Quantifying bone marrow fat using standard T1-weighted magnetic resonance images in children with typical development and in children with cerebral palsy

Excess bone marrow adiposity may have a negative effect on bone growth and development. The aim of this study was to determine whether a procedure using standard T1-weighted magnetic resonance images provides an accurate estimate of bone marrow fat in children with typical development and in children with mild spastic cerebral palsy (CP; n = 15/group; 4–11 y). Magnetic resonance imaging was used to acquire T1-weighted images. It was also used to acquire fat and water images using an iterative decomposition of water and fat with echo asymmetry and least-squares estimation (IDEAL) technique. Bone marrow fat volume and fat fraction in the middle-third of the tibia were determined using the standard T1-weighted images (BMFVT1 and BMFFT1, respectively) and the fat and water images (BMFVIDEAL and BMFFIDEAL, respectively). In both groups, BMFVT1 was highly correlated with (both r > 0.99, p < 0.001) and not different from (both p > 0.05) BMFVIDEAL. In both groups, BMFFT1 was moderately correlated with (both r = 0.71, p < 0.01) and not different from (both p > 0.05) BMFFIDEAL. There was no group difference in BMFVT1 or BMFVIDEAL (both p > 0.05). BMFFIDEAL was higher in children with CP (p < 0.05), but there was no group difference in BMFFT1 (p > 0.05). We conclude that a procedure using standard T1-weighted magnetic resonance images can produce estimates of bone marrow fat volume similar to estimates from the IDEAL technique in children. However, it is less sensitive to variation in the bone marrow fat fraction.

not been validated. Therefore, in this study we sought to test the validity of using standard T1-weighted images to quantify bone marrow fat in children.
The purpose of this study was to determine whether a method using standard T1-weighted magnetic resonance images provides an accurate estimate of bone marrow fat in children who are typically developing, and in children with CP, a population that has been shown to have elevated bone marrow fat compared to their typically developing peers 9 .

Methods
Participants. Thirty children participated in the study. Fifteen were children with spastic CP who were able to ambulate independently or with an assistive device (n = 4 girls, n = 11 boys) and 15 were typically developing children without known neurological disorders and similar in age and sex to children with CP. This study was approved by the Institutional Review Board at the Nemours AI duPont Hospital for Children, Wilmington, DE. All methods were carried out in accordance with relevant guidelines and regulations. Informed consent was acquired from a parent or a legal guardian for study participation and assent was acquired from the participant, if >7 years of age and able, before any data collection was performed.
Anthropometrics. Height and body mass were measured while participants were wearing minimal clothing.
Height was assessed with a stadiometer (Seca 217; Seca GmbH & Co. KG., Hamburg, GER) to the nearest 0.1 cm. Body mass was assessed with a digital weight scale (Detecto 6550, Cardinal Scale, Webb City, MO) to the nearest 0.2 kg. Body mass index (BMI) was subsequently calculated based on the obtained height and body mass. Height, body mass and BMI percentiles were calculated for each participant based on the Centers for Disease Control and Prevention growth charts 19 . Sexual maturity. A physician assistant performed sexual maturity assessments using the Tanner staging technique 20 . Tanner staging is a scale ranging from 1 to 5 with higher numbers indicating more advanced sexual development. In boys, pubic hair and testicular/penile development were assessed. In girls, pubic hair and breast development were assessed.
Gross motor function. The gross motor function classification system (GMFCS) was used to assess gross motor function of children with CP 21,22 . GMFCS is a scale ranging from I to V with higher numbers indicating more compromised motor function. Briefly, a child with the ability to walk without restrictions, but with limitations in more advanced motor skills, was classified as GMFCS I. A child with the ability to walk without assistive devices, but with limitations walking outdoors and in the community, was classified as GMFCS II. A child with the ability to walk with an assistive device, but with limitations walking outdoors and in the community, was classified as GMFCS III. All children with CP included in the current study were classified as GMFCS I-III.
