ARTICLE IN PRESS

Original Investigation

A Machine Learning Algorithm to

Estimate Sarcopenia on Abdominal CT
Joseph E. Burns, MD, PhD, Jianhua Yao, PhD, Didier Chalhoub, MD, PhD, MPH, Joseph J. Chen, BS,
Ronald M. Summers, MD, PhD

Abbreviations Rationale and Objectives: To assess whether a fully-automated deep learning system can accurately
detect and analyze truncal musculature at multiple lumbar vertebral levels and muscle groupings on
SMI abdominal CT for potential use in the detection of central sarcopenia.
skeletal muscle index
Materials and Methods: A computer system for automated segmentation of truncal musculature
L3MI groups was designed and created. Abdominal CT scans of 102 sequential patients (mean age 68 years,
L3 muscle index range 59
81 years; 53 women, 49 men) conducted between January 2015 and February 2015 were
TPA assembled as a data set. Truncal musculature was manually segmented on axial CT images at multiple
total psoas area lumbar vertebral levels as reference standard data, divided into training and testing subsets, and ana-
TPI lyzed by the system. Dice similarity coefficients were calculated to evaluate system performance. IRB
total psoas index approval was obtained, with waiver of informed consent in this retrospective study.

DSC Results: System performance as gauged by the Dice coefficients, for detecting the total abdominal
Dice similarity coefficient muscle cross-section at the level of the third and fourth lumbar vertebrae, were, respectively, 0.953 §
0.015 and 0.953 § 0.011 for the training set, and 0.938 § 0.028 and 0.940 § 0.026 for the testing set.
ICGCC
Dice coefficients for detecting total psoas muscle cross-section at the level of the third and fourth lum-
The International Consensus
bar vertebrae, were, respectively, 0.942 § 0.040 and 0.951 § 0.037 for the training set, and 0.939 §
Group for Cancer Cachexia
0.028 and 0.946 § 0.032 for the testing set.
Conclusion: This system fully-automatically and accurately segments multiple muscle groups at all
lumbar spine levels on abdominal CT for detection of sarcopenia.
Key Words: Adults; CT; Musculoskeletal; Oncology; Computer applications-detection/diagnosis.
© 2019 Published by Elsevier Inc. on behalf of The Association of University Radiologists.

INTRODUCTION sarcopenia, a risk factor associated with physical decline,

decreased quality of life, and increased mortality (4). Computed

S
arcopenia, initially defined as the loss of muscle mass
tomography (CT), the gold standard for muscle mass and body
by Rosenberg, has evolved as a syndrome to include
composition measurement, is the imaging tool of choice for
other muscle function measurements such as strength
opportunistic sarcopenia diagnosis. With over 70 million CT
and performance (1,2). In 2016, the CDC established an ICD-
scans conducted every year in the US alone (5), scans are readily
10 code for sarcopenia, making it a recognized medical condi-
available for sarcopenia screening justifying the need for a fully
tion (3). This crucial step is allowing clinicians to code for
automated deep learning system.
Central sarcopenia is a risk factor for mortality in multiple dis-
Acad Radiol 2019; &:1–10 ease processes, including cancer (6, 7). Varying criteria for deter-
Disclosures: Author Summers receives patent royalties from iCAD, Philips, mination of central sarcopenia on CT have been developed and
ScanMed, PingAn and research support from PingAn and NVIDIA. Author Yao linked to patient outcomes in clinical series. These criteria have
receives patent royalties from iCAD, PingAn, Imbio, Zebra Medical, and is cur-
rently employed in private industry, but the work described in the manuscript in common that they attempt to gauge the mass of truncal mus-
was all performed as part of his employment with the NIH, and is unrelated to cle groups as a presumed measure of patient physiologic
his subsequent private employment.
From the Department of Radiological Sciences, University of California-Irvine
reserve (8). The common metric association of these series is
School of Medicine, Orange, California (J.E.B., J.J.C.); Imaging Biomarkers that worsening patient outcomes are correlated with decreased
and Computer-Aided Detection Laboratory, Radiology and Imaging Sciences, cross-sectional area of tested truncal muscle groups.
National Institutes of Health Clinical Center, Building 10, 1C224, MSC1182,
Bethesda, MD 20892-1182 (J.Y., D.C., R.M.S.). Received February 18, 2019; The most common muscle groups tested in these series
revised March 5, 2019; accepted March 5, 2019. Address correspondence include the total axial cross section of truncal abdominal mus-
to: R.M.S. e-mail: [email protected]
culature also known as skeletal muscle area (SMA) normal-
© 2019 Published by Elsevier Inc. on behalf of The Association of University
Radiologists.
ized for patient height (skeletal muscle index
SMI or MI),
https://doi.org/10.1016/j.acra.2019.03.011 the total psoas muscle cross-section (total psoas area
TPA)

