Automatic cumulative sums contour detection of FBP-reconstructed multi-object nuclear medicine images

https://doi.org/10.1016/j.compbiomed.2017.04.010Get rights and content

Highlights

  • We determined the contour of multiple objects in PET/SPECT images automatically.

  • We applied a modified cumulative sums (CUSUM) scheme in the sinogram.

  • This method can be used for computing body edges for attenuation correction.

  • It is also helpful for eliminating streak artifacts outside the objects.

  • For the phantoms investigated, the rms error lies between 1.14 and 1.69 pixels.

Abstract

The problem of determining the contours of objects in nuclear medicine images has been studied extensively in the past, however most of the analysis has focused on a single object as opposed to multiple objects. The aim of this work is to develop an automated method for determining the contour of multiple objects in positron emission tomography (PET) and single photon emission computed tomography (SPECT) filtered backprojection (FBP) reconstructed images. These contours can be used for computing body edges for attenuation correction in PET and SPECT, as well as for eliminating streak artifacts outside the objects, which could be useful in compressive sensing reconstruction. Contour detection has been accomplished by applying a modified cumulative sums (CUSUM) scheme in the sinogram. Our approach automatically detects all objects in the image, without requiring a priori knowledge of the number of distinct objects in the reconstructed image. This method has been tested in simulated phantoms, such as an image-quality (IQ) phantom and two digital multi-object phantoms, as well as a real NEMA phantom and a clinical thoracic study. For this purpose, a GE Discovery PET scanner was employed. The detected contours achieved root mean square accuracy of 1.14 pixels, 1.69 pixels and 3.28 pixels and a Hausdorff distance of 3.13, 3.12 and 4.50 pixels, for the simulated image-quality phantom PET study, the real NEMA phantom and the clinical thoracic study, respectively. These results correspond to a significant improvement over recent results obtained in similar studies. Furthermore, we obtained an optimal sub-pattern assignment (OSPA) localization error of 0.94 and 1.48, for the two-objects and three-objects simulated phantoms, respectively. Our method performs efficiently for sets of convex objects and hence it provides a robust tool for automatic contour determination with precise results.

Introduction

In nuclear medicine the two prevailing, noninvasive imaging modalities are positron emission tomography (PET) and single photon emission computed tomography (SPECT). These nuclear medicine techniques, which are referred to as emission tomography, have significant clinical and preclinical applications and can be employed to a vast variety of medical fields including oncology, neurology and cardiology.

PET uses the unique decay characteristics of radiopharmaceuticals, such as FDG (18F-2-deoxy-2-fluoro-D-glucose). When FDG is intravenously introduced into the patient's body, it is distributed in tissues in a manner determined by its biochemical properties [1]. Then, PET detectors detect annihilation photons (coincidence gamma rays) that are produced when positrons interact with electrons [2]. On the other hand, in SPECT the intravenously injected tracers, such as technetium (99Tc) labeled with an appropriate agent, radiate single photons. This time, the detector channels count individual photons (γ-ray events) [3].

Image reconstruction is the main goal for almost all inverse problems in imaging, although it is usually adopted in a tomography context, suggesting reconstruction from projection data, i.e. sinograms [4]. There are several reconstruction algorithms which are mainly distinguished as analytic and iterative. The dominant analytic approach for image reconstruction is filtered backprojection (FBP). FBP is based on the inversion of the Radon transform via the central slice theorem [5]. The advantages of FBP amongst other analytic methods are its speed and simplicity. However, its main limitation lies on its formulation which does not incorporate the statistical nature of light; at the same time, it generates “streak artifacts” in the reconstructed images. These artifacts are due to incomplete data measurement, such as angular undersampling, or poor counting statistics [2]. The dominant iterative approach for image reconstruction is ordered subsets expectation maximization (OSEM). In this paper, we will focus on FBP-reconstructed images.

Although OSEM image reconstructions do not exhibit streak artifacts and incorporate the Poisson-statistical nature of light, there are numerous cases where FBP has shown to produce improved quantitative results over OSEM. For example, in several neuroreceptor studies, especially if a low count rate cannot be avoided, FBP reconstruction is preferable to OSEM, in order to estimate the total volume of distribution [6]. In another study, Reilhac et al. [7] suggest that the iterative reconstruction methods are biased at low statistics, especially in the lower part of the image dynamic range and for cold regions. Furthermore, according to Boellaard et al. [8] patient data indicated that for brain, myocardium, and tumor regions of interest (ROIs), OSEM and FBP provide equivalent results.

