The Value of Artificial Intelligence-Assisted Imaging in Identifying Diagnostic Markers of Sarcopenia in Patients with Cancer

Sarcopenia is defined as the loss of skeletal muscle mass and muscle function. It is common in patients with malignancies and often associated with adverse clinical outcomes. The presence of sarcopenia in patients with cancer is determined by body composition, and recently, radiologic technology for the accurate estimation of body composition is under development. Artificial intelligence- (AI-) assisted image measurement facilitates the detection of sarcopenia in clinical practice. Sarcopenia is a prognostic factor for patients with cancer, and confirming its presence helps to recognize those patients at the greatest risk, which provides a guide for designing individualized cancer treatments. In this review, we examine the recent literature (2017-2021) on AI-assisted image assessment of body composition and sarcopenia, seeking to synthesize current information on the mechanism and the importance of sarcopenia, its diagnostic image markers, and the interventions for sarcopenia in the medical care of patients with cancer. We concluded that AI-assisted image analysis is a reliable automatic technique for segmentation of abdominal adipose tissue. It has the potential to improve diagnosis of sarcopenia and facilitates identification of oncology patients at the greatest risk, supporting individualized prevention planning and treatment evaluation. The capability of AI approaches in analyzing series of big data and extracting features beyond manual skills would no doubt progressively provide impactful information and greatly refine the standard for assessing sarcopenia risk in patients with cancer.


Introduction
Sarcopenia was first introduced by Dr. Irwin Rosenberg in 1989, who described it as "age-related loss of skeletal muscle" [1]. It was initially regarded as the progressive decline in skeletal muscle mass, muscle strength, and physical performance associated with aging [2], but the definition and management of sarcopenia have expanded in recent years. In today's broader view, besides associations with aging, the shared risk factors for development of sarcopenia include chronic diseases, nutrition deficiencies, physical inactivity, hormonal changes, insulin resistance, loss of the neurons that stimulate muscle, and fat infiltration into muscle [3]. Among possible comorbidities, malignancy is a major category of diseaserelated sarcopenia. The causes of muscle loss in patients with cancer are multifactorial, especially in older adults [4].
Gender differences have been found in the prevalence of sarcopenia for people younger than 70 years and those older than 80 years; sarcopenia is diagnosed more often in women in those aged <70 years, while among those aged >80 years, more men will have sarcopenia than women [5]. This gender difference is clearly influenced by age, and sarcopenia must be considered when evaluating people of all ages who have cancer.
The etiology of sarcopenia in patients with cancer may vary between different ages and genders and can be associated with genetic predisposition, underlying comorbidities, reduced physical performance, and age-related declines in various hormones. Cancer-induced inflammatory cytokines and anorexia that cause decreased protein intake and synthesis and increased protein degradation may also be markers of sarcopenia in cancer patients. Treatmentrelated causes may include the side effects of chemotherapy, surgery, or radiotherapy [4,6].
Sarcopenia is a prognostic factor for patients with cancer, and confirming its presence helps to recognize those patients at the greatest risk and to guide individualized cancer treatment [7]. The diagnosis of sarcopenia is determined through the assessment of body composition (analysis of adipose and muscle tissue components), and recently, artificial intelligence-(AI-) assisted image measurement is being used to facilitate the detection of sarcopenia in clinical practice [8].
The purpose of this review was to synthesize current information in recent studies addressing AI-assisted imaging assessment of body composition and sarcopenia, particularly to gain a clearer understanding of the mechanism and the importance of sarcopenia in cancer and its diagnostic image markers and interventions for sarcopenia in the medical care of patients with cancer.

Literature Review
We searched the recent literature in PubMed from 2017 to 2021 using ("deep learning"[MeSH Terms] OR ("deep"[All Fields] AND "learning"[All Fields]) OR "deep learning"[All Fields]) AND ("sarcopenia"[MeSH Terms] OR "sarcopenia"[All Fields]). A total of 28 articles addressing AIassisted imaging assessment of body composition and sarcopenia were found, of which 20 reporting DICE coefficients were finally included for review. They are discussed below along with other supportive studies for background, focusing on cancer-related sarcopenia and the current status of AI-assisted imaging in the evaluation of sarcopenia in cancer patients.