1
 ARTICLE IN PRESS
BURNS ET AL Academic Radiology, Vol &, No &&, && 2019

normalized for patient height (total psoas index
TPI), and specialties according to their own metric standards, for use in
the paraspinous muscle group, as measured on axial CT surgical planning and pretreatment risk assessment.
images. The SMI has been used as an imaging biomarker for The purpose of this paper is to assess whether a fully-auto-
mortality in lung cancer (L1 and L3), cirrhosis (L3), and mated deep learning system can be created to opportunistically
trauma patients (L3); blood transfusion and hospitalization in detect and analyze truncal musculature at set vertebral levels and
proximal femoral fracture patients (L4); and therapy monitor- for varied muscle groupings, for potential use in the detection of
ing in pancreatic cancer patients (L3) (9
15). The TPA mus- central sarcopenia using Computed Tomography (CT) scans.
cle group has been previously proposed for risk assessment in
elderly trauma (L3 level), liver transplant (L4), and in pancre-
atic (L3), esophageal (L3), and endometrial cancer (L3) MATERIALS AND METHODS
groups (16
20). The paraspinous muscle group has been dis-
Study Subjects
cussed as a central sarcopenia determinant metric for survival
and surgical outcomes in patients with colorectal cancer (L4), Approval of our Institutional Review Board was obtained,
liver transplant (T12), and general and vascular surgery and our study was Health Insurance Portability and Account-
patients (T12) (15,21,22). ability Act compliant. This retrospective study utilized previ-
The standard for segmentation used in the determination of ously performed imaging studies for system testing and
sarcopenia on CT is manual tracing of muscle group margins on analysis, and informed consent was waived.
axial sections, replicated over many patients. Preliminary work Searches of the National Institutes of Health Clinical
has been performed for automated segmentation of the total Center picture and archiving systems (PACS) database for
abdominal muscle cross section on a manually extracted CT CT examinations of the chest, abdomen, and pelvis, and
image at the L3 level (23). The previously discussed studies con- abdomen and pelvis, with intravenous contrast adminis-
ducted on patients with varying medical conditions, showed tered, were performed. Cases were reviewed sequentially
that different muscle groups assessed at distinct lumbar spine lev- in reverse chronological order, without regard for patient
els are linked to different outcomes. As such, a system which diagnosis. A flow chart of the case accumulation method-
would automatically detect vertebral levels and segment multi- ology is provided in Figure 1. The inclusion criteria for
ple combinations of muscle groups at those levels would allow scans were patient age 59 years or older, intravenous con-
the radiologist to provide directed information to multiple trast administration, and date of service for examinations

Figure 1. Flow chart as process illustration for case accumulation, exclusion, and partitioning as performed in this work.

2
 ARTICLE IN PRESS
Academic Radiology, Vol &, No &&, && 2019 MACHINE LEARNING ALGORITHM FOR SARCOPENIA

between January 2015 and February 2015. The exclusion pedicle), for a total of 10 manually segmented images per
criteria were motion artifact, paraspinous hardware, and patient. On each image, the individual muscle group
significant distortion of the abdominal wall by surgery or margins were marked and categorized into 10 separate
neoplasm. segmentation sets, as left and right posterior paraspinous
A PACS review of the CT examinations was performed, (iliocostalis, longissimus, multifidus, interspinalis), psoas,
and case exclusion criteria (Fig 1) were applied. A total of quadratus lumborum, lateral wall (transversus abdominis,
102 patients were assembled as a case set. obliquus internus, and obliquus externus), and rectus
abdominis muscles (Fig 2). Muscle set segmentation data
were secondarily sorted into three groups by the system
Abdominal Wall and Paraspinous Muscle Identification for stratified analyses for this work: a psoas muscle group
and Description
TPA (left and right psoas muscles) as in Figure 3a, a
posterior paraspinous muscle group (bilateral psoas, ilio-
PACS images were saved in noncompressed DICOM format.
costalis, longissimus, multifidus, interspinalis) as in
Axial CT images with soft tissue reconstruction kernel were
Figure 3b, and a total abdominal cross section group
used by the system for image analysis. The section thickness
(Skeletal muscle area
SMA) which includes all 10 indi-
of the abdominal protocol CT scans used was 2 mm, with an
vidual muscles as in Figure 3c. The percent difference
in-plane image resolution of 0.66
0.98 mm.
between the total muscle pixels generated by the system
After a period of instruction, and margins of the
and the sum of individual muscles’ pixels is 15%. This
abdominal wall and paraspinous muscle groups were elec-
may be explained by the overlap among different muscle
tronically marked and recorded by author J.J.C.
groups which resulted in counting pixels twice at individ-
(a trained medical student) utilizing ITK-SNAP. A pre-
ual muscles’ borders, and once for total abdominal
liminary set of files with coordinates of the electronic
muscles. These manually annotated muscle group data
markings of the muscle margins for all the images was
sets formed the reference-standard for model training and
thus created. Author J.E.B. (a musculoskeletal radiologist
assessment of system performance.
with 11 years’ experience) then independently and sepa-
rately reviewed the electronic coordinates marking the
muscle margins in each image, and corrected the coordi-
Quantitative Image Analysis Methodology
nate ranges where there was perceived mismarking, to
create the ground truth data set. The margins of multiple The system is designed to determine the degree of sarcopenia
muscle groups were electronically marked on two sepa- with a cascaded two-step fully automated processing architec-
rate axial images at the level of each lumbar vertebra (one ture. First, the lumbar vertebrae are detected, segmented, and
image at and one image off the level of the vertebral individually partitioned to form a craniocaudal dimension

