Combined 3D CNN for Brain Tumor Segmentation

A residual 3D U-Net enables multi-scaling with the concatenation of feature maps from different scales. The connectors between different sub-networks assist in fusion of feature maps. A multi-path architecture enables the fusion of feature maps from different scales. In this paper, the combination of two different architectures is proposed to effectively combine the advantages of both the models to solve the problem of complex brain MRIs. To validate our statement, we build a 3D variation of U-Net with residual, inception, and dense connections. After each upsampling process, the early, middle, and late fusion methods are employed to introduce the information from the multi-path model with the residual 3D U-Net. In the proposed work, we use multiple atrous rates on various paths to solve the problem of small receptive field size. To test our algorithm, a publicly available benchmark dataset is used during training and validation. To handle the over-fitting issue, we used the data augmentation techniques. The multi-label dice loss function reduces the chance of class imbalance problem between healthy and unhealthy tissues. The balance between the severely unbalanced classes during training enabling our algorithm to learn efficient features. Our proposed approach gives competitive scores on different evaluations and provides a meaningful solution to understand the problem of sophisticated brain tumor segmentation.