The Definition/Mechanism of Sarcopenia in Cancer
Patients. Complex metabolic pathways are involved in the development process of sarcopenia. Several discriminating metabolites have been identified and investigated as potential biomarkers for the presence of sarcopenia. For example, one study demonstrated that low levels of plasma lysophosphatidylcholine 18 : 2 predict a greater decline of gait speed in older adults [9]. Another study reported that increased asparagine, aspartic acid, citrulline, ethanolamine, glutamic acid, sarcosine, and taurine are found in older adult patients with sarcopenia [10]. As for patients with cancer, a serum and urine metabolomics study found that cancer-related metabolic reprograming may represent a distinct diagnostic model [11].

Diagnostic Image Markers for Sarcopenia.
In clinical practice, assessment techniques for sarcopenia include handgrip strength to measure muscle strength and gait speed and chair stand tests to evaluate physical performance [12]. Bioimpedance analysis and dual-energy X-ray absorptiometry are the most common diagnostic tools for confirmation of muscle quantity and quality [13]. In the field of oncology, the use of abdominal computed tomography (CT) to measure body composition helps to identify sarcopenia in patients with cancer by providing precise and simplified data for describing sarcopenia and its correlation with clinical factors [14]. Thus, the performance of routine abdominal CT at cancer diagnosis, posttreatment evaluation, and regular follow-up provides the means for gauging body composition throughout the course of cancer.
The cross-sectional area (CSA) of muscle tissue at the level of the 3 rd lumbar spine (L3) provides reproducible evaluation of muscle size in cancer patients without the need for additional examinations. The measurements collected from a single slice CT image reveal solid evidence that correlates strongly with whole-body adipose tissue and skeletal muscle [15][16][17]. The common method is to manually draw the total CSA of all muscle groups at L3 or to quantify the CSA using thresholds of Hounsfield units (HU) from -29 to 150 for skeletal muscle using the available software [18]. The third lumbar vertebra, L3, is chosen because it is the current gold standard for quantification of muscle mass by obtaining parameters from the analysis of a single-slice CT scan [19]. The cross-sectional skeletal muscle area (SMA) calculated at the level of L3 can correctly estimate total body muscle mass [17]. A review has shown that attempts to use alternate vertebral levels to L3 (cervical, thoracic, and lumbar CT slices) for evaluating SMA in cancer patients have shown no validation of whole-body skeletal muscle mass in various types of cancer (lung, head, and neck) and a lack of consensus [20]. The skeletal muscle index (SMI, cm 2 /m 2 ) is calculated by dividing the CSA by the square of body height with various cut-off values according to gender and different body mass index (BMI ≧ 25:0 or <25.0) [21]. The formula used was SMI = L3 skeletal muscle CSA ðcm 2 Þ/height 2 ðm 2 Þ. The muscle groups for SMI consist of psoas major, paraspinal muscle, and abdominal wall muscles ( Figure 1). The solitary muscle indices such as psoas muscle index (PMI) and paravertebral muscle index (PSMI) also achieve good performance for sarcopenia evaluation [16,22,23]. The CT-derived measurement of muscle mass is usually evaluated using the method with thresholds of HU from -29 to 150 that will limit the evaluation of myosteatosis (fat infiltrates into muscle) technologically. The patients with higher BMI had greater SMI but lower skeletal muscle density (SMD) [24,25]. In the future, CT-derived measurement of muscle mass (area) and quality (myosteatosis) could be achieved with fully automated segmentation for contouring of muscle groups using deep learning systems [26]. The presence of sarcopenia in older adults may manifest as impaired daily function, disability, increased falls, risk of fractures, loss of independence, poorer quality of life, increased mortality, and high healthcare expenditures [27][28][29][30][31]. In patients with malignancies, sarcopenia is strongly associated with poor oncologic outcomes. A meta-analysis of 4262 participants with ovarian cancer revealed a significant association between the SMI and overall survival (OS) (P = 0:007; hazard ratio (HR): 1.11; 95% confidence interval (CI): 1.03-1.20) [32]. Another meta-analysis of 5497 participants with breast cancer reported similar result (pooled HR: 1.71; 95% CI: 1.25-2.33) [33]. Sarcopenia is also an independent predictor of treatment-related toxicities, including surgical complications, prolonged hospitalization, and more adverse effects of chemotherapy. A cohort study of 234 patients undergoing liver resection for malignant tumors demonstrated that sarcopenic patients had longer hospital stays (10 days vs. 6-8 days; P < 0:001) and more readmission (8.8% vs. 0-7.7%; P = 0:02) than those without sarcopenia [34]. A study of 533 patients with nonmetastatic colon cancer receiving a FOL-FOX regimen reported that lower muscle mass is associated with early discontinuation of chemotherapy (odds ratio (OR): 2.34; 95% CI: 1.04-5.24; P = 0:03), treatment delay (OR: 2.24; 95% CI: 1.37-3.66; P = 0:002), and dose reduction (OR: 2.28; 95% CI: 1.19-4.36; P = 0:01) [35].
Body weight or BMI as an indication of body composition was previously used to predict the clinical outcomes of patients with cancer [36,37]. Emerging evidence suggests that SMI correlates better with negative outcomes and complications than does BMI. A study of 484 patients with pancreatic cancer showed that the changes in BMI during chemotherapy did not have an impact on OS in patients with maintained SMI values (P = 0:750), while decreases in SMI were associated with poor OS in patients with maintained BMI (HR: 1.502; P = 0:002) [38]. This can be explained by the fact that patients with the same BMI may have different SMI values due to different amounts of muscle mass and differences in the level of fat infiltration. Similarly, patients with the same body surface area (BSA) but different SMI value receiving the same dose of chemotherapy may have different severity of adverse effects [39,40].

