Project Description

Multiple Sclerosis (MS) is a chronic, immune-mediated neurodegenerative disease of the central nervous system that often manifests with subtle early symptoms. Early diagnosis is vital but currently relies on invasive, expensive, and time-consuming methods such as brain MRI and lumbar puncture.

This project proposes a non-invasive, AI-powered diagnostic alternative based on retinal imaging. Retinal Optical Coherence Tomography (OCT) provides detailed, cross-sectional views of retinal layers—offering a unique window into neurodegeneration, as the retina shares embryological origins with the brain and spinal cord.

Goals and Innovation

This work aims to establish a computational framework for early MS detection, combining:

  • Multiple Instance Learning (MIL) to enable image-level MS classification using weak (scan-level) labels, accommodating intra-patient variability and noise.
  • Contrastive Learning to enhance representation learning and boost generalisation across diverse data sources.
  • Attention-Based Pooling to highlight the most informative regions in the scan for interpretability and clinical insight.
  • Statistical Harmonisation to address variability introduced by different OCT scanners and acquisition protocols.
  • Federated Learning to train models collaboratively across multiple institutions without centralising data—preserving privacy and legal compliance.

Clinical and Technical Impact

This work will:

  • Enable earlier, less invasive screening for MS, especially for patients with suspected neurodegenerative symptoms but inconclusive MRI or CSF findings.
  • Democratise access to diagnostic tools using widely available OCT devices in clinics and optometry practices.
  • Contribute to interpretable AI in neurology, where understanding the basis of predictions is essential for clinical adoption.
  • Create a scalable, federated framework for deployment across multiple NHS Trusts and global research collaborators.

Technical Stack

  • OCT image preprocessing using retinal layer segmentation and spatial normalisation.
  • MIL architecture with instance-level embedding and global bag-level classification.
  • Supervised contrastive loss and batch harmonisation modules.
  • Federated learning orchestration using secure parameter aggregation protocols.
  • Evaluation across public datasets (e.g., OCTAGON, MS-OCT) and prospective data from partner sites.

Collaborators

  • THRIVE Centre – University of West London
  • Clinical neurologists and imaging specialists (prospective NHS Trust collaborators)
  • External research data consortia contributing OCT datasets