深度学习技术辅助门诊发药实践

Practice of Deep Learning Technology to Assist Outpatient Dispensing of Drugs

目的:研究利用深度学习技术辅助门诊发药的技术可行性,利用计算机视觉技术实现药品类别和数量的自动识别。方法:采集药品外包装图像,利用预处理技术生成训练图像集,建立7层(3C3P1F)卷积神经网络模型进行训练,部署RESTful接口规范的药品图像识别服务。药师利用嵌入药品外包装识别模块的处方发药程序采集药品图像,将其传送至药品图像识别服务,将返回的药品分类与数量结果与HIS中电子处方比对,若发现信息不一致,系统向药师提示报警。结果:通过对56种药品约47万张图像进行3000次迭代训练,训练时长12小时,预测分类准确率达到95.6%。结论:利用深度学习技术,门诊药师可借助药品外包装特征识别技术快速区分易混淆药品,及时发现药品和数量的错误信息,对降低药品错发率具有实际意义。

Objective:To study the technical feasibility of using the deep learning technology to assist outpatient dispensing of drugs,and to realize the automatic recognition of drug categories and quantities by computer vision technology.Methods:A 7-layer(3C3P1F)convolutional neural network model is constructed to train the images of drug packaging,which are preprocessed after collected.The trained model is deployed as the drug recognition service by the standard of RESTful.In the scenario of drug delivery at the drug-issuing window,system gets the drug packaging image through the module of drug packaging identification and transmits it to the drug recognition service.By comparing the drug classification and quantity result returned from the service with the electronic prescription in HIS,error messages will be highlighted to the pharmacist.Results:The model trained about 470 000 images of 56 kinds of drugs,which took about 12 hours after 3000 echoes.The accuracy of drug packaging recognition reaches 95.6%.Conclusion:By the technology of deep learning,the pharmacists can distinguish drugs easily.The pharmacists can find out the wrong drug and wrong quantity in time,which is practical meaningful for reducing risk of drug delivery errors.