透過您的圖書館登入
IP:18.118.23.89
  • 學位論文

基於 YOLO 的肝臟超音波影像病灶偵測系統

Lesion Detection System Based on YOLO in Ultrasound Image of Livers

指導教授 : 黃健興

摘要


在本論文中,我們提出了一個基於 YOLOv4 的醫學影像病灶偵測系統,將影像標記、格式轉換、參數設定、模型訓練與影像驗證等動作標準化。 機器學習最重要的是數據收集和標記,我們與義大大昌醫院合作,通過LabelMe 標記肝臟超音波影像上的病徵,自動將數據轉換為 YOLOv4 所使用的訓練、驗證和測試格式,設計一個循環機制來重複訓練模型。並使用邏輯表示式合併不同模型的辨識結果,進行病理上的語意驗證,如腫瘤就會被解釋為「出現於肝臟上的腫瘤」,提供更人性化的驗證方式。 透過此系統可進行單張影像、多重影像與連續影像的資料驗證。應用於肝臟超音波影像偵測腫瘤、肝血管瘤、射頻燒灼術、膿瘍與轉移性腫瘤,準確率可達95%。

並列摘要


In this thesis, A medical imaging lesion detection system based on YOLOv4 is designed. Which regulates the actions of image labeling, format conversion, parameter setting, model training and image verification. The most important thing in machine learning is data collection and labeling. We collaborated with E-Da Dachang Hospital to acquire and label lesions on liver ultrasound images through LabelMe. Our pipeline then automatically converts this data into a format suitable for train, validation and test in YOLOv4. Finally, we design a supervision andmaintenance mechanisms to reweight the model. At the same time, A logical expressions mechanism is proposed for pathological semantic verification, such as HCC would be regarded as "HCC appear on the Liver". Match humane verification method. Single, multiple and continuous image data verification for tumors, hepatic hemangioma, radiofrequency cautery, abscesses and metastatic, with an accuracy rate of about 95%.

參考文獻


[1] Jung Hee Son, Sang Hyun Choi , So Yeon Kim, Hye Young Jang, Jae Ho Byun, Hyung Jin Won, So Jung Lee, & Young Suk Lim (2019). Validation of US Liver Imaging Reporting and Data System Version 2017 in Patients at High Risk for Hepatocellular Carcinoma. Radioligy, 292(2). https://pubs.rsna.org/doi/10.1148/radiol.2019190035
[2] 肝癌的分期-肝癌懶人包 3 URL: https://www.cmuh.cmu.edu.tw/NewsInfo/NewsArticle?no=7283
[3] 血管瘤-肝臟最常見的良性腫瘤 URL: https://www.mmh.org.tw/gi/index4_3_9.html
[4] 癌症 URL:http://official.kfsyscc.org/cancer/liver-cancer/therapy/rf-ablation/
[5] 肝膿腫 URL: http://cht.a-hospital.com/w/肝脓肿

延伸閱讀