Comparison of four dynamic filtering techniques in the reconstruction of 4D PET data

Complete text of publication follows: Objectives: The most common approach to reconstructing dynamic positron emission tomography (PET) data consists in independently reconstructing each dynamic frame. This approach generally results in very noisy images as the frame duration is often short. Dynamic image filtering techniques have been proposed to obtain more accurate, less noisy images. However they are generally used only post-reconstruction. Here, we study how dynamic filtering techniques can be used within an iterative reconstruction process to better control the noise that increases with the number of iterations. Compared to a conventional reconstruction approach, the reconstruction of each frame benefits from the entire information acquired during the acquisition. This study aimed to provide an overall analysis of the influence of four existing filtering techniques on the reconstruction of 4D PET data. Methods: Realistic simulations of PET data acquired on a SIEMENS Biograph with [¹¹C]PE2i were performed using an analytic procedure previously proposed. Iterative reconstruction of the simulated data was performed using the Customizable and Advanced Software for Tomographic Reconstruction (CASToR, http://www.castor-project.org/) platform. The 4D reconstruction was performed by alternating between the reconstruction of each individual frame and the dynamic filtering of the entire sequence. Four dynamic filters were tested: highly constrained backprojection (HYPR), kinetics-induced bilateral filter (KIBF), vector-based robust anisotropic diffusion (VRAD) and temporal gaussian filter (TGF). Simulated sinograms were corrected for randoms and diffusions. Images were reconstructed using 10 iterations and 16 subsets with an OSEM procedure with attenuation correction. Filtered and non-filtered reconstructions were compared to the Ground Truth using three quantitative criteria: Signal to Noise Ratio (SNR), Structural Similarity Index (SSIM) and Root-Mean-Square Error (RMSE). Results: The average SNR of the images reconstructed with no filter was 2.6324. The average SNR of the images reconstructed with alternating dynamic filtering using HYPR, KIBF, VRAD and TGF were respectively 10.88, 12.17, 11.62 and 9.73. The average SSIM of the images reconstructed with no filter was 0.41 and the average SSIM of the images reconstructed with alternating dynamic filtering using HYPR, KIBF, VRAD and TGF were respectively 0.48, 0.49, 0.51 and 0.43. The average RMSE were respectively 8.3 for the non-filtered reconstructions and 3.2, 2.8, 3.0 and 3.7 for the filtered reconstructions. Conclusion: The use of dynamic filtering within iterative reconstruction of 4D PET data improved the quality of reconstructed images with all four filters compared to classical 3D non filtered OSEM. In our experiments, the most effective filters were KIBF and VRAD, followed by HYPR and TGF. Different behavior were observed and additional data is required to fully study the optimal reconstruction and filtering parameters to achieve the most accurate quantification of dynamic PET data. Nevertheless, these results indicate the potential of the proposed inter-iteration dynamic filtering for the reconstruction of dynamic PET data