Figure 2. Axial CT cross-sections of the abdomen illustrating the elemental muscle and muscle group segmentations as performed at the L1
though L5 lumbar vertebral levels, for the ground truth data set. Color key for muscle group identification is presented at lower right. Patient is
a 65-year-old male.

3
 ARTICLE IN PRESS
BURNS ET AL Academic Radiology, Vol &, No &&, && 2019

Figure 3. Muscle group models (demarked in red) for assessment of sarcopenia from the literature, on axial CT images, tested here. (a) The
total psoas index (TPI) muscle group. (b) The paraspinous muscle group. (c) The skeletal muscle index group (SMI). Patient is a 65-year-old
male. (Color version of figure is available online.)

measurement scale in each patient along the longitudinal axis of Segmentation accuracy was assessed using the Dice Simi-
the spine, forming a scaled reference axis. This reference axis is larity Coefficient (DSC), as a measure of spatial overlap
used by the system for localization to the correct craniocaudal between the manual and automated segmentations (27).
body levels for muscle segmentation. Second, a deep learning The system then calculated the axial muscle area from the
process is used to segment the truncal musculature on axial muscle group segmentations, and values were converted to
image sections. The cross-sectional muscle areas from this seg- lean muscle area using thresholding scripts to calculate the
mentation are then used to determine the degree of sarcopenia. percent infiltration of intramuscular fat within the segmenta-
The coding process chain by which the spine is segmented tion areas, to account for muscle quality. Finally, muscle
and the lumbar vertebrae are partitioned into individual units area values were renormalized by patient height for calcula-
to form the scaled reference axis include intensity threshold- tion of muscle indices (7).
ing and region growing, morphologic operations, and canal The computer system was a Linux based workstation uti-
extraction though acyclic graph search. The vertebrae are lizing an Intel Core i7 processor with 4x Titan X GPUs
then partitioned though aggregated intensity profiles. These (NVidia, Santa Clara, CA) and 12 GB memory per GPU.
processes have been detailed in previous publications (24). Time for training a model based on the training data set of 51
The proposed deep learning
based truncal muscle detec- patients was approximately 4 hours, with time for analysis of
tion system was designed to utilize the U-Net neural net- each testing case noted as less than 1 minute.
work model. The U-Net model was selected for automated
detection of the muscles due to its ease of use in combination
with precise and rapid segmentation (25). The architecture of Statistical Analysis
the U-Net deep learning model used is the same as that of
Ronneberger (25). The image analysis system framework was Baseline characteristics and DSCs were compared
designed using Python (version 2.7; Python Software Foun- between training and testing groups. T-test and Mann-
dation, Wilmington, Del) for algorithm development. Whitney U test were used for normally and non-normally
Uncompressed DICOM images were preprocessed, with con- distributed variables, respectively. Chi-square was calcu-
version to pixel matrix utilizing Python and the Pydicom library lated for categorical variables (Table 1). System perfor-
(version 1.2; Python Software Foundation) to form input data to mance was determined by comparison of the automated
the U-Net system (26). The model was trained de novo, with no (A) with manual (M) muscle group segmentation, using
pretrained elements. Hyperparameter values used during training the following DSC formula as a measure of precision of
include cross entropy for the loss function, and Adam’s method as the automated segmentation:
the optimizer. Initial learning rate was taken as 0.0005, with a 2ð A \ M Þ
decay of 0.9 every 10 epochs. The batch size was determined DSCðA; M Þ ¼ ; where \ is the intersection
ðA þ M Þ
through testing, optimized to a level where computer memory
could support the training, and set at 4. The system was trained DSCs boxplots for all individual muscles as well as paraspi-
for 1000 epochs, and the model with the smallest loss value, nous, TPA, and SMA muscle groups’ images at each of the
determined by the cross entropy, was selected. L1 though L5 vertebrae are presented in Figure 4. Stratifica-
Data, in the form of manual muscle segmentations at two tion by gender was conducted at each of the L1 through L5
axial levels for each lumbar vertebra in the training set, were vertebrae and presented in Figure 5. DSCs are reported for
used to form a model for the deep learning system. One sin- both the training and testing data sets for each grouping and
gle model was trained for all vertebra levels. The system then at each level. Secondary analyses were conducted to derive
performed muscle group segmentations at multiple lumbar Jaccard similarity coefficients (JSCs) from Dice coefficients
vertebral levels. A threshold of 0.5 was applied on the U-net using DSC/(2-DSC). All DSCs and JSCs for individual
output of muscle probability. The muscle area was then com- muscles and muscle groups are reported in Supplemental
puted as the total number of pixels in the segmented regions. Table S1.

