Investigating the Effects of Motion Streaks on pqct-derived Leg Muscle Density and Its Association With Fractures

Transcription

1 Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health, vol. 21, no. 1, , 2018 Copyright 2016 by The International Society for Clinical Densitometry /21: /$ Original Article Investigating the Effects of Motion Streaks on pqct-derived Leg Muscle Density and Its Association With Fractures Adrian C. H. Chan, 1 Jonathan D. Adachi, 2 Alexandra Papaioannou, 2,3 and Andy Kin On Wong*,1,2 1 Osteoporosis Program, University Health Network, Toronto, ON, Canada; 2 Department of Medicine, McMaster University, Hamilton, ON, Canada; and 3 Geriatric Education and Research in Aging Sciences (GERAS) Centre, Hamilton Health Sciences, Hamilton, ON, Canada Abstract Lower peripheral quantitative computed tomography (pqct)-derived leg muscle density has been associated with fragility fractures in postmenopausal women. Limb movement during image acquisition may result in motion streaks in muscle that could dilute this relationship. This cross-sectional study examined a subset of women from the Canadian Multicentre Osteoporosis Study. pqct leg scans were qualitatively graded (1 5) for motion severity. Muscle and motion streak were segmented using semi-automated (watershed) and fully automated (threshold-based) methods, computing area, and density. Binary logistic regression evaluated odds ratios (ORs) for fragility or all-cause fractures related to each of these measures with covariate adjustment. Among the 223 women examined (mean age: 72.7 ± 7.1 years, body mass index: ± 4.97 kg/m 2 ), muscle density was significantly lower after removing motion (p < 0.001) for both methods. Motion streak areas segmented using the semi-automated method correlated better with visual motion grades (rho = 0.90, p < 0.01) compared to the fully automated method (rho = 0.65, p < 0.01). Although the analysis-reanalysis precision of motion streak area segmentation using the semi-automated method is above 5% error (6.44%), motioncorrected muscle density measures remained well within 2% analytical error. The effect of motion-correction on strengthening the association between muscle density and fragility fractures was significant when motion grade was 3 (p interaction <0.05). This observation was most dramatic for the semi-automated algorithm (OR: 1.62 [0.82,3.17] before to 2.19 [1.05,4.59] after correction). Although muscle density showed an overall association with all-cause fractures (OR: 1.49 [1.05,2.12]), the effect of motion-correction was again, most impactful within individuals with scans showing grade 3 or above motion. Correcting for motion in pqct leg scans strengthened the relationship between muscle density and fragility fractures, particularly in scans with motion grades of 3 or above. Motion streaks are not confounders to the relationship between pqct-derived leg muscle density and fractures, but may introduce heterogeneity in muscle density measurements, rendering associations with fractures to be weaker. Key Words: Fractures; image quality; motion artifact; muscle density; peripheral quantitative computed tomography (pqct). Received 08/11/16; Revised 11/30/16; Accepted 12/5/16. *Address correspondence to:andy Kin On Wong, PhD,Toronto General Hospital, 200 Elizabeth St. 7EN-238, Toronto, ON, Canada M5G 2C4. andy.wong@uhnresearch.ca Introduction Peripheral quantitative computed tomography (pqct) can quantify bone and muscle properties of the appendicular anatomy such as the arms and legs (1). Muscle has more recently been examined as related to sarcopenia, risk for fractures, falls, frailty, and disability. Frank-Wilson et al measured a 19% reduced odds for falling per 1 mg/cm 3 130

2 pqct Motion Artifact on Muscle Density and Fractures 131 Fig. 1. Illustration of motion artifact from pqct muscle scans. Note the level to which motion streaks transect the muscle region when motion is more severe. higher leg muscle density (MD) as determined using pqct in community-dwelling older men and women (N = 183, age: yr) (2). pqct-derived lower leg muscle crosssectional area (MCSA) and MD were 32% and 43% lower, respectively, in men with chronic spinal cord injury (N = 70, mean ± SD age: 49 ± 12.0) compared to age-matched controls (3). Although muscle strength has a direct influence on bone through mechanical stimulation (4), it has also been shown that better physical performance and muscle strength could prevent fractures (5). One challenge not addressed in these studies is the influence of motion in pqct muscle quantification, which