以影像為基礎的身體組成分析於肌少症病人的診斷、自動定量及治療反應評估

Image-Based Body Composition Analysis for the Diagnosis, Automatic Quantification and Treatment Response Evaluation in Patients with Sarcopenia

吳志宏(Chih-Horng Wu)

指導教授 ： 施庭芳

共同指導教授 ： 何明志 ； 高嘉宏(Jia-Horng Kao)

國立臺灣大學/醫學院/臨床醫學研究所/博士(2021年)

https://doi.org/10.6342/NTU202103252

摘要

緒論
肌少症一開始主要用於描述老人生理狀況逐漸退化，依照歐洲老年肌少症工作小組定義肌少症包括肌肉量減少、肌肉力氣降低及肌肉表現功能下降。次發性的肌少症和許多慢性疾病有關，包括癌症、肝硬化、慢性腎臟病等等。在過去，營養狀況常以身體質量指數、上臂中圍、血清中白蛋白或生物阻抗分析等來代表。但這些臨床指標在嚴重癌症、肝硬化及慢性腎臟病時會因為皮下水腫及腹水產生的高估的情形。相對來說，斷面影像檢查如電腦斷層及磁振造影，可以對腹部的肌肉及脂肪作準確的分析。腹部肌肉還可分類為腰大肌、豎脊肌、腹壁肌及腹直肌等等。而對於脂肪組織可以分類為臟器脂肪及皮下脂肪。由以上可知使用影像方法來進行肌肉量的計算在肌少症的臨床應用上清楚而客觀。
目前在肌少症的研究有三個未解的問題，一是對肌少症的診斷標準及肌肉量下降的閾值並沒有一致的共識，因此大型及多樣性針對不同性別、種族及疾病族群的肌少症相關研究有其必要性。尤其是早期肌少症的標準由歐洲老年肌少症工作小組所訂，其研究的族群以西方國家為主，西方人的身體組成與東方人有明顯差異，東方國家的研究以日本為主，而且大量使用電腦斷層作為測量腹部肌肉及腰大肌的工具。二是使用斷面影像測量腹部肌肉量的過程對醫師是一項耗時費力的工作，近年來由於機器學習及深度學習在圖形辨識領域的發展，快速且精準的自動切割腹部肌肉及脂肪組織的影像後處理工具變成可能。三是根據以往的經驗及研究，病人的肌少症很難藉由營養的補充或藥物達成，對於生重病的病人使用高強度的鍛鍊也不切實際。因此能否用微創手術的方法來增加食慾並進一步改善病人的肌少症便成為值得探索的話題。
方法
我們以臺大醫院的影像資料庫為材料，收集胰臟癌、晚期肝癌及腹膜透析病人。對胰臟癌的病人我們分別使用西方(俄羅斯研究單位)及東方(臺大醫院)的診斷標準來評估是否為預後的重要因子；對晚期肝癌的病人，我們分別測量腹部肌肉的腰大肌、豎脊肌、腹壁肌及腹直肌，以研究何者為影響存活率的因子；對接受腹膜透析的病人，我們除了測量腹部肌肉及腰大肌的肌肉量，並與淨軟組織量做相關性分析，找出何者較具代表性。然後我們結合前面三個資料集所收集的影像，利用我們已經圈選的遮罩進行機器學習及深度學習訓練自動切割肌肉及脂肪組織的模型。並利用遷移學習的方式將電腦斷層所建立的模型參數轉到磁振造影使用，然後評估兩者的相關性。最後我們利用經頸靜脈肝內門體靜脈支架分流術來治療因肝硬化有靜脈瘤出血及腹水的病人及腎動脈栓塞術來治療多囊腎的病人，並比較治療前及六個月後的影像，能否也同時改善病人肌少症的狀況。
結果
第一部分：經過在俄羅斯的研究單位及臺大醫院對胰臟癌病人進行影像分析後發現對於腹部肌肉、臟器脂肪及皮下脂肪的測量為高度相關(r = 0.974, 0.978, 0.979, p < 0.001)。 但在存活分析中，依照東方的診斷標準才是顯著的存活因子(p = 0.008)，西方的診斷標準不是(p = 0.807)。在對晚期肝癌病人進行個別肌肉測量後顯示對於腹部肌肉，僅腰大肌是顯著影響存活率的因子(p = 0.014)，豎脊肌、腹壁肌及腹直肌不是。在腹膜透析病人進行淨體重估計及腹部影像分析後發現，腰大肌比起全腹部肌肉的測量與淨軟組織量有較高的相關性(r = 0.775 vs. 0.681, p < 0.001)，多因子分析顯示腰大肌是顯著影響全存活率的因子(HR: 2.386, CI: 1.315-4.330)。
第二部分：使用機器學習及深度學習演算法與醫師圈選範圍進行相關性分析，Sobel邊界偵測、AlexNet、VGG、ResNet與基準真相的相關係數分別為0.85, 0.81, 0.91, 0.78 (p < 0.05)。若以VGG的演算法來區分個別腹部肌肉，與豎脊肌、腰大肌、腹壁肌及腹直肌的相關係數分別為(r = 0.85, 0.43, 0.42, 0.02)。接著我們把利用電腦斷層建立的深度模型參數取出，利用遷移學習來對有腹部磁振造影影像的病人進行腹部肌肉自動切割，其精確度相對於隨機參數分別為0.9704及0.9696。最後我們分析使用電腦斷層及磁振造測量腹部肌肉、臟器脂肪及皮下脂肪為高度相關(r = 0.946, 0.969, 0.953)。
第三部分：使用經頸靜脈肝內門體靜脈支架分流術來治療因肝硬化相關的併發症，治療後六個月測量影像上腰大肌的面積明顯大於治療前(15.66±6.99 vs. 13.11±5.47, p = 0.040)，病人的食慾變好。使用腎動脈栓塞術來治療多囊腎，比較治療前及治療後六個月的影像，腎臟的體積明顯縮小(2979.8±2069.6 vs. 4019.4±2438.2, p < 0.001)，腰大肌的面積明顯增大(13.6±7.7 vs. 11.6±6.7, p < 0.001)，病人的腹脹症狀也明顯改善。
結論
肌少症的診斷標準需要依照性別、族群被正確的選擇才能做出正確的診斷。在腹部肌肉當中，腰大肌肌肉量是最重要的預後因子，也是最具代表性的肌肉。對電腦斷層及磁振造影影像，利用深度學習進行肌肉自動切割的正確性高且兩者的值也具一致性。我們可以利用介入放射的方法來治療病人的原始疾病，並有機會同時改善肌少症。

