[Optimization of PET image quality by means of 3D data acquisition and iterative image reconstruction]

In the recent past, several algorithms have been developed in order to transform 3D sinograms acquired at volume PET systems into 2D data sets. These methods offer the possibility to combine the high sensitivity of the 3D measurement with the advantages of iterative 2D image reconstruction. The purpose of our study was the assessment of this approach by using phantom measurements and patient examinations.

The experiments were performed at the latest-generation whole-body PET system ECAT EXACT HR+. For 2D data acquisition, a collimator of thin tungsten septa was positioned in the field-of-view. Prior to image reconstruction, the measured 3D data were sorted into 2D sinograms by suing the Fourier rebinning (FORE) algorithm developed by M. Defrise. The standard filtered backprojection (FBP) method and an optimized ML/EM algorithm with overrelaxation for accelerated convergence were employed for image reconstruction. The spatial resolution of both methods as well as the convergence and noise properties of the ML/EM algorithm were studied in phantom measurements. Furthermore, patient data were acquired in the 2D mode as well as in the 3D mode and reconstructed with both techniques.

At the same spatial resolution, the ML/EM-reconstructed images showed fewer and less prominent artefacts than the FBP-reconstructed images. The resulting improved detail conspicuously was achieved for the data acquired in the 2D mode as well as in the 3D mode. The best image quality was obtained by iterative 2D reconstruction of 3D data sets which were previously rebinned into 2D sinograms with help of the FORE algorithm. The phantom measurements revealed that 50 iteration steps with the optimized ML/EM algorithm were sufficient to keep the relative quantitation error below 5%.

Our measurements show that the image quality in 3D PET can be improved by using iterative reconstruction techniques. The concept of 3D data acquisition and combining the FORE algorithm with 2D ML/EM reconstruction can readily be employed in clinical practice since the computation time is not considerably longer than that in iterative reconstruction of true 2D data.