Gießen 2024 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
HK: Fachverband Physik der Hadronen und Kerne
HK 72: Poster
HK 72.1: Poster
Donnerstag, 14. März 2024, 17:15–18:45, HBR 14: Foyer
Deep-learning for 3D Photon Interaction Position Reconstruction in Monolithic Scintillation Detectors — •Beatrice Villata, Maria Kawula, and Peter G. Thirolf — Department of Medical Physics, Ludwig-Maximilians-Universität München, Germany
Accurate spatial localization of the photons is important for high-resolution medical imaging in PET systems or prompt-gamma Compton cameras. Monolithic scintillation LaBr3 or CeBr3 detectors offer excellent energy and time resolution and can register the light distribution when read out with position-sensitive photosensors. This work presents a supervised deep-learning algorithm to reconstruct the 2D irradiation position on a monolithic crystal, by stacking convolutional layers and creating a CNN. The training dataset includes images of the light patterns acquired by irradiating the detector with a collimated photon source. Addressing the challenge of the depth of interaction (DOI), an additional unsupervised deep-learning algorithm is presented. This architecture combines the data-driven approach of deep learning with the mathematical modeling of the setup, implemented by parameterizing the signal’s evolution in the detector. The unsupervised approach is convenient because the current dataset does not contain information about the ground truth along the third dimension. This architecture would allow the reconstruction without the need to perform a Monte Carlo simulation or an additional irradiation of the setup to obtain a training dataset. Combining these approaches offers a promising approach to achieving precise 3D interaction position information in monolithic crystals.
Keywords: Monolithic Scintillator; Deep Learning; Position Reconstruction; Depth of Interaction; Compton Camera