Medical PET Group - Biological Imaging lgs@ads-lgs.com
the original publication is at Alasbimn Journal Year 4, No. 13, October 2001. The following text is retrieved from a poster presentation.

High quality PET images are a prerequisite for an accurate diagnosis in oncological patients. While the standard filtered backprojection (FBP) method is frequently limited by artefacts due to high localized activity concentrations and/or low number of counts per image, the iterative image reconstruction (IIR) technique provides superior image quality. However, the iterative technique requires a longer reconstruction time as compared to the FBP. We implemented an iterative reconstruction program in a Windows NT environment (Fig. 1). The system for image reconstruction consists of two major parts:

1. a web interface for the input of the reconstruction parameters

2. the reconstruction program module (OS-EM and other methods)

The acquisition data are transferred via ftp from the PET system to the subnet server. The web interface is accessable via computers within the intranet and is used to provide the individual parameters for the image reconstruction, e.g. number of frames, slices, reconstruction method, etc. (Fig. 2). Reconstruction programs are running on dedicated PC systems (double and single processor systems) and checking the subnet server for new reconstruction tasks. The reconstructed PET images are stored on the subnet server for further evaluation on standard PC systems using the software package PMOD (provided by a cooperation with C. Burger, Clinic of Nuclear Medicine, Univ. of Zuerich, Switzerland).

Medical PET Group - Biological Imaging l.strauss@dkfz.de

Fig. 1: System diagram for the iterative image reconstruction and evaluation of PET data.

Medical PET Group - Biological Imaging l.strauss@dkfz.de

Fig. 2: Web interface for the input of reconstruction parameters. The form is accessable via the intranet. The minimum input consists of the patient data folder and the bed position. All parameters of the reconstruction module can be modified. Usually the OS-EM (ordered subsets, maximum likelihood expectation maximization) or OS-WLS (ordered subsets, weighted least squares) algorithm and the MRP (median root prior) correction are used.

Medical PET Group - Biological Imaging l.strauss@dkfz.de

Fig. 3: FBP (upper, left) and IIR (upper, right) image of a recurrent colorectal carcinoma. The lesion is only detectable in the iteratively reconstructed image.

The power spectrum (lower images) demonstrates less noise in the IIR image.

Fig. 4: Small liver metastasis in a patient with ovarian cancer (40-60 min image following FDG injection, theoretical slice thickness 2.425 mm). The FBP image (left) is limited by severe artefacts. Significant improvement of the image quality by the iterative reconstruction method (right).

Medical PET Group - Biological Imaging l.strauss@dkfz.de

Fig. 6: The image quality was assessed with respect to the number of iterations (1, 3, 6, 32 iterations). Parameters: 256*256 matrix, 6 iterations, 4 subsets, MRP = 0.3.

The convergence was best for the OSWLS algorithm, as shown in the diagram below. Actually all methods have a fast convergence and achieve comparable SUVs. OSWLS has found preferential use when only a few iteration steps are used for image reconstruction.

Medical PET Group - Biological Imaging l.strauss@dkfz.de

Fig. 7: The use of the "conventional" attenuation map as provided by the system together with the IIR of the emission data introduces significant artefacts due to the FBP, especially in areas with high attenuation.

The use of both, iteratively reconstructed attenuation maps (on the basis of the transmission data) and the iterative processing of the emission data, provides a superior image quality.

Medical PET Group - Biological Imaging l.strauss@dkfz.de

Fig. 8: Without the use of the MRP correction, the noise is increasing for both high (metastasis) and low uptake (liver parenchyma) areas with the number of iterations.

In general, MRP decreases the noise significantly. Usually a MRP coefficient of 0.1-0.3 is used for the image reconstruction (dependent on the total number of counts).

Medical PET Group - Biological Imaging l.strauss@dkfz.de

Tab. 1: Performance of the IIR according to the PC and operating system. Parameters used: one cross section, 256*256 matrix, 6 iterations, 4 subsets. The performance data are given for one processor. Two double processor systems are used routinely, so the performance is enhanced by a factor of four. The number of additional PCs which can be used for reconstruction is not limited.

Tab. 2: Performance of the IIR according to the reconstruction algorithm. The same parameters as in tab. 1 are used. Generally the ordered subsets (OS) method is used together with all algorithms (EM: maximum likelihood expectation maximization; WLS: weighted least squares; ISRA: image space reconstruction algorithm; SAGE: space alternating generalized expectation maximization).

The iterative image reconstruction can be performed routinely on PC systems using distributed data processing and provides a superior image quality in comparison to the FBP. The comparison of different reconstruction algorithms shows that standard reconstruction parameters should be used to obtain quantitative results, which can be compared for different patients and studies. Based on the quantitative evaluation of high and low uptake areas, the OSWLS has the best performance with regard to convergence and noise. The ordered subsets method enhanced the reconstruction program, however a larger number of subsets was associated with an increase in noise. The median root prior correction was helpful to achieve less sensitivity of the IIR for the reconstruction parameters.