4
 ARTICLE IN PRESS
Academic Radiology, Vol &, No &&, && 2019 MACHINE LEARNING ALGORITHM FOR SARCOPENIA

TABLE 1. Baseline Characteristics of Patients by Training and Testing Dataset

Characteristics Training (N = 51) Testing (N = 51) p Value

Age (years), mean § SD 67.94 § 5.51 65.39 § 5.90 0.01

Gender (male), N(%) 26 (51%) 23 (45%) 0.69
Height (m), mean § SD 1.69 § 0.12 1.67 § 0.11 0.28
Weight (kg), mean § SD 80.1 § 20.4 80.4 § 17.5 0.94
Dice similarity coefficients per muscley
Lateral wall, mean § SD 0.937 § 0.033 0.920 § 0.058 0.001
Paraspinous, mean § SD 0.950 § 0.029 0.933 § 0.030 <0.001
Psoas, mean § SD 0.924 § 0.085 0.914 § 0.088 0.006
Quadratus, mean § SD 0.819 § 0.182 0.803 § 0.190 0.091
Rectus, mean § SD 0.901 § 0.095 0.889 § 0.107 0.07
Dice similarity coefficients per lumbar spine levelz
L1, mean § SD 0.890 § 0.096 0.879 § 0.099 0.140
L2, mean § SD 0.922 § 0.073 0.917 § 0.055 0.008
L3, mean § SD 0.939 § 0.049 0.930 § 0.049 <0.001
L4, mean § SD 0.929 § 0.079 0.913 § 0.093 <0.001
L5, mean § SD 0.850 § 0.184 0.821 § 0.197 <0.001

*statistically significant with p < 0.05; SD: standard deviation.

y DSC for each muscle group averaged over all five lumbar levels.
z Average of all individual muscle groups’ DSCs at each lumbar level.

Determination of Sarcopenia Agreement Among Most muscles were the highest at the L3 (DSCtraining = 0.956;
Commonly Used Muscle Groups, at the L3 Level DSCtesting = 0.940), and L4 levels (DSCtraining = 0.958;
We sought a simple method to correlate for agreement between DSCtesting = 0.943). Quadratus muscles’ DSCs were the high-
the metrics using varied muscle group and spinal levels, for est at L3 for both the training (DSC = 0.945) and testing
patient stratification into potential sarcopenia cohorts. The L3 (DSC = 0.935) sets. For the rectus muscle, DSCs were high
SMI group (28) was used as the reference standard to which at the L1 level with the testing set (DSC = 0.915) performing
alternative metrics were compared. Taking into account the as well as the training set (DSC = 0.9100). The lowest level
patient data set size, a heuristic model of patient stratification of performance for individual muscles was for the quadratus
into lowest quartile cohorts for each metric was used. muscle at L5 with training and testing DSCs of 0.500 and
For each of the L3 muscle groupings, patients falling in the 0.479, respectively.
lowest quartiles were classified as sarcopenic. Using L3SMI as For muscle groups, the system’s highest levels of performance
the referent group, gender specific and overall sarcopenia were for SMA at L3-L4 levels for both the training and testing
kappa statistics for TPA and paraspinous muscle groups were sets, with DSCs ranging between 0.938 and 0.953. For the TPA
calculated (29). All statistics and figures were generated using and paraspinous muscle groups, the highest levels of perfor-
the RStudio statistical package (version 1.1.456). mance were at the L3-L5 and L2-L4 levels, respectively. The
lowest level of performance for the system was at the L1 level,
for all muscle groupings for both the training and testing sets,
RESULTS with DSCs ranging between 0.896 and 0.917 (Table S1).
Patient Cohort
The mean patient age was 66.7 § 5.8 years, with an age range Comparative System Performance for Male and Female
Patients Gauged by Dice Coefficients of Individually
of 59
81 years. There were 53 females and 49 males in the
Segmented Muscles
set. The mean age of the female patient cohort was 66.7 §
5.9 years, with an age range of 59
78 years. The mean age In males and females, system performance for the total
of the male patient cohort was 66.8 § 5.7 years, with an age abdominal group is highest at the L3 and L4 levels. DSCs
range of 59
81 years. The patients were partitioned into across lumbar spine levels were statistically significant. No sta-
subsets as 51 training and 51 testing cases. tistically significant DSCs’ pairwise comparisons by gender
were seen between the training and testing groups.
Muscle Segmentation Precision
Sarcopenia Agreement Among Most Commonly Used
For the posterior paraspinous muscles, mean DSCs were con-
Muscle Groups, at the L3 Level
sistently high across all lumbar spine levels with the lowest
DSCs at L5 for both the training (DSC = 0.926) and testing A variety of cutoffs for sarcopenia classification are noted in
(DSC = 0.900) groups. Mean DSCs of the lateral wall the literature for muscle groups at given spinal levels. Some

