Brain Tumor Volumetric Segmentation in Multimodal MRI Using 3D Convolutional Neural Networks

Complications of the central nervous system in removal brain tumors lie in the differentiation between the neoplasm and healthy tissues. Tumor segmentation is a fundamental task in the detection and clinical treatment planning. In this work, two neural networks architectures based on the U-net network have been designed and trained to automatically segment different tumor substructures using the Multimodal Brain Tumor Segmentation Challenge (BraTS) 2020 training dataset. The first network was obtained by transferring all the operations and 2D layers of the U-net network to their 3D versions and replacing all its skip connections with concatenation blocks, the second architecture is similar to the first by adding a max pooling layer, a transposed convolution layer and a concatenation block. In both architectures the group normalization layer was used instead of batch normalization as a suggestion of the state-of-the-art. Our best model is capable of segmenting gliomas on the validation dataset with a Dice similarity coefficient of 0.85, 0.84 and 0.87 for the whole tumor, tumor core and enhancing tumor core respectively, therefore, the inference process takes 0.21 seconds per patient with a precision of 0.89. Different values of hyperparameters were tested, deeper architectures and different kernel numbers were implemented in the convolution and concatenation blocks, in no case was a significant improvement in our metrics obtained. Our models outperform most state-of-the-art methods in terms of Dice coefficient, sensitivity, specificity and precision.