關鍵字

電腦斷層 ； 磁振造影 ； 肌少症 ； 胰臟癌 ； 肝細胞癌 ； 腎衰竭 ； 深度學習

並列摘要

Introduction
Primary sarcopenia is used to describe aging and progressed with the physiologic decline. According to the European Working Group on Sarcopenia in Older People (EWGSOP), the definition of sarcopenia includes low skeletal muscle mass, decreased muscle strength, or reduced performance. Secondary sarcopenia is associated with many chronic diseases, including cancer, liver cirrhosis, and chronic kidney disease. In the past, nutrition status is evaluated by body mass index, mid-upper arm circumference, serum albumin, and bioelectrical impedance analysis. But these clinical biomarkers will be affected by tissue edema and ascites. In contrast, cross-section imaging, such as computed tomography (CT) and magnetic resonance (MRI), can accurately analyze abdominal muscle and fat. The abdominal muscle can be further divided into psoas muscle, paraspinal muscle, abdominal wall muscle, and rectus abdominis. The adipose tissue can be further categorized into visceral fat and subcutaneous fat. Therefore, muscle mass measurement by image-based post-processing method is a clear and robust method in the application of sarcopenia evaluation.
There are three unmet needs in the field of image-based muscle mass research. First, there are many different diagnostic criteria without consensus in the cut-off of sarcopenia. Therefore, large-scale studies with diverse populations are needed. Besides, the diagnostic standard was made by EWGSOP. There is a significant difference in body shape and composition exist between Western and Eastern people. Most of the Eastern studies came from Japan with extensive usage of CT as a tool to measure abdominal and psoas muscle mass. Second, the image post-processing procedure for muscle mass measurement is time-consuming and labor-intensive. To solve this problem, the advancement in image recognition by machine or deep learning makes rapid and accurate automatic segmentation of muscle and adipose tissue possible. Finally, previous researches have demonstrated sarcopenia is challenging to improve by nutritional supplement and medication. Whether interventional procedures can improve sarcopenia via increase appetite should be further explored.
Materials and Methods
We used the National Taiwan University Hospital (NTUH) image database and collected data on pancreatic cancer, advanced hepatocellular carcinoma, and peritoneal dialysis. For patients with pancreatic cancer, we used Western and Eastern diagnostic criteria to evaluate whether the presence of sarcopenia is a prognostic factor, respectively. For patients with advanced hepatocellular carcinoma (HCC), we measured the muscle mass of psoas, paraspinal, abdominal wall, and rectus abdominis muscles and tried to determine which muscle affects survival. We measured total abdominal and psoas muscle mass for patients receiving peritoneal dialysis and correlated with lean body mass. Then we combined three datasets and used the mask for machine and deep learning training and automatic segmentation model for muscle and fatty tissues. We also used transfer learning to apply in MRI images and then evaluate the correlation between CT and MRI. Finally, we used the transjugular intrahepatic portosystemic shunt (TIPS) to treat refractory variceal bleeding and ascites and renal artery embolization to treat polycystic kidney disease. We compared the pre-operative and 6-month post-operative images to find out whether the intervention can improve sarcopenia.
Results
First part: Image-based body composition analysis was conducted in Western (Federal Medical and Rehabilitation Center, Russia) and Eastern (NTUH, Taiwan) institutes for patients with pancreatic cancer. The correlation coefficients (r) of total skeletal muscle, visceral adipose tissue, and subcutaneous adipose tissue measurements between two institutes were 0.974, 0.978, 0.979 (p < 0.001). However, in survival analysis, patients with sarcopenia diagnosed by Eastern criteria showed a significant difference in overall survival (p = 0.008), but Western criteria didn’t (p = 0. 807). For patients with advanced HCC, only patients with sarcopenia defined by psoas muscle exhibited poorer overall survival than patients without sarcopenia (p < 0.001), but other muscles (paraspinal, abdominal wall, and rectus abdominis muscles) didn’t. In patients receiving peritoneal dialysis, lean soft tissue was more correlated with psoas muscle than total skeletal muscle (r = 0.775 vs. 0.681, p < 0.001). In multivariate analysis, after adjusting clinical and PD-related parameters, only patients with psoas muscle-defined sarcopenia had poorer survival than did those without (hazard ratio [HR]: 2.386, 95% confidence interval [CI]: 1.315 – 4.330).
