Evaluation of edge detection algorithms for nuclear medicine images via ROC and shape analysis with adaptive thresholding

Abstract

Edge detection algorithms are critical to many of the current developments in nuclear medicine, but the distortions produced by applying these algorithms have not been adequately quantified for the low spatial resolution and low information density limit characteristic of nuclear medicine images. In this study, eleven edge detection methods were evaluated in terms of area determination, receiver operating characteristic (ROC) analysis, and shape preservation. Furthermore, edge detection is a multistep process in which allowance can be made for numerous variables. In this study, an adaptive approach was used to allow for variations in background and in information density. Only the nearest neighbor algorithm (NNA) was found to produce generally acceptable results at all information densities studied. At high information densities, the Sobel and Kirsch filters produced acceptable results. Preprocessing of the images by a Gaussian filter did not substantially alter these conclusions.