使用超音波影像紋理分析識別肝纖維化與脂肪化病灶

Abstract

超音波系統因為具有方便性、即時影像、非侵入性且成本相較於其它如核磁共振造影(MRI)或電腦斷層掃描(CT)系統來的低廉等優點，因此廣泛的應用在醫學影像診斷上，更成為第一線臨床診斷工具。有鑑於肝臟疾病目前在臨床診斷上除了做穿刺和切片檢查外，並無一套有效且精準可靠的非侵入式診斷方法，因此本研究以超音波B-mode影像做為切入點，並採用紋理分析方法，實驗的對象分為兩部分：第一部分為大鼠肝臟離體實驗，分為三組，分別以藥物誘發導致肝纖維化和脂肪肝，以及一組正常肝做為對照組，主旨在於以紋理特徵值在隨著藥物使用時間增加導致病情愈發嚴重下的趨勢變化。第二部分為臨床實驗部份，分別以線性探頭和弧形探頭取得的纖維化病例，且有做過切片採Metavir評分由0~4分做嚴重等級區分和對照，觀察各紋理特徵值隨著病情嚴重程度的趨勢變化。 本研究採用的紋理分析方法以Haralick學者於1973年所提出的灰階共生矩陣(GLCM)方法對B-mode影像進行分析並計算其11個特徵值。GLCM為計算一影像紋理在一相對位置下像素間的灰階值關係，排成一機率分佈矩陣，由此矩陣可得知在某相對位置下像素對的灰階值關係為i和j出現的機率大小。再由GLCM算出以下11個特徵值，分別為：能量(Energy)、對比度(Contrast)、相關性(Correlation)、熵(Entropy)、均質性(Homogeneity)、變異數(Variance)、平均和(Sum Average)、變異數和(Sum Variance)、熵和(Sum Entropy)、變異數差(Difference Variance)和熵差(Difference Entropy)。 在大鼠肝臟離體實驗中，結果發現各時期的正常肝在11個特徵值下皆維持一定值，由此可知正常肝臟有其固定的紋理，能做為基準值和肝臟發生病變的紋理特徵值做一比較。而肝纖維化和脂肪肝在各個特徵值下雖有趨勢但走向相同，僅能在數值大小上略做區分，無法單以和正常肝之間的趨勢判定是肝纖維化或脂肪肝。另外在臨床實驗方面，各紋理特徵值隨著病情嚴重程度的趨勢變化在判別無纖維化(Metavir評分0分)和有纖維化(Metavir評分1~4分)效果顯著，但針對纖維化間各程度區分則較為困難。

Due to the advantages of its convenience, instantaneity, non-invasion and low cost than other systems such as MRI or CT, ultrasound system is widely applied to medical imaging diagnosis, and becomes the first choice for clinical diagnosis. Because there is hardly any effective and specific non-invasive way to diagnose the liver diseases except doing biopsy, this research will cut into the point by using B-mode and adopting texture analysis. There are two parts in this experiment: the first part is the in vitro rat experiment which divides the rats into three groups: induced fibrosis liver by medicine injection, induced fatty liver by medicine feeding and the normal liver as the control group. The second part is the cases of clinical experiments. We obtain the fibrosis cases by linear array and convex array transducer, adopting Metavir score from 0~4 as the assessment for the degree of severity after doing biopsy, and investigate the texture features along with the degree of severity. This study uses gray level co-occurrence matrix (GLCM), which was proposed by Haralick in 1973, to analyze B-mode image and calculate the 11 texture features. GLCM is computed for various angular relationships and distances between neighboring pixel pairs on the image. Therefore we can know the probability of pixel pairs of gray level relationship as i and j in GLCM. 11 texture features was extracted from GLCM, including Energy, Contrast, Correlation, Entropy, Homogeneity, Variance, Sum Average, Sum Variance, Sum Entropy, Difference Variance and Difference Entropy. In the in vitro rat experiment, we found that the trend of texture features along with the degree of severity maintained at a stable level in the normal control group, concluding that the normal liver has stationary texture features and could be considered as a control in comparison with diseased livers. Whereas the texture features of the fibrosis and fatty liver groups proceeded in a similar trend. Therefore we can separate the normal livers from fibrosis and fatty livers, but failing to separate fibrosis and fatty liver. In the clinical experiments, we found that the texture features have good performances to separate the non-fibrosis (Metavir score 0) and fibrosis (Metavir score 1 to 4), but they could hardly identify the fibrosis degrees.