Elsevier

Nutrition

Volume 89, September 2021, 111227
Nutrition

Review
Use of artificial intelligence in the imaging of sarcopenia: A narrative review of current status and perspectives

https://doi.org/10.1016/j.nut.2021.111227Get rights and content

Highlights

  • Sarcopenia has an impact on outcomes in various diseases, not only in elderly people.

  • Artificial intelligence could be used to help evaluate sarcopenia.

  • Artificial-intelligence systems can alter the clinical importance of imaging modalities in diagnosing sarcopenia.

  • Automated methods in the evaluation of sarcopenia can potentially give us more information about patients and make clinical practice more tailored.

Abstract

Sarcopenia is a muscle disease which previously was associated only with aging, but in recent days it has been gaining more attention for its predictive value in a vast range of conditions and its potential link with overall health. Up to this point, evaluating sarcopenia with imaging methods has been time-consuming and dependent on the skills of the physician. The solution for this problem may be found in artificial intelligence, which may assist radiologists in repetitive tasks such as muscle segmentation and body-composition analysis. The major aim of this review was to find and present the current status and future perspectives of artificial intelligence in the imaging of sarcopenia. We searched the PubMed database to find articles concerning the use of artificial intelligence in diagnostic imaging and especially in body-composition analysis in the context of sarcopenia. We found that artificial-intelligence systems could potentially help with evaluating sarcopenia and better predicting outcomes in a vast range of clinical situations, which could get us closer to the true era of precision medicine.

Introduction

Artificial intelligence (AI) has gained more and more interest in recent years due to continuous improvements in computer science. Currently it is used in almost every aspect of our lives: social media, web searching, e-mail, online stores, autonomous vehicles, data management, and different fields of science, including medicine. The field of medicine in which AI could be especially valuable is diagnostic imaging. It may facilitate diagnosis of different conditions [1], [2], [3], [4], [5], or assessment of body composition [6], [7], [8], [9], [10], [11], which may be further used to predict health outcomes in different diseases [12], [13], [14], [15], [16], [17], [18]. Body-composition analysis is the area of research that refers to quantitative and qualitative measurements of the tissues that make up the human body [7]. When we talk about it, we mostly focus on the analysis of adipose and muscle tissue components. When we take muscles into consideration, there is one term that lately has been increasingly attracting the focus of body-composition researchers around the world: sarcopenia. Initially, it was defined simply as a loss of muscle in older people, but in the last decade significant improvements have been made in terms of defining and managing this condition [19]. Sarcopenia is considered a severe problem in aging populations; it is estimated that from over 50 million people in 2010, it might increase to 200 million in 2050 [20]. Estimated costs of managing sarcopenia in the United States in 2000 were at the level of $18.5 billion [21]. Unfortunately, more precise estimates, or ones considering other regions of the world, have not been made.

The European Working Group on Sarcopenia in Older People 2 (EWGSOP2) stated in 2018 that muscle quality and quantity are difficult to use as main indicators in defining sarcopenia owing to technical limitations [19]. To better manage the problem of sarcopenia, there is an urgent need to develop simple and accurate methods of assessing and quantitatively evaluating it, and AI can be a possible solution. AI may also help us to take low muscle quantity and quality from sarcopenia research to clinical practice. The primary aim of this review is to present the current status and perspectives of artificial intelligence in the imaging of sarcopenia.

Sarcopenia is defined as a muscle disease rooted in adverse muscle changes; its key characteristic is low muscle strength [19], which can be determined by muscle quality and quantity [20]. As mentioned earlier, at first it was only associated with aging, but in recent years its role has grown exponentially. Now sarcopenia is considered an imaging biomarker able to predict clinical outcomes in different conditions [10,[12], [13], [14], [15], [16], [17], [18],[22], [23], [24]]. It is not surprising that researchers have taken an interest in sarcopenia, given that we are starting to live in the era of precision medicine, which strongly correlates with big data management. One way in which data can be utilized is through AI.

