IC‐P‐145: Image processing pipeline for deformable registration

absolute value of the BTB difference and the standard deviation of the BTB difference (Table 1). As an additional measurement of the reproducibility, a power calculation to determine the group size required to detect a specific treatment effect was completed using a bootstrap simulation based on the Wilcoxon-Mann-Whitney test. Results: Results The 50 percentile, 90 percentile and standard deviation for the combined diagnostic groups shown inTable 1 indicate themajor challenge of comparing theBTB reproducibility at 3Tand 1.5Tthat the shoulders of the distributions aremuch larger than for the often assumed Gaussian distribution. In fact, the 90 percentiles for the BTBdifference distributions are about 3 times larger thanwould be expected for Gaussian distributions with the same values for the 50 percentiles. The large shoulders make it difficult to determine the relative sensitivities of the 3Tand 1.5Tatrophymeasures to treatment effect.While the reproducibility statistics are also broken down by diagnostic groups, the small number of subjects in some groupsmakes the statistics for the diagnostic groups less reliable than the combined group. This is especially true forAD.Thegroup size calculations present a different picture than theBTB spread statistics. The 3T group sizes were consistently about 50% larger than those of the 1.5T. However, it is unclear how sensitive this result is to the choice of the model of treatment effect or to the large shoulders of the reproducibility distributions. Conclusions: The distributions of the BTB differences at both 3T and 1.5T have large shoulders which complicate any comparison of the reproducibility of the two field strengths. The large shoulders may explain why the 50 percentile, rather than the more common standard deviation, is used in the literature to present the reproducibility statistics [2]. However, given the large shoulders of the BTB difference distributions, the 50 percentile alone may be unable to reliably predict the performance of the FSL/Siena atrophy measure in clinical trials. Preliminary calculations of group sizes required to detect a particular treatment effect, indicate that studies at 3T may require substantially larger group sizes. However the confounding factor in these calculations due to the non Gaussian distributions, combined with the relatively small groups in the current analysis, means that further analysis is needed before high confidence in this conclusion can be achieved.