摘要
目的:提高人工智能模型在不同类型眼底相机拍摄图片的准确性和一致性,尤其提高模型在模型训练过程中未见机型上的糖尿病视网膜病变(Diabetic Retinopathy,DR)和年龄的识别性能,降低人工智能模型在临床应用中对相机适配要求。方法:提出一种名为单通道标准差归一化(Single Channel Standard Deviation Normalization,SCSDN)的方法,该方法可以在训练数据机型来源单一的情况下,大幅提升模型在其他未见机型上的性能表现,提升模型在不同人群和不同设备采集数据的一致性。两个独立的眼底图像识别任务被用于检验该一致性,分别是基于年龄预测和DR的诊断分类。结果:相较于传统方法,SCSDN可以将年龄预测的平均绝对误差从5.850岁降低至2.793岁,并将对应的决定系数从0.453提升到0.884。在DR的分类任务中SCSDN可以将外部验证集的AUC从0.875提升到0.938。结论:本方案可以显著提升模型在不同眼底相机图像上的年龄预测和DR分类的识别准确性,减少模型在不同眼底相机拍摄的图片中预测的偏差,提高人工智能模型在临床应用中的适用范围。
Objective:To improve the prediction accuracy and consistency of models on various types of fundus camera,especially on the camera whose images not presented in training process.Methods:This paper proposed the single channel standard deviation normalization method(SCSDN).SCSDN can achieve domain adaptation without using raw data or labels from unseen domains.Comprehensive evaluations were performed on both seen domains and unseen domains;age distribution,image styles,and prediction tasks were considered as possible influences.Two independent evaluation tasks,age prediction and diabetic retinopathy(DR)classification,are used to evaluate the performance of SCSDN.Results:Compared with ordinary normalization methods,SCSDN could decrease the mean absolute error of age prediction from 5.850 to 2.793 years,corresponding coefficient of determinant increasing from 0.453 to 0.884.On the task of DR classification,SCSDN boosts the performance on external test set from 0.875 to 0.938.Conclusion:The proposed methods could significantly improve model performance on age prediction and DR classification.The prediction discrepancy among various cameras could be minimized with SCSDN.
出处
《中国数字医学》
2021年第7期82-87,共6页
China Digital Medicine
关键词
人工智能
域泛化
眼底影像预测
鲁棒性
artificial intelligence
domain adaptation
fundus image prediction
robustness
