Figure 1.

Proposed data processing pipeline. Black arrows show steps necessary for retrieving (filled) and optionally post-processing (dashed) predictions, while orange arrows highlight steps only required for ground-truth creation—i.e., not for making predictions. (1) Data preparation includes processing steps for the extraction of model input such as segmentation and centerline extraction, as well as the extraction of segment information necessary for ground-truth creation. (2) Training involves patch extraction and training of the BRAVE-NET architecture. Patch-wise predictions are used during training for performance assessment and for the reconstruction of whole-brain predictions after training. (3a) Whole-brain predictions can directly be evaluated by voxel-wise scores or (3b) optionally further refined in post-processing via segment washing. For this step, segment information is additionally utilized. (4) Post-processed predictions can be evaluated by both voxel-wise and segment-wise scores.