundus images were procured from the Department of Ophthalmology of Bangladesh Institute of Research and Rehabilitation for Diabetes, Endocrine and Metabolic Disorders for this study. The images were captured between 2010 and 2013. Methodology: Noise was removed from the images using successive Gaussian and median filtering. Green component of the image was used for detection of hard exudates, and red component was used for detection of hemorrhage. To apply Canny’s algorithm, color gradient was calculated, and a threshold was applied to the gradient to select a candidate region. Snakes’ algorithm was applied by scaling the color intensity from 0 to 1, and a threshold color value was chosen to draw the contours. Several filters were applied to the selected region to detect and discard the false-positives. A total of 32 images were used for training purpose. The algorithm was later applied to the rest of the 180 images. Results: For Canny’s algorithm, a threshold color gradient value of 0.30 was chosen to identify the hard exudates, and a value of 0.28 was chosen to identify the hemorrhage regions. For Snakes’ algorithm, a color intensity value of 0.7 was chosen for detection of hard exudates, and 0.83 was chosen for detection of hemorrhage regions. Both Canny’s algorithm and Snakes’ algorithm performed similarly in detection of hard exudates. For detection of hemorrhage regions, generally Canny’s algorithm performed better compared to Snakes’ algorithm. Even in situations where there was poor color contrast, Canny’s algorithm was able to suggest a candidate region, whereas Snakes’ algorithm completely failed to suggest a region. Conclusion: Both Canny’s algorithm and Snakes’ algorithm performs equally effectively in detecting hard exudates. But in detection of hemorrhage regions, generally Canny’s algorithm performs better compared to Snakes’ algorithm.