使用谷歌云Vertex人工智能对视网膜疾病的光学相干断层扫描图像进行自动化机器学习分类
本研究利用公开的1965张来自759例患者的光学相干扫描图像,评估谷歌Vertex AI自动化机器学习在区分年龄相关性黄斑变性、糖尿病黄斑水肿、视网膜前膜、视网膜静脉阻塞和健康对照中的诊断性能。模型总体AUPRC为0.991,灵敏度95.9%、特异度96.9%、阳性预测值95.9%,其中年龄相关性黄斑变性分类表现最佳(AUPRC=0.999)。
要点
- 自动化机器学习模型对五种视网膜疾病及健康对照的分类总体AUPRC达0.991。
- 年龄相关性黄斑变性分类表现最佳,AUPRC=0.999,精确率98.4%,召回率99.2%。
- 新生血管性年龄相关性黄斑变性AUPRC=0.963,优于非新生血管性的0.915。
核心结论
- 自动化机器学习模型能准确分类视网膜疾病的光学相干扫描图像,性能与传统机器学习方法相当或更优。
- 其用户友好设计支持可扩展的人工智能驱动临床整合。
研究设计与样本
- 使用公开验证的光学相干扫描数据集,包含来自759例患者的1965张去标识化图像。
- 图像标记后上传至Vertex AI,训练单标签分类模型,按80%-10%-10%划分训练、验证和测试集。
- 诊断指标包括精确率-召回率曲线下面积(AUPRC)、灵敏度、特异度、阳性预测值和阴性预测值。
- 亚组分析比较新生血管性与非新生血管性年龄相关性黄斑变性。
主要发现
- 模型总体AUPRC=0.991,灵敏度95.9%,特异度96.9%,阳性预测值95.9%。
- 年龄相关性黄斑变性分类AUPRC=0.999,精确率98.4%,召回率99.2%。
- 视网膜前膜AUPRC=0.978,精确率92.9%,召回率86.7%。
- 糖尿病黄斑水肿AUPRC=0.895,精确率81.3%,召回率86.7%。
- 视网膜静脉阻塞召回率80%,精确率100%;新生血管性年龄相关性黄斑变性AUPRC=0.963,非新生血管性为0.915。
关键数值与推荐
- 年龄相关性黄斑变性分类AUPRC=0.999,精确率98.4%,召回率99.2%。
- 视网膜前膜AUPRC=0.978,精确率92.9%,召回率86.7%。
- 糖尿病黄斑水肿AUPRC=0.895,精确率81.3%,召回率86.7%。
- 视网膜静脉阻塞精确率100%,召回率80%。
- 新生血管性年龄相关性黄斑变性AUPRC=0.963,非新生血管性AUPRC=0.915。
临床怎么用
- 原文未据此给出推荐。
局限性
- 原文未报告局限性。
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: Automated machine learning is an artificial intelligence tool that streamlines image recognition model development. This study evaluates the diagnostic performance of Google Vertex AI automated machine learning in differentiating age-related macular degeneration (AMD), diabetic macular edema, epiretinal membrane, retinal vein occlusion, and healthy controls using optical coherence tomography images. METHODS: A publicly available, validated optical coherence tomography data set of 1965 deidentified images from 759 patients was used. Images were labeled and uploaded to Vertex AI. A single-label classification model was trained, validated, and tested using an 80%-10%-10% split. Diagnostic metrics included area under the precision-recall curve (AUPRC), sensitivity, specificity, and positive and negative predictive value. A subanalysis evaluated neovascular versus nonneovascular AMD. RESULTS: The automated machine learning model achieved high accuracy (AUPRC = 0.991), with sensitivity, specificity, and PPV of 95.9%, 96.9%, and 95.9%, respectively. AMD classification performed best (AUPRC = 0.999, precision = 98.4%, recall = 99.2%). Epiretinal membrane (AUPRC = 0.978, precision = 92.9%, recall = 86.7%) and diabetic macular edema (AUPRC = 0.895, precision = 81.3%, recall = 86.7%) followed. Retinal vein occlusion recall was 80% despite 100% precision. Neovascular AMD outperformed nonneovascular AMD (AUPRC = 0.963 vs. 0.915). CONCLUSION: Our automated machine learning model accurately classifies optical coherence tomography images of retinal conditions, demonstrating performance comparable or superior to traditional ML methods. Its user-friendly design supports scalable AI-driven clinical integration.
引用
- Sina EM, Pena J, Zafar S, et al. AUTOMATED MACHINE LEARNING CLASSIFICATION OF OPTICAL COHERENCE TOMOGRAPHY IMAGES OF RETINAL CONDITIONS USING GOOGLE CLOUD VERTEX ARTIFICIAL INTELLIGENCE. Retina. 2025. PMID: 40523306. doi:10.1097/IAE.0000000000004555
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。