CFA FCT/UNL is one of the two scientific groups of Centro de Física Atómica (CFA). Most of its members belong to Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa (FCTUNL).

The Weighed Matching Analysis method and the accuracy of automated retinal image processing

TitleThe Weighed Matching Analysis method and the accuracy of automated retinal image processing
Publication TypeConference Paper
Year of Publication2010
AuthorsAndré, Mora, Pedro Vieira, and José Fonseca
Date PublishedJan
Keywordscomparison, digital, image, images, processing,, segmentation,
AbstractThe assessment of techniques of automatic detection of structures in retinal images is mostly donethrough the comparison between the results produced automatically and the ones produced manuallyby specialists. When the analyses are subjective some disparities are common to appear amongdifferent specialists as well as within the repeated analysis of one specialist. A decisive mechanism istherefore needed to obtain more accurate results.In this article it is presented the weighed matching analysis method, which was developed to be apixel to pixel analysis that uses the statistical significance of the observations to differentiate positiveand negative pixels. It is based on the creation of a probabilities map, which results from thespecialists' markings, followed by the calculus of sensitivity, specificity and kappa coefficient betweenone analysis and the probabilities map.This method was validated with a dataset having 22 retinal images with visible drusen. These weremarked by 8 independent specialists and by the automatic detection method. The results ofsensitivity, specificity and kappa coefficient were calculated using both the weighed matching analysismethod and the binary method, for comparison purposes.It was concluded that this method improved the binary matching analysis, especially on image setsthat contain analysis with significant variability, by automatically removing outlier pixels and by havingrewards and penalizations with different weights based on the probabilities map values when there isno absolute agreement. Also, this method allows not only the method's validation, but also thequantitative comparison between specialists to identify outliers.