could confound the ability to ascertain true clinical differences. Motion artifacts within muscle scans often cause disruption in the imaged cortical shell or manifest in streaks adjacent to the cortical bone that transect the body of the muscle (Fig. 1). These streaks can change muscle s signal intensity, leading to potentially inaccurate MD measurement. The most commonly practiced method for assessing these motion artifacts is through qualitative visual inspection. A rating scale from 1 to 5 is used to grade the severity of streaks and/or breaks in the cortical shell (6), with a higher score being indicative of greater motion severity. Blew et al reported an objective, threshold-based method for segmenting and quantifying motion artifact on pqct lower leg scans of children by measuring the ratio of motion streak area to limb area (%Move) (6). Children who displayed %Move >25% were excluded from further analysis. Although this procedure eliminated the subjectivity of the technician s evaluation, it failed to detect images that were deemed to require rescanning when using the qualitative visual inspection scale. This finding suggests that manual adjustments may be necessary to refine motion streak segmentations. Related to manual segmentations, Wong et al reported lower test-retest precision errors for pqct MD measurement using a watershed-guided manual segmentation method (1.18% 2.01% error) compared to the standard automated threshold-based edge detection algorithm by the manufacturer (1.77% 4.06% error) (7).Leg MD measured by manual segmentation was also larger (70.2 ± 9.2 mg/cm 3 ) than that measured by the fully automated threshold-based method (67.4 ± 10.3 mg/cm 3 ).These results support the use of a watershed method to more precisely quantify motion streaks. Further to the challenge of quantifying motion streaks, it remains unknown whether they contribute to more elevated or more attenuated MD values, as both higher and lower density compartments are present (Fig. 2A). By virtue of having larger motion streaks within scans, MD could be falsely elevated or lowered, and any associations between MD and fractures could be blunted. However, properties of MD itself beyond motion may still be linked to risk for fractures. In this study, it was hypothesized that leg MD, after removing motion streaks, will be significantly different from leg MD including motion streaks. It was also anticipated that the strength of the relationship between pqct-derived leg MD and fractures will remain significant even after correcting for motion streak properties. Methods Study Design and Participants This study was a cross-sectional analysis examining motion streaks (MS) within pqct lower leg images within a subset of women 50 yr of age from the Canadian Fig. 2. Automated threshold-based segmentation for positive movement from muscle. (A) Figure showing both positive and negative motion streaks running through the body of the muscle. (B) Figure showing positive motion streaks and cortical bone segmentation generated from CORTBD analysis with threshold of 110 mg/cm 3.(C) Figure with just cortical bone segmented excluding any motion streaks generated using CORTBD analysis with threshold of 710 mg/cm 3. Positive motion streaks were isolated by subtracting cortical bone (C) from bone plus positive motion streaks (B).

3 132 Chan et al. Multicentre Osteoporosis (CaMos) Bone Quality Study. CaMos is an ongoing (inception 1996), prospective cohort study of community-dwelling, randomly selected men and women 25 years of age living within 50 km of 9 major Canadian cities. The main CaMos objectives, methodology, and sampling framework are described in detail elsewhere (8). Of the 506 study participants who completed a pqct scan of their nondominant leg in 2012 to 2014, 441 were available for analysis at the time of study, and a subset (N = 223) of participants were randomly selected by strata of motion grade (1 5) (described next) for the present analysis examining the effect of MS on muscle and fractures. Additional information on demographics, medical conditions, medication use, total hip areal bone mineral density within the last 3 years, and fracture history from the previous 14 years were obtained from the CaMos database. Fractures were assessed according to 2 categories: fragility (lowtrauma fractures after age 40 due to falling from standing height or less, excluding any fractures of the skull, fingers, and toes), and all-cause fractures (due to any reason, but occurring after 40 years of age). All procedures performed in this study were approved nationally