Research on segmentation of MRI brain tumor image based on improved UNet3+
Yichen Hou · 2022
According to the characteristics of multimodality, complexity and computer-aided segmentation of MRI brain tumor images, an improved UNet3+ brain tumor segmentation network based on full-scale jump connection is proposed. The main work is as follows: the residual structure is integrated into the feature extraction part of the network, and then the data correlation upsampling structure called Dupsampling is integrated into the upsampling part of the network to improve the quality of the feature map and prevent the gradient from disappearing. Finally, the hybrid loss function combined with cross entropy and dice is used to further improve the segmentation accuracy. The method was validated on BraTs2019 partial data, and the dice coefficients of the whole tumor, core tumor and enhanced tumor are 0.9186, 0.8419 and 0.8358 respectively, and the Hausdorff distance is 2.097, 3.672 and 2.008, which are better than most brain tumor segmentation models.