Multi-Dataset Image Reconstruction for Longitudinal and Multi-Tracer Positron Emission Tomography

Student thesis: Doctoral ThesisDoctor of Philosophy


Repeated scanning of subjects is commonplace in clinical and research positron emission tomography (PET), either for the assessment of physiological change, or to provide complementary information from distinct biological processes. In many of these cases, the intrinsically noisy nature of PET data can hinder the interpretation of images, or necessitate the use of higher radioactive doses. Despite the existence of common information between repeated PET scans, they are typically reconstructed individually, representing a sub-optimal use of the available data. This thesis proposes and evaluates methods of transfer-ring information between these multi-dataset PET images during the image reconstruction process, in order to improve the quality of the resulting images. Two approaches for sharing information are investigated, simultaneous recon-struction for longitudinal imaging and guided reconstruction for multi-tracer imaging.
For the simultaneous case, longitudinal datasets are coupled in the image reconstruction optimisation problem by using dierence-image priors, thereby encouraging desired characteristics of the resulting dierence images. This ap-proach was applied to longitudinal PET imaging for tumour monitoring, show-ing reduced image noise and lower reconstruction errors compared to standard image reconstruction methods. Importantly, these methods denoise in the dif-ference image domain, thereby providing longitudinal images that exhibit noise levels comparable to standard PET images reconstructed with higher levels of counts. In general, using the proposed methods to reconstruct S scans results in noise levels similar to those usually achieved by using traditional recon-struction methods with datasets containing S times the number of counts. Experimental results demonstrated how changing the characteristics encour-aged in dierence images can aect the reconstruction of tumour regions in the resulting PET images, highlighting the need for application-specific selection and validation of dierence-image priors.
For multi-tracer PET imaging, guided reconstruction was proposed as a method of transferring information from high quality PET images to lower quality images. Inspired by the literature on anatomically guided PET image reconstruction, a weighted quadratic penalty was applied to [18F]-fluorodeoxy-glucose (FDG)/[11C]-methionine imaging of brain tumours. Image structure information was extracted from the [18F]-FDG images and embedded into the reconstruction of the [11C]-methionine images. Guiding [11C]-methionine image reconstruction with [18F]-FDG information in this way led to edge-preserving noise reduction in the [11C]-methionine images compared to standard recon-struction methods. Furthermore, guiding reconstructions using [18F]-FDG im-ages was observed to outperform anatomically guided PET image reconstruc-tion in some regions. These results suggest that multi-tracer PET imaging could benefit from the use of guided image reconstruction methods, with many possible applications for the utilisation of the guided approach.
Overall, this thesis demonstrates the principle of sharing information be-tween PET images during reconstruction, showing improved image quality when applied to specific contexts. Future work remains in further assessment of the proposed methods in clinical contexts, and testing of the methods in other related areas of multi-dataset PET imaging.
Date of Award2019
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorAndrew Reader (Supervisor) & Julia Schnabel (Supervisor)

Cite this