by the Hamilton Integrated Research Ethics Board and by ethics boards at local jurisdictions. Peripheral Quantitative Computed Tomography A single 2.0 ± 0.5 mm thick slice was acquired at 66% of the tibia length measured proximally from the distal articulating aspect of the medial malleolus toward the medial articulating aspect of the tibial plateau. Scans were acquired on model XCT2000 or XCT3000 pqct scanners (Stratec Medizintechnik, Pforzheim, Germany) with in-plane resolution of 500 μm, a computed tomography scan speed of 20 mm/s, 38 kvp X-ray beam energy, and a tube current of 0.3 ma, reconstructed by filtered back-projection on a matrix size of Hydroxyapatite phantoms were assessed on days when scans were obtained (9). Qualitative Assessment for Motion Artifact Visual inspection determined the severity of MS by using a linear, ordinal scale adapted from Blew et al (6) that ranged from 1 to 5, based on the amount of streaking and/ or breaks in the cortical shell. Briefly, the scale grades are as follows: 1 = negligible motion; 2 = faint thin streaks apparent; 3 = evident high intensity streaks present but without alteration of cortical geometry; 4 = evident high intensity streaks present with some alteration of cortical geometry; and 5 = severe streaking evident with clear break in cortical shell. Quantitative Assessment of MS and MD Semi-Automated Method A single image analyst (author ACHC) blinded to participant demographics and fracture status used the watershed algorithm from SliceOmatic software package (v 4.3, Tomovision, Magog, Quebec, Canada) to manually segment MS from muscle.this algorithm uses threshold configurations (pixel surface and percentage mean intensity difference) to yield a fine network of watershed pools that groups adjacent pixels with similar signal intensities to be tagged in a single click. Pixel surface indicates the size of pixel groupings within which signal intensity was assessed, whereas the percentage mean difference sets an upper limit from which pixel groupings adjacent to each other were allowed to deviate from their mean signal intensity value to be pooled together.watershed parameters that generated larger pools (25 pixel surface and 1.00% mean difference) approximated the bulk of the MS cross-sectional area (MS-CSA), whereas those that generated smaller pools (5 pixel surface and 0.20% mean difference) were used to fine-tune details of MS (Fig. 3). Additional refinements to MS-CSA were performed in manual edit mode. Calibrated values for MS density (MS-D) were also calculated. Muscle was also segmented from bone and subcutaneous fat using the watershed algorithm and labeled separately from MS-CSA (Fig. 3) to compute motion-corrected (c) muscle density (MD C) and total MCSA. Separately, the whole muscle was tagged without segmenting MS to calculate uncorrected (uc) MD (UC). Density values were computed based on linear attenuation calibration equations from pqct scanners and reported in mg/cm 3, while all CSAs were reported in mm 2. Image processing was performed on the same computer screen with gamma settings fixed upon each sitting. Quantitative Assessment of MS Fully Automated Method An automated threshold-based iterative edge detection algorithm was used to segment MS from muscle on Stratec Analysis software (v 6.0. Orthometrix Inc., White Plains, NY) with contour, peel mode settings, as well as threshold selection adapted from Schiferl s Fat vs. Lean Analysis Manual v7.3x (10), also referenced in full by Blew et al (6). Because MS exhibited juxtaposing increased (positive MS) and decreased (negative MS) signal intensity edges, they were segmented in these separate components (Fig. 2). The standard method for cortical analysis (CORTBD) uses a threshold of 710 mg/cm 3, which achieves a clean segmentation of bone without MS. By lowering the threshold to 110 mg/cm 3 (149 mg/cm 3 was initially suggested by Blew (6) and Schiferl in younger participants, but it was not sufficient in capturing positive motion observed in older adults), positive MS can be captured without including muscle. Taking the difference in segmented regions between these 2 CORTBD analyses (bone+positive MS bone) yielded positive MS alone. Negative MS was computed in a 3-step procedure (Fig. 4). First, negative MS and bone marrow area were together captured using bone density analysis (CALCBD) with inner and outer thresholds of 40 mg/cm 3, iterative

