Statistical analysis of PET images: extracting variance from a single scan

Statistical comparisons of images under different conditions (e.g. abnormal vs. normal) require multiple images per condition in order to estimate the variance. This paper presents a new method for calculating the variance from a single scan using image subsampling to remove possible spatial correlation which would bias and reduce the variance. The method uses the mean absolute deviation of the difference image over the whole field of view as a robust estimator of the variance. This single-scan method was compared with the absolute estimator derived from a dynamic PET acquisition of the IEC phantom with different radioactive concentration ratios, acquisition times, spatial extent and threshold levels. The results were quantified by measuring the sensitivity, specificity and detectability of the spherical sources. The new method was found to behave as expected and, for low statistics, perform better than the standard approach.