Complementary Reconstruction for Improving Image SNR with PET Gating

Background: Respiratory and/or cardiac gating is a well-established method to “freeze” physiological motion and improve local image quantitation for positron emission tomography (PET). In particular, it is helpful, or even necessary, for analyzing data from cardiovascular studies. The gated reconstructions have higher quantitative accuracy, but since this technique involves discarding the PET data outside of the frame of interest, they suffer from higher image noise - this effect is amplified when performing combined dual, respiratory and cardiac gating. We investigate in this work complementary frame reconstruction (CFR) for improving the quality of dual-gated PET images.

Methods: CFR is a processing technique designed to improve noise and bias properties of very short-frame PET images by using the statistical support of the longer frame to which it belongs. Instead of extracting and directly reconstructing the PET data from a very short frame, those data are instead removed from the larger frame, reconstructed, and the image is subtracted from that of the long frame (without the counts of interest removed). The result is an indirect image of the frame of interest with noise and bias properties more closely matching that of the long frame reconstructions. CFR was originally introduced as a way to improve quantitative accuracy for short dynamic frames, when high temporal resolution was needed [1]. Here, we extend it to count-limited reconstruction where the frames are defined instead by physiological motion phase, rather than absolute time. The data from fifteen patients, undergoing FDG PET scans of the torso were retrospectively processed. For all patients, respiratory and ECG physio information was recorded concurrently with a 10-minute PET acquisition, and the listmode data were sorted into 32 dual-gated (4 respiratory and 8 cardiac) sinograms and reconstructed both by conventional, independent frame reconstruction (IFR) and CFR. Both IFR and CFR used the same TOF reconstruction with resolution modeling and all reconstructed image matrices were 200 x 200 x 109, with voxel dimensions 4.073 x 4.073 x 2.0313 mm. The images produced by both techniques were compared in terms of image noise and bias.

Results: Although images from both methods where noisy, the CFR yielded slightly lower voxel variance within a spherical volume of interest (VOIs) drawn over uniform regions in the liver. For a typical patient dataset, CFR and IFR yielded a mean ROI standard deviation of 43% and 46%, respectively, across all 32 gated images. Additionally, the noise structure was remarkably different - the CFR produced normally distributed noise, i.e. symmetric about the mean, whereas the IFR resulted in an asymmetric, Poisson-like distribution.

Conclusions: We present here an experiment evaluating performance for complementary frame reconstruction to improve the quality of gated PET images. The most notable finding of this experiment was the noise distribution found in the resulting images. As expected, the IFR produced an asymmetrical distribution - this is due to the sparse data (near-zero) reconstructed with the non-negativity constraint of the MLEM framework. The CFR also used MLEM, but the reconstructions were performed at levels of higher statistical support (farther from zero) and the resulting image noise followed more closely a normal distribution. This may translate to lower bias and higher quantitative accuracy for gated PET image analyses.