Since the 1980s, the problem of determining the contour of an object for attenuation purposes in both PET and SPECT has received reasonable attention in the literature, especially via the sinogram. In cases where no computed tomography (CT) is present, contour detection is a necessary approach for determining the attenuation map for attenuation correction purposes. Nevertheless, even in cases with a CT present, such as PET/CT, the CT delivers single snapshots of the breathing cycle and when used for attenuation correction it may generate edge artifacts in most of the PET respiratory cycle, even though respiratory gating of PET is possible [9].

Several methods have been proposed in order to address the problem of contour detection. It is obvious that the simplest possible method is sinogram thresholding, although this rather subjective method is neither automatic nor adaptive. Other methods include: (i) the maximum slopes algorithm for head contour determination in PET [10], making use of the first derivative of the projections in order to specify the abrupt changes through the derivative maxima, (ii) the automated body contours finding method in SPECT [11] which incorporates thresholding, smoothing and fourth-order Fourier fitting, (iii) the seminal Canny approach to edge detection, encompassing principles from the calculus of variations [12], (iv) the Tomitani method, encompassing differential geometry aspects in contour finding [13], (v) the online brain attenuation correction procedure for head contour determination [14], based on the Tomitani method and a sinogram finite differences scheme, (vi) the method for determination of concave body outlines from SPECT sinograms before reconstruction making use of a fourth-rank tensor projector operator [15], (vii) the sinogram segmentation technique, achieved by surface deformation fitting between the projected model and the input sinogram [16], (viii) the mixed detection and validation methods utilizing the sinogram derivatives and cosine fitting in low count emission images [9], (ix) the novel truncation correction method, utilizing the idea of sinogram decomposition, where sinogram curves are considered as individual image points [17], (x) the sinogram segmentation technique for direct metal suppression in CT [18], (xi) the metal artifact reduction through sinogram segmentation method of lumbar spine CT images [19], (xii) the pre-reconstruction sinogram filtering approach based on three-dimensional (3D) mean-median filters in PET reconstructions, aiming to minimize angular blurring artifacts, to smooth flat regions and to preserve the edges [20], as well as (xiii) the denoising algorithm for cone-beam CT sinograms, encompassing a certain regularization in terms of gradient and Hessian [21].

Most of the above methods correspond to single-object cases, and the first two as well as the fifth focus specifically on the brain and skull regions. Furthermore, all the methods mentioned above are based on the notions of thresholding, sinogram differentiation and segmentation. The method presented here takes into account only the values of the sinogram in a straightforward manner, without computing any threshold or derivative.

The main goal of this study is to develop an automated method for determining the contour of multiple objects in FBP-reconstructed images. This method can be used for obtaining body edges for attenuation correction purposes. More specifically: (a) in PET/CT, body edge detection throughout the PET sequence can improve the single CT scan before attenuation correction [9], (b) in brain SPECT imaging, contour determination is essential for accurate attenuation correction [9] and (c) in PET and SPECT CT-less imaging, it could provide a good starting point for a uniform attenuation map. Furthermore, body edge detection could eliminate streak artifacts outside the body, which could be very useful in the area of compressive sensing reconstruction, where the sinogram and image can be “cleaned” for outside artifacts in every iteration [22]. By applying a simplified cumulative sums (CUSUM) scheme in the sinogram, as originally developed by Page [23], we were able to automatically detect all objects in the image without requiring: (i) prior knowledge of the number of distinct objects in the reconstructed image and (ii) visually selected threshold values, the calculation of noisy sinogram derivatives, or the use of segmentation techniques.

Section snippets

The CUSUM-based algorithm

The CUSUM algorithm was introduced by E.S. Page [23]. It encompasses the notions of adaptive thresholding and complete memory of the information embedded in previous observations [24].

