MRI Radiomics of adult brain tumours

The proposed research sits at an interdisciplinary point between Medical Imaging, Neuroscience and Computer Science and comprises radiomic analysis on a wide range of available datasets from St George's, University of London, in collaboration with the Laboratory of Vision Engineering at Lincoln University. This research aims to provide approaches to further personalised medicine using deep learning architectures to mathematically describe the underlying histopathology of Brain Tumours for the development of artificially intelligent classification systems for use on new and unseen MRI data.

Brain Tumours are incredibly heterogenous in nature, with wide differences in shape, contrast and location between patients. Lower grade gliomas require monitoring in case of transformation into higher grade brain tumours which require immediate administration of treatment. A watch and wait approach is used for low grade glioma, and any indicator of imminent change could prove vital in predicting a deterioration in patient condition.

MRI scanning is a common technique used in the identification of Brain Tumours and can describe varying regions of cancerous tissue, such as oedema, necrosis, and progressive tumour, through the capturing of the textural and shape appearance of these damaged regions. Different MRI images have the potential to visualise each of these areas differently, and a combination of MRI images can lead to a more rounded diagnosis of brain cancer through the development of deep learning architectures trained on a wide range of image types.

MRI can be computationally analysed using deep learning techniques through which tumour regions are processed using image filter operations. This process takes MRI image regions and generates an understanding of tissue subclasses (e.g. normal brain tissue, necrosis, oedema) utilising filtering for the identification of image texture and spatial constructs.

This artificially intelligent learning process takes a logically labelled image containing brain cancer and formulates an understanding of the textural and structural components which comprise the associated cancer tissue. This information is then applied to new images which have not been analysed, and are not labelled for classification into comprising sub-regions of healthy and cancerous tissue.

It is also possible to extract features from trained neural networks for storage external to the deep learning architecture. These large biomarker datasets can be further analysed for a deeper understanding of the statistical significance of information obtained from artificially intelligent architectures through external machine learning investigations. Additional data analysis allows for modularity in system design to explore further research opportunities in relation to tumour histopathology, genetic variants and patient outcome.

This research seeks to build upon existing work in, and where applicable, provide better solutions to:

Tumour segmentation:
The identification of entire tumour regions in new and unknown patient MRI.

Monitoring:
Examine changes in a patient's history where longitudinal data is available, as well as identifying potential changes as a result of administered treatment.

Improved efficiency:
Support radiologists in the identification of regions containing potential grades of cancer tissue, as well as the probability of that grading.

Feature fusion:
Create new features from the combination of different MRI images.

Standardisation:
Attempting to find shared characteristics between MRI generated in different locations and from different scanning systems.

MRC Skill Priority:
-Data analytics and informatics
-Machine learning
-Artificial Intelligence
-Statistics
-Interdisciplinary