Self-supervised deep-learned physics-informed PET image reconstruction for oncology

Aims of the PhD Project

Develop self-supervised deep-learning PET reconstruction methods which do not need ground truth reference data
Develop fast single-step operators for 3D PET reconstruction from sinogram data for any desired objective function
Develop applicability to time-of-flight PET imaging in oncology, improving image quality
Estimate reconstruction uncertainty images

Positron emission tomography (PET) is a medical imaging modality that is able to diagnose cancer and monitor treatment efficacy. PET images are however often limited by noise and low spatial resolution, which can limit the ability to see small regions of disease. Recently, the use of AI within image reconstruction has offered notable improvements in PET image quality, although there can be risks with use of conventional AI methods which draw upon large volumes of data from many other patients for supervised learning.

This project concerns the special case of harnessing advanced AI methodologies especially for the case of using only the acquired data from the unique patient. In tandem with this, use of external data will nonetheless be explored at least for comparison purposes. The goal though is for this project to develop novel self-supervised image reconstruction methods which deliver AI benefits using predominantly only the patient's own data, seeking to avoid some of the pitfalls of conventional supervised deep learning.

The motivation is that PET image quality has an impact on clinical decision making and treatment pathways for patients. AI offers clear benefits, but these need to be robust benefits which rely mainly on a patient's own unique data and where the degree of uncertainty in the reconstructed images needs to be made clear to decision makers.