A machine learning approach for the quantification of tumour microstructure from diffusion-weighted MRI

Brief Description of the context of the research including potential impact
Diffusion-weighted MRI uses the diffusive motion of water to assess tissue microstructure. By sensitising the MR scanner to the restricted diffusion of water, signal contrast can be induced allowing the delineation of tissue microstructure and quantification of properties such as cell size/ density. This information can be used to create biomarkers of tumour grade, prognosis and treatment response which have potential implementation in clinic where they are all critically required.
Simulating MR signal from images of tissue microstructures facilitates the development of methods to quantify the signal. This simulation based approach eliminates time consuming experimental acquisition of MRI data and enables manipulation of scan parameters for optimisation. Simulated data also reduces the likelihood of inconsistencies allowing generated signal to be reproducible and artefact-free.
Aims and Objectives
Traditionally models of tissue microstructure with simplistic geometries have been used to analyse diffusion weighted signals. My aim is to eliminate such models, which can be reductionist, and instead use an alternative approach of machine learning to extract tissue parameters. Monte Carlo simulations will be used to generate Diffusion MR signals from 3D high resolution datasets by measuring molecule displacement from their original starting point over a determined amount of time through 'random walks'. This will be done for a variety of MR pulse sequences to provide multiple MR datasets from a single high resolution tissue microstructure of colorectal tissue.
This synthetic MRI data will be used to train, test and validate a supervised machine learning network to output estimations of tumour microstructure properties dependent on the input vector of MR signal. The training will allow for parameter estimation on previously unseen data.
Novelty of the Research Methodology
Where Monte Carlo simulations have been used for years to synthesise MR signal, the inclusion of blood flow and simulation in complex geometries is a more recent development for which there is currently no available, complete and open library which covers the large range of sequences used in MRI acquisition and is specific to colorectal cancer, the 4th most common cancer in the UK, making it an extremely viable disease to look at.
Whilst machine learning approaches for image classification and segmentation have increasingly widespread in the past decade, their application to medical imaging is still in its infancy and their use in the quantification of MR signals is an even smaller field. In addition, what research has been done into this has almost exclusively focused on application to MR of brain and central nervous tissue whilst cancer microstructure has not yet been researched.
Alignment to EPSRC's strategies and research areas
This project aligns with the EPSRC focus on multi-disciplinary research. The practical aspects involve computer science and artificial intelligence via machine learning. The applications of this project have place in healthcare technologies and look to be the future of assessing diseases such as cancer.