Research theme
Robust everywhere: harmonisation and deployment
Building models that generalise across scanners, sites and clinical settings, including portable devices.
Research / Robust everywhere
The problem
Ultrasound is available almost everywhere, but images differ between scanner makes, settings and sites, and models trained in one hospital often fail in another. Patient data usually cannot leave the hospital, and many clinics use small portable scanners with little computing power.
Our approach
- HarmoniseRemove scanner and site differences so measurements mean the same thing everywhere.
- AdaptDomain adaptation, including source-free methods that need no access to the original training data.
- FederateTrain across hospitals without patient data leaving any of them.
- CompressLightweight models that run in real time on portable scanners.
Key papers
All papers in this theme (13) →- Impaired fetal brain growth and neurodevelopmental deficits at 2 years: deep phenotyping of maternal–fetal pathophysiology
- Clinical Grading of Artificial Intelligence‐Based 3D Fetal Brain Segmentations: A Cross‐Vendor Evaluation of Deep Learning in Fetal Neuroimaging
- Deep learning assessment of fetal brain maturation on 3D ultrasound volumes in early‐onset fetal growth restriction
- UniFed: A unified deep learning framework for segmentation of partially labelled, distributed neuroimaging data
- SFHarmony: Source Free Domain Adaptation for Distributed Neuroimaging Analysis
- Prototype Learning for Explainable Brain Age Prediction
- Challenges for machine learning in clinical translation of big data imaging studies
- STAMP: Simultaneous Training and Model Pruning for low data regimes in medical image segmentation
- INSightR-Net: Interpretable Neural Network for Regression Using Similarity-Based Comparisons to Prototypical Examples
- FedHarmony: Unlearning Scanner Bias with Distributed Data
- Deep learning-based unlearning of dataset bias for MRI harmonisation and confound removal
- Low-Memory CNNs Enabling Real-Time Ultrasound Segmentation Towards Mobile Deployment
- Uncertainty Estimates as Data Selection Criteria to Boost Omni-Supervised Learning