Voxel-based comparison of state-of-the-art reconstruction algorithms for 18F-FDG PET brain imaging using simulated and clinical data
Introduction
Fluorine-18 fluoro-2-deoxyglucose (18F-FDG) positron emission tomography (PET) imaging of the brain is a well-established way to detect changes in neuronal activity induced by a variety of diseases through the alteration in glucose consumption (Varrone et al., 2009). Both filtered backprojection and iterative reconstruction algorithms are currently being used in clinical practice. However, the limited spatial resolution of PET (2.5–4.5 mm for brain imaging (Adam et al., 1997, Wienhard et al., 2002)) in combination with the small cortical thickness (average thickness of 2.5 mm (Fischl and Dale, 2000)) compromises accurate quantification of the tracer uptake in the brain (Hoffman et al., 1979).
Two approaches have been recommended to improve the quantitative accuracy. The first one is to correct for the partial volume effect after reconstruction, by estimating the point spread function and correcting for its effect on the reconstructed activity, given spatially aligned tissue segmentation information. Different partial volume correction methods for brain 18F-FDG PET imaging have been compared in Yang et al. (1996). The second approach is to model the spatial resolution during image reconstruction to sharpen the image (‘resolution recovery’ or ‘resolution modeling’). Even in high-resolution PET systems this can reduce the partial volume effect and improve the quantitative accuracy (Sureau et al., 2008), but it is worth noting that resolution modeling may induce some unwanted secondary effects, such as an increase in inter-voxel correlations and Gibbs artifacts. Hence, an improvement in one figure of merit does not imply an improvement of all possible figures of merit (Alessio et al., 2013). To suppress noise and Gibbs artifacts, resolution modeling is best combined with the use of anatomical prior information during, e.g., a maximum a posteriori (MAP) image reconstruction (Vunckx et al., 2012). In addition, detection accuracy of hypometabolic regions can be significantly improved in this way (Baete et al., 2004a, Goffin et al., 2010).
Unfortunately, segmentation of, e.g., structural magnetic resonance (MR) images does not yield perfect brain tissue classification. Especially subcortical structures are very hard to segment, although increasingly accurate segmentation methods are being developed (Babalola et al., 2009). Partial volume correction (PVC) techniques and anatomical priors that heavily rely on segmentations will automatically translate the segmentation errors into quantification errors and reconstruction artifacts. Therefore, segmentation-free anatomical priors have been proposed as well in the past (Bowsher et al., 2004, Nuyts, 2007, Somayajula et al., 2005).
Because noise suppression through the use of anatomical information was found to be superior with MAP reconstruction compared to with post-processed maximum likelihood reconstruction (Nuyts et al., 2005), we concentrated on the evaluation of reconstruction-based partial volume correction algorithms in this work. Furthermore, diagnosis based on 18F-FDG PET brain images is often performed by looking for locally deviating patterns or small hypometabolic regions. Therefore, we focused on voxel-based rather than region of interest-based image quality evaluation. Also in group analyses, voxel-based statistical parametric mapping techniques are typically used to highlight differences in tracer uptake.
The aim of the present study was to investigate where systematic differences are to be expected when using MLEM with resolution recovery, MAP reconstruction with a segmentation-based prior or MAP with a segmentation-free anatomical prior (three state-of-the-art reconstruction algorithms), and whether the choice of the algorithm influences the outcome of a group analysis. Through the use of a simulation study, results can also be compared to the true activity distribution. The evaluation on clinical data sets enabled us to verify whether similar results are obtained in a realistic setting.
