Abstract
Accurate segmentation of brain tumors is an essential stage in treatment planning. Fully convolutional neural networks, specifically the encoder-decoder architectures such as U-net, have proven successful in medical image segmentation. However, segmenting brain tumors with complex structure requires building a deeper and wider model which increases the computational complexity and may also cause the gradient vanishing problem. Therefore, in this work, we propose a novel encoder-decoder architecture, called Inception Residual Dense Nested U-Net (IRDNU-Net). In this model carefully designed Residual and Inception modules are used in place of standard U-Net convolutional layers to increase the width of the model without increasing the computational complexity. Additionally, in the proposed architecture, the encoder and decoder are connected via a sequence of Inception-Residual densely nested paths to extract more information and increase the depth of the network while reducing the number of network parameters. The proposed segmentation architecture was evaluated on two large brain tumor segmentation benchmark datasets; the BraTS'2019 and BraTS'2020. It achieved a mean Dice similarity coefficient of 0.888 for the whole tumor region, 0.876 for the core region, and 0.819 for the enhancement region. Experimental results illuminate that IRDNU-Net outperforms U-Net by 1.8%, 11.4%, and 11.7% in the whole tumor, core tumor, and enhancing tumor, respectively. Moreover, the IRDNU-Net enables a great improvement on the accuracy compared to comparative approaches, and its ability in the face of challenging problems, such as small tumor regions, with fewer parameters.