Segmentação de disco ótico e vasos sanguíneos em imagens retinais usando wavelets, morfologia matemática e filtragem multiescalas baseada em hessianas

• Ano de defesa: 2017
• Autor(a) principal: Rodrigues, Luiz Carlos Ferreira
• Orientador(a): Marengoni, Mauricio
• Banca de defesa: Lopes, Paulo Batista , Martins, Valéria Farinazzo , Pedrini, Hélio , Costa, Anna Helena Reali
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Presbiteriana Mackenzie
• Programa de Pós-Graduação: Engenharia Elétrica
• Departamento: Faculdade de Computação e Informática (FCI)
• País: Brasil
• Palavras-chave em Português: imagens retinais; morfologia matemática; wavelets; filtros multiescala
• Área do conhecimento CNPq: CNPQ::ENGENHARIAS
• Link de acesso: http://dspace.mackenzie.br/handle/10899/24283

The importance of an accurate and early diagnosis of retinal diseases has motivated the development of vision techniques necessary for a complete automatic evaluation of the retinal system conditions. In this thesis, we present a new algorithm that applies wavelet transform and mathematical morphology for optical disc detection and explore the blood vessels tubular characteristics in Gaussian scale space to segment veins and arteries. Optic disc and the vascular structure are essential for the analysis of the retinal image. Instead of using empirical attempts to choose the best parameters for vessel detection, genetic algorithm (GA) and its sequences of generations and intersections are used. However, the technique of exploring vessel tubular characteristics reaches its limit when vessels are represented by curvilinear sequences and not continuous 1 pixel wide. To overcome this limitation, a new methodology based on the Steger algorithm for segmenting curvilinear structure was created for the vessels not reached by the detector in scale space. The curvilinear detector will identify and attach the lines to the main vessels using the shortest path algorithm developed by Dijkstra, completing the segmentation of wide and thin vessels. The proposed method was developed and tested on two freely available open image databases: DRIVE (Digital Retinal Images for Vessel Extraction) containing 40 annotated retinal images and HRF-DB (High Resolution Fundus Image Database ) that contains 45 annotated images. The experimental results of the method are demonstrated and exhibit an average accuracy of 0.9503 in the DRIVE image database and 0.9445 in the HRF-DB. Such levels of accuracy are close to the state of the art, which are more complex and require one, occasionally two, preprocessing modules, while this method does not require pre or post processing. The developed method has an average processing time, for each image, of 35 seconds and was implemented in a Intel Core i5-3320 processor, 2.60 GHz CPU with 8 GB of RAM.