Segmentation of brain tumor MRI image based on improved attention module Unet network

In recent years, the fully convolutional network represented by Unet has been widely used in the field of medical image segmentation. However, due to the diversity of the shapes of lesions and the differences in the structures of different organs, the segmentation of lesions using only Unet structure cannot meet the requirements of accuracy and speed. Therefore, an improved Unet network for brain tumor segmentation is proposed. To reduce the number of parameters while extracting richer features and improving the accuracy of segmentation, this article introduces the inverted residuals block to replace the convolution module in the encoding and decoding stages of Unet to improve the calculation speed and accuracy; to better combine high-order semantic information with low-order semantic information, improve for the quality of detailed features in the training process, an improved Residuals Convolutional Block Attention Module is added between the encoder and the decoder. Combining the above two points of improvement, this article proposes an improved model based on Unet. Based on the Brats2019 dataset, an ablation experiment was performed on the proposed improved Unet model and compared with the TrUE-Net, ConResNet and OM-Net methods, and the Dice coefficient and Hausdorff distance were used as evaluation indicators to analyze the segmentation effect of the model. The experimental results show that the Dice coefficient of the improved Unet network model proposed in this article is 0.020-0.027 higher than other comparative models on average, and the Haushofer distance is reduced by 2.67-10.06.