Improving TOF Estimation by Integrating Symmetry into Machine Learning Models

Abstract

We present a new learning task formulation for a residual physics-based timing calibration of PET detectors that leverages translational symmetry. We use the supervised machine learning algorithm gradient-boosted decision trees (GBDT) to correct the skewed measured difference in time-of-flight (TOF) given the known offset position of a radiation source during calibration. The previous formulation directly predicted the expected time difference (implicit correction) and required extensive calibration effort. The new formulation predicts an additive correction for the measured time difference (explicit correction). We demonstrate its capability for unbiased predictions that are robust against spatial undersampling of calibration data while achieving performance metrics similar to or better than those of the previous formulation. It eliminates the need for extensive calibration measurements and results in a massive reduction in model size, making the calibration more suitable for complete PET systems. Both formulations could improve the CTR by about 24 % compared to analytical calibration alone. Using explicit correction models, we achieve a CTR of$(282 \pm 3) \text{ps}$. Measurements were conducted with two clinically relevant detectors, each consisting of a$4 \times 4$array of$(3.8 \times 3.8 \times 20) \text{mm}^{3}$LYSO:Ce,Ca crystals allowing light-sharing and coupled to$4 \times 4$Broadcom NUV-MT SiPMs digitized with the TOFPET2 ASIC.

Authors 7

1. L. Lopes de Paiva Aachen Institute of Imaging & Computer Vision

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute of Imaging and Computer Vision (LfB),Aachen,North Rhine-Westphalia,Germany

2. Vanessa Nadig Aachen Institute for Experimental Molecular Imaging Department of Physics of Molecular Imaging Systems

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute for Experimental Molecular Imaging,Department of Physics of Molecular Imaging Systems (PMI),Aachen,North Rhine-Westphalia,Germany

3. Yannick Kuhl Aachen Institute of Imaging & Computer Vision

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute of Imaging and Computer Vision (LfB),Aachen,North Rhine-Westphalia,Germany

4. S. Gundacker Aachen Institute for Experimental Molecular Imaging Department of Physics of Molecular Imaging Systems

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute for Experimental Molecular Imaging,Department of Physics of Molecular Imaging Systems (PMI),Aachen,North Rhine-Westphalia,Germany

5. Florian Mueller Aachen Institute for Experimental Molecular Imaging Department of Physics of Molecular Imaging Systems

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute for Experimental Molecular Imaging,Department of Physics of Molecular Imaging Systems (PMI),Aachen,North Rhine-Westphalia,Germany

6. Volkmar Schulz Aachen Institute of Imaging & Computer Vision

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute of Imaging and Computer Vision (LfB),Aachen,North Rhine-Westphalia,Germany

7. Stephan Naunheim Aachen Institute for Experimental Molecular Imaging Department of Physics of Molecular Imaging Systems

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute for Experimental Molecular Imaging,Department of Physics of Molecular Imaging Systems (PMI),Aachen,North Rhine-Westphalia,Germany