基于人工智能算法STN-GResnet的肝硬化识别  

Liver Cirrhosis Recognition Based on Artificial Intelligence Algorithm STN-GResnet

作　　者：鞠维欣 赵希梅 JU Weixin;ZHAO Ximei(School of Accounting,Qingdao Vocational and Technical College of Hotel Management,Qingdao 266100,China;College of Computer Science and Technology,Qingdao University,Qingdao 266071,China;Shandong Key Laboratory of Digital Medicine and Computer Assisted Surgery,Qingdao 266000,China)

机构地区：[1]青岛酒店管理职业技术学院会计学院,山东青岛266100 [2]青岛大学计算机科学技术学院,山东青岛266071 [3]山东省数字医学与计算机辅助手术重点实验室,山东青岛266000

出　　处：《软件导刊》2023年第12期84-91,共8页Software Guide

基　　金：国家自然科学基金项目(61303079)。

摘　　要：针对传统神经网络泛化性弱、识别率低、参数量大,且特征提取质量不高等问题,提出一种新的模型结构STN-GResnet卷积神经网络。引用部分Resnet18网络结构,加入Ghostmodule模块,并结合空间变换网络,使提取到的特征具有空间不变性,增强卷积特征,同时在网络模型中加入附加角裕度损失函数(ArcFace)对Loss进行训练优化,通过优化类别的特征来增强超声肝CT影像纹理与颗粒度的差别。此外,使用迁移学习预训练模型参数,再通过数据增强来扩充超声肝CT影像中的样本集,避免因样本量不足而造成的模型过拟合现象,最终达到了95.7%的客观识别率,且模型较小,运行效率较高。Aiming at the problems of weak generalization,low recognition rate,large amount of parameters and low quality of feature extrac⁃tion of traditional neural network,a new model structure named STN-GResnet is proposed in this paper.Using part of the Resnet18 network structure and adding the Ghostmodule to enhance convolution features.Combined with the spatial transformation network,the extracted fea⁃tures are space invariant.Meanwhile,the additional angle margin loss function(ArcFace)is added to this network model to train and optimize loss.The difference of texture and granularity of ultrasonic liver CT image is enhanced by optimizing the characteristics of categories.The pre training parameters of transfer learning are used,and then the sample set in ultrasonic liver CT images is improved through data enhancement to avoid the phenomenon of model over fitting caused by insufficient sample size.Finally,the objective recognition rate of the model is 95.7%.What's more,the model is small and the operation efficiency is high.

关 键 词：深度学习 医学图像识别 迁移学习 数据增强 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]