@inproceedings{MAS24a,
title	=	{First Steps towards a Foundation Model for Positioning in Positron Emission Tomography Detectors},
year	=	{2024},
booktitle	=	{2024 IEEE Nuclear Science Symposium (NSS), Medical Imaging Conference (MIC) and Room Temperature Semiconductor Detector Conference (RTSD)},
pages	=	{1--1},
url	=	{https://ieeexplore.ieee.org/document/10655512},
note	=	{ISSN: 2577-0829},
doi	=	{10.1109/NSS/MIC/RTSD57108.2024.10655512},
keywords	=	{Calibration, Detectors, Machine learning algorithms, Data models, Adaptation models, Training data, Topology},
author	=	{Masbaum, Thilo and Lopes de Paiva, Luis and Mueller, Florian and Schulz, Volkmar and Naunheim, Stephan}}
abstract	=	{Lightsharing detectors, e.g., monolithic or semi-monolithic concepts, have raised broad attention in the PET community as combining good performance characteristics with depth of interaction information. Accurate determination of the gamma interaction position within the scintillation crystal can be achieved by machine learning-based algorithms such as gradient tree boosting (GTB). Varying experimental conditions, e.g. adapted surface coatings, still require new time-consuming calibration measurements. In this work, we make first steps towards developing algorithms with a generalized positioning capability. Such algorithms are potentially more robust against varied detector configurations. This implies the possibility of a single, universal model for all the detectors in a PET scanner irrespective of scintillator topology. For the initial investigation of this hypothesis, we simulate the calibration with an external reference (fan beam) and perform planar positioning. The detector configuration consisted of an array of 12 slabs of dimensions 2.66 {textbackslash}mathrmx 32 {textbackslash}mathrmx 19 {textbackslash}mathrm mm{textasciicircum}3, with read-out resembling an array of digital SiPMs with pitch 4 mm. The experimental conditions were varied by utilizing different detector wrappings, including various combinations of ESR foil and black tape. Simulation data labeled with the source position in monolithic direction were used as training data for GTB models. Two types of models were trained: Specialized models only trained on simulation data of one wrapping, and generalized models trained on data from all detectors. The generalized models demonstrated similar performances as the specialized models, with only slight deterioration in mean absolute error (MAE) on the test datasets. This provides an interesting outlook on generalized models that can adapt to different experimental conditions.},