Elsevier

NeuroImage

Volume 32, Issue 1, 1 August 2006, Pages 138-145
NeuroImage

A new improved version of the realistic digital brain phantom

https://doi.org/10.1016/j.neuroimage.2006.03.052Get rights and content

Abstract

Image analysis methods must be tested and evaluated within a controlled environment. Simulations can be an extremely helpful tool for validation because ground truth is known. We created the digital brain phantom that is at the heart of our publicly available database of realistic simulated magnetic resonance image (MRI) volumes known as BrainWeb. Even though the digital phantom had l mm3 isotropic voxel size and a small number of tissue classes, the BrainWeb database has been used in more than one hundred peer-reviewed publications validating different image processing methods.

In this paper, we describe the next step in the natural evolution of BrainWeb: the creation of digital brain phantom II that includes three major improvements over the original phantom. First, the realism of the phantom, and the resulting simulations, was improved by modeling more tissue classes to include blood vessels, bone marrow and dura mater classes. In addition. a more realistic skull class was created. The latter is particularly useful for SPECT, PET and CT simulations for which bone attenuation has an important effect. Second, the phantom was improved by an eight-fold reduction in voxel volume to 0.125 mm3. Third, the method used to create the new phantom was modified not only to take into account the segmentation of these new structures, but also to take advantage of many more automated procedures now available. The overall process has reduced subjectivity and manual intervention when compared to the original phantom, and the process may be easily applied to create phantoms from other subjects.

MRI simulations are shown to illustrate the difference between the previous and the new improved digital brain phantom II. Example PET and SPECT simulations are also presented.

Introduction

Image processing methods need evaluation data sets to characterize, evaluate and optimize their performance. Three main types of evaluation data sets can be distinguished: real acquisitions of subjects, real acquisitions of physical phantoms and simulations from numerical phantoms. By using real acquisitions, the whole acquisition set up is taken into consideration. Although physical phantoms can provide a gold standard because the ground truth of the object is perfectly known, the data obtained have unrealistic complexity due to the limited number of compartments that they include. Unlike physical phantom acquisitions, subject acquisitions have a realistic complexity, but the ground truth is not available. Simulations provide a way of generating data where ground truth is known and where realistic complexity may be taken into account when a realistic numerical phantom used as input to the simulator and the simulator reproduces the physics of data generation. Contrary to the real acquisitions, the contributions of errors from different sources in acquisition can be separated and evaluated independently. Simulations allow the control of various acquisition parameters, whereas real acquisitions are limited to the types of scan acquired (parameters, slice thickness…). Since some subtle real effects may not be included in simulations, these simulations are not sufficient to evaluate image processing methods. However, the known ground truth and the realistic complexity of the simulated data allow the simulations to be an extremely helpful tool during the process of method evaluation.

A set of realistic simulated anatomical brain magnetic resonance imaging (MRI) volumes, known as BrainWeb, is available to the neuroimaging community1. It is also possible for external groups to customize the simulator parameters, run the simulator on our computer system and download the resulting 3D MR simulations along with the digital phantom. These simulations have been used by more than one hundred external groups, e.g. Grabowskia et al. (2000), Arnold et al. (2001), Schnack et al. (2001), Cardenasa et al. (2001) and Tzourio-Mazoyer et al. (2002). These groups have incorporated these simulations to study the performance of techniques such as non-linear co-registration, cortical surface extraction, correction of MRI intensity non-uniformity and tissue classification. For example, Arnold et al. (2001) used the database to study six different algorithms for the correction for MRI intensity non-uniformity.

These MR simulations, and simulations of different modalities as positron emission tomography (PET) and single photon emission computed tomography (SPECT), were generated by varying specific imaging parameters for each tissue type in the simulator. The spatial distribution of these different tissues (gray matter, white matter, cerebrospinal fluid, muscles, skull, skin and fat) was defined on volumetric fuzzy volumes. These fuzzy volumes, where voxel intensity is proportional to the fraction of tissue within the voxel, define the digital brain phantom (Collins et al., 1998).

This digital brain phantom was created by registering and averaging 27 T1-weighted, 12 PD-weighted and 12 T2-weighted MRI scans from a single subject. As a direct result of the high signal-to-noise ratio, these single subject average volumes exhibit fine anatomical details and enable a faithful representation of the brain's complex anatomical structures.

In this paper, we describe the important modifications made to this phantom and its construction. The goal of the first modification is to improve the realism of the phantom and thus of the simulated data. The voxel size of the phantom was decreased to allow the discrimination of finer detail and enable a better classification. To improve the realism of MRI or fMRI simulations, more structures were segmented such as marrow, dura mater and blood vessels. To improve SPECT, PET or CT simulations, the phantom required a better skull class, with cranium and facial bones. These improvements were made possible with computed tomography (CT) and MR angiography (MRA) of this subject, acquired after the creation of the initial phantom in 1998.

The second modification concerns the method used to construct the phantom. This method has been modified not only to take into account the creation of new segmented structures, but also to be mostly automated in order to be easily applied to create phantoms from other subjects.

The following sections detail phantom construction and these improvements.

Section snippets

Data acquisition

The anatomical phantom is derived from high-quality T1-, T2- and PD-weighted images, formed from averages of 27, 12 and 12 scans respectively, of the same normal subject (Holmes et al., 1998) using a 1.5 T Phillips clinical scanner. The acquisition protocol consisted of a series of 20 1 mm3 T1-weighted MR scans acquired using a T1-weighted spoiled GRASS [TR/TE = 18 ms/10 ms, FA = 30°, 256 * 256 matrix, 256 mm field of view, NSA 1], a series of 7 0.78 mm3 T1-weighted MR scans acquired using a

Digital brain phantom II

The average of many single image volumes enhances the quality of MR images. While the process enhances the resolution of the phantom, the resulting volume does not have 0.5 mm3 resolution, but the quality is better than the 1 mm3 data used to create the phantom. The decrease of the voxel size of the phantom could modify partial volume effects at structure borders but allows a better delineation of fine structures (see Fig. 1).

All the fuzzy volumes that define the brain phantom are shown in Fig.

Discussion

The major improvements made to the phantom are decreased voxel size and the addition of four tissue classes: vessels, dura, bone marrow and bone. The final resolution of the phantom is due to the resolution and sampling rate of the original data, the quality of the inter-volume registrations, the interpolation kernel used in the resampling steps and the sampling rate of the final phantom. Interpolation with a kernel that does not have infinite support will decrease the resolution of the

Conclusion

The digital brain phantom was improved using mostly automated techniques. The new tissue classes added (vessels, dura matter and marrow) improve the realism of the resulting simulations. These data are invaluable as it can be used to drive simulators for different modalities including MRI, fMRI, PET, SPECT and CT. As the method proposed to build the phantom is mostly unsupervised, it will be easily used to create phantoms from other subjects.

References (24)

  • A. Evans et al.

    3D statistical neuroanatomieal models from 305 MRI volumes

  • T.J. Grabowskia et al.

    Validation of partial tissue segmentation of single-channel magnetic resonance images of the brain

    NeuroImage

    (2000)
  • Cited by (237)

    View all citing articles on Scopus
    View full text