Improved U-Net with Multi-scale Cross Connection and Dilated Convolution for Brain Tumor Segmentation

Image segmentation plays an important role in the early diagnosis, planning of treatment strategies, monitoring of disease progression and prediction of patient outcome of brain tumors. Manual segmentation of brain tumors in MRI images is a very time-consuming task which depends on the operator's experience. U-Net network has made great contributions to tumor segmentation in brain MRI images. In this work, we propose an improved U-Net structure, whose feature maps of each resolution are extracted in the encoding path, and different scale feature maps are fused by multi-scale cross connection in the decoding path. Besides, cascading dilated convolutions are inserted into the bottom joint between encoding path and decoding path. We evaluate our proposed model on the Multimodal Brain Tumor Image Segmentation (BRATS 2017) dataset, which contains 210 cases with high-grade brain tumor and 75 cases with low-grade tumor. Cross-validation has shown that our method can obtain promising segmentation efficiently. The resulting dice indices for whole tumor, core tumor and enhancing tumors are 0.87, 0.80 and 0.64 respectively.