Deep generative modelling of the imaged human brain
Human-machine symbiosis is a very promising opportunity for the field of neurology given that the interpretation of the imaged human brain is a trivial feat
for neither entity. However, before machine learning systems can be used in
real world clinical situations, many issues with automated analysis must first be
solved. In this thesis I aim to address what I consider the three biggest hurdles
to the adoption of automated machine learning interpretative systems. For each
issue, I will first elucidate the reader on its importance given the overarching
narratives of both neurology and machine learning, and then showcase my proposed solutions to these issues through the use of deep generative models of the
imaged human brain.
First, I start by addressing what is an uncontroversial and universal sign of intelligence: the ability to extrapolate knowledge to unseen cases. Human neuroradiologists have studied the anatomy of the healthy brain and can therefore,
with some success, identify most pathologies present on an imaged brain, even
without having ever been previously exposed to them. Current discriminative
machine learning systems require vast amounts of labelled data in order to accurately identify diseases. In this first part I provide a generative framework that
permits machine learning models to more efficiently leverage unlabelled data for
better diagnoses with either none or small amounts of labels.
Secondly, I address a major ethical concern in medicine: equitable evaluation
of all patients, regardless of demographics or other identifying characteristics.
This is, unfortunately, something that even human practitioners fail at, making
the matter ever more pressing: unaddressed biases in data will become biases
in the models. To address this concern I suggest a framework through which
a generative model synthesises demographically counterfactual brain imaging
to successfully reduce the proliferation of demographic biases in discriminative
models.
Finally, I tackle the challenge of spatial anatomical inference, a task at the centre
of the field of lesion-deficit mapping, which given brain lesions and associated
cognitive deficits attempts to discover the true functional anatomy of the brain.
I provide a new Bayesian generative framework and implementation that allows
for greatly improved results on this challenge, hopefully, paving part of the road
towards a greater and more complete understanding of the human brain.
https://discovery.ucl.ac.uk/id/eprint/10179992/1/phd_ready_for_submission.pdf