5
 ARTICLE IN PRESS
BURNS ET AL Academic Radiology, Vol &, No &&, && 2019

Figure 4. Dice coefficients of individual muscles

and muscle groups per lumbar spine level. (a) Dice
coefficients of individually segmented muscles. (b)
Dice coefficients of segmented muscle groups. Sta-
tistically significant with p < 0.005 after adjusting for
multiple comparisons. Dice coefficients demonstrat-
ing system performance relative to ground truth seg-
mentations for the three most widely used muscle
groups in the literature. Paraspinous muscle
group = psoas, paraspinous; psoas muscle
group = psoas; total muscles group = lateral wall,
paraspinous, psoas, quadratus, rectus; Performance
is presented for both training and testing groups, at
each of the lumbar vertebral levels for the combined
female-male cohort. In a), the individual muscle
groups at the L3 level are provided for illustration.

6
 ARTICLE IN PRESS
Academic Radiology, Vol &, No &&, && 2019 MACHINE LEARNING ALGORITHM FOR SARCOPENIA

Figure 5. Dice coefficients of individually segmented muscles: comparison for male and female patients. Statistically significant with p < 0.005 after
adjusting for multiple comparisons. Results are presented by sex and training and testing groups, at each of the lumbar vertebral levels.

cutoffs are study specific, based on hazard ratios for survival in agreement (Table 2). Cohen’s Kappa statistics for paraspinous
their study groups, while others utilize the lowest 5%, lowest and total psoas indices in comparison with SMI were 0.57
quartile, or below the mean for a particular studied muscle (weak) and 0.62 (moderate), respectively.
index of the study population (6,16,17,20,30).
The SMI group at L3 was used as the reference standard to
Variation in Accuracy of Sarcopenia Assessment with
which the TPI and paraspinous groups were compared. The
Varying Muscle Mass
lowest quartile of the stratified SMI group demonstrates bet-
ter agreement with the lowest quartile of the paraspinous A scatter plot of the DSC versus area of truncal muscle dem-
muscle group than the TPI group, for each tested patient onstrates a trend of decreasing Dice similarity coefficient, and
cohort (female, male, and combined). For the combined so system performance, with decreasing muscle mass (Fig 6).
patient group, agreement of the paraspinous muscle group The r2 value was calculated as r2 = 0.1 with p = 0.002 sugges-
with the SMI was 85.6% as compared to 83.5% for TPI/SMI tive of a very weak association for this modest size data set.

7
 ARTICLE IN PRESS
BURNS ET AL Academic Radiology, Vol &, No &&, && 2019

TABLE 2. Agreement Between the Most Widely used Muscle Groups for Sarcopenia Determination at the L3 Level

L3 SMI Total

Sarcopenia No Sarcopenia

Sarcopenia 17 8 25
TPI
L3

No sarcopenia 8 64 72
Paraspinous

Sarcopenia 18 7 25
L3

No sarcopenia 7 65 72

Total 25 72 97

Correlation between the most widely used muscle groupings in the literature, taken at the L3 level for uniformity. SMI: skeletal muscle index;
TPI: total psoas index.

