摘要
由于临床数据集的建设难度大,且数据构成、数据容量、标注标准化程度要求较高,使用测试集评价人工智能(Artificial Intelligence,AI)产品泛化能力的推广难度较大。作为标准测试集的补充,使用对抗性样本有可能提供有意义的扩展。本研究根据数学物理模型对现有数据进行变换,观测AI的变化。结果发现不同的算法在图像压缩、图像裁剪、图像滤波测试中灵敏度和特异性的表现具有明显差异。本文在模拟对抗上进行了初步的探索,将该方法运用于算法性能评价,结果表明该方法能够从新的角度反应AI产品的性能,辅助对其泛化能力的评价。
Since it is challenging to build high quality datasets with strict requirement on diversity,capacity and standardization,testing methods based on testsets are not feasible for all intended uses.As supplement,simulated data may provide useful extension to regular testsets.In this study,we performed mathematical and physical transform on current data for testing and observed the corresponding changes of AI output.Results showed that the sensitivity and specificity of different algorithms under test showed significant difference after image compression,image cropping and filtering.In this paper,simulated data was used in the evaluation of representative algorithms as an initial attempt,which demonstrated that the value of such a method could reflect the performance of AI from a new perspective,and facilitate the evaluation of generalization ability.
出处
《中国医疗设备》
2018年第12期18-21,共4页
China Medical Devices
基金
国家重点研发计划项目(2016YFC0107100)
体育总局重点课题联合中国红十字基金会燎原基金项目(2015B101)
关键词
糖网
人工智能
对抗测试
数据集
diabetic retinopathy
artificial intelligence
testing by simulated data
datasets