Incorporating anatomical priors into Track-to-Learn
Date:
Tractography has recently been posed as a reinforcement learning (RL) problem [1] so as to leverage the expressiveness of machine learning without the need for hard-to-obtain reference streamlines. Despite their competitive performances, agents trained via this method produced a high rate of false-positive connections as well as low bundle volume compared to both classical and machine-learning based approaches. In this work, we incorporate anatomical priors, which have been used in classical tractography, into the Track-to-Learn framework to alleviate these shortcomings.