AI in radiology holds promise for earlier diagnosis, improved monitoring, and more personalized treatment, but current models are narrow, require large annotated datasets, and do not generalize well across populations. This creates barriers for healthcare providers and risks widening inequalities in access to advanced diagnostics.
Research question: Can a task-agnostic foundation model trained on UK Biobank whole-body MRI enable reliable AI-assisted anomaly detection and provide transferable representations for disease assessment tasks, while reducing the data and computational demands that currently limit wider clinical use?
Objectives:
– Pretraining: Use UK Biobank whole-body MRI as the primary dataset for large-scale self-supervised pretraining, leveraging its unique population coverage to learn normal anatomy, age-related variation, and incidental pathology.
– Core application: Evaluate the model on anomaly detection within UK Biobank, establishing a population-level reference of normal MRI across the life course to reliably distinguish normal from abnormal imaging and support radiologists through triage.
– Extended applications: Fine-tune the pretrained model on complementary pathology-rich datasets (e.g. public and private available imaging cohorts) for disease-specific tasks such as classification, detection, segmentation or image enhancement.
– Impact: Demonstrate that this approach supports earlier detection, better disease monitoring, and more equitable access to advanced AI-driven tools in radiology.