Anatomically guided voxel-based partial volume effect correction in brain PET: Impact of MRI segmentation

https://doi.org/10.1016/j.compmedimag.2012.09.001Get rights and content

Abstract

Partial volume effect is still considered one of the main limitations in brain PET imaging given the limited spatial resolution of current generation PET scanners. The accuracy of anatomically guided partial volume effect correction (PVC) algorithms in brain PET is largely dependent on the performance of MRI segmentation algorithms partitioning the brain into its main classes, namely gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF). A comparative evaluation of four brain MRI segmentation algorithms bundled in the successive releases of Statistical Parametric Mapping (SPM) package (SPM99, SPM2, SPM5, SPM8) using clinical neurological examinations was performed. Subsequently, their impact on PVC in 18F-FDG brain PET imaging was assessed. The principle of the different variants of the image segmentation algorithm is to spatially normalize the subject's MR images to a corresponding template. PET images were corrected for partial volume effect using GM volume segmented from coregistered MR images. The PVC approach aims to compensate for signal dilution in non-active tissues such as CSF, which becomes an important issue in the case of tissue atrophy to prevent a misinterpretation of decrease of metabolism owing to PVE. The study population consisted of 19 patients suffering from neurodegenerative dementia. Image segmentation performed using SPM5 was used as reference. The comparison showed that previous releases of SPM (SPM99 and SPM2) result in larger gray matter regions (∼20%) and smaller white matter regions (between −17% and −6%), thus introducing non-negligible bias in PVC PET activity estimates (between 30% and 90%). In contrary, the more recent release (SPM8) results in similar results (<1%). It was concluded that the choice of the segmentation algorithm for MRI-guided PVC in PET plays a crucial role for the accurate estimation of PET activity concentration. The segmentation algorithm embedded within the latest release of SPM satisfies the requirement of robust and accurate segmentation for MRI-guided PVC in brain PET imaging.

Introduction

Molecular brain imaging using positron emission tomography (PET) has emerged as one of the most promising modalities that steadily gained importance in the clinical and research arenas [1]. Considerable progress has been made to optimize the design of dedicated high resolution PET scanners and to integrate multimodality images to correlate functional findings to anatomy through the use of CT and MRI and to improve the quality and quantitative accuracy of brain PET images, however, emerging clinical and research applications of functional brain imaging promise even greater levels of accuracy and precision and therefore impose more constraints with respect to the information provided to clinicians and research scientists [2]. Since MRI is more suitable than CT for brain imaging owing to its high soft tissue contrast and better spatial resolution, combined PET-MRI systems dedicated for brain imaging have emerged as alternatives to PET-CT [3]. One of the first steps to obtain the best of the various imaging modalities is to coregister functional and anatomical images, and to a pre-segmented atlas if available. Wu et al. [4] has shown that this task can be optimized using non-rigid registration procedures compared to rigid or semi-rigid procedures such as those implemented in the Automated Image Registration (AIR) and the Statistical Parametric Mapping (SPM) packages. These techniques are, however, especially useful when dealing with inter-subject image registration.

Unfortunately PET imaging suffers from many physical degrading effects, partial volume effect (PVE), which is common to all medical imaging techniques owing to the discrete sampling of the image formation process, being one of them. In brain PET imaging, this effect is not negligible owing to the large voxel size, which produces images where high activity regions spillover into low activity regions, potentially leading to erroneous results in the qualification of functional brain imaging [5], [6]. Since neurological PET imaging was developed very early, the first attempts to reduce the impact of the PVE and restore the image content focused on brain imaging [7]. This has been addressed through the calculation of recovery coefficients [8], thus limiting the methods to the use of only PET data, since in these early years multimodality imaging was still not readily available as it is nowadays. The advent of modern multimodality imaging, and particularly brain PET/MR, stimulated the development of new partial volume correction (PVC) techniques which exploit a priori information gathered from anatomical information through partitioning MR images into different compartments, typically gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) [9]. More recently, a novel class of PVC algorithms that do not require segmentation of anatomical images was introduced [10]. This includes very promising approaches such as the wavelet decomposition technique [11] and the Bayesian approach [12]. In both cases, the algorithm is able to find the high frequency information lacking in low resolution PET images at a voxel-level without increasing the noise.

