Realistic Patient-Based E-Phantoms and Simulation of PET-MR Neuroimage Data
Bryant G. Mersereau1, Meher R. Juttukonda1, Hongyu An2, and David S. Lalush1

1Joint Department of Bioengineering, The University of North Carolina at Chapel Hill and North Carolina State University, Chapel Hill, NC, United States, 2Mallinckrodt Institute of Radiology, Washington University in St. Louis, St. Louis, MO, United States

Synopsis

Acquired PET-MR datasets can be problematic to build due to the high logistical and monetary cost associated with recruiting and scanning patients. We propose a new electronic PET phantom (E-phantom) platform to streamline the PET-MR research process. The proposed platform is shown to produce results consistent with acquired PET data reconstructed on manufacturer software and to provide high configurability and flexibility as a simulation tool.

Purpose

To create a realistic PET-MR simulation platform for use in neuro-PET-MR imaging research.

Introduction

The recent introduction of integrated PET-MR scanners has opened up a plethora of new areas of research utilizing multimodal patient data. However, due to practical, logistical, and monetary limitations, the acquisition of real-world tri-modality (PET/MR/CT) data can be problematic. The developed E-phantom platform solves the PET data insufficiency problem for researchers and streamlines the research process without compromising anatomical diversity or PET realism.

Methods

Phantom. The first step of the simulation process is the creation of patient-based PET E-phantoms from co-registered MR and CT images of the same subject. The most fundamental component of each E-phantom is an activity map of the desired tracer distribution. Activity maps are built from tissue-segmented MR (brain) and CT (skull/soft tissue) images with tracer-appropriate relative activity levels assigned by tissue class and literature-based tracer distributions. Activity maps may also be drawn from acquired PET data when available. Additional anatomical structures, such as hot or cold lesions, can also be added to the activity phantom. Lesions may be targeted to brain regions using a registered label atlas or limited to certain tissue classes. Each phantom also requires a true attenuation map created by conversion of CT-derived Hounsfield units into linear attenuation coefficients and, optionally, an attenuation correction map (μ-map).

Simulation. PET data acquisition is simulated by projection of the activity and true attenuation map into sinogram space using the geometry of any PET-MR system. The simulated PET data are then iteratively reconstructed using the industry-standard OSEM method accelerated by GPU. The reconstruction platform is capable of performing attenuation, randoms, and system uniformity corrections on the simulated data but does not support scatter simulation or correction. Reconstructed PET images can be matched to manufacturer reconstruction space, allowing straightforward comparison to acquired PET images.

Validation and capabilities of the simulation platform were demonstrated using tri-modality (18F-Florbetapir PET/MR/CT) data acquired from 28 subjects. The process was validated by comparing simulated to acquired PET images. Acquired PET data were reconstructed using manufacturer software with gold-standard CT-based attenuation correction. These reconstructions were then used as the activity map input to the E-phantom platform and reconstructed using the same gold-standard CT-based attenuation correction. Bulk regional and whole-brain relative error was calculated from the difference of the manufacturer and simulation platform reconstructions.

Results

Mean whole-brain error (±standard deviation) between the reconstruction platforms was 1.54% (±0.46%). Error analysis of MR-based μ-map performance revealed strong linear correlation (slope = 0.90, r2 = 0.93) between errors determined using manufacturer processing software and phantom data reconstructed on the E-phantom platform. The total time required to simulate acquisition of 3.5x108 lines of response (LORs), collapse to 7.3x107 LORs, apply correction factors, and iteratively reconstruct one phantom was 15 minutes on a computer with Intel i7-860 CPU, Nvidia GTX 780 GPU, and 8 GB RAM.

Discussion

The E-phantom simulation platform provides numerous advantages over real-world acquired PET data. Since only bi-modal (MR/CT) data are required to utilize the E-phantom platform, larger research datasets can be constructed than typically available to PET-MR researchers. These large datasets can improve anatomical diversity and statistical power over acquired PET data alone. Additional capabilities of the E-phantom platform include the ability to selectively add region targeted hot and cold lesions and to model different PET tracer distributions, such as PIB, FDG, or Tau. The platform has been demonstrated using MR-based μ-map evaluation; other applications include study of PET dose reduction and MR-driven reconstruction methods.

Acknowledgements

No acknowledgement found.

References

1. Juttukonda MR, Mersereau BG, Chen Y et al. MR-based attenuation correction for PET/MRI neurological studies with continuous-valued attenuation correction for bone through a conversion from R2* to CT-Hounsfield units. Neuroimage. 2015; 112:160-168.

Figures

Figure 1: Simulation components in the E-phantom platform. Figure 1A shows an example E-phantom with hot lesion added (arrow). Figure 1B is the true attenuation of activity. Figure 1C shows activity converted to sinogram space. Figure 1D is the μ-map. Figure 1E is the reconstructed image.

Figure 2: Results of E-phantom PET reconstruction process validation. Acquired PET data were reconstructed using manufacturer software with gold-standard CT-based attenuation correction. These reconstructions were then used as the activity map input to the simulation system and reconstructed using the same gold-standard CT-based attenuation correction.

Figure 3: Evaluation of an MR-based UTE μ-map using manufacturer software and the E-phantom platform. Error maps were generated based on results of reconstructions performed with gold-standard CT-based μ-map and the UTE μ-map. Error distributions from the E-phantom platform were consistent with error distributions from the manufacturer software.

Figure 4: Scatter plot and regression of percent error from simulated E-phantom reconstructions and percent error from manufacturer reconstructed acquired data. The regression analysis suggests a strong linear relationship between error analyses performed in E-phantom space and acquired data space.

Figure 5: Montage of reconstructed phantom 18F-FDG activity distributions created from segmented mixture model tissue maps derived from T1-weighted MR images. White-to-gray matter activity ratios vary between the phantoms, as well as the presence of hot and cold lesions.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
2194