Dedicated 3D list-mode reconstruction for whole-body PET

Although all currently available PET scanners support 3D acquisition protocols, full 3D image reconstruction is only rarely available in clinical PET. Usually, Fourier rebinning is used to reduce the image reconstruction to a set of independent 2D problems. Nowadays, faster computer systems allow to consider real 3D reconstruction as superior approach even for clincial routine, especially for whole body investigations. We report on our work on a dedicated whole-body image reconstruction using a 3D list-mode-based algorithm.

Methodik/Methods:

We implemented a 3D Ordinary Poisson List-mode Ordered Subsets Expectation Maximization algorithm (3D-LMOSEM) with on-the-fly calculation of the system matrix. Matrix elements are considered to be proportional to the intersection volume of voxels with Lines-of-Response (LORs) having a finite cross section. Two assumptions were used in this calculation:
i) the voxel grid is considered to consist of overlapping spheres, where the sphere volume is taken to be equal to that of the corresponding cubic voxel,
ii) LORs are cylinders whose axes connect the centers of the contributing detectors and the radii are calculated from the initial LOR volume.
This procedure reduces the complexity of system matrix computation considerably. The reconstruction is performed simultaneously for all bed positions. This bears the following advantage:
i) count loss in overlap regions is avoided leading to improved image quality,
ii) simultaneous reconstruction of all bed positions enables improved scatter calculation.
The new reconstruction was evaluated by comparison with the standard sinogram-based attenuation weighted OSEM-reconstruction (AW-OSEM) available with our Siemens EXACT HR+ scanner. The evaluation addressed these parameters: i) quantitative accuracy; ii) spatial resolution (FWHM comparison), iii) artifacts and noise characteristics.

Ergebnisse/Results:

Quantitative analysis yielded differences of mean values in selected regions-of-interest below a few percent. 3D-LMOSEM shows much better resolution (4.05 mm in brain and 4.35 mm in whole-body) compared to AW-OSEM (5.65 mm and 5.98 mm, respectively) with less inhomogeneity artifacts and noise. Typical 3D-LMOSEM reconstruction time for a FDG whole-body scan (7 bedpos; 6 min each) is 220 min using 80x2.3 GHz cores.

Schlussfolgerungen/Conclusions:

Compared to AW-OSEM the 3D-LMOSEM provides improved image resolution while decreasing overall noise level. The possibility to reconstruct all bed positions in one run prevents count loss and allows for a more accurate Out-of-FOV scatter correction. Due to the use of multi-threading and distributed computing techniques reconstruction times are approaching a level which will allow use of 3D-LMOSEM in clinical routine in the near future.