- 期刊
人工智慧輔助胸部X光氣胸診斷
Artificial Intelligence Assisted Chest X-ray Pneumothorax Diagnosis
摘要
胸部X光是氣胸最常見的影像診斷方式，當氣胸量較多時，需要放置引流管進行治療。有鑑於近年來深度學習在醫學影像領域中蓬勃發展，本研究將透過遷移式學習，訓練一卷積神經網絡模型，用以自動篩檢出可能需要即時介入處置的氣胸患者。回溯納入6529例胸部X光正位照，包含氣胸993例及非氣胸5536例，年齡範圍從12歲到104歲。資料集以8:1:1分為訓練、驗證、測試子資料集。使用Inception-v4模型於驗證子資料集準確率達98.77%，測試子資料集準確率達97.78%，曲線下面積達0.99。測試子資料集中真陽性為140例、真陰性為1093例、偽陽性為8例、偽陰性為20例。精確性為0.95、召回率為0.88、F1分數為0.91。本研究顯示以本地影像資料所訓練的分類人工智慧模型，接近現有文獻水準。此模型可能作為初步篩檢或輔助診斷之工具，以協助醫師對影像做判讀，並可用於改善處置流程，達到即早通報、及時治療，增進病人安全。
並列摘要
Chest X-ray is the most common diagnostic imaging method for pneumothorax. When the volume of pneumothorax is large, a drainage tube needs to be placed for treatment. In view of the vigorous development of deep learning in the field of medical imaging in recent years, this study will use transfer learning to train a convolutional neural network model for automatical screening of pneumothorax patients who may require immediate intervention. A total of 6529 poster-anterior chest X-rays were retrospectively included, including 993 cases of pneumothorax and 5536 cases of non-pneumothorax, ranging in age from 12 to 104 years old. The data set is divided into training, validation, and testing sub-data sets as 8:1:1. Using the Inception-v4 model, the accuracy rate of the validation sub-data set is 98.77%, the accuracy rate of the testing sub-data set is 97.78%, and the area under the curve is 0.99. There were 140 true positive cases, 1093 true negative cases, 8 false positive cases, and 20 false negative cases in the testing sub-data set. The precision is 0.95, the recall is 0.88, and the F1 score is 0.91. This study shows that the classification artificial intelligence model trained with local image data is state-of-the-art. This model may be used as a tool for preliminary screening or auxiliary diagnosis to assist doctors in the interpretation of images, and can be used to improve the treatment process, achieve early notification, timely treatment, and improve patient safety.