基于多尺度卷积输入和ConvCRFs的非对称U-Net脑肿瘤MRI图像分割

Asymmetric U-Net Segmentation Algorithm of Brain Tumor MR Images Based on Multi-scale Input and ConvCRFs

针对磁共振成像(Magnetic Resonance Imaging,MRI)进行脑胶质瘤病灶边界分割的问题,提出基于多尺度卷积输入和卷积条件随机场(ConvCRFs)的非对称U-Net脑肿瘤MRI图像分割算法。首先,设计了多尺度卷积输入模块作为预处理步骤,以丰富全局上下文语义信息的提取与输入;其次,采用非对称U-Net网络结合ConvCRFs,对分割结果进行判别微调,从而提高肿瘤的分割准确率;最后,为了验证算法的可行性,在Brats2020数据集上进行了实验。实验结果表明,Dice系数达到0.887,表明对脑胶质瘤图像分割算法具有重要的临床引导价值。

In order to solve the problem of Magnetic Resonance Imaging(MRI)for brain glioma focus boundary segmentation,an asymmetric U-Net brain tumor MRI image segmentation algorithm based on multi-scale convolution input and convolution conditional random fields(ConvCRFs)is proposed.First,a multi-scale convolution input module is designed as a preprocessing step to enrich the extraction and input of global context semantic information;Secondly,asymmetric U-Net network combined with ConvCRFs is used to fine-tune the segmentation results to improve the accuracy of tumor segmentation;Finally,in order to verify the feasibility of the algorithm,experiments were carried out on the Brats2020 dataset.The experimental results show that the Dice coefficient reaches 0.887,indicating that it has important clinical guidance value for brain glioma image segmentation algorithm.