First evaluation of a fast full 3D list-mode based image reconstruction for PET

Ziel/Aim:

Despite the fact that all modern PET scanners support 3D data acquisition protocols, up to now only a few of them have supported full 3D image reconstruction. Normally the reconstruction task is performed in sinogram space and reduced to a 2D problem using Fourier rebinning. This implies certain approximations and can degrade image quality. 3D list-mode reconstruction potentially is able to overcome these limitations while allowing at the same time flexible integration of motion correction methods into the reconstruction. The very limited availability of corresponding open-sourced software motivated us to develop our own, platform independent, fully 3D list-mode reconstruction with the final goal of integration of our event-based motion correction into it.

Methodik/Methods:

As the basis for our reconstruction we have taken the Ordinary Poisson List-mode Ordered Subsets Expectation Maximization algorithm (OP-LMOSEM) with on-the-fly system matrix simulation using a ray-tracing technique. The source code (C++) supports multi-threading and allows distributed computing, both of which decreases reconstruction time considerably. It also includes all
necessary corrections (attenuation, normalization, randoms etc.). We use the Single Scatter Simulation algorithm to compensate for Compton scatter and have evaluated the new reconstruction by comparison with the standard OSEM-reconstruction available with our Siemens EXACT HR+ scanner. Phantom measurements were performed in list-mode using the software previously developed in our lab. The evaluation procedure has been divided into three parts: i) quantitative accuracy (ROI's mean value comparison); ii) spatial resolution (FWHM comparison); iii) Signal to Noise Ratio, SNR (ratio of standard deviation to mean value in homogeneous ROIs).

Ergebnisse/Results:

Relative to the standard reconstruction we obtain the following results: quantitatively, images show reasonable concordance with the reference, abstract-dialog | Abstract Management System
the differences are below 8%. The reconstructed spatial resolution is on average 10% better (up to 20% in smaller structures). The mean SNR shows 6% improvement (especially at the axial edges of the field of view). Currently, reconstruction time for a typical 5 minute FDG brain scan is 13 minutes using 64 cores running at 2.3 GHz.

Schlussfolgerungen/Conclusions:

3D list-mode reconstructions are approaching clinical usefulness and prospectively offer the optimal framework for incorporating event-based motion correction methods into the reconstruction. Our implementation has proven to provide results, which already are better then those of the standard reconstruction on our system, although there is still room for much improvement. In the next step we plan to incorporate motion correction into the algorithm.