利用卷積神經網路輔助影像診斷- 急性會厭炎Ｘ光之急診應用

Taipei Medical University Institutional Repository > 醫學院 > 人工智慧醫療碩士在職專班 > 博碩士論文 > Item 987654321/61658

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://libir.tmu.edu.tw/handle/987654321/61658

题名:	利用卷積神經網路輔助影像診斷- 急性會厭炎Ｘ光之急診應用 Using Convolutional Neural Network in neck-lateral-radiograph to diagnose acute epiglottitis in emergency department
作者:	林暘策 LIN, YANG-TSE
贡献者:	醫學院人工智慧醫療碩士在職專班謝邦昌康峻宏
关键词:	急性會厭炎;卷積神經網路;人工智慧;早期診斷;頸部X光;醫學影像;轉移式學習;急診醫學;急診室;誤診 emergency room;acute epiglottitis;early diagnosis;misdiagnosis;convolutional neural network;artificial intelligence;transfer learning;neck-lateral-radiograph;medical image;neck x-ray
日期:	2021-06-24
上传时间:	2022-04-28 22:59:05 (UTC+8)
摘要:	急性會厭炎是一種隨時危及生命的疾病，臨床上若懷疑病人罹患急性會厭炎時，就必須積極治療並密切觀察呼吸道是否阻塞，若延誤診斷則會造成死亡等嚴重後果；而在急診室，最常使用的非侵入性影像工具就是由頸部Ｘ光影像(neck-lateral-radiograph)來判讀會厭附近構造是否腫脹。目前已有許多研究使用預先訓練完成的卷積神經網路(CNN pretrained model) 作轉移式學習(transfer learning)，並應用在醫療影像上的經驗，因此本研究想透過卷積神經網路轉移式學習，提出一個影像診斷模型來協助判讀頸部Ｘ光：可以達到快速診斷的目的，提供一種客觀的診斷標準。本研究蒐集病例源自台灣北部兩家醫療院所，實驗組來自於病例回顧，年齡大於18歲之成年人，由急診入院且出院診斷為急性會厭炎，就診期間經內視鏡( fiberoptic laryngoscopy)確診為會厭炎。對照組來自於年齡大於18歲之成年人之門診頸部Ｘ光，Ｘ光正式報告判讀無急性會厭炎，且病例紀錄沒有急性會厭炎之臨床症狀，將這些蒐集到的影像，輸入不同的卷積神經網路作轉移式學習，最後將資料集進行交叉驗證(5-fold-cross-validation)，得到最後模型對測試資料集之預測成果。實驗組：蒐集急性會厭炎之病人頸部Ｘ光影像共251位，男性佔62.2%，女性佔37.8%，病人平均年齡為46.2±14.4；而對照組之病人頸部Ｘ光影像共938位，男性佔63%，女性佔37%，病人平均年齡為46±15；本研究最後選用InceptionV3作為卷積神經網路模型，並經過交叉驗證後得到的表現：準確度為0.920，F-score為0.918，AUC為0.965。本研究得到的結果顯示，以卷積神經網路作為模型對於急性會厭炎的Ｘ光具有良好的診斷能力。透過卷積神經網路來提升Ｘ光對於急性會厭炎之診斷能力，提供了一項新的非侵入性，準確，且快速的診斷工具；將此輔助診斷模型導入急診醫師的影像診斷流程中，在未來可以急診室系統上，輔助臨床醫師，幫助影像診斷致命的急症，降低診斷錯誤並且加速診斷流程。 Background: Acute epiglottitis is a life-threatening disease at all times. If the patient is suspected of having acute epiglottitis, delay in diagnosis and treatment will cause serious consequences such as respiratory obstruction or even death. The most commonly used non-invasive imaging method to diagnose acute epiglottitis in the emergency room is a neck-lateral-radiograph. Currently, many studies have used pre-trained convolutional neural networks (CNN) to perform transfer learning and assist the analysis of medical images. This research aimed to develop a diagnostic model using CNN-based transfer learning to propose an imaging diagnostic model to assist in the interpretation of neck-lateral-radiograph. This diagnostic model can achieve the purpose of rapid diagnosis of acute epiglottitis and provide an objective diagnostic standard of neck-lateral-radiograph as well. Methods: The cases collected in this study came from two medical institutions in northern Taiwan. The acute epiglottitis came from cases review and met the following conditions: Adults older than 18 years old were admitted to the emergency department and discharged from the hospital with a diagnosis of acute epiglottitis. The diagnosis of acute epiglottitis had been confirmed by fiberoptic laryngoscopy during treatment. The control group came from outpatient neck X-rays of adults over 18 years of age. The official X-ray report judged that there was no acute epiglottitis, and the medical records did not have clinical symptoms of acute epiglottitis. The collected images were input into different pre-trained convolutional neural networks, and finally obtain the prediction results of the final model on the test data set trough cross-validated (5-fold-cross-validation). Results: Overall, 251 lateral neck radiographs of patients with acute epiglottitis and 936 healthy individuals were collected. Inception V3 was used as the pertained model. Based on the average value of the k-fold cross-validation (k=5), the following performance metrics were obtained: accuracy=0.920, F-score=0.918, and area under the curve=0.965. Conclusion: As shown by the results of the five-fold cross-validation performed on the pre-pertained model (Inception V3), using CNN as the pre-trained model can provide a good diagnostic performance of acute epiglottitis based on radiographic images. Analysis of radiographic images using CNN provides a novel non-invasive, accurate, and fast diagnostic method for acute epiglottitis. In the future, this CNN-based diagnosis model can be introduced into the imaging diagnosis process of emergency physicians, and assist the clinicians in imaging diagnosis of fatal emergencies, reduce diagnosis errors and speed up the diagnosis process.
描述:	碩士指導教授：康峻宏指導教授：謝邦昌委員：彭徐鈞委員：黎阮國慶委員：蘇家玉
数据类型:	thesis
显示于类别:	[人工智慧醫療碩士在職專班] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	179	检视/开启

检视Licence

在TMUIR中所有的数据项都受到原著作权保护.

TAIR相关文章

数据加载中.....