基于多期相注意力融合网络的肝脏病灶CT影像分类研究

Research on CT Images Classification of Liver Lesions Based on Multi-phase Attention Fusion Network

肝脏病灶是指肝脏疾病集中的部位或是综合病症、感染的主要部位。由于不同类型的多期相肝脏病灶计算机断层扫描(CT)影像存在异病同影或同病异影的情况,导致同一类型的CT影像结构变化较大,传统方法难以提取丰富的图像特征信息,肝脏病灶分类准确率有待提高。提出一种多期相注意力融合网络MAFNet,使用单期相分支表征单期相病灶图像特征,并在融合分支中采用中期融合的方式,融合单期相分支中提取出的特征映射,从而充分提取图像中不同层次的特征。利用多期相注意力模块提取单期相分支中肝脏病灶的加权特征,重新组织多期相肝脏病灶的特征映射,以保持不同单期相图像信息,表达3个期相影像的时序增强模式,得到更准确的分类结果。实验结果表明,基于该网络的分类方法能充分利用多期相肝脏CT影像的时序特征,有效捕捉同一患者不同期相的信息,肝脏病灶CT影像分类的平均准确率为90.99%。

Liver lesions refer to the concentrated parts of liver diseases or main parts of comprehensive diseases and infections.Owing to different types of multi-phase liver lesions with different diseases or different Computed Tomography(CT)images of the same disease,the structure of the same CT images changes significantly.It becomes difficult to extract rich image feature information using traditional methods,and the accuracy to classify liver lesions needs to be improved.Therefore,in this study,a Multi-phase Attention Fusion Network(MAFNet)is proposed,it uses monophasic branches to represent the features of monophasic focus images.In the fusion branch,a medium-term fusion method is used to fuse the feature maps extracted from monophasic branches to fully extract the features of different levels from the images.The Multi-phase Attention Module(MAM)is used to extract the weighted features of liver lesions in monophasic branches,reorganize the feature mapping of the multi-phase liver lesions,store the information of the monophasic images,express the temporal enhancement mode of the three-phase images,and obtain more accurate classification results.The experimental results show that the classification method based on the proposed network can fully utilize the time sequence characteristics of multi-phase liver CT images and effectively store the information of different phases of the same patient;moreover,the accuracy of the CT image classification of liver lesions is 90.99%.