一种基于深度卷积神经网络的有效识别肝脏超声标准切面的AI模型

An Artificial Intelligence-assisted Recognition Model Based on a Deep CNN for Liver Ultrasound in Standard Plane

目的:结合人工智能技术与超声影像,建立一个有效的识别模型以辅助识别肝脏超声标准切面(LUSP)图像。方法:采集左肝胃底纵切面、左肝-腹主动脉纵切面、肝-下腔静脉纵切面、肝胆纵切面、肝肾纵切面、第二肝门水平高位横切面、第一肝门水平中位横切面、肝胰水平低位横切面、第二肝门高位斜横切面、胆肾水平低位斜横切面、第一肝门中位斜横切面、第一肝门门脉长轴切面、第6-7肋间斜纵切面等13个LUSP的超声图像共14971张,其中11980张用于构建深度卷积神经网络模型(DeepCNN),2991张用于模型验证。以3名长期从事肝脏超声检查及诊断的专家判断一致的肝脏超声标准切面图像作为模型识别的金标准。同时对本模型与VGG16模型识别LUSP的效能进行比较。结果:(1)DeepCNN模型在识别不同LUSP的准确率为0.892。(2)本模型与VGG16实验性能相近(P0.05),但所识别的切面类型更多。结论:DeepCNN模型能有效地分类不同的LUSP图像,对辅助超声医生识别LUSP和进行肝脏超声诊断具有较高价值。

Purpose:To establish an effective recognition model to assist in the recognition of liver ultrasound standard plane(LUSP)images by combing artificial intelligence technology and ultrasound images.Methods:Thirteen sections(14971 images in total)were collected from patients,including left liver and stomach fundus longitudinal sections,left liver-abdominal aorta longitudinal sections,subhepatic vena cava longitudinal sections,liver and gallbladder longitudinal sections,liver and kidney longitudinal sections,high transverse sections at the level of the second porta hepatis,median transverse sections at the level of the first porta hepatis,low transverse sections at the level of the hepatopancreas,high oblique transverse sections of the second porta hepatis,median oblique transverse sections of the first porta hepatis,low oblique cross sections at the level of gallbladder and kidney,longaxis views of the first hepatic portal vein,and 6th and 7th intercostal oblique longitudinal sections.11980 images were used for deep convolutional neural network(Deep CNN)model building and 2991 images were used for model validation.The LUSP images judged by three long-term experts in liver ultrasound examination and diagnosis were used as the standard to compare against.The recognition of LUSP between this model and the VGG16 model was compared.Results:(1)The Deep CNN model achieved an accuracy of 0.892 in identifying different liver ultrasound standard sections.(2)The Deep CNN model showed similar overall experimental performance to the VGG16 model,while more types of sections could be identified(P>0.05).Conclusion:The Deep CNN model can be used to effectively classify different LUSP images,and has high value for ultrasound physicians to identify LUSP and assist indiagnosis of liverdiseases.