Interventions for Sarcopenia within the Medical Care of
Patients with Cancer. The prevalence of sarcopenia in patients with cancer ranges widely from 16% to 71%, depending on the definition in various study settings [7]. The understanding of the presence and the progression of sarcopenia helps to identify high-risk patients and guide the development of treatment plans. Since sarcopenia is significantly associated with treatment-related toxicity [34,35], the dose titration of chemotherapy, the intensity of surgical intervention, and the schedule of postoperative care should be carefully assessed in sarcopenic patients. For the impact of sarcopenia on oncologic outcomes, it also implies the physician about the disease explanation, prognosis expectation, and treatment decision-making.
The interventions for sarcopenia in patients with cancer include nutritional support, resistance exercise, and specific treatments for sarcopenia and the underlying disease [6,[41][42][43][44][45]. Many studies support the use of nutritional supplements [45], pharmacologic agents to increase muscle mass [44], and exercise programs [42]. Some studies show conflicting results for interventions for increasing muscle mass [6,41], and the impact of those interventions on clinical outcomes is still being investigated. Prospective studies on interventions for sarcopenia in patients with cancer are limited.

Medical AI Perspectives in the Diagnosis of Sarcopenia.
The present review identified a total of 20 articles reporting DICE similarity coefficient scores [16,19,[46][47][48][49][50][51][52][53][54][55][56][57][58][59][60][61][62][63]. Table 1 lists the included articles with the population characteristics and segmentation approaches. The reported CT threshold and DICE coefficients of these included studies ranged between 0.93 and 0.98 (Table 2), indicating great promise in the clinical application of AI-assisted imaging. However, as shown in Tables 1 and 2, there is currently no standardized methodology for assessment of sarcopenia. The slicing regions, methods of segmentation, tissues of interest, and ground truth applied varied between the studies. A total of 18 articles used deep learning methods to perform automated segmentation (16 applied fully convolutional networks (FCN) or U-Net, and 2 used ResNet-18). The region of segmentation varied across different systems, but the L3level axial slice was analyzed most frequently due to its strong correlation with whole-body composition [19]. As reference for segmentation (ground truth), 10 studies reported use of a combination of automated or semiautomated commercial segmentation software or cloud-based annotation tool with manual correction; 1 study specified that expert-labeled annotation was used as ground truth; details of the ground truth reference was not specified in the remaining articles (Table 1). Thirteen studies reported CT threshold HU values. However, the CT threshold is likely affected by whether or not contrast medium was used for imaging. Of the 20 articles reporting DICE scores, 10 articles reported DICE coefficients for skeletal muscle only; in the other 10 articles, tissues including visceral adipose tissue, subcutaneous adipose tissue, and intermuscular adipose Figure 1: The muscle groups for the skeletal muscle index consist of psoas major (green), quadratus lumborum (blue), erector spinae (red), and abdominal wall muscles (transversus abdominis muscle, internal and external oblique muscle (yellow), and rectus abdominis (purple)).    (2) Analysis was restricted to a single slice using a 2D U-Net architecture (3) All image labels were performed by two radiologists, and disagreement was solved by consensus, without documenting the disagreement systematically 9