AI and machine learning are very popular terms in recent years. Although they seem similar, they are not synonyms; their relationship is shown in Figure 1. Machine learning can be defined as methods which recognize and use certain patterns in different types of data to make predictions; sometimes it is able to find patterns that are not visible to human eyes. It is based on learning from data provided by researchers and mimics a part of human cognition. Thus, it is a field of AI, which is defined as systems or machines that can mimic human thought processes to achieve certain goals [7]. Machine learning can be divided into two categories: unsupervised and supervised learning [25]. When we give the algorithm manually labeled data—for example, a large number of computed tomography (CT) slices with marked muscles at the L3 level—it can learn from the labels and perform automated segmentation on the new set of data. This is called supervised learning, since supervisors teach the system how to perform certain tasks. When labeled data are not provided, and we let the algorithm work on its own to discover patterns and group them as it wants, it is called unsupervised learning. In medical imaging, supervised learning is more popular [[6], [7], [8], [9],25,26]. There are various types of machine-learning algorithms; the most popular are neural networks, k-nearest neighbors, support vector machines, decision trees, and naive Bayes algorithms [26]. In medical imaging, neural networks—defined as digital models which simulate the behavior of neurons in the human brain [7]—are especially popular. To be exact, convolutional neural networks are the most commonly used. Deep learning is another term connected with AI; it is a subset of machine learning, a special type of neural networks [25] that use multiple layers to extract higher-level features from each layer to give exact results with the best performance, just like human visual cortex. It is highly appreciated because of hardware and time limitations in creating AI algorithms. The use of AI is really promising in medicine, especially in diagnostic imaging, where big data sets are managed. It could help radiologists in their jobs by performing some repetitive and time-consuming tasks that now rest on their shoulders [25]. One such task is muscle segmentation, which is crucial for the process of diagnosing sarcopenia [6,27].

Sarcopenia has been proven to have an impact on outcomes in various diseases, not only in older people [12], [13], [14], [15], [16], [17], [18]. It is associated with elevated mortality rates, functional decline, increased rates of falls, and higher incidences of hospitalization [23]. The predictive value of sarcopenia has been established in people with chronic heart failure [12], gastric cancer [13], rectal cancer [15,17,18], and lung cancer [16]. The prevalence and predictive value of pretherapeutic sarcopenia in oncology patients is highest in the cohort with esophageal and small-cell lung cancer. Furthermore, it is associated with elevated rates of intraoperative complications, chemotoxicity, and poor survival in different types of cancer [14]. The prognostic value of sarcopenia has also been established in people treated with radiochemotherapy for locally advanced esophageal cancer [28] and stage III melanoma [29]. Muscle composition could be used as a predictor for overall survival in advanced ovarian cancer [30]. Moreover, sarcopenia evaluated by imaging methods has predictive value for postoperative formation of pancreatic fistulas after pancreaticoduodenectomy [31], mortality after endovascular aneurysm repair [32], and advanced gastric cancer [33]. It additionally has clinical implications (such as worse outcomes and shorter overall survival) for people undergoing complete resection of non-small cell lung cancer [34] and prognostic value for adults with solid tumors [35]. Further, sarcopenia has an impact on clinical outcomes in breast cancer [36], chronic liver disease [37], orthotopic liver transplantation [38], esophageal cancer [39,40], and proximal femur fractures in older people [41]. If we consider sarcopenia as an imaging biomarker that can alter clinical results, it becomes clear that we need a widely available and easy method for evaluating it.

Currently, the EWGSOP2 recommends an algorithm for case-finding and diagnosis of sarcopenia in which both non-imaging and imaging methods are used [19]. The original European Working Group on Sarcopenia in Older People published in 2010 a definition of sarcopenia that aimed to foster advances in identifying and caring for people with the condition. In early 2018, the group met again (EWGSOP2) to update the original definition to reflect the scientific and clinical evidence that had built up over the previous decade, presenting updated findings. Its objectives were to increase consistency of research design and clinical diagnoses and ultimately improve the care of people with sarcopenia. Sarcopenia was defined as a muscle disease (muscle failure) rooted in adverse muscle changes that accrue across a lifetime; it is common among adults of older age, but can also occur earlier in life. In this updated consensus paper, the EWGSOP2 focused on low muscle strength as a key characteristic of sarcopenia, used detection of low muscle quantity and quality to confirm the diagnosis, and identified poor physical performance as indicative of severe sarcopenia; updated the clinical algorithm that could be used for sarcopenia case-finding, diagnosis and confirmation, and severity determination; and provided clear cutoff points for measurement of variables that identify and characterize sarcopenia. The conclusions of EWGSOP2’s updated recommendations aimed at increasing awareness of sarcopenia and its risks. With these new recommendations, EWGSOP2 called for health care professionals who treat people at risk for sarcopenia to take actions that promote its early detection and treatment. It also encouraged further research in the field of sarcopenia to prevent or delay adverse health outcomes that incur a heavy burden for individuals and health care systems [19].

