Additive Groves with very simple features for brain fibers classification

In this document we describe the entry that came in the third place in the ”Supervised Learning Match Expert” challenge of the ICDM’09 Data Mining Contest. The main method used in our approach is a recently developed tree ensemble algorithm called Additive Groves. Code for Additive Groves is publicly available on the author’s website. 1. Supervised Learning Challenge The data for this competition described scans of three different human brains. Each data point referred to a single fiber in the brain. An expert has labeled eight fiber tracks in one of the brain; the goal was to predict correct labels for the same fiber tracks in the two other brains. 2. Feature Extraction Each fiber was initially represented as a sequence of points in 3D space. The brains were normalized by the organizers of the challenge, so we did not perform any alignment or scaling of the data. We extract simple features to describe location of fibers in the brain. In particular we considered fiber length, starting location, ending location, bounding box of the fiber points and finally the center of mass of all points of the fiber. To compute the length of the fiber, we connected each two adjacent points with a line and summed the lengths of the line segments (1 feature). The starting and ending location are the coordinates of the first and the last points in the fiber (6 features). The bounding box of the fiber points is computed as minimum and maximum for each dimension (x,y,z) among all points in the fiber (6 features). The center of mass is taken as the mean value of fiber point coordinates (3 features). These features describe very rough location of fibers in the brain. This representation is very compact, as we use only 16 real numbers to describe each fiber. Code and extracted features are freely available upon request.