10
Disease Markers tissue were also analyzed. Most of the articles reported training and testing cohort results only; 7 studies performed independent validations (internal or external) ( Table 2). In the evaluation of sarcopenia, abdominal musculature segmentation is accomplished using deep learning with a DICE similarity coefficient of 0.93-0.98 [46,48]. Successful individual segmentation of different muscle groups for SMI are achieved using a DICE similarity coefficient of 0.82-0.95, consisting of psoas major, quadratus lumborum, erector spinae (paraspinal muscle), and abdominal wall muscles (transversus abdominis muscle, internal and external oblique muscle, and rectus abdominis) [47]. The highly accurate segmentation of individual muscle groups provides an opportunity to assess muscle mass and myosteatosis separately. The area of muscular CSA could be reserved for mass evaluation. Using the cut point of CT HU inside the segmented CSA is aimed at assessing myosteatosis [64]. The CT-derived measurement of myosteatosis is associated with cut points of muscle attenuation less than 41 or less than 33 HU, which is consistent with the most common threshold for low-density muscle (0-30 HU) [64]. Knowledge about changes in body composition during cancer treatments and the disease course is currently lacking. The lack of standardized assessment method to determine muscle mass in cancer patients is evident from the varied cut-off values used in different studies, even for the same cancer type (as reviewed by Rier et al. [65] in 2016). The variations in cut-off value between the same cancer types likely have resulted from the different population characteristics between studies including age, BMI, disease severity, and different methods of evaluation [65]. Recent studies have focused on developing reference diagnostic cut-off values among the normal population. For people under 60 years old, the cut-off SMI value ranged between 40 and 45 in male and 30 and 35 in female (Supplement Table 1) [66][67][68][69][70][71][72]. However, the population characteristics were different between these studies, and determination of normal reference cut-off values for different population characteristics using larger series of data via an AI-assisted approach may fasten the development of standardized assessment. AI-assisted body composition measurement would increase the accuracy and efficiency of the sarcopenia evaluation and provides a trend of standardization by which the serial changes in cancer-related sarcopenia are explored [26].

Conclusion
In conclusion, the presence of sarcopenia is represented by prognostic and predictive values in patients with cancer. AI-assisted image analysis is a reliable automatic technique for segmentation of abdominal adipose tissue with the potential to improve diagnosis of sarcopenia and facilitates identification of oncology patients at the greatest risk, supporting individualized prevention planning and treatment evaluation. The capability of AI approaches in analyzing series of big data and extracting high-level abstractions beyond manual skills would no doubt progressively provide impactful information and greatly refine the standard for assessing sarcopenia risk in patients with cancer.

Data Availability
No data were used to support this study.