Assessment of an improved motion correction procedure for high temporal resolution positron emission tomography

Abstract

Introduction: State of the art Positron Emission Tomography (PET) scanners have pushed the limits of molecular imaging for the past decades. During this evolvement towards higher temporal resolutions, new ways of acquiring, processing and correcting PET data were established and recently, the introduction of bolus+constant infusion radiotracer functional PET (fPET) has enabled the assessment of dynamic sequences with frame lengths in the range of seconds. This allows to study brain metabolism, neurotransmitter dynamics or mechanisms related to disease with an unprecedented precision. However, the possibility of functional imaging in PET at such a high temporal resolution has re-introduced an existing challenge for clinicians and researchers alike: compensation of patients’ motion within and between individual dynamic time frames. There are several approaches which employ motion correction before and after image reconstruction. Unfortunately, these procedures are often unsuccessful due to a low signal-to-noise ratio (SNR), especially prevalent in early and very short image frames. Here, the application of deep learning models for PET motion correction may offer a solution.Objective: This thesis sets out to explore whether a recently introduced generative network algorithm, designed to improve the SNR characteristics of early dynamic PET frames, also improves the overall motion correction for high temporal resolution fPET data.Methods: A dataset of 19 subjects, which underwent simultaneous PET and magnetic resonance imaging (MRI), was used to assess the effectiveness of PANDA, an algorithm based on a conditional Generative Adversarial Network (cGAN) that predicts a projection of early-frame activity patterns with a late-frame SNR. The duration of the dynamic PET sequences amounts to only about 3s, substantially limiting the SNR in this data. The artificial high-SNR fPET data were used to retrieve motion parameters that were subsequently used for realigning the original dynamic image sequences. Realignment with the presented prediction framework was incorporated into a preprocessing routine and validated, by comparing realignment parameters as well as the neuronal activation induced by task-performance. Results: An increase of SNR and similarity metrics between early and late frames in the dynamic series was achieved. However, no substantial improvement with regard to the motion correction as defined by the motion parameters was observed. In a respective analysis of neuronal activation, decreases of potential motion artifacts were observed, although the artifacts themselves were below the threshold of statistical significance.Conclusion: PANDA will not be incorporated in future pre-processing pipelines, as its disadvantages, such as high processing time and required resources, outweigh the potential minimal improvements of the overall motion correction.