Optimizing feature-based loss functions for AI Super-resolution of 7T Brain Diffusion MRI

Motivation: The acqusition of data for diffusion MRI analysis like tractography and connectivity in clinical practice is limited by scan time. Deep learning approaches for super-resolution (SR) of angular and/or spatial resolution may alleviate this limitation. Goal(s): In this study, we aim to assess which layers drive the performance of feature-based loss functions for super-resolution models. Approach: We perform an ablation and an isolation study using high resolution 7T DWI. SR-models are using feature-based losses derived from isolated and combined VGG16-layers. Results: Feature-based losses may enable high consistency resolution enhancement for 7T MR diffusion data, if features are derived from shallow layers. Impact: Our results show how feature-based loss functions need to be adapted to work well for SR models targeting MR diffusion data. We further demonstrate how such SR models may be trained using publicly available data, enabling reproducibility and application.