An Overview of Current Evaluation Methods Used in Medical Image Segmentation

Abstract

An important aspect of the development of image segmentation algorithms is the availability of mechanisms to evaluate them. This is necessary in order to estimate how fit a segmentation approach is for the specific task, validate its performance on data and compare it against other approaches. Medical image segmentation is very interesting in this perspective since there is no established evaluation framework. A commonly accepted approach is to compare the segmentation output to some reference results (usually produced with manual segmentation) with a similarity/difference measure. While most authors rely on such methods there is still no agreement concerning the measures used. This is partly because there are no measures that reflect all the important features of a desirable segmentation and the existing measures do not discriminate different segmentation results in an acceptable way. This paper provides a survey of current methods being used for medical image segmentation evaluation.