4 pqct Motion Artifact on Muscle Density and Fractures 133 Fig. 3. Example of the watershed pools used to segment motion streaks from muscle. Larger pools (left column) approximated motion streak area, whereas smaller pools (middle column) were used to further refine its boundaries. Final segmentation result is shown in the right column. A comparison of the image before (top row) and after (bottom row) segmentation is displayed. Red indicates muscle area, and green as motion streak. Total MCSA includes both green and red regions. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.) contour search (mode 31), threshold-based concentric peel (mode 2), and image filter F03F05F05 applied. CORTBD (outer threshold of 710 mg/cm 3, separation mode 4, inner threshold of 101 mg/cm 3 ) was then used to compute bone plus bone marrow areas, which was subtracted from the segmentation above. To account for the bone that was extraneously subtracted, bone area alone was computed using a second CORTBD procedure (outer threshold 710 mg/cm 3, separation mode 4, inner threshold 40 mg/cm 3, filter F03F05F05) and was added back to the previous segmentation. Positive and negative movement masses and areas were summated to yield the overall MS-CSA and the thickness-adjusted quotient, as MS-D. Quantitative Assessment for Motion-Uncorrected Muscle Density Fully Automated Computation of MD UC was performed using a thresholdbased method previously described (9). Briefly, bone and marrow were separated from muscle by applying an outer threshold of 280 mg/cm 3 in CALCBD with contour mode 1 and peel mode 2. Muscle, bone, and marrow were seg- Fig. 4. Automated threshold-based segmentation for negative movement from the muscle tissue. (A) Figure showing both positive and negative motion streaks running through the body of the muscle. (B) Illustration of negative movement alone without bone but including marrow, as generated from CALCBD analysis with inner and outer thresholds of 40 mg/cm 3, contour mode 31, peel mode 2, and F03F05F05 filter applied. (C) Figure of cortical bone with marrow, determined using CORTBD with an outer threshold of 710 mg/cm 3, separation mode 4, and inner threshold of 101 mg/cm 3. (D) Figure of cortical bone without marrow, determined using CORTBD with an outer threshold of 710 mg/cm 3, separation mode 4, inner threshold 40 mg/cm 3, and filter F03F05F05. Marrow from figure B was removed by subtracting bone plus marrow in figure C, but adding back bone using figure D. The final result yields negative motion streaks alone.

5 134 Chan et al. mented from subcutaneous fat with CALCBD using an outer and inner threshold of 40 mg/cm 3 with contour mode 3, peel mode 2, and a smoothing filter of F03F05F05. Mass and area from the above segmentations were subtracted to determine the thickness-adjusted quotient, MD UC. Quantitative Assessment for Motion-Corrected Muscle Density Fully Automated Positive and negative MS masses were subtracted from motion-uncorrected muscle mass to yield motion-corrected muscle mass (MM C). Positive and negative MS-CSAs were similarly subtracted from MCSA to yield a motion-corrected muscle cross-sectional area by which MM C was divided to compute motion-corrected muscle density (MD C). Data Analyses Analysis-reanalysis reproducibility was evaluated for MD, MS-D, and MS-CSA measures obtained using the semi-automated method in a subset of 20 participants. All replicated analyses were performed in a different order, and blinded to pairing of the first analysis. Percentage root mean square coefficients of variation (RMSCV%) and root mean square standard deviations were calculated. A benchmark of 5% RMSCV was used to gauge acceptable reproducibility as per recommendations of Gluer et al (11). A repeated measures analysis of variance determined whether there were significant differences between the semi-automated and the fully automated method for computing MS-CSA, MS-D, MD UC, and MD C. Spearman s rho quantified the correlation between motion grade and MS-CSA derived from the semi-automated and fully automated segmentations. Bland-Altman plots and 95% limits of agreement (LOA) depicted difference patterns across the values of MD C and MD UC for both semi-automated and fully automated segmentation methods. A binary logistic regression analysis was performed to determine the relationship between fragility or all-cause fractures and each of MD UC vs MD C. All models adjusted for age, body mass index, total hip areal bone mineral density, duration of glucocorticoid use, and parental history of hip fractures. Similarly, the relationships between fragility or all-cause fractures and each of MS-CSA or MS-D alone, with and without adjusting for the same covariates described above, were