Improving performance of deep learning models using 3.5D U-Net via majority voting for tooth segmentation on cone beam computed tomography