Influence of reconstruction settings on the performance of adaptive thresholding algorithms for FDG‐PET image segmentation in radiotherapy planning

Abstract

The purpose of this study was to analyze the behavior of a contouring algorithm for PET images based on adaptive thresholding depending on lesions size and target‐to‐background (TB) ratio under different conditions of image reconstruction parameters. Based on this analysis, the image reconstruction scheme able to maximize the goodness of fit of the thresholding algorithm has been selected. A phantom study employing spherical targets was designed to determine slice‐specific threshold (TS) levels which produce accurate cross‐sectional areas. A wide range of TB ratio was investigated. Multiple regression methods were used to fit the data and to construct algorithms depending both on target cross‐sectional area and TB ratio, using various reconstruction schemes employing a wide range of iteration number and amount of postfiltering Gaussian smoothing. Analysis of covariance was used to test the influence of iteration number and smoothing on threshold determination.The degree of convergence of ordered‐subset expectation maximization (OSEM) algorithms does not influence TS determination. Among these approaches, the OSEM at two iterations and eight subsets with a 6–8 mm post‐reconstruction Gaussian three‐dimensional filter provided the best fit with a coefficient of determination R2=0.90 for cross‐sectional areas ≤ 133 mm 2 and R2=0.95 for cross‐sectional areas > 133 mm 2 . The amount of post‐reconstruction smoothing has been directly incorporated in the adaptive thresholding algorithms. The feasibility of the method was tested in two patients with lymph node FDG accumulation and in five patients using the bladder to mimic an anatomical structure of large size and uniform uptake, with satisfactory results.Slice‐specific adaptive thresholding algorithms look promising as a reproducible method for delineating PET target volumes with good accuracy.PACS numbers: 87.57.nm, 87.55.D‐, 87.57.uk

Influence of filter choice on 18F-FDG PET segmentation accuracy determined using generalized estimating equations

“…Previous work has demonstrated that the choice for number of iterations and subsets do not have a marked effect on segmentations (Matheoud et al 2011) and so we feel that, together with the wide range of image acquisition parameters, our results are representative of what one would expect on different scanners/iterative reconstruction algorithms.…”

Joint segmentation of anatomical and functional images: Applications in quantification of lesions from PET, PET-CT, MRI-PET, and MRI-PET-CT images

“…In fixed thresholding, a clinically accepted value of SUV = 2.5 or 40% of the SUV max (i.e., maximum SUV of a predefined region) is used to delineate lesions from the background for a given region of interest (ROI) drawn manually (Nestle et al, 2006). In adaptive thresholding methods, a more optimal thresholding level is searched by examining class uncertainties (Otsu, 1979), by building realistic phantoms (Matheoud et al, 2011; Davis et al, 2006; Brambilla et al, 2008; Schaefer et al, 2008), by applying iterative thresholding based on scanner hardware properties (Drever et al, 2007; van Dalen et al, 2007; Jentzen et al, 2007), or by incorporating local approaches into the threshold selection process (Erdi et al, 2002; Bradley et al, 2004; Ciernik et al, 2005; Koshy et al, 2005). It has been shown in various studies (Fahey et al, 2010) that the lack of optimal threshold levels in these approaches prevents accurate and robust delineation of lesions from the background.…”

Evaluation of diffusion weighted imaging for tumor delineation in head-and-neck radiotherapy by comparison with automatically segmented 18F-fluorodeoxyglucose positron emission tomography

“…T. Schakel, et al Physics and Imaging in Radiation Oncology 5 (2018) [13][14][15][16][17][18] values. Jager et al [5] showed 0.61 for laryngeal cancer using CT and 0.57 using CT/MR; Geets et al reported 0.41 and 0.42 for laryngeal and oropharyngeal GTVs respectively [30]; Mukesh et al reported 0.54 for CTV delineations on CT [31].…”

“…However, the use of PET for segmentation is challenging. The segmentations can depend on the contrast and noise characteristics of the PET images, which, apart from tumor characteristics, originate from different acquisition and reconstruction protocols [15] , [16] . Additionally, a variety of segmentation algorithms have been proposed which result in large variations in target volumes [10] , [17] .…”