Spin-L共催　生理学研究所 感覚認知情報研究部門 部門公開セミナー（第14回）を開催します。

Rethinking diffusion MRI tractography synthesis and analysis using a deep learning representation framework.

Jon Haitz Legarreta

（Postdoctoral fellow, Brigham and Women's Hospital/Harvard Medical School, USA）

Diffusion MRI (dMRI) tractography of the brain provides information about its tissue microstructure and structural connectivity non-invasively. Since its inception in the mid 1990's, tractography has allowed to image the brain white matter pathways in-vivo. Although dMRI tractography methods have evolved to provide increasingly accurate white matter fiber reconstructions, their anatomical fidelity is impacted by inherent ambiguities of the diffusion signal and the fundamentals of the streamline propagation procedure. Several works have shown that conventional streamline tractography produces significant proportions of implausible streamlines, their ability to map a number of white matter pathways being moderate to low. Thus, the in-vivo delineation of the brain white matter architecture remains incomplete. In recent years, much effort has been put into proposing tractography deep learning models to overcome the limitations of conventional tractography synthesis and analysis methods. Deep learning-based tractography methods proposed in scientific literature have shown comparable or superior ability to generate anatomically plausible results compared to conventional methods. In this talk, I will present a family of deep representation learning methods that I have proposed to analyze and synthesize tractography data. By employing tractography streamline data and an  appropriate representation framework, I will demonstrate how the model can be harnessed to accomplish several tasks in tractography without requiring separate training steps. The framework allows to filter and bundle streamlines into anatomically consistent groups, and provides plausible white matter fiber reconstruction capabilities without requiring any streamline propagation. Results show that the proposed framework outperforms conventional streamline propagation and analysis methods on both synthetic and human brain in-vivo data.