Current PVC algorithms that require the segmentation of anatomical images correct functional PET images in the projection space, during the reconstruction process or after their reconstruction (post-reconstruction) at regional level using a region-of-interest (ROI)-based analysis or in a more general way at the voxel level (voxel-based) [13]. Among ROI-based post-reconstruction methods, the most popular techniques use recovery coefficients [7], [8] or the geometric transfer matrix (GTM) method [14], [15] used in our previous work [16]. The principle of ROI-based methods is to calculate the effective activity in different regions assuming that the tracer uptake in each particular region is homogeneous. Naturally, the complexity of the problem increases when the number of considered regions increases, rendering PVC of the whole brains a complex problem. Nevertheless, in a simple case using ROIs has shown promising results [15].

Conversely, voxel-based approaches are not limited to a particular ROI since they attempt to recover the actual activity concentration in the cortex on a voxel-by-voxel basis, though with a priori assumptions about the tracer distribution [10], [17], [18]. These techniques have the advantage of generating corrected image for qualitative assessment and visual interpretation. The principal drawback of voxel-based methods compared to ROI-based methods is that they are quantitatively less accurate and rely on many assumptions. Partition methods are one example of voxel-based methods, the simplest case being to define one unique partition corresponding to brain tissue classes (GM and WM) and to compensate the spillover on non-active regions (CSF) by converting PET intensities from activity per spatial volume to activity per tissue volume [19]. This is achieved by convolving the partition (brain mask) with the point spread function (PSF) of the PET imaging system. This approach was extended to two (GM and WM) [20] and three compartments assuming that the CSF activity is not only the result of spillover from contiguous regions [18].

MR image segmentation is the critical component of MRI-guided PVC in brain PET imaging [14], [17], [21]. In a previous work, we compared the impact of various MR image segmentation algorithms on the GTM algorithm for PVC of 18F-FDG and 18F-FDOPA brain PET data [16]. One of the conclusions of this work was that Statistical Parametric Mapping (SPM2) segmentation software is more suitable for clinical routine examinations owing to the robustness of its normalization algorithm for atypical brains.

In this work, we aim to assess the influence of 4 chronologically successive releases of SPM segmentation software on a voxel-based PVC method proposed by Matsuda et al. [9] in contrast to previous work referenced above.

Section snippets

Brain MRI segmentation algorithms

SPM is among the well established packages used for statistical analysis of neuroimaging data including PET, SPECT and fMRI [22]. It is well-documented, freely available, technically supported by well established brain imaging centers [23] and widely used by the neuroimaging community. The brain MR image segmentation technique implemented in this package considers the three tissue classes of interest for PVC, namely gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF).

A

Comparison of brain MRI segmentation methods

Fig. 1, Fig. 2 show representative slices of a clinical T1-weighted MR study and the corresponding segmentation results when using the 4 segmentation algorithms. Images of GM and WM regions are shown separately. The first observation one can formulate is that non-brain matter is often included into the GM region of segmentations performed using SPM99.

Table 1 shows the relative differences between the 4 segmentation algorithms with respect to GM and WM volume calculation of clinical brain MR

Discussion

It is well established that PVC improves image quality and quantitative accuracy of functional PET imaging where the image degradation resulting from the PVE is higher as a consequence of their poor image resolution. PVC is crucial to avoid misinterpretation of functional brain PET data [5]. Unfortunately, the performance of PVC algorithms is highly dependent on accuracy of the various steps involved in the procedure in a complex way. Overall, with one exception [29], it has been reported that

Conclusion

SPM is a popular software for multi-modal brain image analysis including a powerful toolbox for the challenging task of brain MR image segmentation. The introduction of novel segmentation methods and their implementation in new releases of this package have rendered older versions obsolete [22]. The Bayesian formulation introduced in SPM2 and further consolidated in SPM5 version was, probably, the most relevant improvement into SPM segmentation procedure. The segmentation algorithm embedded

Conflict of interest statement

The authors declare that they have no conflict of interest.

