深度学习模型跨人群检测需转诊糖尿病视网膜病变的泛化性研究
本研究评估深度学习模型跨人群检测需转诊糖尿病视网膜病变(DR)的性能,并检验加入本地数据能否提升土耳其外部验证的稳健性。使用全球数据集(DDR、IDRiD、Messidor)和土耳其真实临床数据集,由三名眼科医生重新分级为需转诊/非需转诊DR,评估五种架构在三种训练-测试设置下的表现。结果显示模型在匹配条件下表现强,但跨人群下降;ConvNeXt-V2-Tiny筛查性能最稳定,支持部署前需本地验证。
要点
- 全局训练-全局测试下Swin V2 CR Small准确率0.938、ROC-AUC 0.985。
- 全局训练-本地测试性能下降,ConvNeXt-V2-Tiny最佳(准确率0.823、敏感度0.940、ROC-AUC 0.913)。
- 混合训练-混合测试中ConvNeXt-V2-Tiny最均衡(准确率0.823、敏感度0.939、NPV 0.948、ROC-AUC 0.875)。
核心结论
- 模型在匹配条件下表现强,但跨人群性能下降。
- ConvNeXt-V2-Tiny显示出最稳定的筛查性能,支持部署前需本地验证。
研究设计与样本
- 使用全球数据集(DDR、IDRiD、Messidor)和土耳其真实临床数据集,图像由三名眼科医生重新分级并二值化为需转诊/非需转诊DR。
- 评估五种架构:EfficientNet-B0、EfficientNet-V2、ResNet50、ConvNeXt-V2-Tiny和Swin V2 CR Small。
- 测试三种设置:全局训练-全局测试、全局训练-本地测试和混合训练-混合测试。
- 性能指标包括准确率、敏感度、特异度、NPV、ROC-AUC和PR-AUC;使用Grad-CAM进行可解释性分析。
主要发现
- 全局训练集包含9,881张图像;两个外部测试集各2,042张图像。
- 评分者间一致性高(ICC[3,1]=0.84;ICC[3,k]=0.96)。
- 全局训练-全局测试中Swin V2 CR Small表现最高(准确率0.938;ROC-AUC 0.985)。
- 全局训练-本地测试中性能下降,ConvNeXt-V2-Tiny表现最佳(准确率0.823;敏感度0.940;ROC-AUC 0.913)。
- 混合训练-混合测试中ConvNeXt-V2-Tiny保持最均衡(准确率0.823;敏感度0.939;NPV 0.948;ROC-AUC 0.875)。
临床怎么用
- 在部署DR深度学习筛查模型前应进行本地外部验证。
- 跨人群应用时需关注性能下降,可能需加入本地数据重新训练。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
PURPOSE: To evaluate cross-population performance of deep learning models for referable diabetic retinopathy (DR) detection and assess whether inclusion of local data improves robustness in a Turkish external validation setting. METHODS: Global datasets (DDR, IDRiD, Messidor) and a Turkish real-world clinical dataset were used. Images were re-graded by three ophthalmologists and binarized as referable/non-referable DR. Five architectures were evaluated: EfficientNet-B0, EfficientNet-V2, ResNet50, ConvNeXt-V2-Tiny, and Swin V2 CR Small. Models were tested under three settings: Global Train-Global Test, Global Train-Local Test, and Mixed Train-Mixed Test. Performance metrics included accuracy, sensitivity, specificity, NPV, ROC-AUC, and PR-AUC; Grad-CAM was used for explainability. RESULTS: The global training set included 9,881 images; two external test sets (2,042 images each) were used. Inter-grader agreement was high (ICC[3,1]=0.84; ICC[3,k]=0.96). Swin V2 CR Small achieved the highest performance in Global Train-Global Test (accuracy 0.938; ROC-AUC 0.985). Performance declined in Global Train-Local Test, where ConvNeXt-V2-Tiny performed best (accuracy 0.823; sensitivity 0.940; ROC-AUC 0.913). In Mixed Train-Mixed Test, ConvNeXt-V2-Tiny remained the most balanced model (accuracy 0.823; sensitivity 0.939; NPV 0.948; ROC-AUC 0.875). CONCLUSION: Models performed strongly under matched conditions but degraded across populations. ConvNeXt-V2-Tiny showed the most stable screening performance, supporting the need for local validation before deployment.
引用
- Engin CD, Selver MA, Köksaldı S, et al. GENERALIZABILITY OF DEEP LEARNING MODELS FOR REFERABLE DIABETIC RETINOPATHY DETECTION: A CROSS-POPULATION STUDY. Retina. 2026. PMID: 42479603. doi:10.1097/IAE.0000000000004924
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。