PyTomography: Advancements in List-mode and Time-of-Flight PET Image Reconstruction

Introduction: In this work, list-mode (LM) and time-of-flight (TOF) PET modeling were developed in the open-source python-based medical image reconstruction library PyTomography. The system modeling includes both forward and back projections, permitting image reconstructions using all the available modality-independent reconstruction algorithms available in the library. While basic backend functionality was developed to make the library directly compatible with GATE simulation data, focus was also placed on support for the more recent PETSIRD LM data format developed by the ETSI initiative (https://etsinitiative.org/), which is both vendor- and scanner-agnostic.

Methods: LM + TOF PET modeling in PyTomography was implemented using the publicly available parallelproj python library, which offers GPU-accelerated 3D Joseph projectors (TOF and non-TOF). The developed system matrix has functionality for attenuation/normalization correction and arbitrary PSF modeling. Input/output functionality was developed for reading raw event data and detector lookup tables in the following formats: (i) GATE (ROOT) data and (ii) PETSIRD datafiles. Additional functionality was developed to aid in the computation of normalization coefficients from simulated calibration scans.

The capabilities of the library are demonstrated for two use cases from data simulated in GATE; events corresponding to phantom scatter and random coincidence were discarded in each case. The first use case considered a generic cylindrical scanner (61.5cm diameter, 12.8cm axial field of view, 283 ps TOF resolution) and acquisition of a Jaszczak phantom with uniform background activity and 6 hot spheres at a 4:1 target-to-background activity concentration ratio; ROOT simulation data was converted to the PETSIRD format using PETSIRD conversion tools before being read by PyTomography. Data were reconstructed with/without TOF using BSREM (100it/1ss) with the relative difference penalty (β=50, γ=0.2 ). The second use case considered acquisition of an ultra-high-resolution PET/MR brain phantom on a scanner with geometry representative of a Siemens mMR geometry, but with 550 ps TOF resolution. In this case, an additional calibration scan using a thin cylindrical shell was simulated to obtain normalization coefficients. Reconstructions were performed in the following 3 ways: (i) using the full dataset with OSEM (4it/20ss), (ii) using 6% of the dataset with OSEM (80it/1ss), and (iii) using 6% of the dataset with BSREM (80it/1ss), where BSREM used the relative difference penalty (β=100 , γ=0.2) , including only 8 nearest neighbours based on the MR image).

Results: The PETSIRD Jaszczak phantom data was successfully read and reconstructed with/without TOF in the PyTomography framework; the mean source-to-background activity ratio in the six spheres was 3.32 (without TOF) and 3.89 (TOF). Similarly, the ultra-high resolution brain phantom had lower levels of bias in grey/white matter when reconstructed with TOF. For TOF reconstructions, high-count OSEM and low-count BSREM yielded similar levels of noise in grey matter (14.3% vs. 13.9%) but low-count BSREM had significantly larger bias (24.6% vs. 28.2%). In both low-count scenarios, OSEM had substantially larger noise than BSREM in grey matter (28.8% vs. 13.9%), but reduced bias (25.0% vs. 28.2%). Plots showing iteration dependent bias/noise are shown in Figure 3 of the supplemental materials.

Conclusions: LM + TOF PET modeling and reconstruction were developed in the open-source python library PyTomography. Use cases were explored for both GATE (ROOT) and PETSIRD raw data using OSEM and BSREM reconstruction algorithms. All code for reconstruction is publicly available on the PyTomography GitHub page.

• Download figure
• Open in new tab
• Download powerpoint