Ellie Thompson & Ziyi Shen - CMIC/WEISS joint seminar series

Speaker: Ellie Thompson

Title: Data-Driven Decomposition of Diffusion MRI for Concurrent Extraction of White Matter Bundles and Grey Matter Networks

Abstract: 
Diffusion MRI is a unique and powerful tool for mapping the white matter connections of the brain, in vivo. However, standard analysis techniques for mapping white matter bundles rely on ROI-based methods, in which a set of masks and logical operations are defined in a template space in order to constrain the streamlines. These masks are not always straightforward to define, for example during early development when the brain is rapidly growing and maturing. In this talk, I will present a novel method for concurrent extraction of white matter bundles and their associated grey matter networks, in an entirely data driven fashion. The method is based on non-negative matrix factorisation, which is inherently suited to the non-negative nature of structural connectivity data. I will demonstrate the validity and reliability of this approach on a publicly available dataset of neonatal dMRI from the Developing Human Connectome project, by comparing the resultant components to tracts obtained using traditional methods, and grey matter networks obtained from resting-state fMRI. I will show how we can use the grey matter networks to generate a robust parcellation of the neonatal cortex, based on structural connectivity, and discuss further potential applications for this approach.

Speaker: Ziyi Shen

Title: Attention-based image reconstruction for degraded images: approaches and datasets

Abstract: 
Image reconstruction, for example, recovering a high-quality image with significant details from a single degraded image, has been an active research area in computer vision. With the increasing use of handle devices, such as cell phones, onboard cameras, and portable instruments, image degradation has become a ubiquitous problem to confront with. Focusing on the complex degradation and reconstruction problem, we propose to synthesize the low-quality data in the real world and solve the problem by discriminating between the foreground and background. By blending the different domain information, we restore the images with more semantic information. In addition, this proposal has been extended to tackle the clinical retinal fundus image enhancement problem.

Chair: Danny Alexander