Post-Reconstruction Image Denoising and Artefact Removal for Low Count Positron Emission Tomography

Student thesis: Doctoral ThesisDoctor of Philosophy


Positron emission tomography (PET) is a powerful medical imaging modality for the brain, for cancer and for the heart. Image reconstruction is a crucial component to the success of PET, and methodology has undergone a number of significant advances, including progression from 2D to 3D PET reconstruction (improving signal to noise ratio), progression from analytical to iterative reconstruction methods (reducing variance) and advanced modelling of the PET data acquisition to improve image quality (improving image resolution). The latter includes resolution modelling (RM), which in PET accounts for effects including the positron range, photon acollinearity and limited detector resolution.

While notable improvements in image quality have been demonstrated, advances are still very much needed, and the aim of this thesis is to deal with noise and artefacts (such as the notorious ringing artefact introduced by RM, as well as partial volume effects) without introducing quantitative errors. Present approaches either leave noise and RM artefacts in the images, or else they compromise the spatial resolution of the end-point images. This thesis proposes novel deep learning (DL) based techniques to reduce noise and resolve artefacts without compromising resolution in a variety of scenarios, including in the case of quantification of small regions (e.g. lesions) and low-count PET imaging.

Furthermore, multimodality scans (specifically PET-MR) are used, where the jointly acquired data provides anatomical information which aids in the reduction of noise and artefacts while increasing resolution, thereby enabling low dose and/or reduced scan durations. The DL techniques developed are also robust enough to cope with highly limited training datasets and have built-in model consistency in order to constrain their outputs. Enforcing such constraints sets a maximum limit on errors in cases when a DL method fails to perform well on test data. A thorough comparison between the current most promising DL proposals for PET is also conducted, with the aim of providing much-needed guidelines for network architecture and design, for a given quantity of available training data.
Date of Award1 Jul 2022
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorAndrew Reader (Supervisor) & Paul Marsden (Supervisor)

Cite this