





Instrumentation & Data Analysis: Image GenerationPET  Reconstruction and Compensation 
^{1} Stanford University, Stanford, California
242^{ }
Objectives: CadmiumZincTelluride (CZT) may be useful for PET,^{ }but its photofraction is low and multiple interactions in the^{ }detectors are frequent. However, a CZT detector can localize^{ }the 3D coordinates of individual photon interactions with high^{ }spatial and energy resolution. A challenge is to correctly estimate^{ }the first interaction, which best defines the correct line of^{ }response. Previously, we had proposed a maximumlikelihood (ML)^{ }approach. In this work, we investigate a maximum a posteriori^{ }(MAP) positioning method.^{ }
Methods: MonteCarlo integration is used to compute the likelihood^{ }of each possible sequence of interaction. An accurate physical^{ }model is used to compute the a priori probability of each sequence^{ }and obtain a posteriori probabilities. This model uses the exponential^{ }photon linear attenuation law and the Klein Nishina formula.^{ }A parameter β defines the relative weights of the likelihood^{ }and the prior model (β=0 for ML). For fast computation,^{ }we are investigating using graphics processing units to parallelize^{ }the algorithm.^{ }
Results: We generated data using MonteCarlo for two simulated^{ }CZTbased PET systems. The highspecs (HS) system has 1x1x1^{ }mm^{3} spatial and 3% fwhm energy resolution. The lowspecs (LS)^{ }system has 1x1x5 mm^{3} spatial and 12% energy resolution. The^{ }method was evaluated on both systems with β varying. The^{ }performance is measured as the success rate in recovering either^{ }the full sequence of interactions or just the first interaction.^{ }
Conclusions: For the HS system, MAP performs worse than ML as^{ }β increases. This indicates that lownoise data should^{ }be trusted more than the prior model. For the LS system and^{ }β=2, MAP outperforms ML by 7% for recovering the correct^{ }interaction sequence and 8% for the first interaction. An accurate^{ }prior model makes the position estimate more robust to noise^{ }in the data.^{ }
Research Support: NIH R01CA119056, R33EB003283, R01CA120474^{ }
