Which Blind Tract Clustering Method Is Most Robust to False Positives?

Mark Drakesmith¹, John Evans¹, Anthony David², Derek Jones¹

The effect of false positives on the performance of blind tract clustering is unclear. From an idealised dataset of 6 pre-defined bundles, various distance metrics and clustering methods were tested across varying FP-rates by substituting a random proportion of true tracts with FPs. Most methods deteriorate gradually with noise. Affinities computed from maximum (Hausdorff) and endpoint distances were most robust to noise. These methods showed FP misclassification were concentrated on the smallest fibre bundle while other methods showed more diffuse misclassification across adjacent bundles. These two distance metrics are therefore best for clustering noisy tractography datasets.