The primary benefit of using digital phantoms and simulated images is that many aspects of the imaging process can be controlled, so that the behavior of an algorithm can be carefully evaluated while yielding the important advantage that the desired answer is known a priori in the validation experiment. Furthermore, unlike physical phantoms, modifications such as identifying particular structures, adding pathologies, and highlighting activations areas are easily performed.
The digital brain phantom described in this paper is based on a high-resolution, high SNR MRI volume of a normal volunteer. In addition to the advantage of being anatomically realistic, it models partial volume effects between tissues, thus avoiding significant errors associated with the use of a discrete model. The processes used to create the average MRI and the digital brain phantom are simple and can be repeated in other labs using publicly available software for registration and classification.
This phantom may serve as a gold standard to measure the performance of image processing algorithms and has been used to generate realistic MR (Fig. 4a), PET (Fig. 4b) and CT (not shown) images that are suitable for validation studies. The availability of a standard digital brain phantom and the associated simulated images will make it possible to evaluate the quality or functionality of procedures developed in different laboratories, judge their applicability for a given purpose and compare them to other existing methods.
The work presented here is the first phase in the creation of a complete digital brain model useful for validation. The next phase will be to incorporate blood vessels, increase the resolution of the phantom and further subdivide the anatomical model. We also plan to model the intensity variations that correspond to fine tracts in the white matter (e.g., occipital tracts). Since our initial design was motivated by the requirements for comparison of classification algorithms, the phantom contains only tissue type information. For use in EEG simulations, we plan to introduce electrical properties. Moreover, we are currently sub-dividing the grey-matter tissue type into specific functional regions so that the phantom can be used in brain-mapping experiments .
The phantom data described here is available from the MNI web site along with a database of simulated MR images  so that any newly developed algorithms can be evaluated with the same data, thus providing a tool for comparison between methods.