Algorithms, atrophy and Alzheimer's disease: cautionary tales for clinical trials

Thompson and Holland (2010) highlight a biologically implausible deceleration of atrophy in results previously published in this journal (Hua et al., 2010); the results were derived using tensor based morphometry on images from the Alzheimer's Disease Neuroimaging Initiative. They speculate that bias may have been introduced due to asymmetric interpolation in global image registration, and/or to the use of a statistically defined region of interest. In their reply, Hua et al. (this issue) acknowledge the presence of a bias, but show that it stems largely from an asymmetry in the local image registration algorithm (an asymmetry common to methods in many published studies using nonlinear registration). Hua et al. demonstrate that the bias can largely be removed using a revised symmetric algorithm. This correspondence raises important issues relating to the lack of ground truth against which image analysis methodologies designed to determine atrophy patterns and rates can be assessed; and the increasing importance of striving to avoid potential biases as these techniques become utilised in clinical trials. In the absence of a "gold standard", we discuss a number of steps against which methodologies designed to quantify atrophy from serial scans can be assessed.