Magnetic resonance imaging. Magnetic resonance imaging (1.5 T; GE, Waukesha, WI) was used to assess bone marrow fat in the midtibia of the more affected limb in children with CP and in the non-dominant limb in controls. The BodyFIX (Medical Intelligence Inc, Schwabmunchen, Germany) was used to secure the children from the waist down to minimize the influence of movement, as previously described 23 . A three-plane localizer was first used to identify the region of interest. Axial images (0.5 cm thick with 0.5 cm spacing) were collected from the tibia plateau to the medial malleolar articular surface using a semiflex long bone array coil (ScanMed, Omaha, NE) and two different protocols. The first protocol (fast spin echo, TR = 650 ms, TE = 14 ms, FOV = 12 cm, NEX = 3, BW = 15.63 kHz, frequency = 512, phase = 256) yielded standard T1-weighted images. The second protocol, which used the IDEAL technique (fast-spin-echo, 2D, 3 echoes, TR = 600 ms, TE = min full, FOV = 12 cm, NEX = 2, BW = 31.25 kHz, echo time = 16.8 ms, frequency = 320, phase = 224), yielded fat and water images. The total number of images collected was based on tibia length and ranged from 24 to 34. Visual description of the procedure to determine bone marrow fat fraction using fat and water images generated from the IDEAL technique and custom software. First, a raw T1-weighted magnetic resonance image (A) was filtered and then used to identify the bone marrow area (yellow region in B). The identified voxels were then applied to fat (C) and water (D) images to determine bone marrow fat area and to calculate bone marrow fat fraction. Bone marrow fat volume was quantified by accounting for the number of images, image thickness and spacing between images.
Total bone marrow volume was determined using the standard T1-weighted images and a custom program developed using Interactive Data Language (IDL; Research Systems, Inc., Boulder, CO). The procedure to determine bone marrow fat fraction (BMFF IDEAL ) and bone fat marrow volume (BMFV IDEAL ) using fat and water images generated from the IDEAL technique and the same custom program is briefly described in Fig. 1. The bone marrow was first identified using the T1-weighted images that were median filtered and separated from other tissues using a fuzzy-clustering algorithm. The T1 and IDEAL images were co-registered and the region (or mask) defined from the T1 images was applied to the corresponding fat and water images. The mean signal intensity (SI) was extracted from the fat and water images. BMFF IDEAL of each slice was calculated using the following equation 24 : IDEAL BMFF IDEAL was multiplied by the bone marrow area identified from T1-weighted images to estimate bone marrow fat area. The bone marrow fat volume for the midtibia was then calculated.
The procedure to determine bone marrow fat fraction (BMFF T1 ) and bone marrow fat volume (BMFV T1 ) using T1-weighted images and the open source software ImageJ 25 is briefly described in Fig. 2. The raw images were first imported to ImageJ and then the threshold that best matched the segmentation of subcutaneous fat area was determined. The threshold was used to estimate bone marrow fat area from which BMFV T1 for the midtibia was calculated. BMFF T1 was calculated for each image by dividing bone marrow fat area by bone marrow area and multiplying by 100. The average BMFF T1 for all images evaluated is reported. The test-retest reliability for bone marrow fat fraction in the midtibia was previously determined in a sample of four children with CP and four typically developing children (5 to 11 years of age) tested twice, 6 months apart. The intraclass correlation was > 0.96 and the coefficient of variation was 0.04% 9 .
Statistics. Data were analyzed using SPSS version 24.0 (IBM Corp, Armonk, NY). Physical characteristics data in typically developing children and in children with CP were checked for normality. If the data were normality distributed, group comparisons were made using independent t-tests. If data were not normally distributed, group comparisons were made using the Mann-Whitney test.
To determine the accuracy of using T1-weighted images to estimate bone marrow fat volume and concentration, data from each group were first compared separately. Differences between BMFV T1 and BMFV IDEAL and between BMFF T1 and BMFF IDEAL were determined using paired t-tests in each group. In addition, Pearson correlation analysis was performed to determine the association of bone marrow fat yielded from the two techniques. Moreover, for each group, bone marrow fat fraction and area estimated from T1-weighted magnetic resonance images were compared to estimates from fat and water images and the IDEAL technique using Bland-Altman plots 26,27 . Finally, between groups analyses for BMFF T1, BMFF IDEAL , BMFV T1 and BMFV IDEAL were also performed using independent t-tests. All data were presented as mean ± SD in tables. Data were reported as mean ± SE in figures. The magnitude of the effects was determined by Cohen's d (d) whenever applicable, with 0.2, 0.5 and 0.8 representing small, moderate and large effect sizes, respectively 28 .

Results
All physical characteristics are summarized in Table 1. No between group differences were detected between typically developing children and children with CP except for height percentile and body mass percentile, which were significantly lower in children with CP (both p < 0.05).
