An ensemble of 5 randomly initialized 2D U-Net models trained on augmented data. In the decoding path, max-pooled versions of the original images are reintroduced to highlight WMH information. During inference, original images flipped versions (flipped along the x-axis, y-axis, and both) processed by the five ensembles and aggregated.

Description of the method.
Docker container of this method.
Journal paper of the method.