| 摘要: | 緒論
冠狀動脈疾病（coronary artery disease, CAD）是全球主要的死亡原因之一。本研究旨在研究無症狀健康篩檢個案，比較六種類胸腔X光影像，透過卷積神經網絡（convolutional neural network, CNN）二分類預測有無冠狀動脈鈣化。
材料與方法
本研究通過新光財團法人新光吳火獅紀念醫院人體試驗委員會審查，回溯性蒐集2017年11月1日至2022年12月31日期間接受健檢的無症狀成人個案。收集其胸腔X光正面照、胸腔X光側面照和電腦斷層冠狀動脈鈣化檢查影像。經影像前處理和資料增強後，分別以六種類影像透過預訓練VGG19、ResNet101和InceptionV3模型，進行訓練和5折交叉驗證，以最佳模型進行外部測試。以Grad-CAM（gradient-weighted class activation mapping）視覺化方式分析熱圖。比較人工智慧與專科醫師預測的結果。
結果
六種類影像輸入類型，均有分類預測有無冠狀動脈鈣化之能力，但未達顯著統計差異。其中，胸腔X光正面照經由VGG19網絡預測，在外部測試集，操作者特徵曲線下面積達0.858，正確率達0.786，且特異性達0.903。
結論
在亞裔無症狀健康篩檢族群，通過卷積神經網絡，常規胸腔X光影像，通過卷積神經網絡，具有預測個案冠狀動脈鈣化之能力。影像本身或可做為預測因子之一，且由於高特異性，現階段該模型可能做為初步篩檢，避免非必要電腦斷層冠狀動脈鈣化檢查以減少輻射劑量暴露。
Introduction
Coronary artery disease (CAD) is one of the leading causes of death worldwide. This study aims to examine asymptomatic individuals undergoing health screenings and compare six types of chest X-ray images to predict the presence of coronary artery calcification (CAC) using convolutional neural networks (CNNs).
Materials and Methods
This study was approved by the Institutional Review Board of Shin Kong Wu Ho-Su Memorial Hospital. Asymptomatic adults who underwent health screenings from Nov. 1, 2017, to Dec. 31, 2022, were retrospectively collected. Chest X-ray poster-anterior (PA) and lateral views and computed tomography (CT) images for CAC assessment were collected. After image preprocessing and data augmentation, the six types of images were used to train and validate the models using pre-trained VGG19, ResNet101, and InceptionV3 networks with 5-fold cross-validation. The best-performing model was tested on external dataset. Grad-CAM (gradient-weighted class activation mapping) was used to visualize the heatmaps. Comparison of results between artificial intelligence and radiologist.
Results
All six types of image inputs have the ability to classify and predict the presence of coronary artery calcification, but none reached statistical significance. Among them, the chest X-ray PA view, predicted by the VGG19 network, achieved an area under the receiver operating characteristic curve (AUC) of 0.858, an accuracy of 0.786, and a specificity of 0.903 in the external test set.
Conclusion
In asymptomatic Asian health screening populations, routine chest X-ray images, analyzed through CNN, have the ability to predict coronary artery calcification. The images themselves could serve as one of the predictive factors. Due to the high specificity, this model could currently be used for initial screening to avoid unnecessary CT coronary artery calcification scans, thereby reducing radiation exposure. |
