摘要
糖尿病视网膜病变是进展性可致盲眼病,近年出现了基于眼底照片的人工智能糖网软件,但尚无针对性的安全有效性评价方法,本文在现有深度学习算法监管要求框架基础上,针对算法性能和泛化能力,提出了安全有效性的评价方法,主要包括训练数据质控、人机配合、算法性能评估、性能影响因素评估、注意事项说明、网络安全6方面内容。评价方法有效指导相关产品上市申报,并对类似产品的评价起到借鉴作用。
Diabetic retinopathy is a progressive blinding eye disease.In recent years,artificial intelligence diabetic retinopathy diagnostic software has been developed.However,there is no corresponding security effectiveness evaluation method,based on the existing regulatory framework of deep learning algorithm,this paper proposed evaluation method for the safety and effectiveness of the algorithm performance and generalization ability,including quality control on training data,human-computer cooperation,algorithm performance evaluation,performance influencing factor evaluation,notes for attention,and cyber security,warning and attention.The evaluation method can effectively guide the registration work of related products and provide reference for similar products.
作者
刘枭寅
梁宏
郭兆君
贺伟罡
LIU Xiaoyin;LIANG Hong;GUO Zhaojun;HE Weigang(Center for Medical Device Evaluation,National Medical Products Administration,Beijing 100081,China;Yangtze River Delta Region Center for Medical Device Evaluation and Inspection,National Medical Products Administration,Shanghai 201203,China)
出处
《中国医疗设备》
2021年第12期22-24,34,共4页
China Medical Devices
基金
科技创新2030-“新一代人工智能”重大项目(2020AAA0105000)。
关键词
人工智能辅助诊断软件
人机配合
医疗器械技术审评
安全有效性评价
糖尿病视网膜病变
artificial intelligence aided diagnosis software
human-computer cooperation
technical evaluation for medical device
evaluation for safety and effectiveness
diabetic retinopathy
