Enhancing PET/CT imaging for preclinical research using mathematical modeling and AI

This thesis aims at augmenting Positron Emission Tomography (PET)/Computed Tomography (CT) imaging in preclinical research to allow for potentially a new myocardial disease model in rats and improve the reliability and time-cost of the xenografted subcutaneous tumor model in mice. We hypothesized that these goals could be attained by leveraging mathematical modeling and artificial intelligence solutions. The presented work spans Positron Range Correction (PRC) in PET imaging, the development of annotated datasets, and the application of Artificial Intelligence (AI) for automated tumor segmentation in (CT) scans of mice.

In Study I, we investigated the use of PRC to enhance the spatial resolution of cardiac 82Rb PET/CT imaging in rat models. Using phantom studies and proof-of-concept animal experiments, we demonstrated that PRC could significantly improve image quality and reduce spillover effects, though at the cost of increased noise. The Tissue-Dependent Spatially-Variant (TDSV) PRC approach showed the best improvement in spatial resolution but also introduced more artifacts. These findings suggest that PRC could enable the use of 82Rb PET/CT for studying cardiac infarction and perfusion in rat models, providing
a cost-effective alternative that does not require an on-site cyclotron.
Study II focused on establishing a comprehensive CT database of mice with subcutaneous, where we collected 452 CT scans across 10 datasets and had them labeled by three independent annotators. The study quantified inter-annotator variance, reporting a Fleiss’ Kappa of 0.903 ± 0.046 and a mean volume estimation deviation of 6.9% ± 4.7%. This database provides a valuable resource for developing and validating automated tumor segmentation tools.

In Study III, we leveraged the database from Study II to develop and validate TumSeg, an automated deep learning system for tumor segmentation. The model achieved a Dice score of 0.934 ± 0.036 and estimated tumor volumes with 1.5% ± 9.8% error. We also implemented an uncertainty quantification module capable of predicting segmentation accuracy and generating uncertainty heat maps. The system demonstrated successful transfer learning capabilities when adapted to Magnetic Resonance Imaging (MRI) scans, achieving a Dice score of 0.890 ± 0.012 after fine-tuning.

Together, these studies advance the field of preclinical imaging by improving image quality through physics-based corrections, establishing robust datasets for algorithm development, and creating reliable automated analysis tools. The work provides both immediate practical solutions and foundations for future technological developments in medical imaging.