Leveraging transfer learning for the super-resolution reconstruction of QSM with limited data for the study of the cerebrovasculature

Motivation: Lengthy acquisitions are needed to produce quantitative susceptibility mapping (QSM) containing enough details to reliably identify the vascular network and capture accurate values of oxygen saturation. Goal(s): To fine-tune a deep learning model for single image super resolution reconstruction o f QSM maps from 1mm to 0.5mm resolution. Approach: Transfer learning was applied on a previously trained 3D densely-connected super resolution network (DCSRN) model, and the vascular network was segmented from the reconstructed susceptibility maps. Results: Transfer learning applied on a DCSRN model demonstrated substantial improvements in the reconstruction of small vessels and reduced partial volume effects. Impact: By demonstrating the effectiveness of transfer learning with a 3D Densely Connected Super-Resolution Network (DCSRN) model, this study provides a practical approach for researchers to improve the resolution of their own QSM data, even with limited resources.