Towards artificial data generation for accelerated PET detector ML-based timing calibration using GANs

Abstract

In PET systems, TOF scanners require dedicated time skew calibration of the individual detectors, which is commonly performed by analytical approaches. In previous research, we demon-strated that performing timing calibration of detectors with a machine learning (ML) approach yields superior results. However, this method demands an extensive data acquisition due to its reliance on supervised learning. Our current research introduces initial advancements in the data gathering approach by augmenting measurement data obtained from detectors using generative artificial intelligence (GenAI). GenAI excels at learning the data distribution of complete multidimensional samples, thereby preserving correlations between features and ensuring high-quality representation of individual feature distributions. We compare our approach with an analytical baseline, where we linearly interpolate the data between different measured positions. In our study, we leverage an adapted version of the conventional Fre’ chet inception distance (FID) score, a dissimilarity measure, to evaluate the quality of the generated data. Utilizing generative adversarial network (GAN) as our GenAI model, we benefit from its well-documented efficiency in tabular data setups. The adjusted FID score demonstrates a notable enhancement, with interpolated data scoring 12.05 and generated data scoring 1.56, while the comparison of subsets of measurement data results in 0.02. The cumulative Kullback-Leibler divergence (KLD) of generated data with GAN shows slightly worse performance with a value of 7.68 in comparison to linearly interpolated data of 3.75, with measurement data scoring 0.04. However, the influence of FID score improvement significantly overcomes the decrease in the KLD performance, as we show strikingly similar correlations between the features of GAN-generated data and measurement data, in contrast to predominantly zero correlations observed in linearly interpolated data.

Authors 5

1. Kostiantyn Lavronenko 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,Aachen,North Rhine-Westphalia,Germany

2. L. Lopes de Paiva 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,Aachen,North Rhine-Westphalia,Germany

3. 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,Aachen,North Rhine-Westphalia,Germany

4. Volkmar Schulz 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,Aachen,North Rhine-Westphalia,Germany

5. 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,Aachen,North Rhine-Westphalia,Germany