Automated supervised segmentation of anatomical fiber tracts using an AdaBoost framework

Diffusion Magnetic Resonance Imaging (DMRI) can be used to reconstruct the main neural pathways in the brain, herein termed `fibers'. Segmentation of anatomical tracts out of full brain fiber-sets acquired by tractography can lead to a better understanding of white matter diseases. Hence, we developed an automatic segmentation tool based on a renowned, supervised, machine-learning framework, called Viola-Jones [1]. We applied the algorithm for tract segmentation, using simple physical, statistical and geometrical features that can characterize the tract's fibers such as length, location, variance, FFT coefficients etc. An AdaBoost based learning framework was applied in order to select the most discriminative set of features for the classification of fibers to different anatomical tracts. Linear combinations of such features were used to construct classifiers. Those classifiers were arranged in a cascade which efficiently filters out fibers that do not belong to the desired tract. The algorithm was applied on a training set consisting of brains obtained from the Human Connectome Project. A cascade was learned for three different tracts. Performance evaluation was done by calculating the dice coefficients for each of the tested brains, yielding a result of around 90% for the tracts under evaluation, indicating a successful segmentation of the tracts of the entire brain.