This paper introduces Masked V-Net architecture, a variant of the recently introduced V-Net[13] that reformulates the residual connections and uses a ROI mask to constrain the network to train only on relevant voxels. This architecture allows dense training on problems with highly skewed class distributions by performing data sampling on the output instead of in the input. We use Masked V-Net in the context of brain tumor segmentation and report results on the BraTS2017 Training and Validation sets.

Peer Reviewed