Partial volume estimation from MRF acquisition using a deep learning approach

Motivation: Partial volume effects in MRI limit the accuracy of quantitative parameters like T1 and T2, especially when a voxel contains mixed tissue types. This research seeks to improve partial volume estimation process. Goal(s): The study aims to enhance partial volume estimation by introducing a self-supervised neural network, MRF-PVMixer, designed to process MRF signals directly for more accurate mapping. Approach: Applied deep learning framework, MRF-PVMixer extends the MRF-Mixer model with another U-Net decoder to estimate partial volume fractions. This method was validated on synthetic datasets created with radial-undersampling. Results: MRF-PVMixer showed promising results in PV estimation as seen in both visual and quantitative results. Impact: This study's self-supervised deep learning approach for partial volume estimation directly from MRF signals could improve diagnostic accuracy and streamline quantitative MRI processes. It opens avenues for real-time, artifact-resilient tissue characterization, potentially transforming clinical workflows and supporting patient-specific imaging studies.