Synthetic Medical Images for Robust, Privacy-Preserving Training of AI: Application to Retinopathy of Prematurity Diagnosis

Purpose: Developing robust artificial intelligence (AI) models for medical image analysis requires large quantities of diverse, well-curated data, which can prove challenging to collect due to privacy concerns, disease rarity, or diagnostic label quality. Collecting datasets for training diagnostic models for retinopathy of prematurity (ROP), a potentially-blinding disease, suffers from all of these challenges. Progressively-growing generative adversarial networks (PGANs) may help, as they can synthesize highly-realistic images that may increase both the size and diversity of medical datasets. Design: Diagnostic validation study of convolutional neural networks (CNNs) for plus disease detection, a component of severe ROP, using synthetic data. Participants: 5842 retinal fundus images (RFIs) collected from 963 preterm infants. Methods: Retinal vessel maps (RVMs) were segmented from RFIs. PGANs were trained to synthesize RVMs with normal, pre-plus, or plus disease vasculature. CNNs were trained, using real or synthetic RVMs, to detect plus disease from two real RVM test datasets. Main Outcome Measures: Features of real and synthetic RVMs were evaluated using Uniform Manifold Approximation and Projection (UMAP). Similarities were evaluated at the dataset and feature level using Frechet Inception Distance (FID) and Euclidean distance, respectively. CNN performance was assessed via area under the receiver operating characteristics (AUROC) curve; AUROCs were compared via bootstrapping and Delong’s test for correlated ROC curves. Confusion matrices were compared using McNemar’s 𝜒2 and Cohen’s κ. Results: The CNN trained on synthetic RVMs had a significantly higher AUROC (0.971, p = 0.006, p = 0.004) and classified plus disease more similarly to a set of eight international experts (κ = 0.922) than the CNN trained on real RVMs (AUC = 0.934, κ = 0.701). Real and synthetic RVMs overlapped, by plus disease diagnosis, on the UMAP manifold, showing that synthetic images spanned the disease severity spectrum. FID and Euclidean distances suggested that real and synthetic RVMs were more dissimilar to one another than real RVMs were to one another, further suggesting that synthetic RVMs were distinct from the training data with respect to privacy considerations. Conclusions: Synthetic medical datasets may be useful for training robust medical AI models. Furthermore, PGANs may be able to synthesize realistic data for use without protected health information concerns.研究目的：构建用于医学图像分析的鲁棒人工智能（Artificial Intelligence, AI）模型，需要大量多样化且经过精心整理的数据集，但由于隐私顾虑、疾病罕见性或诊断标签质量问题，这类数据的采集往往极具挑战。针对潜在致盲性疾病早产儿视网膜病变（Retinopathy of Prematurity, ROP）的诊断模型训练用数据集的采集，就面临上述所有难题。渐进式生成对抗网络（Progressive-growing Generative Adversarial Networks, PGANs）或可解决这一问题，因其能够合成高度逼真的医学图像，从而扩充数据集的规模与多样性。 研究设计：本研究为采用合成数据开展的诊断验证研究，旨在评估卷积神经网络（Convolutional Neural Networks, CNNs）对Plus病变（plus disease）的检测性能，而Plus病变是重症早产儿视网膜病变的核心亚型之一。 研究对象：从963名早产儿中采集的5842张视网膜眼底图像（Retinal Fundus Images, RFIs）。 研究方法：从视网膜眼底图像中分割出视网膜血管图（Retinal Vessel Maps, RVMs）；训练渐进式生成对抗网络（PGANs），使其能够合成具有正常、Pre-Plus病变或Plus病变血管形态的视网膜血管图；分别采用真实或合成的视网膜血管图训练卷积神经网络，使其能够从两个真实视网膜血管图测试数据集中检测Plus病变。 主要结局指标：采用均匀流形近似与投影（Uniform Manifold Approximation and Projection, UMAP）评估真实与合成视网膜血管图的特征；分别采用弗雷歇初始距离（Frechet Inception Distance, FID）与欧氏距离，在数据集层面与特征层面评估二者的相似性；通过受试者工作特征曲线下面积（area under the receiver operating characteristics, AUROC）评估卷积神经网络的性能；采用Bootstrap法与DeLong相关ROC曲线检验比较各组AUROC值；采用McNemar’s 𝜒2检验与Cohen’s κ系数比较混淆矩阵。 研究结果：基于合成视网膜血管图训练的卷积神经网络，其AUROC值（0.971，p=0.006，p=0.004）显著高于基于真实视网膜血管图训练的模型（AUC=0.934），且其Plus病变分类结果与8名国际专家的共识一致性更高（κ=0.922，而真实数据训练模型的κ=0.701）。根据Plus病变诊断结果，真实与合成视网膜血管图在UMAP流形上存在重叠，表明合成图像覆盖了疾病严重程度的全谱系。FID与欧氏距离分析显示，真实与合成视网膜血管图之间的相似度低于真实视网膜血管图组内的相似度，进一步说明从隐私保护角度而言，合成视网膜血管图与训练用真实数据集存在差异。 研究结论：合成医学数据集可用于训练鲁棒性更强的医学AI模型；此外，渐进式生成对抗网络能够合成无需顾虑受保护健康信息（protected health information）的逼真医学数据。