Fast Acquisition for Diffusion Tensor Tractography

Diffusion tensor tractography is a powerful method for in-vivo white matter mapping. Its implementation involves long scanning sessions to capture local diffusion orientations, followed by tedious post-processing to generate accurate tracts. While some initial research was conducted to reduce the number of required gradient directions, the current state-of-the-art still considers acquisition protocol acceleration and automatic tract segmentation as two separate tasks. We aim at optimizing the whole workflow, from acquisition to tract segmentation. We propose a collaborative neural framework for diffusion-encoding color map denoising and white matter tract segmentation. It generates high-quality white matter tracts using DWI acquired for a small number of diffusion-encoding gradient directions (GDs), thus minimizing acquisition and post-processing time. The proposed method is first validated on the high-angular resolution (270 GDs) HCP dataset using a novel spherical k-means method to select a subset of 16 quasi-uniformly distributed GDs. Further validation is provided for a prospective clinical dataset of 10 cases acquired at both 16 and 64 GDs. Encouraging experimental results are obtained using several state-of-the-art neural architectures and training loss functions.