Nested Dilation Networks for Brain Tumor Segmentation Based on Magnetic Resonance Imaging

Aim: Brain tumors are among the fatal cancers in the world. Diagnosing and manually segmenting tumors are a time-consuming task clinically and the performance is highly depended on the experience of a doctor. Therefore, automatic quantitative analysis and accurate segmentation of brain tumors are greatly needed in cancer diagnosis. Methods: This paper presents an advanced three-dimensional multi-modal segmentation algorithm named Nested Dilation Networks (NDNs), which is inspired by the U-Net structure and has been modified to achieve better performance for brain tumor segmentation. We propose residual blocks nested with dilations (RnD blocks) in encoding part to enrich the low-level features and use Squeeze-and-Excitation (SE) blocks in both encoding and decoding parts to boost the significant features. For the purpose of proving the reliability of the network structure, our results are compared with standard U-Net and its transmutation networks. Then, different loss functions are tried to cope with class imbalance problems and maximize brain tumor segmentation results. A cascade training strategy is employed finally to run the Nested Dilation Networks for coarse-to-fine tumor segmentation. The strategy decomposes the multi-class segmentation problem into three binary segmentation problems and then trains each task sequentially. Furthermore, various augmentation techniques are utilized to increase the diversity of data and to avoid overfitting. Results: This approach achieves the dice similarity scores of 0.6652, 0.5880, 0.6682 for edema, non-enhancing tumors and enhancing tumors respectively, in which the training in a single-pass way uses dice loss. After cascade training, the dice similarity scores rise up to 0.7043, 0.5889, 0.7206. Conclusion: Experiments show that the proposed deep learning algorithm outperforms other U-Net transmutation networks in brain tumor segmentation. Moreover, the cascade training applied to Nested Dilation Networks make a better performance than other existing methods. The findings of this study provide considerable insight into the automatically accurate segmentation for the brain tumor.