Purchase this article with an account.
P. Kasae, A. Ruggeri; Automated Identification of Optic Disc in Retinal Images Using Local Entropy. Invest. Ophthalmol. Vis. Sci. 2010;51(13):5344.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Identification of optic disc (OD) in retinal fundus images is an important step for computer aided diagnosis. Qualitative and quantitative comparison of OD images over time can measure the disease progression. Identification of OD would facilitate the automatic detection of disease signs in retinal images, e.g. bright lesions. Moreover localization of OD helps to determine the location of macula and fovea.
The green channel is at first extracted from the image. In a pre-processing step, after contrast enhancement, the background variability is smoothed with a high-pass filter and the "salt and pepper" noise is reduced by median filtering. Candidate OD areas are then identified by intensity segmentation where the optimal threshold value is determined by maximizing the local entropy of the resulting image. A post-processing stage is then applied, where most of the "false" candidates are removed, considering the fact that OD is a roundish object with approximately 0.2 mm radius. The largest of the surviving objects is then considered as the main part of the OD. Among the other objects, the ones with distance more than 0.2 mm from the selected one are removed, the others are consolidated into the OD area. Finally, the center of this conglomeration is considered as the center of the OD.
The results on a dataset containing 239 images, including the full DRIVE and STARE public databases, report a 92.05% success rate (i.e., identified OD center was located inside the manually segmented OD), with 97.04% success on the 135 healthy images and 85.58% success on the 104 diseased images.
This study presents a novel, robust approach for the automatic identification of OD. The method proved to be very accurate and reliable also on images affected by different pathologies. At variance with many alternative methods, it is not based on vessel segmentation, which itself is a complicated task. Moreover, it exploits the concept of entropy, which has not been adequately considered so far in the literature.
This PDF is available to Subscribers Only