Developing an Expert System to Improve Lesion Quantification for Personalized PET Imaging

The measurements of lesion standardized uptake value (SUV) in clinical PET studies are affected in a complicated way by many aspects of the data, including-but not limited to-count level, lesion size, lesion shape, lesion location, background level and structure, and patient size. Optimized reconstruction algorithms and their parameters can provide substantial improvements in lesion quantification for personalized PET imaging. Full optimization of reconstruction parameters can be too complicated/tedious for human beings, thus we propose an expert system to optimize these parameters and to improve lesion quantification. Three core techniques are explored: 1) synthetic lesion embedding technique, 2) system local impulse response (LIR) modeling, and 3) bootstrapping for uncertainty estimation. We experimentally acquire list-mode data from physical radioactive spheres of known activity concentration with different sizes at different locations using the same PET scanner as was used for acquiring the patient data. We can then synthetically embed the spheres by merging the sphere list-mode data into patient data. We also use statistical bootstrapping method to generate multiple replicates of patient data with embedded lesions to determine the variance of the estimated SUV. We also use the LIR to characterize the local properties and responses of an imaging system, which can be used as a criterion for optimizing the reconstruction parameters. We have performed simulated phantom studies to confirm the use of bootstrapping of clinical data to determine the variance of SUV, and the use of LIR to determine the recovery coefficients and the bias of SUV of each lesion with different size. We are currently integrating the expert-system tools with a graphical user interface to allow the execution and visualization of ensembles of lesions with the use of different reconstruction algorithms and parameters. The end result will be patient- and lesion-specific corrected SUVs with estimated uncertainty.