Miyoshi Toru, Higaki Akinori, Kawakami Hideo, Yamaguchi Osamu
Department of Cardiology, Ehime Prefectural Imabari Hospital, Imabari, Japan.
Department of Cardiology, Pulmonology, Hypertension and Nephrology, Ehime University Graduate School of Medicine, Toon, Japan.
Open Heart. 2020 May;7(1). doi: 10.1136/openhrt-2019-001177.
Coronary angioscopy (CAS) is a useful modality to assess atherosclerotic changes, but interpretation of the images requires expert knowledge. Deep convolutional neural networks (DCNN) can be used for diagnostic prediction and image synthesis.
107 images from 47 patients, who underwent CAS in our hospital between 2014 and 2017, and 864 images, selected from 142 MEDLINE-indexed articles published between 2000 and 2019, were analysed. First, we developed a prediction model for the angioscopic findings. Next, we made a generative adversarial networks (GAN) model to simulate the CAS images. Finally, we tried to control the output images according to the angioscopic findings with conditional GAN architecture.
For both yellow colour (YC) grade and neointimal coverage (NC) grade, we could observe strong correlations between the true grades and the predicted values (YC grade, average r=0.80±0.02, p<0.001; NC grade, average r=0.73±0.02, p<0.001). The binary classification model for the red thrombus yielded 0.71±0.03 F-score and the area under the receiver operator characteristic curve was 0.91±0.02. The standard GAN model could generate realistic CAS images (average Inception score=3.57±0.06). GAN-based data augmentation improved the performance of the prediction models. In the conditional GAN model, there were significant correlations between given values and the expert's diagnosis in YC grade but not in NC grade.
DCNN is useful in both predictive and generative modelling that can help develop the diagnostic support system for CAS.
冠状动脉血管内镜检查（CAS）是评估动脉粥样硬化改变的一种有用方法，但图像解读需要专业知识。深度卷积神经网络（DCNN）可用于诊断预测和图像合成。
分析了2014年至2017年在我院接受CAS的47例患者的107张图像，以及从2000年至2019年发表的142篇被MEDLINE收录的文章中选取的864张图像。首先，我们开发了一个血管内镜检查结果的预测模型。接下来，我们制作了一个生成对抗网络（GAN）模型来模拟CAS图像。最后，我们尝试使用条件GAN架构根据血管内镜检查结果控制输出图像。
对于黄色分级（YC）和新生内膜覆盖度（NC）分级，我们都能观察到真实分级与预测值之间有很强的相关性（YC分级，平均r=0.80±0.02，p<0.001；NC分级，平均r=0.73±0.02，p<0.001）。红色血栓的二元分类模型的F值为0.71±0.03，受试者操作特征曲线下面积为0.91±0.02。标准GAN模型可以生成逼真的CAS图像（平均Inception分数=3.57±0.06）。基于GAN的数据增强提高了预测模型的性能。在条件GAN模型中，给定值与专家对YC分级的诊断之间存在显著相关性，但与NC分级无关。
DCNN在预测和生成建模中都很有用，有助于开发CAS的诊断支持系统。
