Predicting 3D photon interaction in a hexagonal positron emission tomography detector: a deep learning approach.

• Leandro A. Hidalgo-Torres ^[1] ; David Pérez-Benito ^[1] ; Rigoberto Chil ^[1] ; Pedro Macias ^[1] ; Juan José Vaquero ^[2]
  1. [1] Universidad Carlos III de Madrid

     Universidad Carlos III de Madrid

     Madrid, España

     
  2. [2] Instituto de Investigación Sanitaria Gregorio Marañón

     Instituto de Investigación Sanitaria Gregorio Marañón

     Madrid, España

     
• Localización: XXXVIII Congreso Anual de la Sociedad Española de Ingeniería Biomédica. CASEIB 2020: Libro de actas / Roberto Hornero Sánchez (ed. lit.), Jesús Poza Crespo (ed. lit.), Carlos Gómez Peña (ed. lit.), María García Gadañón (ed. lit.), 2020, ISBN 978-84-09-25491-0, págs. 401-404
• Idioma: inglés
• Enlaces
  • Texto Completo Libro
• Resumen

  • Spatial resolution is considered one of the main limitations of Positron Emission Tomography (PET). Most of the commercial scanners use pixelated array crystal scintillators on their detectors, and its pixel size determines the detector’s intrinsic spatial resolution. Alternative developments have exploited the characteristics of monolithic crystal scintillators to further improve spatial, energy and timing resolution. The major challenge of monolithic crystals is the estimation of the gamma photon position of interaction due to the broad scintillation light distribution. In this work, we explore the use of artificial neural networks to estimate the 3D position of the γ-photon interaction in a non-conventional hexagonal detector with a side length of 17.6 mm, 10 mm thick scintillator and 61 hexagonal silicon photomultipliers channels with a side length of 2.25 mm. The selected model was successfully implemented and trained with Monte Carlo simulated data including optical interactions in GATE v8.0. The best results were achieved using a multilayer perceptron architecture with 5 hidden layers of 128 hidden nodes. The obtained 2D and 3D average spatial resolution were 0.65 mm and 0.70 mm. The obtained results with the proposed model confirm that ANN approaches could provide spatial resolution comparable to state of art methods tested on simulated data. Hence, this study supports the suitability of artificial neural networks as a positioning algorithm for PET detectors with monolithic crystal scintillators