Physics Contribution
Accurate Automatic Delineation of Heterogeneous Functional Volumes in Positron Emission Tomography for Oncology Applications

https://doi.org/10.1016/j.ijrobp.2009.08.018Get rights and content

Purpose

Accurate contouring of positron emission tomography (PET) functional volumes is now considered crucial in image-guided radiotherapy and other oncology applications because the use of functional imaging allows for biological target definition. In addition, the definition of variable uptake regions within the tumor itself may facilitate dose painting for dosimetry optimization.

Methods and Materials

Current state-of-the-art algorithms for functional volume segmentation use adaptive thresholding. We developed an approach called fuzzy locally adaptive Bayesian (FLAB), validated on homogeneous objects, and then improved it by allowing the use of up to three tumor classes for the delineation of inhomogeneous tumors (3-FLAB). Simulated and real tumors with histology data containing homogeneous and heterogeneous activity distributions were used to assess the algorithm's accuracy.

Results

The new 3-FLAB algorithm is able to extract the overall tumor from the background tissues and delineate variable uptake regions within the tumors, with higher accuracy and robustness compared with adaptive threshold (Tbckg) and fuzzy C-means (FCM). 3-FLAB performed with a mean classification error of less than 9% ± 8% on the simulated tumors, whereas binary-only implementation led to errors of 15% ± 11%. Tbckg and FCM led to mean errors of 20% ± 12% and 17% ± 14%, respectively. 3-FLAB also led to more robust estimation of the maximum diameters of tumors with histology measurements, with <6% standard deviation, whereas binary FLAB, Tbckg and FCM lead to 10%, 12%, and 13%, respectively.

Conclusion

These encouraging results warrant further investigation in future studies that will investigate the impact of 3-FLAB in radiotherapy treatment planning, diagnosis, and therapy response evaluation.

Introduction

Although most clinical applications of positron emission tomography (PET) rely on manual and visual analysis, accurate functional volume delineation in PET is crucial for numerous oncology applications. These include the use of tumor volume and associated determination of semiquantitative indices of activity concentration for diagnosis and therapy response evaluation (1) or the definition of target volumes in intensity-modulated radiation therapy (IMRT) (2). Subjective (1) and tedious manual delineation cannot perform accurate and reproducible segmentation, particularly when considering complex shapes and nonhomogeneous uptake. This results from the low quality of PET images due to statistical noise and partial volume effects (PVE) (3), arising from the scanner's limited spatial resolution.

Most of the previously proposed methods for PET volume definition are semiautomatic and threshold-based, using either fixed (30%–75% of the maximum activity) 2, 4, 5 or adaptive approaches incorporating the background activity 6, 7, 8, 9, 10. Unfortunately, these approaches often require additional a priori information and are user- and system-dependent. They require manual background regions of interest (ROIs), and their performance depends on parameters requiring optimization using phantom acquisitions for each scanner and reconstruction. Finally, all of these approaches are strictly binary and were not validated considering heterogeneous volumes.

Numerous works have addressed PET lesion segmentation using more advanced image segmentation methodologies 11, 12, 13, 14, 15, 16, 17, 18, 19. However, the majority of these approaches often depend on pre- or postprocessing steps such as deconvolution or denoising, are often binary only, and are validated on phantom acquisitions or clinical data without rigorous ground truth.

We previously developed an algorithm for PET volume definition by combining a fuzzy measure with a locally adaptive Bayesian-based classification (FLAB) that has been shown to perform better with respect to fixed thresholding, fuzzy C-means (FCM), or fuzzy hidden Markov chains (FHMC) for PET volume definition, as far as homogeneous spheres or slightly heterogeneous and nonspherical tumors are concerned (20). Preliminary results show that FLAB is also robust with respect to variability of the acquisition and reconstruction parameters (24).

