透過您的圖書館登入
IP:3.229.124.236

摘要


Background and Purpose: Chest X-Ray (CXR) imaging is an essential first-line diagnostic tool for pneumothorax, a virtually life-threatening condition. The insertion of a chest tube to drain air is the main therapeutic method for pneumothorax. This research aimed to develop an artificial intelligence (AI) model for identifying pneumothorax on CXR. Furthermore, we created an auxiliary AI model for recognizing an inserted chest tube. Although pneumothorax can be life-threatening, the presence of the chest tube means the pneumothorax is under treatment. In the clinical scenario, a fresh pneumothorax is emergent, but pneumothorax after chest tube treatment is not. Therefore, we proposed a combined AI model to distinguish the newly onset emergent condition from the condition after treatment. Method: We trained a neural network based on MobileNet V2 to do the binary classification of whether a pneumothorax has occurred or not. The dataset was labeled by three radiologic technologists and reviewed by a radiologist. The same techniques were also applied to the AI model for the chest tubes. Results: The accuracy for pneumothorax and chest tube identification on the test dataset was 92.19% and 98.22%, respectively. The areas under the receiver operating characteristic curve were 0.9638 (pneumothorax) and 0.9968 (chest tube). The inference time of each image ranged 0.21-0.6 second. Discussions: The AI models achieved satisfactory results and can be further integrated into the Hospital Information System (HIS) in the future to assist in the early detection of emergent pneumothorax.

並列摘要


背景及目的:Chest X-Ray(CXR)影像為診斷氣胸最重要之第一線工具,氣胸會對生命造成威脅,而放置胸管來引流空氣為目前治療氣胸的主要方式。本研究之主要目的為建立AI模型來判讀CXR影像上有無氣胸,並且建立輔助的模型偵測胸管,雖氣胸會對生命造成威脅,但胸管的存在則表示已接受治療。在臨床上,影像判讀出有胸管則表示氣胸的症狀已非緊急,因此本研究提出一結合的AI模型來分辨新形成及接受治療過後的氣胸。本研究旨在使用AI(Artificial Intelligence)作為胸部影像判讀的輔助工具,使準確率提昇及品質均一化,與提昇診斷的效率。方法:本研究基於MobileNet V2的類神經網路方式將有無氣胸做二元分類,輸入資料庫是由放射師做標註並由放射科醫師審查,胸管亦是以同樣方法標註。結果:對氣胸及胸管的辨識準確度可達92.19%及98.22%,AUC值分別為0.9638及0.9968,每張影像的判讀時間更僅只有0.21至0.6秒。討論:本AI模型達到令人滿意的結果並能在未來導入醫療資訊系統,以用於協助緊急氣胸之早期偵測。

並列關鍵字

類神經網路 深度學習 胸部 X光 氣胸 胸管

延伸閱讀