Estimation of image noise in PET using the bootstrap method

The bootstrap method applied to PET data was evaluated as a technique to determine regional image noise in PET. To validate the method, 250 scans (5 min each) of a uniform cylinder filled with ⁶⁸ Ge was acquired and reconstructed using filtered backprojection (FBP). A single 5 min list mode scan was also acquired. From the list mode data, 250 bootstrap replicates were generated by randomly drawing, with replacement, prompt and random events. In each replicate, the total numbers of prompt and random events were kept identical to the number in the original list mode data set. The 250 individual scans, and the bootstrap replicates, were reconstructed using FBP and OSEM. Mean and standard deviation (SD) images were generated from the reconstructed images. Mean and SD were also calculated in a central region of the image sets. Visual inspection showed no appreciable difference between the SD images derived from the repeated scans and the bootstrap replicates. Profiles through the images, showed no significant difference between image sets. Using an increased number of bootstrap replicates produced less noise in the standard deviation images. ROI analysis showed, that the standard deviations derived from the bootstrap replicates were very close to the ones derived from the repent scans, independent of reconstruction algorithm. The results indicate that the bootstrap method can accurately estimate regional image noise in PET. This could potentially provide a method to accurately compare image noise in phantom and patient data under various imaging and processing conditions, without the need for repeat scans