[Random sample size in ROC analyses].

Abstract

ROC analysis has proved useful in assessing diagnostic efficiency. Large x-ray film series of thoracic images intended for cardiovascular diagnosis were used for studying diagnostic efficiency in relation to random sample size. Exploration of random samples from two classes of findings requiring diagnostic differentiation, with the same number of x-ray films in each class, showed satisfactory convergence between the radiologists' rating and the ROC curve if the size of the random sample groups was about 200 images each or larger. The smaller the random sample size (less than 200 images), the greater the scatter of the rating performance. In another series the rating ability of experienced radiologists was compared with that of a semi-automatic classificator. The semi-automatic classificator attained almost the same performance index as the low-performance evaluator.