Segmentação automática do disco óptico e de vasos sanguíneos em imagens de fundo de olho

Abstract

The efficient segmentation of the optic disc and the blood vessels enables the development of an array of retinal images diagnostic tools, thus contributing for the medical diagnosis of pathologies such as Hypertensive and Diabetic retinopathy, Retinopathy of Prematurity, Glaucoma, retinal lesions caused by the Zika virus, etc. However, the step of segmentation of the optic disc and the blood vessels is considered by many one of the most difficult tasks in image processing, especially in colour fundus photography, which possesses lower contrast due to the prevalence of red in all the retina. Thus, the objective of this work is to segment the main physiological structures of the retina, those being the optic disc and the blood vessels. In order to do so, it was proposed (i) an algorithm based on a multiscale energy filter with a Hough transform to rapidly locate the disc and a decision criterion in the disc location, respectively; (ii) an algorithm based on 3D roughness index combined with mathematical morphology and the use of the Atanassov Intuitionistic Fuzzy Set representation to segment the optic disc; (iii) a multilayer perceptron neural network in which the differential is the setting of the pre-processing step based on Contrast Limited Adaptive Histogram Equalisation (CLAHE) and aWiener filter to segment the blood vessels. As for results, it is observed that (i) the disc location method achieved 100% effectiveness with the HRF and DRIVE databases, 98,5% with the Messidor database and 94,38% with the DIARETDB1 database; (ii) in the automatic segmentation of the optic disc with the Messidor database the accuracy achieved was of 99,59% and the sensitivity was of 91,56%; (iii) in the segmentation of the blood vessels, it is possible to fine-tune the pre-processing step allowing for a superior network performance; it was obtained an accuracy of 94,87% for the first observer, and 95,31% for the second observer in the DRIVE database, such figures exceed even more complex technologies.