New Ect Reconstruction Algorithm with L0 Regularizer
Wei Zheng,Jiahan Zhang,Si Li,Xueying Zeng,C. Schmidtlein,David Feiglin,Yuesheng Xu,Andrzej Krol
2016-01-01
Abstract:1968 Objectives To study performance of fixed-point proximity algorithm with Moreau envelope of l_0 norm (L0-FPPA) gradient regularization in SPECT reconstruction. Methods In this study, we considered an objective function for SPECT image reconstruction that was nonconvex and discontinuous, which is well known to be very difficult to compute. However, we found an approach to link this problem to a convex one. Using the fixed-point proximity operator, we derived a novel algorithm (L0-FPPA) with the Moreau envelope of l_0 norm of the image gradient as a regularizer for SPECT reconstruction. Further, we proved its convergence to a local minimum under well-defined conditions. To evaluate the performance of L0-FPPA for SPECT, we created a digital cylindrical phantom that consisted of lumpy warm background with randomly distributed large Gaussian blobs and a set of 6 hot spheres with Gaussian activity distribution (128 × 128 × 128 voxels, voxel volume = 2.4 × 2.4 × 2.4 mm^3). The hot spheres had radii (FWHM) 3, 4, 5, 6, 7, and 8 mm and the maximum-activity-to-mean-background ratio for hot spheres of 4:1 and 7:1, respectively. The SIMIND Monte Carlo simulation package was used to model a Siemens E.Cam gamma camera with low-energy parallel-hole (LEHR) parallel-beam collimators. Up to third order scatter photons were considered in the simulation. We set the detector element size to 2.4 × 2.4 mm^2. The radius of rotation was set at 11 cm, and 120 projections were simulated for each phantom. The main energy window and scatter energy window were set at 127-155 keV and 123-127 keV, respectively. A total of 1.0 × 10^9 photon histories per view were simulated to create approximately “noise-free” data. Poisson noise was added to the simulated projection data. Four noise levels corresponding to 45,000, 89,000, 178,000 and 357,000 counts per view, respectively, were considered. A hundred noise realizations at each noise level were created for each phantom. Poisson noise was added to both photopeak-window and scatter-window counts according to the count level. Scatter correction was implemented by adding estimated scatter counts in forward projection at each iteration. Scatter counts were estimated using the scatter-window data. The reconstruction performance of L0-FPPA was compared with EM algorithm with Huber-regularizer (H-EM) and preconditioned alternating projection algorithm with isotropic TV regularizer (iTV-PAPA). The optimized parameters were qualitatively selected for all three algorithms. Results We obtained bias vs. noise, contrast-to-noise (CNR) vs. noise curves, channelized Hotelling observer (CHO) detection performance vs. sphere size, local spatial resolution and local noise power spectra for all noise levels and contrast ratios for all three algorithms. We also evaluated qualitatively the obtained images. Conclusions When the relative contribution of l_2 norm is larger than l_0 norm term in the Moreau envelope of l_0 norm the images obtained by L0-FPPA resemble images reconstructed by H-EM. However, when l_0 norm term is dominant, the L0-FPPA images resemble more TV-PAPA images including staircasing artifacts. L0-FPPA outperforms TV-PAPA in terms of lesion detectability but TV-PAPA exhibit better noise performance. Both algorithms significantly outperform H-EM. $$graphic_EA315DA9-AB6F-481B-8EA6-6D77BEDF8F64$$