Orthogonal-nets: a large ensemble of 2D neural networks for 3D brain tumor segmentation

We propose Orthogonal-Nets consisting of a large number of ensembles of 2D encoder-decoder convolutional neural networks. The Orthogonal-Nets takes 2D slices of the image from axial, sagittal, and coronal views of the 3D brain volume and predicts the probability for the tumor segmentation region. The predicted probability distributions from all three views are averaged to generate a 3D probability distribution map that is subsequently used to predict the tumor regions for the 3D images. In this work, we propose a two-stage Orthogonal-Nets. Stage-I predicts the brain tumor labels for the whole 3D image using the axial, sagittal, and coronal views. The labels from the first stage are then used to crop only the tumor region. Multiple Orthogonal-Nets were then trained in stage-II, which takes only the cropped region as input. The two-stage strategy substantially reduces the computational burden on the stage-II networks and thus many Orthogonal-Nets can be used in stage-II. We used one Orthogonal-Net for stage-I and 28 Orthogonal-Nets for stage-II. The mean dice score on the testing datasets was 0.8660, 0.8776, 0.9118 for enhancing tumor, core tumor, and whole tumor respectively.