Developing next-generation, AI-enabled, medical image processing in the spinal cord for MS clinical trials and routine care

1) Brief description of the context of the research including potential impact

Multiple sclerosis (MS) is a demyelinating disease in which the immune system attacks the central nervous system, causing physical and cognitive disabilities. Anatomical images taken from the brain and spinal cord can provide insights into morphometric changes during disease progression, with the spinal cord being a main driver of disability in this disease. However, imaging data acquired in a clinical setting are often heterogeneous, and poorer quality compared to clinical trials or research data.

Widely used deep learning models like convolutional neural networks (CNNs) work well in segmentation tasks, but their performance depends on the contrast and resolution of the images they are trained on and, even with conventional preprocessing and data augmentation techniques, these models have difficulties generalising to other image contrasts and resolutions. These models also require ground truth labels such as manual annotations for model training and validation, and because this constitutes a very time-consuming task, it is considered a bottleneck in AI-model developments. Existing methods for spinal cord segmentation fail to accurately process lower quality, routine care spinal cord MRI scans.

2) Aims and Objectives

The main objective of this project is to develop a deep-learning based model for spinal MRI data that does not rely on a significant amount of ground truth labels, is MRI-contrast agnostic and allows segmentation of different structures at any level of the spine fully automatically. The outcome of this project will aim to enable the analysis of spinal cord morphometry in clinical trials and real-world data.

3) Novelty of Research Methodology

The project will leverage the access to extensive MS patient datasets from our partner hospital (UCLH Trust / National Hospital for Neurology and Neurosurgery) and collaborating hospitals from across the UK.

In relation to the methodology,

- Deep learning models that allow a larger proportion of unlabelled data for training and minimise the ground truth label generation process will be used in this project, in particular, self-supervised learning models.

- The developed model will also allow segmentation of the spinal cord structures at any level of the spine present in the MRI scan (i.e. cervical, thoracic or lumbar). This will be achieved with some mechanisms that optimise the learning focus of the model, paying special attention to the spinal cord structures regardless of the vertebral level they have been acquired. This will be approached using attention mechanisms.

- Finally, synthetic spinal cord data will be used to complement the data for model training and improve model generalisation to other domains, such as different MRI contrasts and image resolutions. These synthetic images will be produced from MRI images using a generative model that will be developed during the first year as part of the MRes research project.

4) Alignment to EPSRC's strategies and research areas

This project perfectly aligns with the EPSRC mission of making a world-class impact to transform health and healthcare by developing next generation technologies in medical imaging. The project will facilitate the analysis of spinal cord clinical data and have the potential to contribute to large-scale studies that focus on tracking disease progression in MS, make an impact in drug development and patient care by helping to choose the right treatment for the right person or simply leverage large datasets for processing from hospital archives that, while not actively used for patient management, hold significant importance for MS research.

5) Any companies or collaborators involved

IXICO Technologies Ltd, University College London Hospital NHS Foundation Trust, Imperial College Healthcare NHS Trust and University Hospital of Wales.