A Multi-path Decoder Network for Brain Tumor Segmentation

The identification of brain tumor type, shape, and size from MRI images plays an important role in glioma diagnosis and treatment. Manually identifying the tumor is time expensive and prone to error. And while information from different image modalities may help in principle, using these modalities for manual tumor segmentation may be even more time consuming. Convolutional U-Net architectures with encoders and decoders are state of the art in automated methods for image segmentation. Often only a single encoder and decoder is used, where different modalities and regions of the tumor share the same model parameters. This may lead to incorrect segmentations. We propose a convolutional U-Net that has separate, independent encoders for each image modality. The outputs from each encoder are concatenated and given to separate fusion and decoder blocks for each region of the tumor. The features from each decoder block are then calibrated in a final feature fusion block, after which the model gives it final predictions. Our network is an end-to-end model that simplifies training and reproducibility. On the BraTS 2019 validation dataset our model achieves average Dice values of 0.75, 0.90, and 0.83 for the enhancing tumor, whole tumor, and tumor core subregions respectively.