Clinical tumors may be characterized by heterogeneous uptake, thus demanding a nonbinary approach for an accurate segmentation that may have a significant impact in defining biological target volumes for dose painting (21). The goals of this work were to (1) improve the FLAB model by incorporating the use of three hard classes and three fuzzy transitions and (2) evaluate its accuracy on real (with known diameter measured in histology) and simulated (with known ground truth) data sets containing inhomogeneous tumors.

Section snippets

Three-class fuzzy Bayesian segmentation (3-FLAB)

The 3-FLAB algorithm is an extension of our previous work considering only a binary segmentation (20). FLAB automatically estimates parameters of interest from the image, maximizing the probability of each voxel to belong to one of the considered classes. This probability is estimated for each voxel as a function of its value and the values of its neighbors relative to the voxels' statistical distributions in the image, which corresponds to an estimation of the noise within each class. Hence,

Results

Figure 5 contains one axial slice of the segmentations obtained on three simulated tumors of Data Set 1 and one tumor of Data Set 2. Figure 6a contains the mean classification errors and standard deviation obtained by all the methods on the 20 tumors of Data Set 1. FLAB (binary only) performed well on homogeneous tumors but failed as expected on strongly heterogeneous lesions, leading to overall errors of 15% ± 11%. 3-FLAB, in contrast, produced segmentation maps closer to the ground truth,

Discussion

Functional volume delineation represents an area of interest for multiple clinical applications (routine and research) of PET. Such areas include response to therapy studies and the use of biological tumor volumes in radiotherapy treatment planning. Although several fully automatic algorithms have recently been proposed 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, segmentation methodologies currently used in clinical practice are based on the use of fixed and adaptive thresholding 4, 5, 6, 7, 8, 9,

Conclusion

A modified version of the FLAB algorithm has been developed to include the estimation of three hard classes and three fuzzy transitions. This automatic approach combines statistical and fuzzy modeling to address specific issues associated with 3D-PET images, such as noise and PVE. Its accuracy has been assessed on both simulated and clinical images of complex shapes containing inhomogeneous activities and small regions. The results demonstrate the ability of 3-FLAB to delineate such lesions,

Acknowledgments

This work was supported by the Brittany Region grant program (Grant No. 1202-2004), the French National Research Agency (Grant Nos. ANR-06-CIS6-004-03 and ANR-08-ETEC-005-01), and Cancéropôle Grand Ouest (R05014NG).

References (39)

  • P.H. Jarritt et al.

    The role of PET/CT scanning in radiotherapy planning

    Brit J Rad

    (2006)
  • M. Soret et al.

    Partial-volume effect in PET tumour imaging

    J Nucl Med

    (2007)
  • Y.E. Erdi et al.

    Segmentation of lung lesion volume by adaptive positron emission tomography image thresholding

    Cancer

    (1997)
  • U. Nestle et al.

    Comparison of different methods for delineation of 18F-FDG PET-positive tissue for target volume definition in radiotherapy of patients with non–small cell lung cancer

    Jour Nucl Med

    (2005)
  • J.A. Van Dalen et al.

    A novel iterative method for lesion delineation and volumetric quantification with FDG PET

    Nucl Med Commun

    (2007)
  • C.J. White et al.

    A semi-automatic approach to the delineation of tumour boundaries from PET data using level sets. Society of Nuclear Medicine 52nd Annual Meeting

    (June 2005)
  • P. Tylski et al.

    18F-FDG PET images segmentation using morphological watershed: A phantom study

    IEEE Neurosci Symp Conference Record

    (2006)
  • W. Zhu et al.

    Automation segmentation of PET image for brain tumours

    IEEE Neurosci Symp Conference Record

    (2003)
  • D.W.G. Montgomery et al.

    Fully automated segmentation of oncological PET volumes using a combined multiscale and statistical model

    Med Phys

    (2007)
  • Cited by (157)

    • K-nearest neighbor driving active contours to delineate biological tumor volumes

      2019, Engineering Applications of Artificial Intelligence
    View all citing articles on Scopus

    Conflict of interest: none.

    View full text