基于注意力机制的三维肿瘤图像分割算法  

3D Tumor Image Segmentation Algorithm Based on Attention Mechanism

作　　者：刘晓丽 程晓荣[1] LIU Xiaoli;CHENG Xiaorong(School of Control and Computer Engineering,North China Electric Power University,Baoding 071003)

机构地区：[1]华北电力大学控制与计算机工程学院,保定071003

出　　处：《计算机与数字工程》2025年第1期228-233,239,共7页Computer & Digital Engineering

基　　金：中央高校基本科研业务费专项资金(编号:2020MS122)资助。

摘　　要：脑肿瘤严重危害人类身心健康,从核磁共振图像中自动分割出三肿瘤区域可以有效辅助医生诊疗。针对2D系列网络只关注到局部信息,忽略了不同模态空间一致性问题,提出了3D端到端的U型网络分割方法。首先,为充分利用多尺度特征信息,在编码器部分加入通道空间混合域注意力机制以实现特征提取增强;其次,通过改进上采样和下采样操作机制以防止信息丢失;最后,为增强网络训练,解码器部分引入了深度监督。基于BraTS2021数据集的实验结果表明,Dice相似系数、灵敏度和HD95距离分别达到了85.92%,92.04%和17.47,在多模态肿瘤分割和边缘轮廓方面都有较好的准确性。Automatic segmentation of brain tumor regions from multiple modes of MRI can effectively assist doctors in diagno⁃sis and treatment.A 3D end-to-end U-shaped network segmentation method is proposed to solve the problem that the 2D series net⁃work only focuses on local information and ignores the spatial consistency of different modalities.Firstly,a channel space mixed do⁃main attention mechanism is added in the encoder to realize feature extraction enhancement.Secondly,the operation mechanism of up-sampling and down-sampling is improved to prevent information loss.Finally,in order to fuse global information,balance the loss of different levels and speed up network training,depth supervision is introduced to the decoder part.The experimental results based on BraTS2021 data set show that Dice's similarity coefficient,sensitivity and HD95 distance reached 85.92%,92.04%and 17.47,respectively,which shows good accuracy in multi-modal tumor segmentation and edge contour.

关 键 词：注意力机制 3D Unet网络 核磁共振图像 深度监督算法 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]