肝臟超音波影像之腫瘤區別診斷：
影像特徵效能受系統參數影響之研究

摘要根據行政院衛生署統計資料，目前台灣地區的十大死亡原因，以因惡性腫瘤死亡的人數佔首位。如果只計算男性時，則在惡性腫瘤中，因肝臟的惡性腫瘤死亡者最多，並且因其他慢性肝疾病死亡者又佔十大死亡原因的第六位。由此可見肝臟惡性腫瘤在台灣對全民的健康構成了重大的威脅。肝臟超音波影像對肝疾病是一種既經濟又有效的非侵入性診斷方式。然而在判讀肝臟超音波影像之良性的肝血管瘤與惡性的肝細胞癌時，往往因為醫師的經驗不同或同一醫師在不同時間做出不同的診斷。為了降低人為因素對於肝臟超音波影像之良惡性腫瘤判讀的影響，具有客觀且高再現性的電腦輔助診斷技術逐漸受到重視。其基本的想法則是將視覺上所得的概念量化，以避免對於同一影像，因人因時而異的描述，以及獲取較視覺所能處理之更為複雜的影像特徵。雖然傳統的影像特徵，常常被證明對特定的影像可得到不錯的分類效能，但實際上，卻極易受到系統參數的影響。本研究主要的目的即在探討系統參數對於傳統影像特徵之分類效能的影響。本研究採用傳統的 Gray Level Co-occurrence Matrix與Gray Level Run Length Matrix 為超音波影像特徵的基礎，使用決策樹分類器為特徵區別判斷的工具，採取leave-one-out cross-validation 為分類效能之驗證，並以窮舉法進行特徵選取，來挑出最佳的傳統型影像特徵。為了降低系統參數對分類效能的影響，有別於傳統型特徵，我們將 ROI 內的灰階值作分佈調整後，再擷取其數值特徵，稱之為分佈型特徵。而為了瞭解系統參數對於傳統影像特徵的影響，我們將圖形作 gamma 修正或倍率修正，以模擬系統參數的改變，觀察傳統型特徵與分佈型特徵之分類效能的變化。實驗資料為包含49張肝血管瘤與45張肝細胞癌影像。產生的傳統型特徵，搭配決策樹分類器，以窮舉法找到之最佳特徵組合，其分類正確率可達到92.6%。但當隨機改變gamma 修正與倍率修正的參數時，分類正確率為66%及75.5% 均在80%以下，由此可知傳統型特徵極易受到gamma 修正與對比度改變(倍率修正)的影響。對於分佈型特徵，最高的分類正確率達86.2%。雖然其正確率低於傳統型特徵，但由實驗結果顯示，隨機地改變gamma 修正或倍率修正的參數時，分類正確率都在80.9%。由此可知，分佈型特徵具有抑制系統參數影響的功能。其主要原因是將影像ROI內的灰階值作分佈調整的過程中，系統對比度的因子會被消除。所以雖然分佈型特徵亦受到系統參數的影響，但受影響的幅度相對小於傳統型特徵。關鍵字：超音波影像，肝血管瘤，肝細胞癌，電腦輔助診斷，分類，決策樹，特徵擷取，特徵選取，窮舉法，Gamma修正，對比度。

關鍵字

對比度； Gamma修正；窮舉法；特徵選取；特徵擷取；決策樹；分類；電腦輔助診斷；肝細胞癌；肝血管瘤；超音波影像

並列摘要

Abstract Neoplasm has been the leading cause of death in Taiwan in recent years based on the statistics provided by the Department of Health, Executive Yuan, Taiwan R.O.C. If only males are taken into account, hepatic tumors have been the number-one cause of death among all kinds of cancers. Furthermore, chronic liver diseases rank the sixth of the top ten causes of deaths. It is therefore clear that hepatic tumors have become a tremendous threat to the health of citizens in Taiwan. Hepatic sonography is a cost-effective non-invasive approach to the diagnosis of liver diseases. Nevertheless, it happens frequently that for the same ultrasound images, different diagnoses are made by medical doctors with different experiences or by the same medical doctor but at different times for the differential diagnosis of hemangiomas and hepatomas. To minimize the human effects on the interpretation of hepatic sonograms for differentiating hemangioms from hepatomas, computer aided diagnosis (CAD) techniques featuring objectiveness and high reproducibility have received significantly more attentions than ever. The basic idea of a CAD approach is to convert the abstract concepts embedded in visual perception into quantitative descriptors, which not only can alleviate the influence of human factors but also can extract the image features more complicated than those manageable by visual perception. Although it has been repeatedly reported that reasonably high classification accuracy may be achieved by using conventional image features, the performances of these features are theoretically dependent on the settings of ultrasound imaging systems. To reveal this fact, this study aims to investigate the influence of the system settings on the performance of conventional image features. Two classes of conventional image features are adopted in this study, namely, Gray Level Co-occurrence Matrix and Gray Level Run Length Matrix. The classifier for differential diagnosis is a decision tree. All performances are evaluated based on leave-one-out cross-validation. To ensure the fairness of this study, the best conventional features are selected by using enumeration approach from all candidate features. In the attempt to minimize the influence of system settings on the classification performance, a new set of image features have also been investigated, called distribution-adjusted features. Unlike the conventional features, these features are derived from the normalized ROIs with all histograms stretched to cover the entire dynamic range of available gray levels. To study the effect of system settings on the conventional features and distribution-adjusted features, all images are randomly modified by different gamma curves or contrast enhancement to simulate potential variations of system settings. The materials for this study include 49 hemangioma images and 45 hepatoma images. For the best conventional features selected by enumeration approach, the best classification accuracy achieved by a decision tree is 92.6%. However, as the gamma curve and contrast enhancement are randomly applied to the images, the classification accuracy degrades rapidly down to as low as 66% and 75.5%, respectively. It suggests that the conventional features be easily affected by gamma corrections and contrast settings. On the other hand, for the distribution-adjusted features, the best classification accuracy is 86.2%. Although the best performance is worse than the conventional features, the experimental results show that as the ROIs are randomly modified by different gamma curves and contrast settings, all classification accuracies are about 80.9%. It suggests that the distribution-adjusted features are more robust to the influence of system settings than their counterparts. It is mainly because the contrast factor is eliminated as the histograms of the ROIs are stretched to the entire dynamic range. Key Words: Ultrasound Images, Hemangioma, Hepatoma, Computer Aided Diagnosis, Classification, Decision Tree, Feature Extraction, Feature Selection, Enumeration Method, Gamma Correction, Contrast.

並列關鍵字

Contrast. ； Gamma Correction ； Decision Tree ； Feature Extraction ； Feature Selection ； Enumeration Method ； Computer Aided Diagnosis ； Classification ； Hepatoma ； Ultrasound Images ； Hemangioma

參考文獻

[14] E-Liang Chen, Pau-Choo Chung, Ching-Liang Chen,

[1] M. M. Galloway, “Texture analysis using gray level run

172-179, June 1975.

value distribution of run lengths for texture analysis,”

Pattern Recognit. Lett., vol. 11, pp. 415-420, June 1990.

被引用紀錄

朱書志（2007）。肝硬化之超音波影像分析〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU.2007.02269

國際替代計量

肝臟超音波影像之腫瘤區別診斷：影像特徵效能受系統參數影響之研究

全文下載

主題瀏覽