Non-local means methods using CT side information for I-131 SPECT image reconstruction

Abstract

Recently, non-local means (NLM) methods for both image denoising and inverse problems have shown promising results in image processing and medical imaging. Moreover, some researchers have also shown that using additional information with low noise and/or high resolution for these problems can improve the image quality further. We investigated several NLM methods including NLM filters and NLM regularizers with and without CT side information for 1-131 SPECT image reconstruc tion. We compared two different ways to incorporate CT side information in the NLM filtering methods. We also propose a new way of incorporating side information in the NLM reconstruction method. XCAT simulation results with two uniform tumors (72 cc , 11 cc) show that the NLM filtering method with CT side information decreased the root mean square error (RMSE) up to 18.6% as compared to unregularized method. Our proposed NLM-based regularizers in iterative image reconstructions with CT side information also yielded up to 29.2% better RMSE and up to 18.1 % better recovery coefficient than un regularized reconstruction. I. I NTRODUCTION NON-LOCA L means (NLM) methods exploit the self similarity of small patches in an image for denoising [1] and for regularization [2]. These methods yield superior de noising results as compared to conventional local filters such as a Gaussian filter. Moreover, some researchers have also shown that using these NLM methods with high-quality side information can improve the image quality further [3]-[7]. Multi-modal imaging systems such as PET-CT, PET-MR, or SPECT-CT can benefit from these state-of-the-art methods to reconstruct better-quality images.SPECT-based dosimetry in 1-131 radioimmunotherapy and radioiodine therapy can be a natural application for these NLM methods using additional side information since high resolution CT image is available from SPECT-CT system. Improved accuracy in SPECT quantification for both total activity and distribution using these methods can potentially improve dose-response correlations so that the efficacy and toxicity of treatments can be better-assessed.In this paper, we evaluate various NLM filters and re construction methods with and without side information for an 1-131 SPECT-CT system, and compare the results with ).that of conventional ordered-subset expectation-maximization (OSEM) [8] with and without Gaussian post-reconstruction smoothing. We also propose a new way of incorporating side information in the NLM methods and show some promis ing preliminary results. 3D XCAT phantom simulation was performed to evaluate NLM post-reconstruction filtering and NLM reconstruction methods with and without high-resolution CT information. II. M ETHOD A. Statistical image reconstruction for SPECTThe SPECT image x can be reconstructed iteratively from X £ ar gm i nL(y l x) ",::: 0(1)where y is a measured sinogram data, L denotes a negativePoisson log-likelihood function (2) Y i is the ith element of the measurement y , and A denotes the system model and S i is a scatter component for t...

PET Image Reconstruction Using Kernel Method

“…Among different regularization methods [3]–[8], [10]–[13], the Bowsher method [9], which adaptively chooses neighboring pixels for each pixel in the image estimate using information from a prior image, was found to utilize anatomical prior information better than others in terms of performance and computational complexity [16]. The nonlocal regularization [17] can incorporate anatomical weights through a methodology similar to the neighborhood selection in the Bowsher method [14], [15] to improve emission reconstruction [18]–[20]. Alternatively, the prior image constrained compressive sensing (PICCS, [21]–[23]), an iterative reconstruction approach introduced in the context of dynamic computed tomography (CT), explores sparsity on the difference image between the image estimate and a composite image.…”

“…Besides the regularized image reconstruction methods, post-reconstruction denoising methods can also incorporate prior information to improve PET image quality [20], [27], [31], [32]. The highly constrained back-projection (HYPR, [25], [26]), which was originally developed in magnetic resonance imaging (MRI) for image reconstruction from under-sampled data, is a method of using a composite image prior and has been applied to denoising dynamic PET images [27].…”

A Novel Kernel-Based Regularization Technique for PET Image Reconstruction

“…The NLM algorithm is a nonlinear spatial filter that uses a weighted average of the intensity of adjacent pixels to adjust the intensity value of each pixel [20,21]. While comparisons of neighboring pixels can be performed between pixels separated in the image by any distance, for the benefit of computation speed, comparisons are typically limited to a narrow search window around the pixel undergoing value adjustment.…”

“…Alongside regular reconstruction methods, noise smoothing can be performed in the pre-reconstruction phase of the sinogram domain [11][12][13][14][15], in the post-reconstruction phase of the image [16][17][18], or during the iterative process of statistically-based reconstruction [19]. An alternative method that incorporates prior knowledge into de-noising images is non-local means (NLM) [20,21]. In general, applying denoising methods correctly and accurately post-reconstruction is more difficult than within the reconstruction.…”

Anatomy-guided brain PET imaging incorporating a joint prior model

“…Furthermore, Kazantsev et al (2011) presented a non-local algorithm for the weights estimation, which was included into the modified Bowsher prior in order to improve the robustness of the method to non-correlated information between activity and anatomy. Chun et al (2012) proposed a non-local means methods to incorporate side information in SPECT/CT reconstruction, while Nguyen and Lee (2013) proposed a new approach to incorporating prior anatomical information within PET reconstruction using the non-local regularization method, designed to selectively take into account the anatomical information only when it is reliable. Overall, all these methods try to modify the weight of the potential function (commonly quadratic function) based on the anatomical information rather than the PET image itself.…”