Comparison of software and human observers in reading images of the CDMAM test object to assess digital mammography systems

European Guidelines for quality control in digital mammography specify minimum and achievable standards of image quality in terms of threshold contrast, based on readings of images of the CDMAM test object by human observers. However this is time-consuming and has large inter-observer error. To overcome these problems a software program (CDCOM) is available to automatically read CDMAM images, but the optimal method of interpreting the output is not defined. This study evaluates methods of determining threshold contrast from the program, and compares these to human readings for a variety of mammography systems. The methods considered are (A) simple thresholding (B) psychometric curve fitting (C) smoothing and interpolation and (D) smoothing and psychometric curve fitting. Each method leads to similar threshold contrasts but with different reproducibility. Method (A) had relatively poor reproducibility with a standard error in threshold contrast of 18.1 ± 0.7%. This was reduced to 8.4% by using a contrast-detail curve fitting procedure. Method (D) had the best reproducibility with an error of 6.7%, reducing to 5.1% with curve fitting. A panel of 3 human observers had an error of 4.4% reduced to 2.9 % by curve fitting. All automatic methods led to threshold contrasts that were lower than for humans. The ratio of human to program threshold contrasts varied with detail diameter and was 1.50 ± .04 (sem) at 0.1mm and 1.82 ± .06 at 0.25mm for method (D). There were good correlations between the threshold contrast determined by humans and the automated methods.