We use a 3D RU-Net which the decoder part is focus on ROI. Before training the network, we resampled all MR images into a fixed size of 384กม384กม32.