There was no difference between BMFV T1 and BMFV IDEAL in children with CP (d = 0.144, p = 0.978) or in typically developing children (d = 0.041, p > 0.99). The scatter plot in Fig. 3A shows that BMFV T1 was Figure 2. Visual description of the procedure to determine bone marrow fat fraction using standard T1weighted magnetic resonance images and the ImageJ software. First, a raw image (A) was used to determine a segmentation threshold that best matched the subcutaneous fat area (green outer ring in B). The image was then binarized and bone marrow fat voxels identified at the same threshold as the subcutaneous fat area threshold were identified (C) and used to estimate bone marrow fat area. Bone marrow fat volume was quantified by accounting for the number of images, image thickness and spacing between images.
www.nature.com/scientificreports www.nature.com/scientificreports/ highly correlated with BMFV IDEAL (r > 0.99, p < 0.001 for both groups). The Bland-Altman plot in Fig. 3B confirms the validity of BMFV T1 when compared to BMFV IDEAL showing no bias between the two measures (mean difference = 0.04 cm 3 ) and a very small degree of variability in the difference between them (SD of the difference = 0.16 cm 3 ).
There was no significant difference between BMFF T1 and BMFF IDEAL in children with CP (d = 0.055, p = 0.881) or in typically developing children (d = 0.234, p = 0.527). The scatter plot in Fig. 3C shows that BMFF T1 was moderately correlated with BMFV IDEAL (r = 0.71, p = 0.003 in both groups). The Bland-Altman plot in Fig. 3D suggests that there was no bias between BMFF T1 and BMFF IDEAL (mean difference = 0.17%); however, there was more variability between them (SD of the difference = 1.97%) than the variability shown between BMFV T1 and BMFV IDEAL .

Discussion
The primary finding was that bone marrow fat volume estimates based on standard T1-weighted magnetic resonance images agreed extremely well with estimates based on fat and water images generated from the IDEAL technique in both typically developing children and in children with CP. However, bone marrow fat fraction estimates based on standard T1-weighted images were only moderately correlated with estimates based on fat and water images from the IDEAL technique. Moreover, the technique using standard T1-weighted images did not detect a higher bone marrow fat concentration in children CP compared to typically developing children, which was detected using fat and water images from the IDEAL technique. The finding suggests that the technique using standard T1-weighted images is less sensitive to variations in fat concentration in bone marrow. The source of errors for the BMFV T1 and BMFF T1 can probably be attributed to both partial volume effects and the use of a single threshold for bone marrow fat segmentation 29 .
The assessment of bone marrow fat by the IDEAL technique has been validated using MRS. Excellent agreement between bone marrow fat by the IDEAL technique and by the MRS technique has been observed using a phantom 30 and human vertebra 31 . MRS is the most commonly used technique to quantify bone marrow fat in humans, especially at the vertebral site [32][33][34][35] . However, compared to MRS, the IDEAL technique has several advantages. One advantage is its shorter scan time 36 because it does not require a long shimming time to make the magnetic field more homogeneous. This is particularly important when imaging clinical populations who may have a difficult time holding still over an extended period of time, such as children with spastic CP. In addition, compared to the most widely used single-voxel MRS method, the IDEAL method can provide much greater coverage because it is not limited by the size of the voxel. Due to these advantages, the IDEAL method may be a better choice than the MRS method when assessing bone marrow fat for clinical and research purposes, especially when assessing children.
To our knowledge, this is the first human study to assess the validity of bone marrow fat fraction estimates based on standard T1-weighted images. Previous studies have quantified bone marrow fat using T1-weighted magnetic resonance images 17,18,29 . Shen et al. 17 found that pelvic bone marrow fat volume and total body bone marrow fat volume determined using T1-weighted magnetic resonance images similar to the images described in the present study were both inversely and significantly associated with total body BMD in Caucasian women. This observation supports the link between high bone marrow fat and low bone density. Similar associations between bone marrow fat volume from T1-weighted images and total body BMD was also observed in both younger and older adults 18 . Another study tested the validity of using T1-weighted magnetic resonance images to quantify bone marrow fat at the third lumbar vertebra and the femoral neck in post-menopausal women using the MRS and the DIXON methods for comparison 29 . Bone marrow fat volume from T1 images was strongly related to bone marrow fat fraction from the MRS method (r = 0.88, p < 0.001 for vertebra; femoral neck MRS measurement was CP (n = 15) Con (n = 15) d p www.nature.com/scientificreports www.nature.com/scientificreports/ not acquired) and from the DIXON method (r = 0.79, p < 0.001 for vertebra and r = 0.86, p < 0.001 for femoral neck). However, the relationship between bone marrow fat fraction from T1-weighted images and the other methods was not reported. Moreover, to date, no studies have examined the validity of using T1-weighted magnetic resonance images to quantify bone marrow fat in children.