Second part: We used machine and deep learning for correlation analysis with the region of interest made by radiologists. The correlations coefficients for Sobel edge detection, AlexNet, VGG and ResNet algorithms were 0.85, 0.81, 0.91 and 0.78 (p < 0.05). Then we used VGG for further subgroup analysis. The correlation coefficients in paraspinal, psoas, abdominal wall, and rectus abdominis muscles are 0.85, 0.43, 0.42, 0.02. Then we extracted the pre-trained parameters with the model made from CT images and performed transfer learning to MRI images for abdominal muscle automatic segmentation. The accuracy for transfer learning was higher than the random state (0.9704 vs. 0.9696). Finally, the CT images for total skeletal muscle, visceral adipose tissue, and subcutaneous adipose tissue were strongly correlated with MRI (r = 0.946, 0.969, 0.953).
Third part: We used TIPS to treat decompensated liver cirrhosis and found the 6-month post-TIPS psoas muscle area was significantly larger than the pre-operative evaluation (15.66±6.99 vs. 13.11±5.47, p = 0.040). The patients with liver cirrhosis increased appetite dramatically after TIPS. We also used renal artery embolization to treat polycystic kidney disease. After comparison between pre-operative and 6-month post-operative images, the total kidney volume decreased in size significantly (2979.8±2069.6 vs. 4019.4±2438.2, p < 0.001), and the psoas muscle area increased significantly (13.6±7.7 vs. 11.6±6.7, p < 0.001). In addition, the patients with polycystic kidney disease reduced abdominal distention after renal artery embolization.
Conclusion
Our study demonstrated that different diagnostic criteria for sarcopenia should be selected correctly according to different populations. Among the abdominal muscle groups, sarcopenia defined by psoas muscle mass was an independent predictor of poor prognosis for advanced HCC and was the most representative in patients receiving peritoneal dialysis. For CT and MRI images, we can perform automatic muscle segmentation accurately and concordantly with deep learning. We also can use interventional procedures to treat the underlying disease and improve sarcopenia simultaneously.

並列關鍵字

Computed tomography ； Magnetic resonance imaging ； Sarcopenia ； Pancreatic cancer ； Hepatocellular carcinoma ； Renal failure ； Deep learning

參考文獻

1. Cruz-Jentoft AJ, Baeyens JP, Bauer JM, Boirie Y, Cederholm T, Landi F, et al. Sarcopenia: European consensus on definition and diagnosis: Report of the European Working Group on Sarcopenia in Older People. Age Ageing. 2010 Jul;39(4):412–23.

2. Bekfani T, Pellicori P, Morris DA, Ebner N, Valentova M, Steinbeck L, et al. Sarcopenia in patients with heart failure with preserved ejection fraction: Impact on muscle strength, exercise capacity and quality of life. International journal of cardiology. 2016 Nov 1;222:41–6.

3. Prado CMM, Lieffers JR, McCargar LJ, Reiman T, Sawyer MB, Martin L, et al. Prevalence and clinical implications of sarcopenic obesity in patients with solid tumours of the respiratory and gastrointestinal tracts: a population-based study. Lancet Oncol. 2008 Jul;9(7):629–35.

4. Hanai T, Shiraki M, Nishimura K, Ohnishi S, Imai K, Suetsugu A, et al. Sarcopenia impairs prognosis of patients with liver cirrhosis. Nutrition. 2015 Jan 1;31(1):193–9.

5. Pereira RA, Cordeiro AC, Avesani CM, Carrero JJ, Lindholm B, Amparo FC, et al. Sarcopenia in chronic kidney disease on conservative therapy: prevalence and association with mortality. Nephrol Dial Transplant. 2015 Oct;30(10):1718–25.