Analýza dynamiky optického disku u predčasne narodených detí na základe retinálnych obrazov

Abstract

Retinopathy of prematurity (ROP) is a serious disease that affects the visual system of premature infants, where monitoring the stability of the optic disc (OD) plays a key role during screening. Early diagnosis using retinal images is essential for preserving vision. Automation of detection and analysis of OD variability has enormous potential for increasing screening efficiency and objectivity of the examination. For the implementation of the diploma thesis, a database of retinal images was created from the Clarity RetCam 3 and Phoenix ICON data sets, which contain images of premature infants with ROP. The selected images were manually segmented, with the aim of creating gold standards. Two automatic segmentation methods were tested: U-net (deep convolutional neural network) and the active contour method. The performance of both methods was assessed using evaluation metrics (sensitivity, transferability, DICE and others) and image fusion. Based on quantitative evaluation, the optimal segmentation method was the active contour method, which achieved better results in accuracy (99.81%) and stability of OD segmentation. This method was subsequently used to analyse the variability of the optic disc. The assessment of OD variability was performed based on geometric (area, equivalent diameter, circumference ...) and intensity (mean value, median, variance ...) parameters of the images within the follow-up of patients over time. The results showed that some parameters showed the smallest percentage changes during the follow-up and thus the highest stability. These were mainly the geometric parameters equivalent diameter (20%) and circumference (21%). This indicates that geometric parameters are more reliable and can be used as stable references in ROP screening. The diploma thesis confirmed that active contours represent a robust approach to OD segmentation in premature infants. The parameters equivalent diameter and circumference can be considered stable references for the detection of pathological changes in ROP. These results contribute to the development of objective assessment in clinical practice and support the use of automated methods in pediatric ophthalmology.