In modern clinical practice, non-imaging methods are sufficient to assess sarcopenia and begin intervention [19]. These methods include questionnaires, muscle physical performance tests, anthropometric measures, bioelectrical impedance analysis, and serum or urinary biomarkers [20]. They are usually cheaper than imaging methods, but unfortunately they are not always as accurate as we expect. The EWGSOP2 recommends the SARC-F questionnaire to identify patients who could develop or have already developed sarcopenia, but it has low to moderate sensitivity and it will mostly detect only severe cases [19]. Furthermore, it strongly depends on neurologic health—in cases of dementia or other states that lower individuals’ cognition, it might be challenging to fill out the SARC-F questionnaire. Physical performance tests such as grip strength, the chair stand test, gait speed, and the Short Physical Performance Battery are effective but have their own downsides, for instance false positive results due to orthopedic or neurologic disorders [20]. Bioelectric impedance analysis is used for confirmation of sarcopenia, but since it is not an imaging method, it is described in this section. It is a straightforward and popular method for calculating muscle and fat mass through the different resistances of various tissues in the human body. Regrettably, diverse factors can bias the results, such as hydration status, exercise, and food intake [42]. Lastly, anthropometric measurements are cheap and can be performed in every situation, but their correlation with actual muscle mass is poor [19].

Another group of methods which at present are used mainly for confirmation of sarcopenia are imaging modalities such as dual X-ray absorptiometry, ultrasonography, magnetic resonance imaging (MRI), and CT. Except for dual X-ray absorptiometry, they can provide information on muscle quality, including the highly relevant problem of myosteatosis [43]. Dual X-ray absorptiometry has been widely used in clinical trials for the evaluation of lean and fat mass [20]. It is the most frequently used radiologic technique for assessing body composition [22], because of its low radiation exposure, low cost, and short time of image acquisition [42], but it also has limitations. It does not measure muscle quality, and it might be influenced by hydration status [22]; it also cannot be used for quantitative evaluation of sarcopenia. Moreover, accurate measurement of the trunk muscles and adipose tissue is impossible [42]. It is not as precise as CT or MRI, which are regarded as a gold standard for the assessment of muscle quantity and quality.

The next method, ultrasonography, may be considered an effective tool for evaluating muscle mass and composition because of its reasonable costs, portability, lack of radiation exposure, and ability to provide a large amount of data on muscles. On the other hand, it requires experience, and its accuracy depends on the specific operator and recipient; in consequence, its reproducibility is poor [20,22,42]. Moreover, CT and MRI may be used to assess both muscle quantity and quality, with similar accuracy, but MRI provides more information about muscle composition. The most significant advantages of CT over MRI are that it is more frequently performed and cheaper. The biggest disadvantage of CT is high radiation exposure and complex postprocessing. By contrast, MRI is not linked to radiation exposure, but postprocessing is as complicated as in CT, and there are more contraindications for MRI than for CT. Both modalities depend on cross-sectional or 3D imaging, and neither has established cutoff points for diagnosing sarcopenia [20,22,42]. Although nowadays these methods are used to confirm sarcopenia, thanks to the rapid development of AI research they could potentially become used in the future in clinical practice for finding and assessing sarcopenia.

