A quantitative image reconstruction platform with integrated motion detection for total-body PET

Introduction: The ability to accurately measure radiotracer concentration across the body is a hallmark of PET. To do so requires physics corrections applied during image reconstruction - corrections for attenuation, scatter, randoms, deadtime, normalization, and decay - as well as motion compensation during the scan. Quantitative bias from motion artifacts is accentuated in total-body PET compared to conventional PET with shorter axial FOV, since the wider axial acceptance angle distributes the quantitative bias from mu-map mismatch over a wider image volume compared to conventional PET. Our goal is to develop an image reconstruction platform for total-body PET that delivers accurate quantitation by combining iterative image reconstruction with frame-based motion compensation across the body. Here, we show a preliminary evaluation of our image reconstruction and motion detection using phantom experiments and human subjects.

Methods: An in-house listmode TOF-OSEM reconstruction algorithm was developed incorporating all physics corrections. Besides normalization factors that are provided in the raw data, the in-house reconstruction code and correction factors are independent from those provided in the vendor’s software. A Monte Carlo simulation (based on SimSET) algorithm was developed to estimate block-pair scatter sinograms by scaling the CT-derived Hounsfield units to water-equivalent materials with variable density and simulating the scatter distribution using the prior iteration image as the trues input. Paralyzable deadtime and randoms from delayed coincidences are estimated from block-pair singles and coincidence count rates. Motion is detected throughout the body by computing the TOF-weighted centroid-of-distribution (COD) within each 5 x 5 x 5 cm3 volume in the FOV with 0.1 s temporal sampling. Motion detection was performed for 8 anatomical regions by computing the average COD traces from the inclusive (x,y,z) grid-CODs in each respective drawn ROI. A statistical change-point detection algorithm is used to detect motion in each ROI from the average COD traces.

Results: Total image-derived activity bias (max vs min) in a 15 cm x 210 cm uniform pipe phantom and an initial activity of 173 MBq in the FOV was 1% over a 60-min dynamic reconstruction (2 - 300-s frame lengths). Low image noise was achieved with the in-house software, while there are differences in image contrast compared to matched reconstructions obtained with the vendor’s software especially in the lungs and lower abdomen. The motion detection survey suggests constant motion frequency throughout the duration of the 20 min PET acquisition, with an average rate of approximately 4-6 motion time-points (MTPs) per minute, and an average of ~100 unique MTPs for the whole body.

Conclusions: The custom image reconstruction platform developed for total-body PET provides flexibility to support future research (kernel-based reconstruction, deep-learning, parametric imaging), as well as an independent benchmark for validating novel reconstruction algorithms, or as a comparison tool for human subject imaging alongside the vendor’s software. Integration of frame-based image registration for automated motion compensation across the whole body is in development, including the implementation of joint activity and attenuation reconstruction (MLAA) to obtain motion-matched mu-maps.

• Download figure
• Open in new tab
• Download powerpoint