Abstract
574
Objectives: The iterative image reconstruction in emission tomography is heavily influenced by the noise induced by limited data statistics. Regularization by quadratic prior has been successful in reducing the noise in the data, but with the washout of useful features from the image, due to inability of the quadratic prior to distinguish between the real features and noise. On the other hand, it can be noticed during iterative reconstruction (e.g., using OSEM) that the low frequency component of the image is being recovered in the early iterations, while medium and high frequency components (i.e., organ boundaries and lesions need more iterations to resolve), and noise contains majorly high frequency components and has less impact to early iterations. Here, we propose and investigate the evolution-scaled quadratic prior (ESQP) where the penalty strength is being modulated by the evolution of the image components as these are being reconstructed iteratively. The performance of the proposed technique is evaluated in application to PET, and compared to standard OSEM, as well as classical quadratic prior regularization.
Methods: The main idea behind the proposed ESQP is to utilize the evolution map of the changing regions generated in early iterations of the OSEM reconstruction. The evolution map is defined as the difference between two image estimates that were generated during early iterations of OSEM, thus it should be noise-free. The evolution map can then be inverted into the map for scaling the beta parameter in quadratic prior. Large values of the evolution map (regions that are still recovering) would mean that the beta parameter need to be reduced (reducing the penalty), and small values of the evolution map (already converged, flat regions) would indicate the need to increase the beta parameter (increasing the penalty to reduce any potential noise). We compare the performance of the ESQP to that of unregularized OSEM, as well as to standard Quadratic prior with constant β, using the data from Philips Vereos PET/CT with 5 mm diameter inserted lesions.
Results: The evolution-based scaling of the quadratic prior promoted the recovery of the lesions and boundaries that were still converging when the feature image has been generated. At the same time, it has penalized the low frequency (background) regions which were already largely converged in order to prevent noise amplification. Therefore, the ESQP was able to achieve both competitive contrast recovery of the inserted small lesions, as well as to mitigate the noise in the background.
Conclusion: The proposed approach demonstrates the improvement in noise and contrast over the standard OSEM, as well as OSEM penalized with standard quadratic prior. The approach demonstrates similar edge-preservation potential to other edge-preserving priors, however, with the benefit of the parameters tuned to the intrinsic object properties as they evolve during reconstruction. Future work is needed to refine and verify the optimal study-independent rules for the generation of the evolution scaling map, for example, rules for consistent selection of the iteration difference in such way so that even weaker lesions and boundaries can be highlighted. Research Support: This work was supported by Philips Healthcare.