Magnetic Resonance Imaging (MRI) is a powerful, non-invasive medical imaging
modality adept at showcasing soft-tissue contrast and well-suited to imaging most body
parts. However, MRI is overwhelmingly used to produce qualitative images whose
individual voxel values carry no diagnostic value. Instead, information is primarily
derived from analysing the contrast between regions of interest.
Challenges persist when it comes to downstream analyses predicated on using images
acquired under different conditions. The first is that models are prone to lack generalisability
to domains which they were not made privy to during training, and the second is
the lack of standardisation when extracting biomarkers, as models typically cannot divorce
perceived contrast from the true underlying anatomy.
This thesis addresses the generalisability and standardisation problem by designing
self-supervised segmentation networks that are cognizant of the physics underpinning the
acquisition process. These networks are trained using simulated MR images boasting a
wealth of contrasts, thus enabling a breadth of generalisability and granting them the ability
to innately account for and standardise MR images, regardless of the sequence parameters
used to acquire them. This is followed by iterating over the initial designs, enhancing
generalisability and robustness and reducing the pre-processing time by modifying various
aspects of the training pipeline. Further, uncertainty modelling is incorporated into the
models to allow for additional levels of safety and introspection. Additionally, we demonstrate
that despite their simulation-based training, our models generalise to real-world data,
and so too does their internal modelling of the interplay between contrast and sequence parameters.
Lastly, an unsupervised, heteromodal framework for translating typical qualitative images
into quantitative tissue maps is proposed, the first of its kind.
The hope is that the work contained herein will benefit the standardisation community
and that its concepts will be translated into a greater variety of sequences and body part
images.