PDF version of this entire document

Sensitivity

Figure  shows the results of applying sensitivity analysis to the validation study. These demonstrate that Specificity is more sensitive (i.e. is able to detect smaller misregistrations) than the overlap-based approach, which is in turn more sensitive than Generalisation. Note from the error bars that these differences are statistically significant. Maximum sensitivity is achieved with a shuffle radius of $1.5$ or $2.1$. Figure  also suggests that the most sensitive generalised overlap measure is obtained using label-complexity weighting. The choice of the Tanimoto Coefficient is consistent through all the experiments and each fuzzy overlap measure is based on it. The Tanimoto Coefficient is monotonically related to the popular Dice Similarity Coefficient (DSC), which would therefore produce similar results.

Figure: as a function of misregistration is shown for a number of datasets created by adding noise to a set the brain images (0-10%). The steepness of the curve is affected by high levels of noise, but the behaviour of the curve remains consistent.