measurement landmarks. The current build of this system

is for use in fully-automated opportunistic detection and
monitoring of sarcopenia.
Quantitative data generated by our system has potential to
decrease interobserver variability in sarcopenia detection.
Detailed information regarding overall severity of sarcopenia
in addition to detailed localized muscle atrophy data predicts
future ability to provide repeatable, timely, and detailed eval-
uation of central sarcopenia, and to allow extension to other
surgical applications such as outcome predictors for spinal
fusion surgery. A combined risk model for sarcopenia and
osteopenia, modularly integrating the current system with a
previously designed and tested system to assess bone density,
may find application in generating predictive features and risk
Figure 6. Variation of automated segmentation precision with vari- profiles for treatment of osteopenic trauma patients for pro-
ation in degree of sarcopenia. Scatterplot of the L3MI (L3 Muscle gressive fracturing, through a combined risk matrix (4,31).
Index) versus DSC (Dice Similarity Coefficient) for each patient, Operating at multiple vertebral levels and segmenting multi-
showing trend of decreasing DSC with increasing sarcopenia ple combinations of muscle groups, the system can provide the
(decreasing L3MI). Solid blue line indicates linear regression. Dashed
lines indicate confidence bands. The linear regression model fits the
radiologist with data fitting the sarcopenia metrics of a variety
data with [DSC = 0.001 £ L3MI (cm2/m2) + 0.91] and r2 = 0.1 of referring medical and surgical specialties. Additionally, to
(p = 0.002), suggestive of a very weak association. date, no central organizing framework known to the authors
has been developed to determine interchangeability and corre-
Calculation of Sarcopenia for the Studied Patient Cohort lation between these metrics. This system can provide an auto-
mated platform to facilitate large scale clinical trial testing and
The muscle areas were renormalized to lean muscle area for sar- validation in the creation of a standardized metric for determi-
copenia determination, by detection and exclusion of the intra- nation of sarcopenia, potentially unifying the variability in for-
muscular fat via thresholding. The fractional muscle fat content mulation of sarcopenia currently seen.
was found to be 25.8% § 13.9% and 17.3% § 7.8% in the sar- Although the assessment of sarcopenia is currently just
copenia and nonsarcopenia groups, respectively. The Interna- imaging-based as there is no universally defined reference
tional Consensus Group for Cancer Cachexia (ICGCC) defines standard, L3MI (28) currently appears as the most commonly
the sarcopenia cutoff for the L3MI (total axial muscle area at used metric in the literature, and was used as reference for
L3/patient height2), as <39 cm2/m2 for women and <55 cm2/ comparison on this small scale data set. The system accurately
m2 for men (28). Using the ICGCC standards for the L3MI determined the L3 axial total muscle area, as a building block
group, sarcopenia was present in 39.6% (21 of 53) of female and of the L3MI, with a Dice coefficient of 0.938 § 0.028. This
73.5% (36 of 49) of male patients of the cohort studied here. level of performance, as determined by the Dice coefficients,
represents the current state-of-the-art for fully-automated
muscle segmentation. There was better agreement between
DISCUSSION the L3MI and paraspinous group than between L3MI and
This system autonomously determines the axial cross-sec- TPI for lowest quartile patient stratification. This difference
tional area of multiple truncal muscle groups on CT studies was thought related to a closer statistical sample size of seg-
at standardized craniocaudal levels using lumbar vertebrae as mentation data points between the L3MI and paraspinous

8
 ARTICLE IN PRESS
Academic Radiology, Vol &, No &&, && 2019 MACHINE LEARNING ALGORITHM FOR SARCOPENIA

Figure 7. Axial CT images demonstrate variation in muscle quality with muscle mass; in particular, increasing ill-definition of the muscle mar-
gins on sarcopenia patients. (a) Sarcopenic 62 y/o female with L3 MI of 21.4. (b) Nonsarcopenic 65 y/o female with L3 MI of 44.2.

