Optimization of Image quality and noise reduction by regulating beta penalty function of BSREM reconstruction algorithm

Introduction: The aim of the study is to determine the optimum penalization factor of the Block sequential regularized expectation maximization (BSREM) reconstruction algorithm and comparing with standard PET reconstruction OSEM on the basis of both quantitative and qualitative analysis for its clinical use in Ca breast patients undergoing F18 Fdg PET CT. Accuracy of SUV calculation is essential for quantification of data acquired and noise reduction is essential to apply during image reconstruction to improve the overall quality of the image. OSEM is a fast and one of the most applied PET reconstruction algorithms. However, image noise increases with subsequent iterations, and the algorithm is stopped before the images become too noisy. This could lead to inaccuracies for quantitative assessment as the algorithm does not reach full convergence. Q.Clear is a block sequential regularized expectation maximization (BSREM) algorithm for PET, which allows full convergence of measured and estimated data without noise amplification during image reconstruction.

Methods: We included the retrospective study of 20 patients' data with F18 FDG PET-positive ca breast scans. These patients underwent a PET CT scan with an average injected dose of 4 MBq per kg weight with 60 minutes uptake period and scan time of 1min per bed. Patient data were acquired on GE Discovery IQ 5 rings PET CT. PET emission data were reconstructed using conventional OSEM and BSREM with standard correction of decay, scatter, random, dead time, attenuation, normalization with the image matrix of 256x256. OSEM reconstruction was performed by selecting VPHD with 12 subsets 2 iterations and 4mm FWHM. BSREM reconstructions were performed with Q.Clear where penalized likelihood objective function of varying beta values of 200,250,300,350,400, 450,500, 550 and 600 were used . Reconstructed images were analyzed on PET CT review tool on advantage workstation and region of interest were drawn around lesions using a segmentation threshold of 40% and ROI was propagated in all images for calculating SUVmax, SUV mean and right lobe of the liver is taken for background ROIs where background SUVmean and standard deviation SUVstd was calculated. Quantitative evaluation was done by calculating noise level, signal to noise ratio SNR of the lesion, and Signal to background ratio SBR whereas qualitative analysis was done by visual scoring for image quality, lesion detectability, and background noise.

Results: Total 21 lesions were identified for the study in the size range of 1.2 to 2.5 cm. It was observed that SNR of the lesion was increased with increasing beta values as noise levels were decreased with increasing beta values, whereas SBR was decreased with increasing beta values as SUVmax of the lesion was also decreasing when noise was decreased. Reconstructed images with a beta less than 300 have higher background noise and images with a beta greater than 400 have lower background noise. Visual scoring suggests 350 beta value provides better lesion detection whereas 400 beta value gives better background noise on comparing with OSEM.

Conclusions: This study gives an indication that in the BSREM algorithm Q.Clear beta value of 350 is optimum for SUV calculations with overall better image quality in Ca breast positive F18Fdg PET CT studies. Beta penalty function of 300 to 400 is well documented optimum range for F18fdg PET torso studies which are further validated in our experimental study. On comparing with OSEM Q.Clear definitely provides higher signal and lesser noise without compromising image quality however as the penalty function of BSREM is a user-dependent variable to the algorithm this should be optimized depending upon the anatomical region scanned and the radiopharmaceutical used.

• Download figure
• Open in new tab
• Download powerpoint