Classification accuracy by lesion and image type.

BackgroundAccurate interpretation of chest radiographs requires years of medical training, and many countries face a shortage of medical professionals to meet such requirements. Recent advancements in artificial intelligence (AI) have aided diagnoses; however, their performance is often limited due to data imbalance. The aim of this study was to augment imbalanced medical data using generative adversarial networks (GANs) and evaluate the clinical quality of the generated images via a multi-center visual Turing test.MethodsUsing six chest radiograph datasets, (MIMIC, CheXPert, CXR8, JSRT, VBD, and OpenI), starGAN v2 generated chest radiographs with specific pathologies. Five board-certified radiologists from three university hospitals, each with at least five years of clinical experience, evaluated the image quality through a visual Turing test. Further evaluations were performed to investigate whether GAN augmentation enhanced the convolutional neural network (CNN) classifier performances.ResultsIn terms of identifying GAN images as artificial, there was no significant difference in the sensitivity between radiologists and random guessing (result of radiologists: 147/275 (53.5%) vs result of random guessing: 137.5/275, (50%); p = .284). GAN augmentation enhanced CNN classifier performance by 11.7%.ConclusionRadiologists effectively classified chest pathologies with synthesized radiographs, suggesting that the images contained adequate clinical information. Furthermore, GAN augmentation enhanced CNN performance, providing a bypass to overcome data imbalance in medical AI training. CNN based methods rely on the amount and quality of training data; the present study showed that GAN augmentation could effectively augment training data for medical AI.

背景：胸部X光片（chest radiographs）的精准解读需历经多年医学培训，诸多国家均面临难以匹配此类需求的医疗专业人员短缺困境。近年来人工智能（artificial intelligence, AI）技术的进步为疾病诊断提供了辅助手段，但受数据不平衡问题制约，其性能往往受限。本研究旨在利用生成对抗网络（generative adversarial networks, GANs）对不平衡医疗数据进行增强，并通过多中心视觉图灵测试（visual Turing test）评估生成影像的临床质量。方法：本研究纳入6个胸部X光片数据集（MIMIC、CheXPert、CXR8、JSRT、VBD及OpenI），通过starGAN v2生成携带特定病理特征的胸部X光片。来自3所大学附属医院的5名具备至少5年临床经验的委员会认证放射科医师，通过视觉图灵测试对影像质量开展评估。此外，本研究还设置了额外评估环节，以探究GAN数据增强是否可提升卷积神经网络（convolutional neural network, CNN）分类器的性能。结果：在鉴别GAN生成影像为人工合成样本的任务中，放射科医师的识别敏感度与随机猜测无显著统计学差异（放射科医师结果：147/275（53.5%） vs 随机猜测结果：137.5/275（50%）；p=0.284）。GAN数据增强可使CNN分类器的性能提升11.7%。结论：放射科医师可有效通过合成胸部X光片对胸部病理进行分类，这表明生成影像包含充足的临床信息。此外，GAN数据增强可提升CNN的性能，为解决医疗AI训练中的数据不平衡问题提供了可行路径。基于卷积神经网络的方法依赖于训练数据的数量与质量，本研究证实GAN数据增强可有效为医疗AI扩充训练数据集。