IV. RESULTS
A. System Matrix Calculation
The subset of system matrix from the ring difference 0 to 87 was pre-computed based on the proposed 3D system model with a 1.5mm voxel. The calculation time was about 26 hours on a single Pentium 3.2GHz PC. The size of the subset is reduced when the maximum ring difference (MRD) is restricted. Fig. 4 shows the relationship between the size of the subset and the MRD. The total number of LORs is also plotted with a relative scale. When the MRD increases, the size of the subset increases rapidly while the total number of LORs increases slowly, because DRFs of a large MRD become broader. It should be noted that the size of the subset for the non-DOI case becomes larger because broadness of DRFs directly affects the ratio of non-zero system matrix elements. In this paper, we implemented 3D OSEM on a single Itanium 1.6GHz PC with 16Gbyte memory installed. In order to store the memory with all the subset of system matrix, the MRD was restricted to 54. Loss of counts was not significant because the reduction of total number of LORs was only 10%.
3386
0.010.020.030.040.050.060.001020304050607080900.00.20.40.60.81.0System matrix size [
Gbyte]Maximum ring differenceTotal number of LORs (
relative)DOICnon-DOItotal number of LORs0.010.020.030.040.050.060.001020304050607080900.00.20.40.60.81.0System LORs
Fig. 4. The subset size of the pre-computed system matrix.
B. Human Brain Imaging
A normal volunteer (21 years old male, 60 kg) was scanned for 40 min (100 min past 104MBq FDG injection). 3D OSEM (8 subsets, 8 iterations, 1.53mm3 voxel) was applied to DOI-compressed data with MRD=54. Any other compression such as span was not applied. Reconstructed images are shown in Fig. 5. These images show the excellent imaging performance of the jPET-D4: clear demarcations of thin gray matter from underlying white matter as well as fine visualization of the deep structures such as caudate nucleus, putamen and thalamus.
Fig. 5. Human brain images of the jPET-D4.
V. CONCLUSIONS
In this work, we present software strategies for fully 3D image reconstruction and imaging performance of the jPET-D4 prototype. The 142Pbyte system matrix was compressed into 13.4GB by reducing zero elements, applying the 3D-expanded DOIC method, factorizing with respect to ring differences and restricting the MRD to 54. Histogram-based 3D OSEM based on accurate system modeling was implemented on a single Itanium 1.6GHz PC with 16Gbyte memory. The first human brain images show the excellent imaging performance of the jPET-D4. At this development stage, calculation time was 7.5 hours per iteration. We think that speedup is possible by modifying program cording because random access to data overflowing system cash memory would be a bottleneck.