Figure 7. Two dimensional Isomap embedding of the lesion data (T2 dataset with zero-filled lesions).

The algorithm attempts to preserve proximity in high dimensions in the 2 dimensional embedding shown here. Similar lesions are therefore likely to cluster together. The colours index the logarithm of the asymptotic p value from a one-tailed, two-sample K-S test of the difference between the cost function masking and enantiomorphic methods. Each data point represents an individual lesion. Successive plots zoom in on the central area to reveal it in more detail. The absence of a clear overlap between spatial and p value clustering, except for a tendency for the effects to be somewhat less significant for small lesion sizes, suggests there is no clear set of lesion features for which the enantiomorphic method is consistently inferior.