Bonn 2020 – wissenschaftliches Programm
Die DPG-Frühjahrstagung in Bonn musste abgesagt werden! Lesen Sie mehr ...
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
ST: Fachverband Strahlen- und Medizinphysik
ST 2: Postersession - Medical Physics
ST 2.1: Poster
Dienstag, 31. März 2020, 17:00–18:00, Grotte
Photon interaction position determination in monolithic scintillators via Neural Network algorithms — •Maria Kawula1, Tim Binder1,2, Silvia Liprandi1, Katia Parodi1, and Peter G. Thirolf1 — 1Ludwig-Maximilians-Universität München — 2KETEK GmbH München
Monolithic scintillators are an attractive alternative to pixelated crystals as a part of multiple-component photon detectors like Compton cameras. We propose a novel algorithm for determining the position of γ-ray interactions in a monolithic scintillation crystal, based on Supervised Machine Learning involving Convolutional Neural Networks (CNN). The new method is an alternative to well-established algorithms such as "k-Nearest Neighbours" (kNN), which suffers from long computation time and high memory requirements. Two crystals, LaBr3:Ce and CeBr3, of size 50 mm×50 mm×30 mm were examined. The spatial resolution of the CNN algorithms was tested for three energies of the initial γ quanta: 662 keV (137Cs), 1.17 MeV and 1.33 MeV (60Co). A spatial resolution of the algorithm of 1.04 (± 0.04 stat. ± 0.2 sys.) mm at 662 keV and 0.90 (± 0.02 stat. ± 0.2 syst.) mm at 1.3 MeV for LaBr3:Ce and CeBr3, respectively, was achieved. The new reconstruction scheme is compatible with CPUs and GPUs and can reconstruct up to 2· 104 events/s, which is four orders of magnitude faster than the kNN. Memory requirements are reduced by ≈1/1000. This work was supported by the DFG Cluster of Excellence MAP (Munich Centre for Advanced Photonics) and the Bayerische Forschungsstiftung.