Applying adaptive convolution to brain data – Making use of transfer learning

Motivation: Deep learning approaches for QSM-based dipole inversion lack generalizability towards acquisition parameters. Goal(s): Our aim was to address data scarcity by integrating known information in the network model and investigate the feasibility of transfer learning. Approach: The acquisition parameters (voxel size, FOV orientation) were integrated with manifold learning. The models were pre-trained on large-scale synthetic data sets and fine-tuned on in-vivo brain data in a second step. Results: The use of manifold learning increased generalizability, while transfer learning substantially improved the quality of computed susceptibility maps. Impact: While this study demonstrates the feasibility of cross-domain knowledge transfer in deep learning approaches for QSM, it also points to the potential of fine-tuning network parameters to scanner-specific data in general, boosting the performance of neural networks therewith.