Deep learning strategies for acquisition, processing, and quantification of dynamic susceptibility contrast perfusion MRI

Abstract

Perfusion MRI is an imaging technique to assess hemodynamic status of a patient. The routine clinical protocol for perfusion MRI is a contrast agent based imaging technique called dynamic susceptibility contrast MRI (DSC MRI). The current DSC MRI imaging technique has a few weaknesses, including its requirement of contrast agent injection, inaccurate approximation of perfusion parameters, a lack of standardized artefact correction methods, and variability in AIF selection. This dissertation mainly focuses on addressing the aforementioned weaknesses by applying deep learning solutions. First, a method to acquire perfusion information from dynamic angiographic data is proposed, resolving the requirement of additional contrast agent injection while also making clinical protocol for stroke diagnosis shorter. Second, a simulation-based physics-informed neural network is proposed, which addresses the inaccuracy in approximating perfusion parameters while also being accurate across a wider range of noise levels and requiring shorter processing time. Third, various generative artificial intelligence methods are tested to recover DSC MRI data in various artefacts cases. Lastly, a simulation-based deep learning method is proposed for automatic AIF selection that is more accurate and reduce the need for manual selection. With these technical developments, DSC MRI analysis can be done quickly with minimal human intervention but still reliable, and robust for real clinical application.