Abstract
195
Objectives: Quantitative nuclear-medicine (NM) imaging offers a non-invasive approach to perform dosimetry for treatment planning for targeted radionuclide therapy. However, due to the high noise and limited resolution in NM imaging, optimization of the imaging protocol (IP) for reliable performance in the task of quantifying organ and tumor activity uptakes is essential. Ideally this optimization must be performed with patient data, but that requires a gold standard for reference, which is typically unavailable. No-gold-standard (NGS) techniques have been developed to evaluate quantitative imaging methods using patient data1-4, but these require the same patient to be scanned using all the IPs, which is impractical. Our objective was to develop a practical NGS evaluation approach that overcomes this requirement.
Methods: A no-patient-overlap-NGS (NO-NGS) technique was developed to evaluate IPs based on how reliably they measure the true mean activity uptake. The technique assumes a linear relationship, characterized by a slope, bias, and a noise standard deviation term, between the true and measured mean activity uptake values for each IP. This linear relationship has been observed to hold3 and is justified by the linearity of the NM-imaging system and the quantitation operation and approximate linearity of robust reconstruction methods. Intuitively, a protocol with the highest bias/standard deviation will be most inaccurate/imprecise. Assuming that the mean activity uptake for patients scanned using each imaging protocol have been sampled from the same unimodal distribution, the distribution of activity uptake values using the different protocols is mathematically derived to be completely characterized by the unknown linear-relationship and unimodal distribution parameters. Thus, using a maximum-likelihood technique, these parameters are estimated for the different IPs. To validate the NO-NGS technique, SPECT imaging of an I-131 radioisotope distribution modeling the uptake of an I-131 labeled anti-CD20 antibody used for radio-immunotherapy of non-Hodgkin’s lymphoma was simulated. The object database consisted of a 42-patient digital-phantom population. The phantoms were based on the anthropomorphic NCAT phantom with organ sizes and uptakes based on patient data. The projection data were simulated using highly realistic and previously validated Monte Carlo simulation methods. The patient population was split into two halves, where, for each half, the projection data was reconstructed using two different methods. The NO-NGS technique was used to rank these two IPs based on how reliably the activity uptakes in the different organ VOIs were estimated. 50 trials of this process were conducted, where, in each trial, different subset of patients were allotted to different IPs, thus simulating population variability. The entire process was repeated for 50 different noise realizations of the data, to study robustness. Thus 2500 (50 X 50) trials of the NO-NGS technique were conducted.
Results: The biases and standard deviations estimated using the NO-NGS technique yielded the same rankings as the true rankings of the IPs on the basis of accuracy and precision for more than 90% of the 2500 trials, as shown in Fig. 1. The estimated (true) bias using the NO-NGS technique for the two IPs were 2.95±0.09 (2.93) and 0.27±0.09 (0.23), respectively. Similarly, the estimated (true) standard deviation for the two IPs were 3.86±0.07 (3.93) and 1.91±0.03 (2.01), respectively.
Conclusion: A NO-NGS technique developed to evaluate imaging protocols (IPs) for quantitative imaging when different sets of patients are scanned using different IPs was shown to estimate figures of merit to quantify the accuracy of the different IPs for quantitative SPECT in the absence of any gold standard.