examined. Because individuals with small amounts of motion dilute the overall effect of motion-correction on impacting MD C and fracture relationships, moderation analyses were performed to examine the effect of motioncorrection on the MD-fracture relationship when motion grade was 3 or< 3. All statistical analyses were performed on SAS v9.3 (SAS Institute Inc, Cary, NC). Results A total of 441 lower leg scans from the CaMos study were visually inspected and graded. Characteristics of the larger cohort of women were similar to those in the randomly chosen subset of quantitatively analyzed scans (N = 223). Of the 223 women, 32 (14.6%) had experienced at least 1 fragility fracture within the last 14 years, whereas 107 (48.9%) have sustained an all-cause fracture within the same period (Table 1). Within the cohort of 20 participants from which analysis-reanalysis reproducibility was assessed, the median motion grade was 3.0 with an interquartile range of 2.0 to 4.0. Mean MS-CSA was significantly larger and mean MS-D was significantly lower when computed with the semiautomated vs the fully automated algorithm (p < 0.001) (Table 2). However, mean MD values for the semiautomated method were significantly higher both before and after motion-correction compared to those determined by the fully automated algorithm. (p < 0.001). Correcting for MS by excluding MS segmentations from MD calculations resulted in significantly lower MD values consistently across motion grades and segmentation methods, with the largest impact observed in the motion grades 4 5 category. Although there was a significant difference in motion corrected and uncorrected MD within grades 1 3 images computed using the semi-automated algorithm, this difference was only 0.22 mg/cm 3. Qualitatively assessed motion grade showed a higher correlation with semiautomated MS-CSA (Spearman s rho = 0.90, p < 0.01) Table 1 Descriptive Characteristics for Participants by Motion Grade Categories Motion grades Characteristic All (N = 223) 1 3 (N = 177) 4 5 (N = 46) Age (SD), yr (7.07) (6.86) (7.36) BMI (SD), kg/m (4.97) (5.00) (4.89) Total hip abmd (SD), T-score 1.38 (0.95) 1.49 (0.92) 1.32 (1.04) Prior fragility fracture, N/Total (%) 32/223 (14.35) 24/177 (13.56) 8/46 (17.39) All-cause fractures age 40, N (%) 107/223 (47.98) 85/177 (48.02) 22/46 (47.83) Fell within the past 12 months, N (%) 71/223 (31.84) 58/177 (32.77) 13/46 (28.26) Abbr: abmd, areal bone mineral density; BMI, body mass index; SD, standard deviation.

6 pqct Motion Artifact on Muscle Density and Fractures 135 Table 2 Comparison of Motion Streak Cross-Sectional Area (MS-CSA), Motion Streak Density (MS-D), and Muscle Density (MD) Mean (Standard Deviation) Values for Semi-Automated and Fully Automated Algorithms Organized According to Motion Grade Semi-automated algorithm Fully automated algorithm Motion grade All All MS-CSA (mm 2 ) ( )* (836.18)* ( )* (226.22) (148.28) (280.99) MS-D (mg/cm 3 ) (17.07)* (19.59)* (7.90)* (83.27) (94.08) (32.11) MD UC (mg/cm 3 ) (4.10)* (4.14)* (3.98)* (3.91) (3.95) (3.67) MD C (mg/cm 3 ) (4.06)*, ** (4.10)*, ** (3.85)** (4.49)** (3.81)** (5.96)** Abbr: MS-CSA, motion streak cross-sectional area; UC, motion-uncorrected; C, motion-corrected. *Statistically significant difference between semi-automated and fully automated values (p < 0.001). **Statistically significant difference between MD UC and MD C values (p < 0.01). Table 3 Analysis-Reanalysis Reproducibility Values for Motion Streak Cross-Sectional Area (MS-CSA), Motion Streak Density (MS-D), and Muscle Density (MD) Measurements Obtained Using the Semi-Automated Segmentation Algorithm Output measure RMSCV (%) RMSSD LSC MS-CSA mm mm 2 MS-D mg/cm mg/cm 3 MD UC mg/cm mg/cm 3 MD C mg/cm mg/cm 3 Abbr: C, motion-corrected; LSC, least significant change; RMSCV, root mean square coefficient of variation; RMSSD, root mean square standard deviation; UC, motion-uncorrected. compared to fully automated MS-CSA (Spearman s rho = 0.65, p < 0.01). Analysis-reanalysis RMSCV values for MS-CSA and MS-D measured using the semi-automated algorithm were slightly above the 5% benchmark (Table 3). Lower leg MD values with and without removing MS as measured using the semi-automated method all had acceptable precision error under 5%. The LOA between MD UC vs MD C using semi-automated segmentation was 3.18 to 2.35 mg/cm 3, and 6.32 to 0.35 mg/cm 3 using the fully automated algorithm (Fig. 5A and B). The larger discrepancies in MD values before and after correcting for motion using either segmentation method can be seen in the 2 Bland-Altman plots where motion grade 