Whole-body positron emission tomography - computed tomography (PET-CT) imaging is routinely used worldwide in many clinical indications. Nuclear imaging is now a key procedure in the diagnosis, (re)staging, and assessment of treatment response in many cancer types and inflammatory diseases. However, due to the high cost and radioactive exposure, PET scans are only performed when there are strong indications, potentially missing many early stage tumors. Especially children and pregnant women have very limited access to PET scans.
Recent papers have shown that it is possible to enhance PET scans using deep learning techniques. Typically, this is accomplished by recording the list-mode data of the scanner, i.e. a list of radioactive events that are measured during the PET scan. In clinical practice, 100% of these events are used to reconstruct a full 3D image of the patient. Low-dose scans can be simulated by using only e.g. 25% of the measured events. However, in clinical practice, data is not recorded in list-mode and consequently hospitals need to change their acquisition protocol. Moreover, the synthetic generation of low-dose scans from high-dose scans, would enable to use only standard dose scans which are widely available. The aim of this project is to deploy state of the art of diffusion based probabilistic models to learn a mapping from standard dose images to low-dose PET scans. Initially, the synthetic data will be used to train a denoising model in two stages. If successful, the student should explore the possibility of combined training of a denoising and noise-adding module. All implementations should be done using the PyTorch framework.

Python
Deep Learning
PyTorch