Automated interpretation of the coronary angioscopy with deep convolutional neural networks

Background 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. Methods 107 images from 47 patients, who underwent coronary angioscopy in our hospital between 2014 and 2017, and 864 images, selected from 142 MEDLINE-indexed articles published between 2000 and 2019, were analyzed. 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. Results For both yellow color (YC) grade and neointimal coverage (NC) grade, we could observe strong correlations between the true grades and the predicted values (YC grade, average r value = 0.80 ± 0.02, p-value <0.001; NC grade, average r value = 0.73 ± 0.02, p < 0.001). The binary classification model for the red thrombus yielded 0.71 ± 0.03 F1-score and the area under the ROC curve (AUC) 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 and NC grade. Conclusion DCNN is useful in both predictive and generative modeling that can help develop the diagnostic support system for CAS.