Year: 2023
Type: dissertation
Abstract: In the human brain, the hippocampus is an internal structure which plays an essential role in the formation of memory, before those are migrated within the general cortex for longer term storage.As visible from an MRI, the hippocampus and the gray-matter cortical ribbon remain relatively constant during the mid-life, even though at the population level, a slight trend of decline in size can be revealed by statistical analysis.However, in late aging, the hippocampus may atrophy at accelerating, measureable pace.This pace increases with some forms of dementia.It is therefore desirable to accurately quantify the size of the gray matter structures, both for research into the causal factors separating a healthy aging from a dementia outcome, but also for early detection of potential health risk at the level of an individual.In this dissertation, I will describe some specific techniques that I recently implemented to quantify structures in MRI images of the brain, in the context of large cohorts.The analyses will focus specifically on two points, reflecting two published articles: the efficient segmentation of the hippocampal formation, and the parcellation of the wider brain cortex into multiple regions of interest.For this, I will rely on the Convolutional Neural Networks (ConvNets), recent machine-learning methods particularly adapted for image processing.Beyond the necessary accuracy and consistency property, I will show that ConvNets have the useful advantages of robustness, an important concern when analysing large datasets, and of fault-tolerance, necessary to train them from imperfect data.Further, I will discuss how Neural Networks allow for transferlearning, which in my case, refers to the ability to leverage information from one context and apply it into another context.After a first chapter introducing, from my point of view, some relevant general concepts, I will detail, sometimes technically, the two models of hippocampal segmentation and cortical parcellation, based on annotating my published manuscripts.Finally, a brief fourth chapter will conclude. * * *Where this work stands This work is situated at the intersection of several fields of interest to me.Neuroscience, Medical Imaging, and Machine Learning.The work presented here focuses on efficient and robust image analysis methods of structural images, for the eventual goal of providing recent progress of image analytics to every potential person who had a brain scan.This is a continuation of previous work focusing on population imaging in the context of small and large brain cohorts.While the work focus on structural image, that constitutes only one particular aspect of brain image analysis.In fact, my previous experience of cohorts with richer MRI analysis, including fast functional-MRI, which reveals the brain functions and connectome via sophisticated time series modeling; or the high-dimensional diffusion sequences that can highlight the internal structure of brain connectivity; has been showing what possible to do with MRI, while at the same time revealing the gap between research and practical application.A gap that could be reduced, thanks to the wider acceptance of computational analytics or Artificial Intelligence (data mining) in the medical field.
Author: BENJAMIN, THYREAU
Institution: Tohoku University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7145728642
