A Generalized Linear modeling approach to bootstrapping multi-frame PET image data

Abstract

PET imaging is an important diagnostic tool for management of patients with cancer and other diseases. Medical decisions based on quantitative PET information could potentially benefit from the availability of tools for evaluation of associated uncertainties. Raw PET data can be viewed as a sample from an inhomogeneous Poisson process so there is the possibility to directly apply bootstrapping to raw projection-domain list-mode data. Unfortunately this is computationally impractical, particularly if data reconstruction is iterative or the acquisition protocol is dynamic. We develop a flexible statistical linear model analysis to be used with multi-frame PET image data to create valid bootstrap samples. The technique is illustrated using data from dynamic PET studies with fluoro-deoxyglucose (FDG) and fluoro-thymidine (FLT) in brain and breast cancer patients. As is often the case with dynamic PET studies, data have been archived without raw list-mode information. Using the bootstrapping technique maps of kinetic parameters and associated uncertainties are obtained. The quantitative performance of the approach is assessed by simulation. The proposed image-domain bootstrap is found to substantially match the projection-domain alternative. Analysis of results points to a close relation between relative uncertainty in voxel-level kinetic parameters and local reconstruction error. This is consistent with statistical theory.

Mapping18F-FDG Kinetics Together with Patient-Specific Bootstrap Assessment of Uncertainties: An Illustration with Data from a PET/CT Scanner with a Long Axial Field of View

“…2 ). Numeric studies ( 10 , 11 ) have shown that image-domain DGP bootstrapping matches the accuracy of the much more computationally intensive list-mode bootstrapping approach of Haynor and Woods ( 19 ).…”

“…NPRM approximates the voxel-level residue by the positive linear sum-of-basis elements that have been selected by a cross-validation–guided analysis of a comprehensive collection of time courses produced by segmentation of all the available data in the study ( 10 , 18 ). Individual basis elements are of the form for k of 1, 2,…, K. Here, R k is the basis element residue and Δ k is its associated delay factor.…”

“…For any delay, the optimal set of α coefficients is found by quadratic programming. A crude grid search is used to optimize delay ( 10 ).…”

Quantitation of Dynamic Total-Body PET Imaging: Recent Developments and Future Perspectives

“…Building on (17), an image-domain bootstrap data generation process can be defined by the spatial and temporal patterns of model residuals [135,136]. It has been used to assess the uncertainty (standard errors) of parametric imaging [137].…”

PET AIF estimation when available ROI data is impacted by dispersive and/or background effects

“…Based on the analysis provided in (Gu et al 2021 ) the optimal statistical combination of separate estimators is the covariance-weighted average of the individual estimators, i.e. where is the covariance of the j th estimator—this might evaluated using a suitable bootstrapping process (O’Sullivan et al 2021 ). The current implementation uses a more simplified weighted averaging process.…”