基于改进U-Net模型的脑肿瘤MR图像分割

Segmentation Method of Brain Tumor MR Image Based on Improved U-Net Model

针对全卷积神经网络在医学图像分割中信息丢失、分割精度低等问题,提出了一种基于改进U-Net模型的脑肿瘤分割方法。首先使用深度残差模块替换U-Net结构中原有的卷积块,能够提取更多特征信息并防止网络退化;其次在U-Net的每个跳跃连接之间加入注意力机制,把注意力集中到对分割有用的特征,抑制冗余特征;最后采用改进的混合损失函数以缓解类不平衡的问题。使用BraTS提供的脑肿瘤MR图像数据集对改进模型进行验证,用Dice系数评估分割效果,在整体肿瘤区域、核心肿瘤区域和增强肿瘤区域的平均Dice值分别为:0.90、0.85、0.81。实验结果表明,本文提出的改进模型能够提高脑肿瘤MR图像分割精度,具有良好的分割性能。

A segmentation method of brain tumor based on improved U-Net model is proposed to solve the problem of information loss and low segmentation accuracy of full convolutional neural network in medical image segmentation.Firstly,the original convolution block in U-Net structure is replaced by deep residual module,which can extract more feature information and prevent network degradation.Secondly,an attention mechanism is added between each skip connection of U-Net to focus attention on features useful for segmentation and inhibit redundant features.Finally,the problem of class imbalance is alleviated by an improved hybrid loss function.The improved model is verified by the brain tumor MR image data set provided by BraTS(The Brain Tumor Image Segmentation Challenge),and the segmentation effect is evaluated by Dice coefficient.The average Dice value in the overall tumor area,core tumor area,and enhanced tumor area is 0.90,0.85 and 0.81,respectively.The experimental results show that the proposed model with its good segmentation performance is capable of improving the segmentation accuracy of brain tumor MR image.