An automated segmentation model based on CBAM for MR image of glioma tumors

Abstract

As a serious disease endangering human life, the incidence of glioma is increasing in recent years. A semantic segmentation model of glioma based on the deep separable convolution of the attention mechanism is proposed. The model uses an encoder-decoder structure, where the encoder part uses an improved Xception backbone network. In the improved Xception backbone network, CBAM is added after each convolutional layer, thereby improving the segmentation accuracy. In the entire network structure, the Mish activation function is used instead of the ReLU activation function to ensure a smooth gradient descent during training and optimize network performance. The segmentation results of magnetic resonance image slices obtained based on the BraTS2019 data set show that the joint intersection is 83.68%, the Kappa coefficient is 90.74%, and the Dice coefficient is 0.9111, which is better than mainstream semantic segmentation models. The semantic segmentation model proposed in this paper has a high accuracy rate for glioma segmentation. This work can effectively alleviate the complex recognition work of doctors on tumors, and is of practical significance to the medical diagnosis process.