Motivation: Existing methods for arterial input function (AIF) selection in DSC-MRI, such as deep learning, criterion-based, and clustering-based approaches, are either inaccurate, exhibit low reproducibility, or rely on specific clinical data for training. Goal(s): Our goal was to predict the true ground-truth AIF from baseline simulated DSC-MRI time series data without the need for specific AIF pixel selection or training with clinical data. Approach: We utilized a physics-informed neural network (PINN) approach using a simulation-based dataset. Results: Our model predicted a more accurate AIF compared to existing methods, resulting in perfusion maps that highlighted perfusion anomalies which were otherwise difficult to detect. Impact: Our model was trained on a large amount of simulation data without requiring clinical data. Moreover, Input data consisted solely of baseline DSC-MRI, eliminating the need for AIF selection, whether manual or automatical.
Asaduddin et al. (Tue,) studied this question.