5 data points lied beyond the LOA. This pattern of discrepancy toward higher motion grades was more obvious in data generated from the semi-automated method than from the fully automated method. Association With Fractures Neither MS-CSA nor MS-D as computed using semiautomated or fully automated methods was, by itself, or after adjusting for covariates, associated with either fragility or all-cause fractures (data not shown). Neither MD UC nor MD C by itself was associated with fragility fractures overall when derived from either method (Table 4, overall column). Motion grade was a significant moderator for the relationship between MD and fragility fractures with and without motion-correction using any method. In particular, motion corrected but not uncorrected MD was significantly associated with fragility fractures when motion grade was 3. Adjusting motion uncorrected MD associations for MS- CSA as a covariate did not have the same effect. The relationship between MD derived from the fully automated algorithm and fragility fractures remained significant regardless of whether or not motion-correction was applied (Table 4). In contrast to the lack of a relationship between MD C and fragility fractures overall, there was a significant association with all-cause fractures. In addition, MD UC computed from the semi-automated method was related to all-cause fractures before and after adjusting for MS-CSA (Table 5, overall column). Unlike analyses examining fragility fractures, motion grade 3 was not a significant moderator for relationships between MD and all-cause fractures. In either case, within individuals with motion grade 3, there was a significant relationship between MD and all-cause fractures regardless of motion-correction using the semi-automated method. No significant associations between any MD variable and all-cause fractures were found with or without motion-correction when the fullyautomated method was used (Table 5). Discussion Summary of Results In this study of 223 women (mean age: ± 7.07 yr, mean body mass index: ± 4.97 kg/m 2 ), pqct-derived leg motion streak-corrected muscle density measures were significantly lower than those without motion streak correction applied using either semi-automated or fully automated methods. This finding suggests that motion

7 136 Chan et al. Fig. 5. Bland-Altman analysis comparing the difference between muscle densities (MD) measured before (MD UC) vs after (MD C) excluding motion streaks measured using the semi-automated method (A) and the fully automated method (B). Upper and lower dashed lines indicate 95% limits of agreement. Solid thick line indicates the mean of differences between MD UC and MD C. Plotted points are darker shades as motion grade is higher. streaks are on average higher in density than muscle. The magnitude of MD difference before and after motioncorrection was larger for those with a higher qualitative grade of motion. The hypothesis that MD would still relate to fractures despite motion streak correction was supported by the fully automated method of motion streak quantification and correction within motion grades 3 to 5 images. Within semi-automated methods, this relationship was strengthened after motion-correction for both fragility and all-cause fractures. Effect of Motion Streaks on Muscle Density Mean MD C was smaller than mean MD UC for both segmentation methods, indicating that motion streaks tended to falsely elevate MD in general, despite the fact that both higher and lower density edges appear adjacent to the cortical shell. To our knowledge, only 1 other study by Blew et al (6) evaluated motion streaks in pqct scans in 506 healthy girls (9 13 yr) using Move%. Among children who had rescans due to movement, images with 25% Move% (corresponding to motion grade 3.33) had significantly higher MDs (1.19 ± 1.1 mg/cm 3 larger, p < 0.01) before compared to after repeated scanning. This observation supports the higher MD observed in motion-uncorrected images in the present study compared to after removing motion streaks using either semi-automated or fully automated segmentation techniques. Although MD before and after motion-correction for grades 1 3 semi-quantitatively assessed images were significant, the absolute difference (0.22 mg/cm 3 ) was much smaller than previously reported clinically least significant change values of 3.96 mg/cm 3 for MD yielded from a similar semi-quantitative assessment method (7). The large sample size (N = 176) resulting in a small standard error (0.22 mg/cm 3 ) likely contributed to this apparent significant difference.

