Software reliability and algorithm validation for medical imaging: performance of common edge detection methods in nuclear medicine.

Abstract

Ten common edge detection algorithms plus the NNA (nearest-neighbor algorithm) developed by the authors were evaluated under the conditions of low ID (information density) and low spatial resolution commonly found in nuclear images. Both phantoms and clinical images were used, and isolated regions as well as overlapping regions with variable backgrounds were analyzed at various ID's. Threshold criteria were also adaptively varied as a function of local ID. Performance was quantitated in terms of (a) accuracy of area determinations, (b) receiver operating characteristic operating points, and (c) shape preservation. With adaptive thresholding at high ID's, several of the methods performed well on isolated regions and adequately on overlapping regions, but only the NNA performed consistently at low ID's as well as at high ID's.