Hybrid PET-MR List-Mode Kernelized Expectation Maximization Reconstruction// Confounding in age-period-cohort models

Daniel Deidda & Thomas Higgins, University of Leeds. Part of the statistics seminars series.

Deidda: Ordered subsets expectation maximization (OSEM) has been widely used in positron emission tomography (PET) imaging. Although Bayesian algorithms have been shown to perform better than OSEM, they are still not widely used in clinical practice due to the difficulty of choosing appropriate and robust regularization parameter values. The recently introduced kernelized expectation maximization (KEM) method has shown promise for different applications. Therefore, we propose and investigate a list-mode-hybrid KEM (LM-HKEM) reconstruction algorithm, implemented in the open source Software for Tomographic Image Reconstruction. The proposed algorithm uses both the magnetic resonance (MR) image and PET update image to create the kernel matrix. One phantom and two clinical datasets were acquired with the Biograph Siemens mMR: the phantom to validate the algorithm and two carotid artery patient studies, with two different tracers, to show possible applications. The reconstructed images are assessed and compared for different LM algorithms implemented in STIR: OSEM, the Bayesian version of OSEM with median root prior, LM-KEM and LM-HKEM. The results show better contrast and bias for LM-HKEM with promising quantification performances even for the low-count images. There is around 4% bias for LM-HKEM compared to 8% for LM-KEM and over 10% for the other techniques for the phantom and similar results for the patient data. Our results show that the proposed technique can be used to improve quantification at different noise levels. This shows promising stability indicating that it is unnecessary to change parameter values for different datasets with comparable count-levels.

Higgins: An individual\'s life starts at birth and ends at death. Given their date-of-birth (“cohort") and the current date (“period"), their age is determined as period minus cohort. For a population of individuals, we record the cohort, period-at-death and age-at-death of each individual. For convenience, we round these three numbers down to the nearest integer. Usually data collectors count deaths in the current integer period and categorise by integer ages. Lifetables are a prime example. However, given age and period, cohort can now take two possible values. This age-by-period rounding can only approximately capture the effects of age and cohort.

The effects of age, period and cohort on the risk of death could be quite different when studied under age-by-period rounding and the ideal age-by-cohort rounding. Our paper explores a problem of misinterpretation when the age distribution is independent of cohort and the number of births is rapidly changing with cohort. An artefact suggesting a shortening of the age distribution appears under age-by-period data that is not apparent under age-by-cohort data. The use of age-by-period rounded data is widespread and so we explore the significance of this artefact for a number of case studies.

Daniel Deidda & Thomas Higgins, University of Leeds