Section snippets
MLEM with resolution modeling
Some of the resolution lost due to blur during the detection process can be recovered by modeling the resolution of the detector during iterative reconstruction. In these iterative algorithms, the currently estimated activity distribution Λold = [λ1 … λJ] (with J as the number of reconstruction image voxels) is gradually improved based on a comparison between the measurement Q = [q1 … qI] (with I as the number of detector pixels) and the forward projected Λold, which can be written as (with
Bias with respect to true activity and differences between algorithms
The results of the comparison studies are shown in Fig. 1, Fig. 2, Fig. 3, Fig. 4. The colored voxels indicate the percentage difference in reconstructed activity, but only in those voxels in which these differences were considered to be significant. For the mutual comparison of the reconstruction algorithms (Fig. 2, Fig. 3, Fig. 4), three voxel-based analyses were performed, i.e. on reconstructions of simulated normal brain data in native space, of simulated data after spatial normalization
Discussion
All reconstruction methods under investigation include resolution recovery, which makes that even MLEM yields a strong partial volume correction. Because resolution recovery requires many iterations to reach convergence, the MLEM images are too noisy to be useful in clinical practice. Therefore, its quantification accuracy was only investigated after post-smoothing, which obviously resulted in a great loss of the recovered resolution, apparent as a significant underestimation of the GM activity
Conclusion
Three state-of-the-art iterative algorithms for PET image reconstruction, i.e. (post-smoothed) MLEM with resolution recovery, and two MAP algorithms using either a segmentation-based (A-MAP) or a segmentation-less (AsymBowsher) anatomical prior, have been compared for their use in human 18F-FDG PET brain imaging using voxel-based statistical analysis techniques. In the first study, the algorithms were compared to the true activity and to each other. Significant differences were mainly found at
Acknowledgments
K. Vunckx is a postdoctoral researcher of the Research Foundation — Flanders (FWO Vlaanderen). K. Van Laere is a Senior Clinical Investigator of the Research Foundation — Flanders.
References (31)
- et al.
Effect of anatomical variability, reconstruction algorithms and scattered photons on the SPM output of brain PET studies
NeuroImage
(2008) - et al.
An evaluation of four automatic methods of segmenting the subcortical structures in the brain
NeuroImage
(2009) - et al.
Evaluation of anatomy based reconstruction for partial volume correction in brain FDG-PET
NeuroImage
(2004) - et al.
Correlations of interictal FDG-PET metabolism and ictal SPECT perfusion changes in human temporal lobe epilepsy with hippocampal sclerosis
NeuroImage
(2006) - et al.
Performance evaluation of the whole-body PET scanner ECAT EXACT HR+ following the IEC standard
IEEE Trans. Nucl. Sci.
(1997) - et al.
Point/counterpoint. Resolution modeling enhances PET imaging
Med. Phys.
(2013) - et al.
Anatomical based FDG-PET reconstruction for the detection of hypometabolic regions in epilepsy
IEEE Trans. Med. Imaging
(2004) - et al.
Small focal cortical dysplasia lesions are located at the bottom of a deep sulcus
Brain
(2008) - et al.
Using MRI information to estimate F18-FDG distributions in rat flank tumors
- et al.
Design and construction of a realistic digital brain phantom
IEEE Trans. Med. Imaging
(1998)
Acetazolamide vasoreactivity in vascular dementia: a positron emission tomographic study
Eur. Neurol.
Measuring the thickness of the human cerebral cortex from magnetic resonance images
Proc. Natl. Acad. Sci. U. S. A.
Anatomy-based reconstruction of FDG-PET images with implicit partial volume correction improves detection of hypometabolic regions in patients with epilepsy due to focal cortical dysplasia diagnosed on MRI
Eur. J. Nucl. Med. Mol. Imaging
Quantitation in positron emission computed tomography: 1. Effect of object size
J. Comput. Assist. Tomogr.
Accelerated image reconstruction using ordered subsets of projection data
IEEE Trans. Med. Imaging
Cited by (14)
Approximating anatomically-guided PET reconstruction in image space using a convolutional neural network
2021, NeuroImageCitation Excerpt :Moreover, it reduces the detection of small structures with low contrast. To overcome these challenges, especially in PET brain imaging, many techniques for anatomically-guided PET reconstruction using high-resolution structural MR images have been developed in the past (see e.g. Baete et al., 2004; Bland et al., 2017; Bowsher et al., 2004; Comtat et al., 2002; Ehrhardt et al., 2016; Knoll et al., 2016; Lipinski et al., 1997; Loeb et al., 2015; Mehranian et al., 2018; Nuyts, 2007; Schramm et al., 2017; Vunckx et al., 2012; Vunckx et al., 2014; Vunckx and Nuyts, 2010). These techniques have received increased interest recently due to the introduction of whole-body, clinical PET/MRI scanners during the last decade, which effectively removed all logistical and practical constraints for procuring the required high-resolution MR images.
AI for PET image reconstruction
2023, British Journal of RadiologyFeasibility of dose reduction for [18F]FDG-PET/MR imaging of patients with non-lesional epilepsy
2022, Nuklearmedizin - NuclearMedicineComparison of an in-house acquired brain F-18 FDG PET normal database with commercially available normal data
2021, Nuclear Medicine CommunicationsQuantitative Rodent Brain Receptor Imaging
2020, Molecular Imaging and Biology