Fig. 4.
Segmentation results in the training dataset. From top to bottom are axial slices of four different subjects. From left to right are FLAIR image, ground truth, the results from the standard U-Net, and the results from the proposed U-Net with multi-scale HF method. Yellow voxels indicate WMHs segmented using each method. Color boxes show our approach’s improvement of false-positives (orange border) or false-negatives (green border) observed in the standard U-Net.