Skip to Main Content
Iterative image reconstruction algorithms may suffer from bias in case of poor statistics. The purpose of this study is: (1) to implement various existing and new iterative methods and (2) to evaluate the applicability of these methods for quantitative 3D brain PET studies. Various OSEM and weighted least squares (WLS) algorithms were implemented. WLS and WLS with OSEM start image (OS-EM-WLS) were implemented with and w/o nonnegativity constraints. Accuracy and precision were assessed using phantom and simulations studies. Clinical evaluation was performed for 3 tracers using 9 dynamic brain PET studies including arterial sampling. ROI analysis was performed to evaluate contrast, noise and bias. Simulations were used to generate variance images at various noise levels. Dynamic human brain scans were analysed using parametric Logan plot analysis. Phantom data showed that all methods with nonnegativity constraint showed bias up to a factor 4 in low activity areas. Algorithms that allowed negative values within sinogram and image space provided smallest bias. Simulations showed that WLS w/o constraint, but with an OSEM start image, provided lowest variance amongst all tested reconstruction methods. OS-EM-WLS was more accurate and showed lower variance than other methods tested. Finally, parametric Logan analysis showed that OS-EM-WLS agreed most with FBP. OS-EM-WLS provided most accurate results and is therefore a promising candidate for reconstruction of 3D brain PET studies.