- 期刊
利用深度學習從胸部X光中自動化鑑別新冠肺炎感染
Automatic Differentiation of COVID Infection from Chest X-ray Using Deep Learning
摘要
胸部X光為廣泛應用的第一線診斷工具，能夠快速的識別胸腔範圍可能的病徵，特別是在識別肺炎上十分有效，然而，新冠病毒與其他原因導致之肺炎在影像上的病徵仍缺乏可遵循的條件區別，為幫助臨床即早從胸部X光診斷可能的新冠肺炎，本研究利用深度神經網路建立一套鑑別輔助工具。本研究從公開資料庫取得6446例非新冠及6329例新冠肺炎胸部X光，分別使用80%及20%的影像訓練及驗證建置之殘差神經網路，並計算鑑別新冠肺炎之準確率、靈敏度、特異性、精確率以量化評估其鑑別非新冠及新冠肺炎之效能，此外，亦評估使用濾器數量對殘差神經網路鑑別效能的影響。驗證結果顯示，在僅使用25%濾器數量之下，殘差神經網路之鑑別準確率、靈敏度、特異性及精確率平均皆高於98%，若使用50%以上的濾器數量，四項指標平均可進一步提高至99%以上。本研究建置之殘差神經網路可望應用臨床上，協助從胸部X光偵測鑑別可能的新冠肺炎。
並列摘要
Chest X-ray is a widely used first-line diagnostic tool. It can quickly identify possible disease features in the thoracic cavity, especially in identifying pneumonia. However, the features of pneumonia caused by COVID and other diseases are still lacking. In order to help clinical diagnosis of possible COVID pneumonia from chest X-ray, this study uses a deep neural network to establish a differential diagnosis tool for distinguishing non-COVID and COVID pneumonia. In this study, 6,446 cases of non-COVID and 6,329 cases of COVID chest X-ray images were obtained from a public database. 80% and 20% of the images were used for training and validation of a built residual neural network, respectively. Four indices, including accuracy, sensitivity, specificity, and precision, were calculated to quantitatively evaluate the performance of the network in distinguishing between non-COVID pneumonia and COVID pneumonia from chest X-rays. In addition, the impact of the number of applied filters in the network on the performance is also evaluated. The validation results show that when only 25% of the filters are used, the average accuracy, sensitivity, specificity, and precision of the residual neural network were higher than 98%. If more than 50% of the filters are used, the four indices can be further improved to higher than 99% on average. The residual neural network built in this study might be applied clinically to help identify possible COVID pneumonia from chest X-ray detection.