The ability to assess bone marrow fat in vivo is of particular interest to many researchers because of its well-established connection with osteoporosis. Furthermore, bone marrow fat has been suggested to play a role in the pathogenesis of human metabolic risks. However, the mechanisms underlying the effect of bone marrow fat on metabolism are complicated and yet to be elucidated. An increase in bone marrow fat could potentially lead to a reduced number of available hematopoietic stem cells by simply taking over the available spaces, as well as by negatively influencing the microenvironment of these cells 37 . This is important because hematopoietic stem cells only utilize glucose (glycolysis) as energy resources 38 . Thus, those with elevated bone marrow fat may have a higher risk of developing type 2 diabetes, and bone marrow fat may serve as a surrogate marker for early signs of glucose dysregulation; although more studies are needed to confirm this notion.
The significant correlations found in the current study between BMFV T1 and BMFV IDEAL in both typically developing children and children with CP indicate that bone marrow fat as estimated with T1-weighted magnetic resonance images can yield good volumetric measures in both groups. However, in the pediatric population, bone marrow may keep expanding simply due to growth and maturation. Therefore, bone marrow fat fraction compared to bone marrow fat volume may be a more appropriate measure of bone marrow fat infiltration in cross-sectional and longitudinal studies involving children 9,39 . The scatter plot shows strong agreement between bone marrow fat volume estimated from the standard T1weighted images (BMFV T1 ) and bone marrow fat volume estimated using fat and water images from the IDEAL technique (BMFV IDEAL ). (B) The Bland-Altman plot shows the level of agreement between BMFV T1 and BMFV IDEAL . (C) The scatter plot shows moderate agreement between bone marrow fat fraction estimated from the standard T1-weighted magnetic images (BMFF T1 ) and bone marrow fat volume estimated using fat and water images (BMFF IDEAL ). (D) The Bland-Altman plot shows the level of agreement between BMFF T1 and BMFF IDEAL . The dotted lines in panels A and C represent the lines of identity. The dotted lines in panels B and D indicate the mean difference ± 2 SD between estimates from the standard T1-weighted and the fat and water images. The solid lines indicate no difference for estimation between the two methods.
Unfortunately, the correlations between BMFF T1 and BMFF IDEAL observed in typically developing children and in children with CP were only moderate. More importantly, the between group comparison of BMFF T1 was not significant and the effect size was only small to moderate (d = 0.409). On the other hand, BMFF IDEAL was significantly higher in the children with CP than in the typically developing children and the effect size was large (d = 0.816). It is estimated that 58 participants per group would be needed to detect a significant between group difference in BMFF T1 ; whereas, only 15 participants per group were needed to detect a difference in BMFF IDEAL . Together, the correlational data and the group comparison data suggest that when compared to the IDEAL technique, the technique based on T1-weighted magnetic resonance images may be less sensitive to variations in the bone marrow fat fraction.
The limitations of the study must be considered. It is assumed that the bone marrow fat fraction estimates based on the fat and water images generated using the IDEAL technique provided more accurate estimates of bone marrow fat fraction than the technique based on standard T1-weighted images. Although the accuracy of any in vivo technique has limitations, bone marrow fat fraction estimates from fat and water images and the IDEAL technique have been validated against the MRS technique 30,31 , which is considered the most accurate in vivo method for assessing bone marrow fat fraction 30 . In addition, our intent was to capture the fat content within the entire bone marrow area, and due to the shape of bone marrow, such a goal is hard to achieve with the single-voxel MRS technique.

Conclusion
A procedure using standard T1-weighted magnetic resonance images can produce estimates of bone marrow fat volume similar to estimates from fat and water images generated using the IDEAL technique in children. However, it is less sensitive to variation in the bone marrow fat fraction. . Bone marrow fat volume (A) and fat fraction (B) estimated by the procedure using standard T1-weighted images (BMFV T1 and BMFF T1 , respectively) and by the procedure using fat and water images generated from the IDEAL technique (BMFV IDEAL and BMFF IDEAL , respectively) in children with cerebral palsy (CP) and in typically developing children (Con). Values are means ± SE. *Group difference, p < 0.05.