Direct3γ PET: A Pipeline for Direct Three-gamma PET Image Reconstruction

Abstract

Direct3{\gamma}PET is a novel, comprehensive pipelinefor direct estimation of emission points in three-gamma (3-{\gamma})positron emission tomography (PET) imaging using \b{eta}+ and {\gamma}emitters. This approach addresses limitations in existing directreconstruction methods for 3-{\gamma} PET, which often struggle withdetector imperfections and uncertainties in estimated intersectionpoints. The pipeline begins by processing raw data, managingprompt photon order in detectors, and propagating energy andspatial uncertainties on the line of response (LOR). It thenconstructs histo-images backprojecting non-symmetric Gaussianprobability density functions (PDFs) in the histo-image, withattenuation correction applied when such data is available. Athree-dimensional (3-D) convolutional neural network (CNN)performs image translation, mapping the histo-image to radioac-tivity image. This architecture is trained using both supervisedand adversarial approaches. Our evaluation demonstrates thesuperior performance of this method in balancing event inclu-sion and accuracy. For image reconstruction, we compare bothsupervised and adversarial neural network (NN) approaches.The adversarial approach shows better structural preservation,while the supervised approach provides slightly improved noisereduction.