Data-driven fiber tractography with neural networks

Abstract

Elaborate expert modeling has been the predominant approach to fiber tractography. It attempts to invert the measurement process of diffusion-weighted MRI to reconstruct fibers. We present a purely data-driven neural network regression model for fiber tractography. The model sequentially takes as input a local block of data and the incoming direction of the fiber. From this input, the neural network predicts the outgoing direction. The training data can be provided by either automatic or human supervision. On both real, and synthetic data we observe that our model produces smoother and more accurate fibers than its supervisor. The performance of the model is scored with the Tractometer tool, where it consistently improves the supervisor baseline as well as the state-of-the-art in data-driven tractography. We show that our approach benefits from additional data, which can be incorporated easily, even from different supervisors. In experiments, the model is robust to noise and variation in the data, while being simple to use.