User:Jadrian Miles/Streamline clustering

Tubegen generates an easy-to-parse .nocr file specifying points on streamlines.

1. Pick a good dataset (Diffusion_MRI#Collaboration_Table) -- $G/data/diffusion/brown3t/cohen_hiv_study_registered.2007.02.07/patient120
2. Run tubegen on it with modified parameters so it doesn't cull anything---this will result in ~100k curves, with an average of ~70 points per curve.
3. Write a python script to divide the computation of the curve-to-curve distance matrix among many computers.
   • Try max and mean minimum point-to-curve distance in overlapping region as inter-curve distance measure. Or exponentially weighted mean a la cad.
     • See also cad's /map/gfx0/tools/linux/src/embed/utils/fast_distance_computing/src/ICurveDist/test
   • The per-curve script should return the assigned matrix line as well as a list of curves sorted by distance and annotated by the distance, for fast clustering.
   • After computing the upper half of the matrix, create an ordered list of curve-to-curve distances annotated with the curve pairs. Distributed w:quicksort? [1]
4. Build up clusters until some termination condition: satisfactory number of non-singleton clusters, satisfactory median size of non-singleton clusters, etc. Or just run until you get one huge cluster, but store the binary cluster tree. It may be really skewed but maybe a tree rebalancing algorithm could help in post-processing.
   • Initialization: each curve is a singleton cluster.
   • A curve's distance to a cluster is the minimum distance to any curve in that cluster.
   • In each iteration, with lowest minimum curve-to-curve distance to its closest cluster.