Advancing Quantitative PET Imaging with Machine Learning

Abstract

Medical imaging with positron emission tomography (PET) plays an important role in the detection, staging, and treatment response assessment in cancer, neurological and cardiovascular conditions, inflammation and infection. PET imaging is based on measuring the distribution of injected radioactive tracers, designed to follow specific biological pathways. One of the main advantages with PET is that it allows, not only visualization of regional tracer uptake, but also quantification of the underlying biological process.

There are many challenges associated with PET imaging, which, unless accounted for, may reduce the accuracy and precision in PET-based quantification. This thesis addresses the impact of imaging artifacts and subject motion on static PET-based tumor quantification and on machine-learning-based prediction models. Furthermore, the challenge of arterial blood sampling, required for quantification in dynamic PET is addressed. To this end, four papers are presented, suggesting methodology for improved PET-based quantification.

In the first two papers, the impact of imaging artifacts (Paper I) and respiratory motion (Paper II) on tumor quantification is investigated in two lung cancer PET/magnetic resonance imaging cohorts. This knowledge is important, not only for PET-based quantification on a patient level, but also as a pre-processing step in machine-learning-based prediction models.

In the next two papers, a non-invasive machine-learning-based approach is proposed to replace the arterial input function, required for tracer kinetic modelling in dynamic PET applications. The approach is evaluated in a pre-clinical mouse PET cohort (Paper III) and in a human clinical PET cohort (Paper IV). The proposed methodology may considerably simplify the acquisition and analysis workflow in future pre-clinical and clinical dynamic PET studies, by avoiding the need for invasive blood sampling.