groups; however, further studies with larger data sets would performance on noncontrast examinations may require fur-
be helpful to exclude an alternative etiology of statistical sam- ther optimization through the addition of noncontrast train-
pling variance due to small patient data set size. ing cases. Finally, as intuitively expected, the fractional
The difference in precision of the system for segmentation as muscle fat content was found to be higher in the sarcopenic
assessed by the DSC was not statistically significant between the group than the nonsarcopenic group, compatible with
total abdominal, TPA, and paraspinous muscle groups, compar- increased muscle wasting with intra- and perimuscular fat
ing on a level-by-level basis, and for the same data sets. The infiltration (Fig 7). A trend of decreasing precision of agree-
lowest level of performance for the system for all muscle group- ment between the automated and manual segmentation was
ings for both the training and testing sets was at the L1 level, found with increasing degree of muscle wasting, presumably
likely due to increased muscle complexity in this region of chest due to increasing ill definition of muscle marginations, but
wall muscle and rib overlap. The performance difference the association was very weak. Although result could suggest
between the training and testing sets was not statistically signifi- decreased accuracy of sarcopenia classification due to
cant for each specific muscle group at a particular lumbar level, decreased accuracy of segmentation with higher degrees of
consistent with generalizability of the system to other data sets. muscle wasting, this would be unlikely to affect classification
As qualitatively expected a priori, no statistically significant dif- of borderline cases and more likely to exert preferential effect
ference between the DSCs was seen for the female and male on profound cases well within the classification zone.
patient groups, again implying generalizability of the system.
Prior work for automated assessment of abdominal muscle
segmentation includes work on deep learning segmentation CONCLUSION
of the total abdominal wall muscle area on a manually prese- We have created and validated an automated computer sys-
lected image at the L3 level (23). Compared to this effort, we tem to detect and quantify the truncal musculature on CT
have designed a system to autonomously analyze the CTs by studies, and detect central sarcopenia. This system is presented
detecting the spine, selecting spinal levels, and automatically as a proof of concept assembly for surgical planning and pre-
assessing multiple muscle groups at each level for determina- treatment risk assessment.
tion of multiple sarcopenia indices.
One system limitation is patient selection bias, introduced
with an inclusion criterion of 59 years and older for patient FUNDING
selection. Relatedly, there is potential for overtraining in
This research was supported in part by the Intramural
osteopenic patient muscle groups; however, this age group is
Research Programs of the National Institutes of Health Clini-
the expected most probable patient age group for system
cal Center (grant 1Z01 CL040004) (RMS, JY, DC).
application. Additionally, the system was tested on studies
without significant motion artifact or spinal hardware, and all
patients had IV contrast. In real time applications, patients REFERENCES
with these exclusion factors or without IV contrast will of 1. Rosenberg I. Summary comments: epidemiological and methodological
course be encountered; however, the current form of the sys- problems in determining nutritional status of older persons. Am J Clin
tem is not the expected final distribution form, and further Nutr 1989; 50:1231–1233.
2. Rosenberg IH. Sarcopenia: origins and clinical relevance. J Nutr 1997;
refinements are planned before clinical implementation. For 127(5 Suppl):990S–1S. doi:10.1093/jn/127.5.990S. Epub 1997/05/01.
instance, the system was trained and tested on 2 mm thick sli- 3. Anker SD, Morley JE, von Haehling S. Welcome to the ICD-10 code for sar-
ces, and although it is expected to work as well on 2.5 mm copenia. J Cachexia Sarcopenia Muscle 2016; 7:512–514. doi:10.1002/
jcsm.12147. Epub 2016/11/29.
slices, location and segmentation accuracy may be affected 4. Santilli V. Clinical definition of sarcopenia. Clin Cases Miner Bone Metab
when implementing it on 5 mm thick slices. Additionally, 2014. doi:10.11138/ccmbm/2014.11.3.177.

