An intercomparison of three quantitative data-processing techniques using mathematical tumours and receiver operating characteristic analysis

The technique of superimposing computer-simulated tumours onto normal clinical scans as a means of assessing image processing techniques has been suggested previously. The technique is shown to be applicable to gamma-camera scintigraphy. A set of 100 normal brain scintigrams was selected and 50 mathematical tumours simulated and added at various positions, no more than one to a view. Three quantitative data processing techniques were chosen for comparison. (1) Quantitation of difference: Lesions are first detected using an unsharp masking technique (the Canterbury filter) and then quantified by comparing observed and expected values in the region of the suspected lesion. (2) Step filter: This method compares a circular foreground element with a concentric background ring and outputs a signal-to-noise ratio (S/N). (3) Maximum convexity: This method is effectively a one-dimensional step filter which measures S/N in several directions and outputs (S/N)sub(max). By using the detection criteria appropriate to each method, a consultant chose the most likely tumour position on each scintigram unless he was satisfied that a lesion was ''almost definitely not present''. The results were compared by forming LROC curves based on the quantitative estimate given by each method. On the basis of purely quantitative considerations, quantitation of difference was shown to be best. By using a five-category rating system, the other two methods were compared with grey-scale output for the raw data and the Canterbury filter. All processing methods were shown to be superior to raw data and the step and Canterbury filters, when considered this way, wer shown to be comparable with quantiation of differences. (author)