腹部電腦斷層掃描骨骼肌影像之穩健判別與區分方法及臨床應用研究

Abstract

骨骼肌指標(Skeletal Muscle Index, SMI)、骨骼肌密度指標是放射科醫師常作為判斷病人身體狀況的重要指標。臨床上，醫師經常需要利用生理測量指標衡量患者是否適合接受手術或化療，其骨骼肌指標相較於其他生理指標較客觀且具有代表性。

人體的骨骼肌含量可經由多種方法測量，如：利用生物電阻抗(BIA)原理檢測身體組成之InBody身體組成分析儀、磁力共振成像(MRI)、利用X射線對人體各器官和組織有不同的穿透率之電腦斷層掃描(CT)，其中以影像分析的用途最為廣泛且結果也最為精準。

針對腹部電腦斷層掃描橫切面影像之骨骼肌判別與區分，已有一套由學者許(2017)[11]與學者謝(2018)[10]所提出的自動判別與區分方法。由於其判別方法容易於腹肌部分圈選到緊鄰腹肌之肝臟與其他臟器，於腰肌部分則是容易圈選到緊鄰腰肌之腎臟或僅圈選到部分腰肌；其區分方法則是容易於區分腹直肌與腹壁肌部分因部分影像之腹肌厚度變化不明顯而造成錯誤之腹直肌與腹壁肌分界點，於區分腰大肌與豎脊肌部分則是切點不夠精確，圈選出之腰大肌易包含部分豎脊肌；故在骨骼肌外圈、腰肌內圈、腹肌內圈以及區分不同部位之骨骼肌仍有許多改善空間，因此本研究延續採用腹部電腦斷層掃描橫切面影像，除了改善現有判別方法，對於內臟與骨骼肌相鄰、腹水之案例，正確的將骨骼肌區域判別出，同時將骨骼肌不同部位的區分進行更精確的修正，再利用自動化判別與區分結果進行臨床研究分析。

本研究參考現有的骨骼肌自動判別與區分流程，使用Canny邊緣偵測，由外向內搜尋，依序定義皮膚層、骨骼肌外圈、骨骼肌內圈，最後在骨骼肌外圈、內圈之間取得的骨骼肌範圍。不同於現有之判別方法，本研究修改現有搜尋骨骼肌內圈之方法，透過邊緣位置與長度特性重新定義腰肌內圈邊緣，透過腹肌厚度特性與邊緣位置重新定義腹肌內圈邊緣。根據穩健判別出之骨骼肌對稱性，提出影像品質指標，並據以評估有效性及骨骼肌品質特性。此外，本研究不同於現有之區分方法，於穩健判別腹部骨骼肌區域後，透過大津演算法結合骨骼肌之位置、構造特性自動區分出不同部位之骨骼肌，將骨骼肌區分為腹直肌、腹壁肌、豎脊肌以及腰大肌四部分。

針對深度學習骨骼肌影像之區分，由於深度學習在區分不同區域之骨骼肌影像時需進行多次影像標識(image labeling) ，是一項耗時且耗人力的工作，故本研究採Canny邊緣偵測，由內向外搜尋，依序定義脊椎、骨骼肌內圈、骨骼肌外圈，接著沿用腹部電腦斷層掃描橫切面影像之骨骼肌區分方法，再利用自動化區分結果進行臨床研究分析。

為了驗證本研究之穩健判別與區分骨骼肌方法，首先利用臺大醫院提供150筆人體腰椎第三節(L3)之CT橫切圖分別與醫師手繪區域內骨骼肌與骨格肌四個部位面積進行迴歸相關性分析。結果顯示本研究改善後的穩健骨骼肌判別與區分方法之R2皆較原方法高。利用同組案例驗證骨骼肌不同部位區分方法，發現骨骼肌指標與腹壁肌指標最為相關，其次為豎脊肌指標。當整體骨骼肌總面積增加一單位(〖mm〗^2)，19%來自腰大肌，34%來自豎脊肌，11%還自腹直肌，36%來自腹壁肌。除了比較不同部位骨骼肌之關係，使用MedCalc軟體中的Kaplan-Meier curve針對台大醫院所提供的150筆資料進行評估術後健康狀況，分析整體骨骼肌面積、指標與進行標靶治療之肝癌病患之存活率相關性，結果顯示進行標靶治療之肝癌患者中，同時患有肌少症之患者之存活率較低，非肌少症患者之存活率較高，且以十八個月為存活設限之肌少症與非肌少症患者之存活曲線差異達統計上顯著。 最後針對電腦斷層掃描影像之穩健判別與區分骨骼肌方法與深度學習區分骨骼肌分法，本研究使用Visual C#建制一套程式，介紹操作方式與程式介面。

