Andrew LaBella’s second paper “Convolutional Neural Network for Crystal Identification and Gamma Ray Localization in PET” is accepted for publication in IEEE Trans Rad Plasma Med Sc.
In this paper, we use convolutional neural networks (CNNs) for 3D gamma ray localization in PET detector arrays with multicrystal scintillators. We trained and tested our CNN on Monte Carlo simulated PET data. Our results demonstrate that our CNN can correctly identify the crystal where gamma ray absorption takes place with over 99% accuracy throughout the entire detector array, including at the edges and corners where crystal identification is most problematic. In addition, our CNN achieves 2.75 mm FWHM DOI resolution, which is similar to the performance of conventional DOI localization in dual-ended readout detectors. Preliminary results suggest our CNN may be able to overcome the spatial limitations of PET detectors and achieve sub-scintillator resolution. Additional studies will be performed with our CNN on experimental data to validate its utility in practice. Our CNN can be used to perform 3D gamma ray localization in cost-effective PET detector modules with high accuracy.