Acknowledgments

HZ and KL are supported by the Swiss National Science Foundation under grants SNSF 31003A-135576, 33CM30-124114 and 33CM30-140337. FA is supported by Fonds Cognitive Memory, NAC 08-025 and the Swiss National Science Foundation under grants 320030_138163/1 and SPUM 33CM30-124115.

Daniel Gutierrez received his Master in Physics in 2000 and his Minor in Infographics in 2003 from the University of Geneva to finally obtain his PhD in Life Sciences (Medical Imaging) in 2008 from the University of Lausanne, Switzerland. He is currently a Postdoc research fellow at the PET Instrumentation & Neuroimaging Laboratory (PINLab) at Geneva University Hospital (Switzerland). His research interests include medical image processing, dose optimization and multimodality imaging techniques.

References (32)

  • W.-D. Heiss

    The potential of PET/MR for brain imaging

    Eur J Nucl Med Mol Imaging

    (2009)
  • M. Wu et al.

    Quantitative comparison of AIR, SPM, and the fully deformable model for atlas-based segmentation of functional and structural MR images

    Hum Brain Mapp

    (2006)
  • C.C. Meltzer et al.

    Does cerebral blood flow decline in healthy aging? A PET study with partial-volume correction

    J Nucl Med

    (2000)
  • H. Zaidi et al.

    Correction for image degrading factors is essential for accurate quantification of brain function using PET

    Med Phys

    (2004)
  • E.J. Hoffman et al.

    Quantitation in positron emission computed tomography. 1. Effect of object size

    J Comput Assist Tomogr

    (1979)
  • R.M. Kessler et al.

    Analysis of emission tomographic scan data: limitations imposed by resolution and background

    J Comput Assist Tomogr

    (1984)
  • Cited by (35)

    • PET/MRI Hybrid Systems

      2018, Seminars in Nuclear Medicine
      Citation Excerpt :

      The accuracy of the coregistration is greatly enhanced by simultaneous data acquisition based on the fact that the anatomical data match the physiological state of the functional PET data, and no errors due to nonrigid motion of organs are introduced.150,166,168 Furthermore, segmentation of anatomical data in uniform compartments (eg, gray matter, white matter, and cerebrospinal fluid) is greatly enhanced by using MRI data because of its excellent soft tissue contrast.166,169 Particularly in brain imaging, MRI can provide detailed segmentation of brain compartments, enabling accurate segmentation and coregistration superior to CT.150

    • Cross-sectional variations of white and grey matter in older hypertensive patients with subjective memory complaints

      2018, NeuroImage: Clinical
      Citation Excerpt :

      After normalization to the adaptive template, 18F-FDG PET images were smoothed with a Gaussian filter (FWHM 8 × 8 × 8 mm3) and normalized to through intensity ratios relative to mean cerebellar activity (Van Der Gucht et al., 2015). Thereafter, the PET images were corrected for partial volume effect using the grey matter volume segmented from co-registered MR images for each patient (Gutierrez et al., 2012). The SPM linear regression models, used for correlating 18F-FDG metabolism from the brain grey matter voxels with the ADCmean values from overall white matter as well as for each lobe, were obtained: 1) at a threshold (voxel-level significance) of p < 0.005, 2) with a correction for cluster volume and using the expected voxels per cluster provided by SPM, in order to avoid type II errors as recommended (Lieberman and Cunningham, 2009) and 3) by using age and gender as covariates.