9
 ARTICLE IN PRESS
BURNS ET AL Academic Radiology, Vol &, No &&, && 2019

5. Brenner DJ. Slowing the increase in the population dose resulting from oesophageal cancer patients. Interact CardioVasc Thorac Surg 2016;
CT scans. Radiat Res 2010; 174:809–815. doi:10.1667/RR1859.1. Epub 24:13–19. doi:10.1093/icvts/ivw274.
2010/08/25. 20. Kuroki LM, Mangano M, Allsworth JE, et al. Pre-operative assessment of
6. Shachar SS, Williams GR, Muss HB, et al. Prognostic value of sarcope- muscle mass to predict surgical complications and prognosis in patients
nia in adults with solid tumours: a meta-analysis and systematic review. with endometrial cancer. Ann Surg Oncol 2014; 22:972–979. doi:10.1245/
Eur J Cancer 2016; 57:58–67. doi:10.1016/j.ejca.2015.12.030. s10434-014-4040-8.
7. Rier HN, Jager A, Sleijfer S, et al. The prevalence and prognostic value 21. Lee CS, Cron DC, Terjimanian MN, et al. Dorsal muscle group area and
of low muscle mass in cancer patients: a review of the literature. The surgical outcomes in liver transplantation. Clin Transpl 2014; 28:1092–
Oncologist 2016; 21:1396–1409. doi:10.1634/theoncologist.2016- 1098. doi:10.1111/ctr.12422.
0066. 22. Canvasser LD, Mazurek AA, Cron DC, et al. Paraspinous muscle as a pre-
8. Wagner D. Role of frailty and sarcopenia in predicting outcomes among dictor of surgical outcome. J Surg Res 2014; 192:76–81. doi:10.1016/j.
patients undergoing gastrointestinal surgery. World J Gastrointest Surg jss.2014.05.057.
2016; 8:27. doi:10.4240/wjgs.v8.i1.27. 23. Lee H, Troschel FM, Tajmir S, et al. Pixel-level deep segmentation: artificial
9. Montano-Loza AJ, Duarte-Rojo A, Meza-Junco J, et al. Inclusion of sar- intelligence quantifies muscle on computed tomography for body morpho-
copenia within MELD (MELD-Sarcopenia) and the prediction of mortality metric analysis. J Digital Imaging 2017; 30:487–498. doi:10.1007/s10278-
in patients with cirrhosis. Clin Transl Gastroenterol 2015; 6:e102-e. 017-9988-z.
doi:10.1038/ctg.2015.31. 24. Jianhua Y, O’Connor SD, Summers RM. Automated Spinal Column
10. Kaplan SJ, Pham TN, Arbabi S, et al. Association of radiologic indicators Extraction and Partitioning. In: 3rd IEEE International Symposium on Bio-
of frailty with 1-year mortality in older trauma patients. JAMA Surg 2017; medical Imaging: Macro to Nano, 2006: IEEE; 2006:390–393.
152:e164604doi:10.1001/jamasurg.2016.4604. 25. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for
11. Kim EY, Kim YS, Park I, et al. Evaluation of sarcopenia in small-cell lung Biomedical Image Segmentation. Lecture Notes in Computer Science:
cancer patients by routine chest CT. Supportive Care Cancer 2016; Springer International Publishing, 2015. p. 234–241.
24:4721–4726. doi:10.1007/s00520-016-3321-0. 26. Mason D, Su E. Pydicom: an open source DICOM library. Med Phys
2011; 38:3493.
12. Kim EY, Kim YS, Park I, et al. Prognostic significance of CT-determined
27. Zou KH, Warfield SK, Bharatha A, et al. Statistical validation of image
sarcopenia in patients with small-cell lung cancer. J Thoracic Oncol
segmentation quality based on a spatial overlap index. Acad Radiol
2015; 10:1795–1799. doi:10.1097/jto.0000000000000690.
2004; 11:178–189. Epub 2004/02/21.
13. Fogelman DR, Holmes H, Mohammed K, et al. Does IGFR1 inhibition
28. Fearon K, Strasser F, Anker SD, et al. Definition and classification of cancer
result in increased muscle mass loss in patients undergoing treatment
cachexia: an international consensus. Lancet Oncol 2011; 12:489–495.
for pancreatic cancer? J Cachexia Sarcopenia Muscle 2014; 5:307–313.
doi:10.1016/s1470-2045(10)70218-7.
doi:10.1007/s13539-014-0145-y.
29. McHugh ML. Interrater reliability: the kappa statistic. Biochem Med
14. Chang C-D, Wu JS, Mhuircheartaigh JN, et al. Effect of sarcopenia on clini-
(Zagreb) 2012; 22:276–282.
cal and surgical outcome in elderly patients with proximal femur fractures.
30. Prado CM, Lieffers JR, McCargar LJ, et al. Prevalence and clinical impli-
Skeletal Radiol 2017; 47:771–777. doi:10.1007/s00256-017-2848-6.
cations of sarcopenic obesity in patients with solid tumours of the respi-
15. Deng C-Y, Lin Y-C, Wu JS, et al. Progressive sarcopenia in patients with
ratory and gastrointestinal tracts: a population-based study. Lancet
colorectal cancer predicts survival. Am J Roentgenol 2018; 210:526–
Oncol 2008; 9:629–635. doi:10.1016/S1470-2045(08)70153-0.
532. doi:10.2214/ajr.17.18020.
31. Burns JE, Yao J, Summers RM. Vertebral body compression fractures
16. Leeper CM, Lin E, Hoffman M, et al. Computed tomography abbreviated
and bone density: automated detection and classification on CT images.
assessment of sarcopenia following trauma. J Trauma Acute Care Surg
Radiology 2017; 284:788–797. doi:10.1148/radiol.2017162100. PubMed
2016; 80:805–811. doi:10.1097/ta.0000000000000989.
PMID: 28301777.
17. Englesbe MJ, Patel SP, He K, et al. Sarcopenia and mortality after liver
transplantation. J Am Coll Surg 2010; 211:271–278. doi:10.1016/j.jam-
collsurg.2010.03.039.
18. Peng P, Hyder O, Firoozmand A, et al. Impact of sarcopenia on out-
comes following resection of pancreatic adenocarcinoma. J Gastrointest SUPPLEMENTARY MATERIALS
Surg 2012; 16:1478–1486. doi:10.1007/s11605-012-1923-5.
19. Park SY, Yoon J-K, Lee SJ, et al. Prognostic value of preoperative Supplementary material associated with this article can be found
total psoas muscle area on long-term outcome in surgically treated in the online version at doi:10.1016/j.acra.2019.03.011.

10