8 pqct Motion Artifact on Muscle Density and Fractures 137 Table 4 Odds Ratios for Fragility Fractures Associated With Differences in Muscle Density (MD) With and Without Motion-Correction Using Semi-Automated or Fully Automated Segmentation Algorithms Independent variable Overall (N = 223, 32Fx) p Int motion 3 (N = 79, 9Fx) motion < 3 (N = 144, 23Fx) Per SD +/ (mg/cm 3 ) Semi-automated MD C 1.04 (0.67, 1.60) (1.05, 4.59) 0.68 (0.38, 1.21) 3.91 MD UC 0.95 (0.61, 1.50) (0.82, 3.17) 0.64 (0.36, 1.17) 3.95 MD UC + MS-CSA 0.94 (0.60, 1.48) (0.82, 3.10) 0.64 (0.35, 1.17) 3.95 Fully automated MD C 1.03 (0.66, 1.62) (1.06, 5.34) 0.72 (0.43, 1.21) 3.32 MD UC 1.14 (0.75, 1.74) (1.16, 5.08) 0.74 (0.42, 1.29) 3.75 MD UC + MS-CSA 1.15 (0.76, 1.76) (1.02, 4.64) 0.70 (0.39, 1.24) 3.75 Binary logistic regression models computed odds ratios (OR) for fragility fractures (FFx) based on differences in lower leg muscle density (MD) with and without motion-correction. Motion-uncorrected MD was also examined for fracture associations after adjusting for motion-streak cross-sectional area (MS-CSA) as a covariate. All models were adjusted for age, body mass index, total hip areal bone mineral density, duration of glucocorticoid use, and parental history of hip fracture. Moderation analyses examined the interaction of motion grade and MD variables on FFx. ORs were expressed per standard deviation (SD) increase (+) or decrease ( ) in the primary variable overall, in those with motion grade 3, and in those with motion grade < 3. Abbr: C, motion-corrected; CI, confidence interval; Fx, fractures; Int, interaction; MD, muscle density; MS-CSA, motion streak crosssectional area; UC, motion-uncorrected. All models had nonsignificant Hosmer-Lemeshow p values. Bold value indicates statistically significant association between fragility fractures and differences in muscle density at the 95% confidence level. Table 5 Odds Ratios for All-Cause Fractures Associated With Differences in Muscle Density (MD) With and Without Motion-Correction Using Semi-Automated Or Fully Automated Segmentation Algorithms Independent variable Overall (N = 223, 107Fx) p Int motion 3 (N = 79, 36Fx) motion < 3 (N = 144, 71Fx) Per SD +/ (mg/cm 3 ) Semi-automated MD C 1.49 (1.05, 2.12) (1.10, 3.45) 1.29 (0.86, 1.93) 3.91 MD UC 1.47 (1.03, 2.09) (1.02, 3.05) 1.29 (0.85, 1.95) 3.95 MD UC + MS-CSA 1.45 (1.02, 2.06) (1.01, 3.03) 1.28 (0.85, 1.94) 3.95 Fully automated MD C 1.21 (0.87, 1.68) (0.98, 3.60) 1.05 (0.73, 1.50) 3.32 MD UC 1.29 (0.92, 1.80) (0.89, 2.52) 1.17 (0.79, 1.74) 3.75 MD UC + MS-CSA 1.31 (0.93, 1.83) (0.92, 2.84) 1.21 (0.81, 1.80) 3.75 Binary logistic regression models computed odd ratios (OR) for all-cause fractures (AllFx) based on differences in lower leg muscle density (MD) with and without motion-correction. Motion-uncorrected MD was also examined for fracture associations after adjusting for motion-streak cross-sectional area (MS-CSA) as a covariate. All models were adjusted for age, body mass index, total hip areal bone mineral density, duration of glucocorticoid use, and parental history of hip fracture. Moderation analyses examined the interaction of motion grade and MD variables on AllFx. ORs were expressed per standard deviation (SD) increase (+) or decrease ( ) in the primary variable overall, in those with motion grade 3, and in those with motion grade < 3. Abbr: C, motion-corrected; CI, confidence interval; Fx, fractures; MD, muscle density; Int, interaction; MS-CSA, motion streak crosssectional area; UC, motion-uncorrected. All models had nonsignificant Hosmer-Lemeshow p values. Bold value indicates statistically significant association between all-cause fractures and differences in muscle density at the 95% confidence level.