Skeletal Muscle Index (SMI) and Skeletal Muscle Density Index are widely used as important indexes to evaluate health condition of patients by radiologists. Clinically, radiologists often use physiological measures to decide whether a patient’s health condition is acceptable for undergoing an operation. Skeletal Muscle Index is more objective and representative than other physiological indicators.

The skeletal muscle contained in the human body can be measured by a variety of methods, such as InBody composition analyzer, Magnetic Resonance Imaging (MRI), Computerized Tomography (CT), Dual-energy X-ray Absorptiometryetc (DXA), etc. Among all common evaluation methods, image analysis provides more accurate measurement results and is widely adopted. Hsu (2017) and Hsieh (2018) proposes an automatic segmentation and differentiation method of Skeletal Muscle on abdominal CT images. However, this method still has deficiencies, such as misidentification of the abdominal muscles, problems of segmenting the inner circumference of the lumbar muscles and problems of differentiating abdominal and lumbar muscles. Moreover, for outcome of referencing a deep learning model, it is tedious to do image labeling repeatedly. Therefore, this research is to improve this method for better differentiation of the skeletal muscle region and effective clinical applications.

This research follows the steps of Hsieh’s automatic segmentation and differentiation method to identify the margins of tissue layers from the outer part gradually towards the inner part of the human body. That is, the outer margin of the skin layer is first defined by the high contrast between the human body portion and the non-human body portion on the CT images. Then, the search continues to find the outer and inner boundaries of the skeletal muscle. The region of the skeletal muscle is segmented using a proposed edge detection method adapted from the well-known Canny Edge Detection method.

Different from Hsieh’s method, this research proposes refining the existing method of searching the inner boundary of skeletal muscle by improving the detection algorithm of skeletal muscle inner edge. According to the symmetry of the skeletal muscle symmetry, the image quality index was proposed and the effectiveness and skeletal muscle quality characteristics were evaluated. After improvement of abdominal skeletal muscle segmentation, in this research Otsu algorithm are used to improve distinguishing the skeletal muscles of different parts and divide the skeletal muscle into four parts: the rectus abdominis, the abdominal wall muscle, the erector spinae and the psoas muscle.

Patients with newly diagnosed liver cancer who underwent targeted therapy at National Taiwan University Hospital are used to demonstrate the proposed segmentation method. Body composition is assessed using cross-sectional CT images to calculate the total skeletal muscle (TSM) area and index. The results of automatic segmentation of the skeletal muscle were compared to the results of the skeletal muscle manually circumscribed by a radiologist. Regression analysis is performed and the resulted R2 reaches 0.92 higher than the original method. Using the same case, With the differentiation of skeletal muscle parts, it is found that the skeletal muscle index was most correlated to the erector spinae index, followed by the abdominal wall muscle index. When the total skeletal muscle area is increased by one unit ( 2 mm ), it is found 19% from the psoas muscle, 34% from the erector spinae, 11% from the rectus abdominis, and 36% from the abdominal wall muscle.

In addition to comparing the relationship between subparts of skeletal muscles, KaplanMeier curve with different index of the skeletal muscle are used to compare the effect on overall survival (OS). The results show that among patients with liver cancer who underwent targeted therapy, the survival rate of patients with sarcopenia was low, the survival rate of patients with non-sarcopenia was higher. For 18 months of survival, the difference in survival curves for patients with sarcopenia was statistically significant.

Finally, for the robust discrimination of computerized tomography images and the distinction between skeletal muscle method and deep learning to distinguish skeletal muscle division method, this study uses Visual C# to build a program to introduce the operation mode and program interface.