以肝臟超音波影像紋理特徵區別肝硬化之研究

蔡漪琳(I-Lin Tsai)

指導教授 ： 陳中明

共同指導教授 ： 楊培銘

國立臺灣大學/工學院/醫學工程學研究所/碩士(2005年)

https://doi.org/10.6342/NTU.2005.00272

摘要

本論文是不受機器參數影響的超音波肝臟影像分類之研究，為了描述肝臟影像的紋理特徵，我們透過watershed transform等影像處理步驟，去找出如cell number、與cell size等影像的cell-based特徵，另外還有region size、以及常見的grey-level co-occurrence matrix特徵值。在這些特徵值的設計上，我們都會以脾臟的特徵作為肝臟特徵的正規化參考，以消除不同人之間的潛在差別，並使不同機器參數設定下的影像之間得以正規化。而最後所找到的所有特徵值，我們會繼續透過forward selection的處理去找到一個能夠表現最佳分類效能的特徵值組合作為最後分類的數據。而我們是利用這些特徵值來將影像分為兩類-肝硬化影像、與正常肝臟影像。
本研究的分類方法上，是採用leave-one-out 的logistic regression function來產生分類器。我們所使用的資料一共有273筆，皆是由台大醫院所提供。其中有134筆是肝硬化影像、另外的139筆是正常肝臟影像。最後得到的分類結果有93%的正確率；且在本研究中，我們另外提出兩種方法(Ogawa的方法、與nonseperable wavelet decompositions)來進行結果的比較，而我們所提出的方法顯現了最好的分類正確率。為了驗證採用脾臟作為正規化參考的功效，在本研究中去計算了沒有脾臟特徵來作為參考時的特徵組合的分類結果，其正確率為91%。因此得出沒有脾臟作為正規化參考的分類結果較有脾臟作為正規化參考的結果來得不好的結論。也可用來說明本研究所提出的特徵值是需要以脾臟作為正規化參考的。

關鍵字

肝硬化 ； 超音波 ； 分類 ； 迴歸分析函數 ； 特徵值

並列摘要

The classification of ultrasonic liver images is studied, making use of the features of cell-based algorithms, region size and grey-level co-occurrence matrix features to describe the texture characterization problem. We use the ultrasonic texture of the same patient’s spleen image as the reference in reading that of the liver image for the purpose of normalization the texture variation caused from different system settings. The resulting features are further processed with forward selection to obtain a combination of features which represents the most discriminating pattern space for classification. The extracted features are used to classify two sets of ultrasonic liver images-normal liver, and cirrhotic liver.
The leave-one-out logistic regression function classifier is employed to evaluate the performance of these features. The proposed method was evaluated by using 273 clinical ultrasound images collected at National Taiwan University Hospital, including 139 normal livers and 134 cirrhotic livers. The classification accuracy was 93%. The proposed method was compared to other texture description methods using only images from livers (Ogawa’s method and nonseperable wavelet decompositions). The proposed method gave the highest classification rate, showing its applicability for the approach based analysis of the large class of natural textures. We also modified the features to be without spleen reference to verify the effect of normalization by reference to spleen images. And the classification accuracy was 91%. Obviously the proposed method has better classification performance when it referenced to the spleen images.

並列關鍵字

Liver Cirrhosis ； Ultrasound ； Classification ； Logistic Regression Function ； Texture Features

參考文獻

2. Yang PM, Huang GT, Lin JT, et al. Ultrasonography in the diagnosis of benign diffuse parenchymal liver diseases: A prospective study. J Formosan Med Assoc 1988; 87: 966-977.

3. Goyal AK, Pokharna DS, Sharma SK. Ultrasonic diagnosis of cirrhosis reference to quantitative measurements of hepatic dimensions. Gastrointest Radiol 1990; 15: 32-34.

4. Vilgrain V, Lebrec D, Menu Y, Scherrer A, Nahum H. Comparison between ultrasonographic signs and the degree of portal hypertension in patients with cirrhosis. Gastrointest Radiol 1990; 15: 218-222.

5. Joseph AEA, Saverymuttu SH, Al-Sam S, Cook MG, Maxwell JD. Comparison of liver histology with ultrasonography in assessing diffuse parenchymal liver disease. Clinical Radiology 1991; 43: 26-31.

6. Ferral H, Male R, Cardiel M, Munoz L, Ferrari FQ. Cirrhosis: Diagnosis by liver surface analysis with high-frequency ultrasound. Gastrointest Radiol 1992; 17: 74-78.