An ensemble of 2d convolutional neural network for 3d brain tumor segmentation

Abstract

We propose an ensemble of 2D convolutional neural networks to predict the 3D brain tumor segmentation mask using the multi-contrast brain images. A pretrained Resnet50 and Nasnet-mobile architecture were used as an encoder, which was appended with a decoder network to create an encoder-decoder neural network architecture. The encoder-decoder network was trained end to end using T1, T1 contrast-enhanced, T2 and T2-Flair images to classify each pixel in the 2D input image to either no tumor, necrosis/non-enhancing tumor (NCR/NET), enhancing tumor (ET) or edema (ED). Separate Resent50 and Nasnet-mobile architectures were trained for axial, sagittal and coronal slices. Predictions from 5 inferences including Resnet at all three orientations and Nasnet-mobile at two orientations were averaged to predict the final probabilities and subsequently the tumor mask. The mean dice scores calculated from 166 were 0.8865, 0.7372 and 0.7743 for whole tumor, tumor core and enhancing tumor respectively.