    • Magnetic Resonance–Derived Improvements in PET Imaging

      2017, Magnetic Resonance Imaging Clinics of North America
    View all citing articles on Scopus

    Daniel Gutierrez received his Master in Physics in 2000 and his Minor in Infographics in 2003 from the University of Geneva to finally obtain his PhD in Life Sciences (Medical Imaging) in 2008 from the University of Lausanne, Switzerland. He is currently a Postdoc research fellow at the PET Instrumentation & Neuroimaging Laboratory (PINLab) at Geneva University Hospital (Switzerland). His research interests include medical image processing, dose optimization and multimodality imaging techniques.

    Marie-Louise Montandon received the MS degree in Psychology and a PhD in Neuroscience from the universities of Geneva and Lausanne within the lemanic doctoral program in Neurosciences. She is an active member of the PET Instrumentation and Neuroimaging Laboratory (PINLab) of Geneva University. She is actively involved in developing imaging solutions for neuroscience research and clinical diagnosis. Her research centres on the development of improved methods for quantification of functional three-dimensional brain PET images using statistical image analysis tools. She is recipient of many awards and distinctions among which the 2005 best PhD thesis award given by the Medical School of Geneva University and the Research Grant for the Advancement of Women 2004 given by the Advisory Board for the Advancement of Women, National Center of Competence in Research CO-ME (Computer Aided and Image Guided Medical Interventions).

    Frédéric Assal received his MD in 1988 then a medical doctorate thesis in plasticity of the visual system. He completed his medical training in neurology and a fellowship in cognitive & behavioural neurology at the UCLA Alzheimer disease research center. He is currently leading the neuropsychological unit in the department of clinical neuroscience, Geneva. His research work cover dementia, MCI and related disorders.

    Karl-Olof Lovblad studied medicine at the University of Geneva, Switzerland. He was trained in Radiology and Neuroradiology at the university hospital in Bern. He also worked as a Research Fellow in MRI at the Beth Israel Deaconness Medical Center in Boston. In 2010, he was appointed Chairman and Professor of Neuroradiology at Geneva University Hospitals.

    Habib Zaidi is Chief Physicist and head of the PET Instrumentation & Neuroimaging Laboratory at Geneva University Hospital and faculty member at the medical school of Geneva University. He is also a Professor of Medical Physics at the University Medical Center of Groningen (The Netherlands) and visiting Professor at Ecole Nationale Supérieure d’Electronique et de ses Applications (ENSEA, France). He received a PhD and habilitation (PD) in medical physics from Geneva University for dissertations on Monte Carlo modeling and quantitative analysis in positron emission tomography. Dr. Zaidi is actively involved in developing imaging solutions for cutting-edge interdisciplinary biomedical research and clinical diagnosis in addition to lecturing undergraduate and postgraduate courses on medical physics and medical imaging. He was guest editor for 5 special issues of peer-reviewed journals and serves as Editor-in-Chief for the Open Medical Imaging Journal, and Associate editor for leading journals in medical imaging. His academic accomplishments in the area of quantitative PET imaging have been well recognized by his peers and by the medical imaging community at large since he is a recipient of many awards and distinctions among which the prestigious 2003 Young Investigator Medical Imaging Science Award given by the IEEE, the 2004 Mark Tetalman Memorial Award given by the Society of Nuclear Medicine, the 2007 Young Scientist Prize in Biological Physics given by the International Union of Pure and Applied Physics and the prestigious 2010 Kuwait Prize of Applied Sciences (known as the Middle Eastern Nobel Prize) given by the Kuwait Foundation for the Advancement of Sciences. Dr. Zaidi has been an invited speaker of many keynote lectures at an International level, has authored over 330 publications, including ∼140 peer-reviewed journal articles, conference proceedings and book chapters and is the editor of three textbooks on Therapeutic Applications of Monte Carlo Calculations in Nuclear Medicine, Quantitative Analysis in Nuclear Medicine Imaging and Multimodality Molecular Imaging of Small Animals. More details about this work can be found at http://pinlab.hcuge.ch/.

    View full text