CMIC Seminar: Noise level effects on dMRI parameter inference by synthetically trained deep neural networks

Speaker: Prof. Yoshitaka Masutani, Hiroshima City University
UCL Contact: Dominique Drai (Visitors from outside UCL please email in advance).
Date/Time: 21 Nov 18, 13:00 - 14:00
Venue: 1.02

Abstract

Diffusion MRI (dMRI) is a powerful tool for characterizing local properties of microstructures in living bodies, based on parameters of various signal models.

Recently, instead of conventional fitting, machine-learning approaches have been employed for the model parameter inference by regression, which include deep regression neural networks. In addition, a training approach with only synthetically generated data is expected to overcome the limitations of real data training. During the synthetic training data generation, dMRI signal values are often contaminated by artificial noise for realistic data synthesis. In this talk, the importance of noise level adjustment for the synthetic training of deep neural networks is mainly introduced, that is, the amount of noise should be matched between training data and test data to obtain optimal robustness. Several experimental results for DTI, DKI, and NODDI are shown for synthetic and real datasets. An original software demonstration for dMRI analysis is also planned.

Prof. Yoshitaka Masutani

Prof. Yoshi Masutani currently conducts Medical Imaging Lab. in Hiroshima City Univ. Graduate School of Information Sciences, Japan. He obtained Ph.D. degrees for biomedical engineering field in 1997 and for medical science in 2010, both at University of Tokyo, Japan. He has joined several multi-disciplinary research institutes where clinicians, scientists and engineers collaborate, such as Univ. Hospital of Hamburg-Eppendorf, Univ. of Chicago Hospital, and Univ. of Tokyo Hospital. Prof. Masutani’s research specialty covers biomedical image analysis and software development, especially for dMRI.