Comparative evaluation of image reconstruction methods for the siemens PET-MR scanner using the stir library

Abstract

Abstract-With the introduction of Positron Emission Tomography -Magnetic Resonance (PET-MR) scanners the development of new algorithms and the comparison of the performance of different iterative reconstruction algorithms and the characteristics of the reconstructed images data is relevant . In this work, we perform a quantitative assessment of the currently used ordered subset (OS) algorithms for low-counts PET-MR data taken from a Siemens Biograph mMR scanner using the Software for Tomographic Image Reconstruction (STIR, stir.sf.net). A comparison has been performed in terms of bias and coefficient of variation (CoV). Within the STIR library different algorithms are available, such as Order Subsets Expectation Maximization (OSEM), OS Maximum A Posteriori One Step Late (OSMA-POSL) with Quadratic Prior (QP) and with Median Root Prior (MRP), OS Separable Paraboloidal Surrogate (OSSPS) with QP and Filtered Back-Projection (FBP). In addition, List Mode (LM) reconstruction is available. Corrections for attenuation, scatter and random events are performed using STIR instead of using the scanner. Data from the Hoffman brain phantom are acquired, processed and reconstructed. Clinical data from the thorax of a patient have also been reconstructed with the same algorithms. The number of subsets does not appreciably affect the bias nor the coefficient of variation (CoV=11%) at a fixed sub-iteration number. The percentage relative bias and CoV maximum values for OSMAPOSL-MRP are 10% and 15% at 360 s acquisition and 12% and 15% for the 36 s, whilst for OSMAPOSL-QP they are 6% and 16% for 360 s acquisition and 11% and 23% at 36 s and for OSEM 6% and 11% for the 360 s acquisition and 10% and 15% for the 36 s. Our findings demonstrate that when it comes to low-counts, noise and bias become significant. The methodology for reconstructing Siemens mMR data with STIR is included in the CCP-PET-MR website (www.ccppetmr.ac.uk).

Hybrid PET-MR list-mode kernelized expectation maximization reconstruction

“…This is of interest because it would have the potential to allow the examination to more patients [10,19,35]. It has previously been shown that OSEM and penalized algorithms show negative bias under low-count circumstances [1,12,13,36,70]. Positive bias is associated with the frequently observed positivity constraint in iterative reconstruction algorithms.…”

“…(n) j is the feature vector that is calculated from the nth PET update image, and σ m , σ p , σ dm and σ dp are scaling parameters for the distances in (13) and (14). For each voxel of the PET image the corresponding feature vectors, z (n) j and v j , are extracted from the local neighborhood of the voxel from the PET update image and MR image, respectively.…”

Performance evaluation of the Q.Clear reconstruction framework versus conventional reconstruction algorithms for quantitative brain PET-MR studies

“…The FBP reconstruction is not available for clinical use on the GE SIGNA PET-MR scanner; hence, OSEM reconstructions have been used for processing brain studies. In smaller regions, such as the ones that can be found in the brain, the convergence rate of OSEM process must be stopped early in order to not compromise image quality due to excessive noise [ 4 , 5 ]. Although OSEM is being used for processing of both whole-body and brain scans, studies such as the ones conducted by Reilhac et al [ 6 ] and Walker et al [ 7 ] have reported a positive bias in regions with low activity and a negative bias in regions of high activity in low-count scans which had been reconstructed with this algorithm.…”

Hybrid PET-MR list-mode kernelized expectation maximization reconstruction for quantitative PET images of the carotid arteries

“…For this reason, this work aims to propose a Kernel Method, which makes use of anatomical information, able to improve image reconstruction in the clinical environment while avoiding the aforementioned problems related to regularization. In this work emphasis is given on quantification for low-count condition, which has shown to be challenging due to bias, noise and contrast losses [5,25,26,27,28].…”