Techniques for Measuring Brain Deformation Derek Hill Professor of Medical Imaging Sciences Centre for Medical Image Computing University College London Derek.hill@ucl.ac.uk Introduction Image analysis provides tools to measure change in the size and shape of the brain. These changes can be caused by disease processes (eg: atrophy caused by Alzheimer’s disease), therapies, normal aging or normal development (most dramatic in-utero and in neonates), neurosurgery, or changes in physiological parameters such as blood gas mixture or hydration. The most widely used techniques for measuring brain deformation of all these types are based on longitudinal imaging with 3D gradient echo volume scans, normally T1 weighted. This tutorial reviews the methods that are widely used, and their applications. It also discusses the difficulties caused by image artefact. Methods for Quantification of the change in brain size or shape The oldest technique for measuring changes in brain size from MR scans is volumetry. Brain volumetry, at its most basic, involves using a mouse or similar device to draw around structures of interest in the brain, one slice at a time, at all the time points of interest, and count voxels within the boundaries to determine volume change between time-points. The limitations of this approach are firstly, the effort required from a skilled operator (can be several hours per brain), and secondly the subjectivity. The subjectivity can cause poor reproducibility, as inter-observer variability can be high. The limitations of volumetry lead researchers to devise more computationally sophisticated approaches to reduce the interaction time and hence make the techniques more widely applicable, and/or to increase the precision to make the techniques more sensitive to change. Fully automatic brain extraction 1 can be used for volumetry, but these techniques have not yet been shown to have sufficient precision to quantify the subtle changes over time that are of interest. Alternative approaches can use intensity information to increase the precision. The Boundary shift integral 2 makes use of accurate segmentions obtained interactively, but rather than just calculating volumes, registers the repeat scans back to the baseline using rigid registration, and uses this transformation to transform the repeat segmentation into base-line coordinates. The algorithm then creates a “between borders” region using an XOR operation on the two segmentations in baseline coordinates. Differences in the scans that lie in a defined intensity range are then integrated over this “between borders” region, approximating the volume traversed by the brain/ CSF boundary over time. The technique is less sensitive to segmentation errors than volumetry as it ignores differences where the paired intensities are