Thesholding methods review for the location of the Optic disc in retinal fundus color images

This work compares Triangle, Maximum Entropy and Mean Peak thresholding methods to locate the optic disc in color fundus images. Localizing the optic disc is a significant task in an automated retinal image analysis process as is used on most vessel segmentation, disease diagnostic, and retinal recognition algorithms. The DIARETDBv1 dataset includes 89 retinal images are used to evaluate the 3 thresholding methods. To analyze the effect of image conditions on the performance of the thresholding methods, three set of images are created based on their contrast and quality: high contrast, low contrast, and poor quality; with 58, 16, and 15 images respectively. As green channel in the fundus image provides best contrast, it is used to extract the Optic disc. We analyze the effect of applying a previous preprocessing technique, each thresholding method is applied to 2 versions of the image: the original green channel and the original green channel with CLAHE. In terms of OD location, CLAHE preprocessing shows a great improvement for all thresholding methods. The average Method performance is evaluated with the percentage of intersected pixels with the groundtruth. From all results Triangle thresholding method performs consistently better than Maximum Entropy and Mean Peak, achieving 70.78% mean overlap with groundtruth in the best case and locating OD in 89/89 images.