Let f^ij denote the sinogram, i.e. the tabulated, matrix version of either the Radon transform (PET) or the attenuated Radon transform (SPECT) of the two-dimensional (2D) radioactivity distribution function f(x,y), as in [25]. The sinogram is a matrix with Nθ rows (angular dimension, θ) and Nρ columns (signed

Results

The execution time for contour determination depends on the number of total distinct objects (N) and is clearly less in cases were N=1. More specifically, in all N=1 cases (the simulated IQ phantom, the simulated concave phantom, the real NEMA phantom as well as the clinical thoracic phantom) the execution time was less than 4.3 s (4.21 s, 4.12 s, 4.29 s and 4.02 s respectively) and in the simulated two-ellipses phantom and three-ellipses phantom cases the execution time was much larger, more than 22

Discussion

In this study we developed a method to automatically determine the contour of multi-object FBP-reconstructed nuclear medicine images, without prior knowledge of the total number of objects involved. In order to accomplish this goal, a CUSUM-based algorithm has been implemented and tested in simulated and real phantoms. Given a sinogram, we perform sinogram edge detection and thus create a sinogram mask. By mapping the masked sinogram into the image space, we create an image mask, and

Conclusions

An automatic method for the contour detection of FBP-reconstructed, multi-object nuclear medicine images has been developed and tested in both simulated and real phantoms. Our method, based on cumulative sums, automatically detects all objects in the image without the requirement of prior knowledge of the total number of distinct objects. Through contour determination, we are able to efficiently compute body edges. In future studies, we intend to implement our method in STIR as well as to

Conflict of interest

None declared.

Acknowledgements

This work was partially supported by the research programme “Inverse Problems and Medical Imaging” (200/842) of the Research Committee of the Academy of Athens. G.A. Kastis acknowledges the financial support of the “Alexander S. Onassis Public Benefit Foundation” under grant No. R ZL 001-0. N.E. Protonotarios would like to thank Assoc. Prof. P. Tsiamyrtzis of Athens University of Economics and Business for his valuable suggestions and comments. George M. Spyrou holds the Bioinformatics ERA

George A. Kastis is an Associate Researcher Professor at the Research Center of Pure and Applied Mathematics of the Academy of Athens, as well as the Acting Director of the Center. He holds a PhD in Optical Sciences from the University of Arizona which was obtained under the supervision of Regents Professor Harrison H. Barrett. His main research interests include small-animal imaging and analytic image reconstruction in PET and SPECT.

References (36)

  • R. Boellaard et al.

    Experimental and clinical evaluation of iterative reconstruction (OSEM) in dynamic PET: quantitative characteristics and effects on kinetic modeling

    J. Nucl. Med.

    (2001)
  • M. Bergström et al.

    Determination of object contour from projections for attenuation correction in cranial positron emission tomography

    J. Comput. Assist. Tomogr.

    (1982)
  • M. Hosoba et al.

    Automated body contour detection in SPECT: effects on quantitative studies

    J. Nucl. Med.

    (1986)
  • J. Canny

    A computational approach to edge detection

    IEEE Trans. Pattern Anal. Mach. Intell.

    (1986)
  • T. Tomitani

    An edge detection algorithm for attenuation correction in emission CT

    IEEE Trans. Nucl. Sci.

    (1987)
  • C. Michel et al.

    Online brain attenuation correction in PET: towards a fully automated data handling in a clinical environment

    Eur. J. Nucl. Med.

    (1989)
  • J. Case, M. King, D.-s. Luo, E.J. Soares, M.S. Rabin, et al., Determination of concave body outlines from SPECT...
  • V. Elangovan, R.T. Whitaker, From sinograms to surfaces: a direct approach to the segmentation of tomographic data, in:...
  • Cited by (0)

    George A. Kastis is an Associate Researcher Professor at the Research Center of Pure and Applied Mathematics of the Academy of Athens, as well as the Acting Director of the Center. He holds a PhD in Optical Sciences from the University of Arizona which was obtained under the supervision of Regents Professor Harrison H. Barrett. His main research interests include small-animal imaging and analytic image reconstruction in PET and SPECT.

    Nicholas E. Protonotarios is a doctoral researcher at the Research Center of Pure and Applied Mathematics of the Academy of Athens and a PhD candidate at the Department of Mathematics of the National Technical University of Athens (NTUA). He holds an MEng in Naval Architecture and Marine Engineering as well as an MBA in Engineering Economics, both from NTUA. His current research interests include analytic reconstruction techniques for PET and SPECT.

    George M. Spyrou is a senior research scientist at the Center of Systems Biology, Biomedical Research Foundation, Academy of Athens. His main aim is the application of his knowledge and skills in Physics, Mathematics and Informatics to medically and biologically relevant issues, mainly on the ’mining’ of the information either in direct medical data (e.g. medical images) or in genomes and proteomes relevant to medical problems.

    View full text