In recent years, CT has become a routine diagnostic tool for different diseases, and almost every oncology patient has at least one CT scan performed. Up to this point, some part of the data from these examinations has not been analyzed, such as quantitative and qualitative data on muscle and adipose tissue, but as previously discussed, their impact can be bigger than we thought. Apart from simple muscle quantification, CT may be used to get information about the quality of muscle by assessing muscle attenuation, which correlates with myosteatosis. The most frequently analyzed regions for sarcopenia evaluation in CT scans are the midthigh and the lumbar spine region, especially the L3 vertebra level total abdominal muscle area and the psoas muscle area [22,42]. Sarcopenia assessment is performed by segmentation of the muscles and adipose tissue in these specific regions, along with measurement of their cross-sectional area and sometimes also muscle attenuation [6,8,9,11,22,27,43]. As a more objective measure, the skeletal muscle index is used, calculated as the ratio of cross-sectional area to the square of the individual's height [22].

Moving to details, it was discovered that the cross-sectional area of muscle and adipose tissue compartments at the L3 vertebra level strongly correlate with total body composition [19], and a single-slice CT analysis allows an accurate assessment of sarcopenia and body composition [27]. Nonetheless, there is a problem with a variety of sarcopenia cutoff values and segmentation procedures. Researchers have not reached consensus as to which thresholds and methods are best for sarcopenia evaluation [42], but the most common CT cutoff values of skeletal muscle index range from 52 to 55 cm2/m2 for men and from 39 to 41 cm2/m2 for women [22]. In addition, body-composition analysis can be done on MRI scans. MRI is considered the most sophisticated method for evaluating muscle quality, as it provides information about myosteatosis similarly to CT but also allows for the detection of myofibrosis, abnormal edema, and inflammation of the muscle [20,42].

There are a few different ways segmentation can be done. One is manual segmentation, which yields good results but is highly time-consuming and depends on the skills of the reader; its reproducibility is therefore poor. Image thresholding and region growing can also be used; they are the oldest semiautomated methods for image segmentation [7]. They must be supervised by a radiologist and tend to differ from patient to patient. Multiatlas segmentation is another possible approach, more frequently applied to MRI. In this method, atlases with defined anatomic labels are manually created, then transferred to new subject images, where the algorithm classifies muscle and adipose tissue by a voting scheme. It is a fast and accurate method, but anatomic variations might cause problems with atlas fitting [44]. The next technique uses artificial intelligence, mostly machine-learning algorithms, especially the U-net model [6], [7], [8], [9]. Segmentation can also be performed by fast graph-based algorithms [11] or other methods of digital image processing [45], [46], [47], [48], [49], [50]. Figure 2 shows automated segmentation of visceral and subcutaneous adipose tissue on CT scans [11].

As mentioned earlier, sarcopenia has an impact on various conditions and diseases, and therefore it is not surprising that researchers are trying to find better and faster ways to identify and confirm sarcopenia. It is also not surprising that they are using artificial intelligence to do so. Some research concerning the use of artificial intelligence in assessing body composition or sarcopenia has already been published.

Section snippets

Methods

We thoroughly searched the PubMed database (last search on July 11, 2020) using the following keywords: “sarcopenia AND artificial intelligence,” “automated AND sarcopenia,” “automated AND body composition analysis,” “automated AND muscle segmentation,” and “automated AND fat infiltration.” We analyzed only original articles in English from the last 5 y that presented methods for automated evaluation of sarcopenia or techniques that might be used for automated assessment of sarcopenia (as

Results

We found 23 articles: 13 described methods based on CT examination [6,[8], [9], [10],48,[51], [52], [53], [54], [55], [56], [57], [58]], 2 on low-dose CT examination [59,60], and 8 on MRI examination [[45], [46], [47],49,50,[61], [62], [63]]. As expected, studies using CT were more common, probably due to more frequent use of this imaging modality in clinical practice. Eleven articles described solely methods of skeletal muscle segmentation [6,9,52,53,55,[57], [58], [59], [60], [61],63]; the

Discussion

As we can see, models that are capable of automated evaluation of muscles and adipose tissue have already been designed, and although this is a novel research interest, it has already obtained very promising results. It requires multidisciplinary cooperation between diagnostic-imaging specialists, clinicians, and machine-learning experts. Criteria that are used nowadays in assessing sarcopenia are straightforward and do not require great financial input. Some may say that this is sufficient and

Conclusions

Sarcopenia was formerly considered a simple geriatric syndrome, but recently it has been discovered that its essence may be more complex than we previously thought. It has an impact on the course of various diseases, from cardiovascular to the oncological, and even on overall morbidity. The cause of sarcopenia has not been completely understood yet, but we can already take it into consideration when planning clinical actions. Thanks to the use of artificial intelligence, the process of

References (70)

  • JS Lee et al.

