應用機器學習法對無顯影電腦斷層肝影像進行分類之研究

To Study The Extracted Classification Model From Non-Enhanced Computed Tomography Liver Image By Machine Learning

田昀靈(Yun-Ling Tien)

指導教授 ： 黃詠暉

共同指導教授 ： 陳泰賓(Tai-Been Chen)

義守大學/電機資訊學院/資訊工程學系/碩士(2018年)

摘要

前言：肝癌是十大癌症死因排行第二。利用肝臟電腦斷層影像執行臨床肝癌診斷；其中肝臟電腦斷層造影過程需經由靜脈注射對比劑以及三相掃描設定方能得到具有診斷肝癌之影像。因此，對對比劑具有不良反應患者，不建議執行此項掃描，為擴大造福人群。本研究利用無顯影電腦斷層肝影像特徵進行正常肝臟和肝癌之分類研究。
材料與方法：本研究採用回顧性分組實驗設計，收集2015/6/26-2017/7/31期間的肝臟電腦斷層無顯影影像共計139例。記錄生理參數與影像診斷。定義影像特徵包括影像全肝體積、平均值、標準差、最小值、最大值、訊雜比以及每一立方公分平均一年全肝影像品質變化性。採用機器學習法進行分類模型之研究，包括邏輯斯特迴歸(Logistic Regression, LR)及支持向量機(Support Vector Machine, SVM)。將139例分成對照組為正常肝臟68例、實驗組為肝癌71例進行10折交叉驗證。驗證指標包括一致性統計量、曲線下的面積(Area Under the Curve, AUC)、準確度；用當P值小於0.05時代表具有統計顯著性。
結果：Albumin與SAV結合值在LR、LR訓練組、LR交叉驗證組、SVM訓練組、SVM交叉驗證組之準確度分別為0.705、0.669、0.662、0.698、0.676；AUC分別為0.742、0.742、0.711、0.700、0.679。
結論：本研究發現重要生理參數為Albumin，顯著特徵為SAV，其中SVM可做為分類之模型。未來仍需要更多案例以及合理影像特徵，同時考慮人工智慧，演算法進行分類研究。

關鍵字

無顯影電腦斷層 ； 邏輯斯特迴歸 ； 支持向量機

並列摘要

Purpose: Hepatocellular carcinoma (HCC) is the second leading cause of cancer death. Contrast-enhanced three phase computed tomography (CT) of the liver is the gold standard for HCC diagnosis. However, contrast medium can have side effects on patients. The study reports a model for differentiating normal liver from HCC-affected liver through nonenhanced CT.
Materials and Methods: The study retrospectively included 139 patients with clinical suspicion of HCC referred between June 26, 2015, and July 31, 2017, for three-phase CT of the liver. The physiological parameters and diagnostic reports were recorded. The defined image features included total liver volume, mean, standard deviation, minimum, maximum, signal to noise ratio (SNR), and liver image quality variability per age per volume (10K imesSNR/Age/Vol, SAV). Machine learning methods were used to study the classification model, including Logistic Regression (LR) and Support Vector Machine (SVM). In the classification model, the 139 cases were divided into a control group, with 68 normal liver cases, and experimental group, with 71 HCC cases10-Fold Cross Validation. Validation indicators included consensus statistics, area under the curve (AUC), and accuracy; A p value below 0.05 was considered statistically significant.
Results: LR, LR with training, LR with cross-validation, SVM with training, SVM with cross-validation classification results using predictions established by Albumin and SAV. Accuracy of them were 0.705, 0.669, 0.662, 0.698, 0.676, respectively. AUC of them were 0.742, 0742, 0.711, 0.700, 0.679, respectively.
Conclusion: Albumin and SAV are significant parameters. Furthermore, SVM can be used as a classification model. Studies including larger sample sizes and suitable image features are warranted. An artificial intelligence algorithm may be used for classification research.

並列關鍵字

Non-Enhanced Computed Tomography ； Liver Imaging ； Hepatocellular Carcinoma ； Logistic Regression ； Support Vector Machine ； Accuracy ； Area Under The Curve

參考文獻

[ 1] World health statistics 2018: monitoring health for the SDGs, sustainable development goals. Geneva: World Health Organization; 2018.

[ 2] Global Health Estimates 2016: Deaths by Cause, Age, Sex, by Country and by Region, 2000-2016. Geneva, World Health Organization; 2018.

[ 3] Global Health Estimates technical paper series: WHO methods and data sources for country-level causes of death, 2000-2016. Geneva, World Health Organization; 2018.

[ 4] Global Hepatitis Report 2017. Geneva: World Health Organization; 2017.

[ 5] Zhang W, Wang X, Wang Y, Zhao X, Duan W, Wang Q, Wu X, Kong Y, Ma H, You H, OuX, Jia J. Effective viral suppression is necessary to reduce hepatocellular carcinoma development in cirrhotic patients with chronic hepatitis B: Results of a 10-year follow up. Medicine (Baltimore). 2017;96(44):e8454.