基于深度学习的住院部口服药分类模型的构建  

Construction of classification model for oral medication in inpatient departments based on deep learning

作　　者：王茜玉 李南欣 向凡 李杨[2] 唐良友 向军莲[2] 张俊然[1] WANG Xiyu;LI Nanxin;XIANG Fan;LI Yang;TANG Liangyou;XIANG Junlian;ZHANG Junran(College of Electrical Engineering,Sichuan University,Sichuan 610065 China)

机构地区：[1]四川大学电气工程学院,四川610065 [2]德阳市人民医院

出　　处：《护理研究》2024年第6期948-954,共7页Chinese Nursing Research

基　　金：国家自然科学基金“数学与医疗健康交叉”重点专项项目,编号:12126606;四川省科技计划项目,编号:2023YFS0201;德阳科技(揭榜)项目,编号:2021JBJZ007;琶洲实验室(黄埔)研发项目,编号:2023K0605。

摘　　要：目的:基于深度学习法构建针对住院部的口服药分类模型。方法:模拟实际应用场景,采集95类药丸图片构建数据集,并对其进行图片预处理操作;以MobileNet V2网络为基础架构建立药丸分类模型,并在模型中嵌入注意力机制以增强网络特征通道间的依赖关系;融合迁移学习的方法,利用自建药丸数据集对模型进行训练和测试,通过模型分类准确率和模型参数量指标检测模型性能。结果:本研究构建的模型在自然环境中采集的口服药丸图片分类方面表现卓越,通过使用包含95类药丸、总计728张图片的自建数据集进行训练和测试,模型分类准确率为95.8%,分别比MobileNet V2、ShuffleNet V2、ResNet50高11.6%、14.3%、11.3%。模型参数量为2.55 M,约为ResNet50的1/10。结论:本研究构建的模型可以较好地平衡模型的复杂度和分类准确率,为药房等场景下涉及的药丸自动分类系统提供技术路线和效果验证,对于提升药房发药、病房分药等具体情形的护理自动化水平具有一定的理论和实际应用价值。Objective:To construct an oral medication classification model for inpatient pharmacies based on deep learning.Methods:Actual scenes were simulated,95 types of pill picture were collected to construct dataset,and pictures in dataset were preprocessed.Classification model for pills was constructed based on MobileNet V2 network,and Squeeze-and‐excitation networks were embeded in the model to enhance the network′s feature channel dependencies.Method of transfer learning was applied,using the pill dataset built by researchers to train and test the model,and the performance of the model was tested through the classification accuracy and the parameter quantity of the model.Results:The constructed model in this study demonstrates outstanding performance in classifying oral pill images collected in natural environments.Trained and tested on a self‐built dataset comprising 95 categories and a total of 728 images,the accuracy of model classification was 95.8%.This outperforms MobileNet V2,ShuffleNet V2,and ResNet50 by 11.6%,14.3%,and 11.3%,respectively.The amount of model parameters was 2.55 M,which was approximately 1 out of 10 of ResNet50.Conclusions:The model constructed in this paper better balances the complexity and classification accuracy of the model,and provides a technical route and effect verification for the automatic pill classification system involved in pharmacies and other scenarios.It has certain theoretical and practical application value for improving the level of nursing automation in specific situations such as pharmacy dispensing and ward drug distribution.

关 键 词：药房 口服药 图像处理 分类模型 深度学习 MobileNet V2网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] R95[医药卫生—药学]