Building a Probabilistic Anatomical Brain Atlas for Multiple Sclerosis
Alexandre Guimond, Xizhang Wei · 2002
We present a completely automatic method to build an average anatomical brain model of Multiple Sclerosis (MS) using a set of magnetic resonance images. Brain shape variations between subjects were identified using segmented images where White Matter Signal Abnormalities (WMSA) and Normal Appearing White Matter (NAWM) jointly defined the White Matter (WM) anatomical extent. Differences due to linear transformations were excluded, resulting in the quantification of pure morphological differences. The result is an average shape image representing MS brains characteristics. Shape variations around this mean are produced, allowing statistical analysis of the extent of brain configuration changes in MS. Introduction An important tool used to diagnose abnormal anatomical variations are medical atlases. Computerized atlases comprising information in a more practical and quantitative manner than paper atlases are becoming available. They usually include information obtained from a large set of subjects and are sometimes built using methods that enable the quantitative analysis of shape variations in a given population. While the construction of computerized atlases is now under investigation by several groups and more disease-specific models are becoming available, there is a lack of resources dedicated to the construction of computerized anatomical atlases for pathologies where MR images comprise WMSA. Though there are various reasons for this, our group's primary obstacle has been the presence of WMSA which bias the registration procedure at the base of our atlasing method. This work is a first attempt at circumventing this problem and produce an atlas of average brain shape and variations for MS. Methods Our method to build anatomical brain atlases is based on a fully automatic intensity-based registration procedure (l). It can be summarized as follows: I) Affine registration between a set of MR scans from a group of subjects and a reference image corrects for positioning and global shape differences due to translation, rotation, scaling and shearing; 2) Local deformable intensity-based registration is then used to evaluate residual variations due to pure morphological differences and produce images having the same shape as the reference image for every subject; 3) Averaging the residual deformations and the locally registered images yields an average deformation and an average intensity image, respectively; 4) The average deformation is then applied to the average image to produce the model. It presents an average shape corresponding to the characteristics of the set of MR scans under investigation. While this method is well suited to deal with images without signal abnormalities, images containing WMSA, such as is the case in MS, bias the deformable registration procedure (step 2) and thus the characteristics of the resulting probabilistic models. To circumvent this problem, we applied the algorithm to segmented images where WMSA and NAWM were considered as 1 tissue class. The registration is then performed on segmented images containing only normal tissue types (Cerebro-Spinal Fluid (CSF), White Matter (WM), Gray Matter (GM)). The resulting models are thus not influenced by bad correspondences due to lesions, while still reflecting anatomical shape variations. R*feTercnc?