透過您的圖書館登入
IP:44.200.26.112
  • 學位論文

針對腹部超音波影像利用巢式ResNet方法建立脂肪肝分類模型

Classified Fatty Livers of Abdominal Ultrasound Tomography by Using Nested ResNet Method

指導教授 : 陳泰賓
共同指導教授 : 黃詠暉(Yung-Hui Huang)

摘要


臨床診斷超音波脂肪肝影像具主觀性診斷,目前電腦輔助診斷系統(Computer-Aided Diagnosis, CAD)分析脂肪肝影像分類需要手動預處理分割並進行特徵萃取。近年利用人工智慧辨識影像類別,已證實具有合理性及前瞻性。因此本研究採用小樣本下具有良好表現之ResNet進行脂肪肝影像分類。 本研究採用回顧性分組實驗設計,蒐集2016/01/01至2017/11/30腹部B-mode超音波肝臟影像,共計416筆正常肝臟案例及311筆脂肪肝案例;脂肪肝分成三類包括輕度(n=108)、中度(n=101)、重度(n=102)。採用ResNe 演算法建立正常與非正常、輕度與非輕度(即,中度與重度脂肪肝影像)、中度與重度共三組狀模型(Nested Model);其中訓練組、驗證組及測試組各佔比為70%、20%、10%;模型評估採用測試組之分類準確度、特異性、靈敏度及Kappa一致性統計量。 測試組結果顯示正常與非正常超音波肝臟影像分類結果最佳準確度為91.7%;輕度與非輕度脂肪肝影像分類結果最佳準確度為90.3%;中度與重度脂肪肝影像分類最佳結果準確度為95%,三組ResNet分類都具有趨近90%的準確度;巢狀模型對四種分類準確度為70%;四組分類結果之Kappa一致性統計量為0.6,表示分類結果具有中度一致性。 本研究針對正常、輕度、中度與重度分類表現準確度具有合理性;未來可以探討不同人工智慧影像辨識模型如DenseNet、NASNet-Mobile、NASNet-Large等,期能提升腹部B-mode超音波脂肪肝影像預測分類可靠性及穩定性。

並列摘要


In clinics, the diagnosis of fatty liver via ultrasound images is subjectively. Meanwhile, Computer-Aided Diagnosis (CAD) was adopted to classify stages of fatty liver through manually pre-processing and extracting features of images. Now a day, the artificial intelligent with deep learning algorithms are developed to identify or classify the categories of image. In this study, the ResNet, performed with small sample, was utilized to classify the stages of fatty liver. A retrospective designed with multiple groups were enrolled in this study. The liver ultrasound images were collected from 2016/01/01 to 2017/12/31. Total of 416 and 311 cases were normal and fatty liver respectively. There were three categories of fatty liver, including mild (n=108), moderate (n=101) and severe (n=102) groups. The ResNet algorithm was used to construct a nested model which was combined three binary classification models. One was for classification between normal and abnormal. Second was for classification between mild and non-mild fatty liver. Third was for classification between moderate and severe fatty liver. The percentage of training, validation and testing set were 70%, 20% and 10% of overall samples. The classified performance was evaluated by the accuracy, specificity and sensitivity on testing set. The Kappa statistics was applied to exam the agreement of classified results with supervised diagnostic judgments. The accuracy among three binary classified models were 91.7%, 90.3%, and 95% with respectively to normal and abnormal, mild and non-mild, and moderate and severe groups. The accuracy of classified 4 groups via presented nested model was 70%. Meanwhile, the Kappa statistics provided by presented nested model was 0.6 which was moderate agreement with clinical judgments. The presented model was demonstrated feasibility for classified 4 groups. In the future, the DenseNet, NASNet-Mobile, and NASNet-Large might be adopted to create classified models for liver ultrasound tomography with expect to improve the reliability and stability.

參考文獻


[1] Adams LA, Harmsen S, St Sauver JL, Charatcharoenwitthaya P, Enders FB,Therneau T, Angulo P. Nonalcoholic fatty liver disease increases risk of death among patients with diabetes: a community-based cohort study. Am J Gastroenterol. 2010;105(7):1567-73.
[2] Adams LA, Anstee QM, Tilg H, Targher G. Non-alcoholic fatty liver disease and its relationship with cardiovascular disease and other extrahepatic diseases.Gut. 2017;66(6):1138-1153.
[3] Byra M, Styczynski G, Szmigielski C, Kalinowski P, Michałowski Ł, Paluszkiewicz R, Ziarkiewicz-Wróblewska B, Zieniewicz K, Sobieraj P, Nowicki A. Transfer learning with deep convolutional neural network for liver steatosis assessment in ultrasound images. Int J Comput Assist Radiol Surg. 2018;13(12):1895-1903
[4] Sharma P, Malik S, Sehgal S, Pruthi J. Computer aided diagnosis based on medical image processing and artificial intelligence methods. Int. J. Inf. Comput. Technol. 2013;3(9):887-892.
[5] Tang J, Agaian S, Thompson I. Guest editorial: computer-aided detection or diagnosis (CAD) systems. IEEE Syst. 2014;8(3):907-909.

延伸閱讀