Foundation Models and Generative AI for Accelerated Multimodal Medical Image Reconstruction, Data Quality Enhancement, and Multicentre Harmonisation

Medical imaging is at the heart of modern healthcare, playing a crucial role in the early detection, diagnosis, treatment planning, and monitoring of diseases such as heart disease, cancer, and neurological conditions. Across the NHS and healthcare systems globally, imaging methods like MRI, PET, CT, and advanced microscopy are essential. However, acquiring high-quality images is often time-consuming, resource-intensive, and varies from one hospital or scanner to another. Long scan times can cause discomfort for patients, reduce the number of scans performed daily, and increase healthcare costs and waiting lists. This fellowship project aims to address these challenges by developing advanced artificial intelligence (AI) methods—in particular, generative AI, foundation models, and federated learning—to dramatically speed up image reconstruction, improve image quality, and harmonise data from multiple centres and devices. Here: (a) Generative AI refers to computer programs that can learn from large amounts of data and then create realistic new examples of that data—in this case, generating high-quality images from incomplete or noisy scans, as well as improving image clarity and detail; (b) Foundation models are large AI models that are trained on huge and diverse datasets so they can adapt to a variety of tasks, such as improving images from different scanners and centres without needing to retrain for each new situation; and (c) Federated learning is a privacy-preserving technique that allows AI models to be trained using data from many hospitals without moving the data from its secure location. Instead, models are shared and improved collaboratively, while patient information stays protected. Key Challenges: -Long scanning times that limit patient throughput and place strain on healthcare resources. -Differences in data quality between centres and devices, making it hard to compare results or conduct large-scale studies. -Missing or incomplete data that can reduce the reliability of clinical decisions. -The need for automated analysis tools that can quickly and accurately process medical images and highlight important findings. Research Goals: -To develop generative AI methods that accelerate image reconstruction for MRI, PET, CT, photon counting CT, and microscopy, while enhancing clarity and detail. -To create data enhancement tools that ensure consistency and reliability of biomedical images across different hospitals and scanner types. -To develop foundation model-based methods for synthesising and completing missing data to make better use of patient records. -To design and implement federated learning systems that train AI models across multiple centres securely and fairly. -To build fully automated pipelines that can rapidly process medical images and extract important indicators of disease, making clinical decisions faster and more consistent. Benefit and Impact: -This fellowship project will greatly improve clinical workflows by reducing scan times and image analysis from hours to minutes. Patients will benefit from faster, more comfortable scans and quicker results. Clinicians will save time, reduce manual work, and receive more consistent, reliable information to help guide treatment. -Healthcare resources will be used more efficiently, enabling hospitals to scan more patients without additional equipment or staff. This will help reduce waiting lists and overall costs. -The wider socio-economic impact is significant. Earlier and more accurate diagnosis will lead to better patient outcomes, fewer hospital admissions, and long-term savings for the healthcare system. In addition, enabling large-scale, multi-centre research will accelerate the discovery of new treatments and personalised care strategies, benefitting patients and healthcare providers around the world.