Medical Image - Atlas Registration Using Deformable Models for Anomaly Detection

We introduce a system that automatically segments and classifies structures in brain MRI volumes. It segments 144 structures of a 256x256x124 voxel image in 18 minutes on an SGI computer with four 194 MHz R10K processors. The algorithm uses an atlas, a hand segmented and classified MRI of a normal brain, which is warped in 3-D using a hierarchical deformable registration algorithm until it closely matches the subject. This customized atlas contains the segmentation and classification of the subject’s anatomical structures. The system has processed 198 MRIs of normal brains, and 3 MRIs and 1 CT of brains with pathologies. Quantitative evaluations yield high segmentation accuracy. Combined with domain knowledge, the registration algorithm is able of detecting asymmetries and abnormal variations in the subject’s data that indicate the existence and location of pathologies. 1. Anomaly Detection Medical experts detect anomalies by comparing a particular subject to normal cases. In Figure 1, the normal brain is approximately symmetric across the center line, e.g. the pair of lateral ventricles have symmetric shapes and sizes. However, the pathological brain lost this symmetry, which is an indication of the existence and location of an anomaly, i.e. a bleeding. Our research assists the diagnosis by automating the comparison process. Our reference is an atlas, a magnetic resonance imaging (MRI) volume of a normal brain. It contains 123 slices, and each slice is a 256x256 pixel matrix. An expert spent 8 months to manually segment 144 anatomical structures, and gave each structure a unique label. The labels of the structures were color-coded to illustrate the segmentation, Figure 2. We developed a 3D hierarchical deformable registration algorithm, which constructs a mapping from the atlas to a subject. By applying the mapping to the color-coded segmentation of the atlas, we acquire a customized segmentation of the subject’s anatomical structures, Figure 2. We combine the registration result and the segmentation information with domain knowledge to aid anomaly detection. 2. Problem definition The data we work with is T1 weighted MRI. An MRI volume is a series of parallel cross-sections along one of three principal axes, see Figure 4. The atlas may differ from a subject in two ways. Genetic and environmental factors cause variations in the shape, size, density, and location of anatomical structures, i.e. variations that are local and intrinsic. The lack of standards in the data acquisi1. This research has been sponsored in full or in part by Office of Naval Research (ONR) under Contract N00014-95-1-0591. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of ONR or the U.S. Government. Lateral Ventricles