    Prognostic significance of CT-determined sarcopenia in patients with advanced gastric cancer

    PLoS One

    (2018)
  • Y Suzuki et al.

    Clinical implications of sarcopenia in patients undergoing complete resection for early non-small cell lung cancer

    Lung Cancer

    (2016)
  • SS Shachar et al.

    Prognostic value of sarcopenia in adults with solid tumours: a meta-analysis and systematic review

    Eur J Cancer

    (2016)
  • F Rossi et al.

    Evaluation of body computed tomography-determined sarcopenia in breast cancer patients and clinical outcomes: a systematic review

    Cancer Treat Res Commun

    (2019)
  • YX Yang et al.

    Automated assessment of thigh composition using machine learning for Dixon magnetic resonance images

    MAGMA

    (2016)
  • S Dabiri et al.

    Muscle segmentation in axial computed tomography (CT) images at the lumbar (L3) and thoracic (T4) levels for body composition analysis

    Comput Med Imaging Graph

    (2019)
  • P Blanc-Durand et al.

    Abdominal musculature segmentation and surface prediction from CT using deep learning for sarcopenia assessment

    Diagn Interv Imaging

    (2020)
  • MT Paris et al.

    Automated body composition analysis of clinically acquired computed tomography scans using neural networks

    Clin Nutr

    (2020)
  • R Barnard et al.

    Machine learning for automatic paraspinous muscle area and attenuation measures on low-dose chest CT scans

    Acad Radiol

    (2019)
  • EJ Limkin et al.

    Promises and challenges for the implementation of computational medical imaging (radiomics) in oncology

    Ann Oncol

    (2017)
  • DY Yoon et al.

    Comparison of low-dose CT and MR for measurement of intra-abdominal adipose tissue: a phantom and human study

    Acad Radiol

    (2008)
  • G Wang et al.

    Interactive medical image segmentation using deep learning with image-specific fine tuning

    IEEE Trans Med Imaging

    (2018)
  • KH Cha et al.

    Bladder cancer segmentation in CT for treatment response assessment: application of deep-learning convolution neural network—a pilot study

    Tomography

    (2016)
  • S Trebeschi et al.

    Deep learning for fully-automated localization and segmentation of rectal cancer on multiparametric MR

    Sci Rep

    (2017)
  • M Kallenberg et al.

    Unsupervised deep learning applied to breast density segmentation and mammographic risk scoring

    IEEE Trans Med Imaging

    (2016)
  • B Wang et al.

    Artificial intelligence in the evaluation of body composition

    Semin Musculoskelet Radiol

    (2020)
  • AD Weston et al.

    Automated abdominal segmentation of CT scans for body composition analysis using deep learning

    Radiology

    (2019)
  • PM Graffy et al.

    Deep learning-based muscle segmentation and quantification at abdominal CT: application to a longitudinal adult screening cohort for sarcopenia assessment

    Br J Radiol

    (2019)
  • EM Cespedes Feliciano et al.

    Evaluation of automated computed tomography segmentation to assess body composition and mortality associations in cancer patients

    J Cachexia Sarcopenia Muscle

    (2020)
  • PD Lopez et al.

    Low skeletal muscle mass independently predicts mortality in patients with chronic heart failure after an acute hospitalization

    Cardiology

    (2019)
  • JR Lieffers et al.

    Sarcopenia is associated with postoperative infection and delayed recovery from colorectal cancer resection surgery

    Br J Cancer

    (2012)
  • DR Dolan et al.

    The relationship between sarcopenia and survival at 1 year in patients having elective colorectal cancer surgery

    Tech Coloproctol

    (2019)
  • AJ Cruz-Jentoft et al.

    Sarcopenia: revised European consensus on definition and diagnosis

    Age Ageing

    (2019)
  • RD Boutin et al.

    Sarcopenia: current concepts and imaging implications

    AJR Am J Roentgenol

    (2015)
  • I Janssen et al.

    The healthcare costs of sarcopenia in the United States

    J Am Geriatr Soc

    (2004)
  • Cited by (0)

    View full text