Objektivizace tortuozity u nedonošených dětí s využitím automatizovaného segmentačního algoritmu

Abstract

This paper deals with the classification of ROP plus form using tortuosity sign in premature infants from retinal images taken with the RetCam 3 fundus camera based on an automated segmentation algorithm. The result of this work is the design, implementation and quantitative testing of a comprehensive segmentation and classification system for the identification of retinal vessels and pathological vessels of the ROP plus form. The work includes an extensive search of recent literature to verify the state of research in segmentation techniques with a review of all available and private retinal databases and an analysis of tortuosity, which is a manifestation of ROP plus form. The proposed automated segmentation algorithm makes use of morphological operations, relying on the optimization of this algorithm based on the randomized setting of input parameters in a certain interval. Included are procedures for algorithm testing and objective evaluation on retinal images of adult subjects from the freely available DRIVE database with an accuracy of 0.9442, STARE with an accuracy of 0.9522, and the private RetCam 3 database provided by the Center for Children with Visual Impairment at FNO with an accuracy of 0.9740. This work includes a comparative analysis of convolutional neural networks GoogLeNet, ResNet, DenseNet with different parameter settings with the aim of finding the optimal combination of these parameters for the most efficient classification of ROP plus forms. The proposed classification system based on ResNet convolutional neural network to classify the retinal vasculature into physiological vessels and vessels with ROP plus form achieves an accuracy of 0.9750.