基于U-Net改进模型的多模态脑肿瘤分割方法

Multi-model Brain Tumor Segmentation Method Based on Improved U-Net Model

诊断脑肿瘤时,如果能从多种模态的核磁共振成像(MRI)图像中精准分割出脑肿瘤区域,将有助于医生快速和准确的诊断。针对分割脑肿瘤时出现的边界分割不精准问题,该文提出了一种基于U-Net改进模型的脑肿瘤分割方法。该方法将U-Net每级编码器的特征图保留,来捕获分割目标的边界细节信息,进而对保留的特征图采用自注意力模块计算通道级别注意力,加强分割目标的边界空间信息提取,最后使用尺度融合模块统一特征图的尺度和通道数,来融合分割目标的边界信息,作为解码器的输入,从而提高分割性能。该方法在BRATS2017数据集上进行训练和测试,在Dice、SE、SP三个评估指标的参考下,通过消融实验证明了融合多尺度模块和自注意力机制的有效性,并与U-Net、ResUNet、SGNet、RelayNet四种网络模型进行对比实验,表明由于融合了分割目标边界的细节和空间信息,该模型得到的分割区域更加接近真实脑肿瘤区域。

When diagnosing brain tumors,if the brain tumor area can be accurately segmented from multiple modal MRI images,it will help doctors make a quick and accurate diagnosis.Aiming at the problem of inaccurate boundary segmentation when segmenting brain tumors,we propose a brain tumor segmentation method based on U-Net model.This method retains the feature map of each level of U-Net encoder to capture the boundary detail information of the segmentation target,uses the self-attention module to calculate the channel-level attention of the retained feature map,and strengthens the boundary space information extraction of the segmentation target.Finally,the scale fusion module is used to unify the scale and the number of channels of the feature map to fuse the boundary information of the segmentation target as the input of the decoder,thereby improving the segmentation performance.The proposed method is trained and tested on the BRATS2017 data set.Under the reference of the three evaluation indicators of Dice,SE and SP,the effectiveness of the fusion of multi-scale modules and the self-attention mechanism is proved through ablation experiments.Compared with the four network models of U-Net,ResUNet,SGNet and RelayNet,the segmentation area obtained by the model is closer to the real brain tumor area,due to the integration of the details and spatial information of the segmentation target boundary.