Relative Difference Prior with Dynamic Edge Preservation Parameter

Objectives: Effective noise control in the iterative reconstruction is challenging due to the statistical nature of emission tomography. Regularization by using edge preserving priors such as Relative Difference Prior (RDP) has been successful in reducing the noise in the image while preserving the contrast. The task of setting optimal parameters for such priors is still challenging. Often, even for hand-optimized parameters, there is still a risk that certain noise fluctuations may get past the edge preservation threshold, get amplified during the iterative reconstruction and appear as “real” features (hot-spots) in the final image. Since the common property of the typical iterative image reconstruction algorithm (e.g., OSEM) is that noise is amplified towards later iterations, we propose to use regularized reconstruction with dynamically adjustable, iteration-dependent edge preservation parameter (such as the gamma parameter in RDP). We investigated the performance of the proposed approach using RDP for real PET data.

Methods: We utilize the following feature of the iterative reconstructions. During the iterations of OSEM, large components (lower frequency) of the image converge faster. At later iterations, the high-frequency components increase which most likely represent the noisy component of the image. The main idea of this work is to allow stronger edge preservation in earlier iterations (lower penalty to allow for unrestricted recovery of real features before the noise starts to build up). As the regularized reconstruction goes towards later iterations to pursue convergence, the edge preservation parameter is relaxed (leading to gradual increase of the penalty) to prevent the noise amplification and ensure convergence. We compared the performance of the proposed method to the standard RDP with fixed parameters optimized either for improved contrast or maximum noise and hot-spot removal. We used NEMA IQ phantom data to test the proposed method.

Results: In one particular reconstruction using NEMA IQ phantom data, with a standard RDP penalty with settings (i.e., gamma=1.0) optimized for contrast preservation, in one of the slices there is a hot-spot artifact. This hot spot is easily removed when the contrast preservation threshold gamma is set lower (gamma=0.5) but at the cost of greatly diminished contrast for the 10 mm sphere. Whereas, the combination of RDP with dynamic edge preservation parameter (starting initially from a high gamma value of 1.5 and gradually reduced to 0.5 over the course of 20 iterations) allows the reconstruction to both suppress hot-spot development and to preserve the majority of contrast in the 10 mm sphere.

Conclusion: The proposed RDP reconstruction with dynamically changing iteration-dependent parameter demonstrated the improvement in noise-contrast tradeoff over the standard OSEM. The approach demonstrates the benefit of reduced chance for artificial hot-spots with similar edge-preservation potential compared to RDP with fixed parameters. Research Support: This work is supported by Philips.

• Download figure
• Open in new tab
• Download powerpoint