9 138 Chan et al. The smaller motion streak areas measured by the fullyautomated algorithm compared to the semi-automated method suggest that automated thresholds may underestimate motion streak areas, given that the semi-automated method is based on careful visual inspection of the motion streak boundaries. Although threshold-based methods compute segmentations more quickly and reliably (RMSCV < 5%) compared to region growing, a global threshold applied across all individuals is less accurate in defining the true boundaries between muscle and motion streaks (smaller correlation against qualitative motion grade assessment), which is often difficult to discern due to the minimal contrast between these features. It is possible that the fully utomated threshold-based method did capture as much lower-density (negative) motion streaks, thus resulting in a larger apparent difference in MD before and after motion-correction (Table 2). Previous studies using high-resolution (HR)-pQCT examined how motion artifact affects the measurement of various bone parameters (12). In repeated distal radius and tibia scans of 248 men and women, the presence of movement caused larger systematic errors in the measurement of microarchitecture followed by densitometric parameters. These results suggest that density parameters including muscle density may be more resilient to motion compared to finer structural elements. To achieve acceptable measurement precision, Pauchard et al advised that analyzed HR-pQCT scans generally should not exceed a manual motion grade of 3 (13). Based on results of motion streak exclusion from MD measurements (Table 3), it also appears reasonable to repeat scans with pqct leg scan motion grades above 3, as supported by the notably smaller MD value between correction and no correction in motion grades 4 and 5. This recommendation is also supported by Blew et al who suggested that grades 1 to 3 motion images are not predicted to impact analysis (6). Although these recommendations by Blew et al were based on prior experience, no other pqct studies have provided further indications on when to repeat scans, particularly those completed for muscle analysis. Precision of Motion Streak Measurements RMSCV% values for the semi-automated segmentation of motion streaks were above the 5% benchmark of acceptable precision (MS-CSA = 6.44%, MS-D = 8.17%), but analysis-reanalysis reproducibility values for MD with and without motion-correction were more favorable (MD UC = 1.46%, MD C = 1.15%). The subjectivity required when deciding which pixels along the fuzzy boundary between muscle and motion streak to include likely contributed toward the lack of precision in segmenting motion streaks using the semi-automated method. Despite this challenge, precision of MD using the same semi-automated (watershed) method proved to be acceptable due to clearly defined boundaries between muscle and subcutaneous fat, thus resulting in RMSCV% for MD of well within 2%. In fact, Wong et al showed a similar level of precision for MD values (RMSCV% = 1.18% 2.01%) across different cohorts (younger adults [18 30 years old], older adults [>60 years old], and adults with spinal cord injury [8 45 years old]) using the semi-automated watershed algorithm (7). Until more standardized segmentation approaches are developed, the precision of motion streak area values may remain less desirable than MD. Motion Streak s Effect on Muscle s Association With Fractures The hypothesis that properties of muscle beyond motion still relate to fractures was supported by the persistent relationship between motion streak-corrected MD and both fragility and all-cause fractures when examining those individuals with grade 3 motion. Although this relationship was true even when including all individuals for all-cause fractures, the effect of motion-correction was stronger in those who have more motion in their scans. Indeed, motion grade was a significant moderator to the effect of motion-correction on strengthening the MDfracture relationships shown here. Together, lower MD itself may be a true risk factor for fractures independently of motion. The fact that fully automated software-derived MD associated with fragility fractures in higher motion grades regardless of motion-correction may suggest that the threshold-based algorithm was already sufficient in excluding some aspects of the motion streaks responsible for obscuring the relation between MD and fractures. Using an inner threshold of 40 mg/cm 3 in the basic muscle analysis removes intermuscular fat regions, but may also exclude the lower density (negative) motion streaks. Therefore, by using the basic muscle analysis alone, it is possible that some aspects of negative motion are already removed, thus limiting motion s effect on obscuring the MD-fracture relationship. In contrast, the semi-automated method captures the full muscle region without excluding intermuscular fat. Therefore, efforts to exclude motion in this semi-automated method may show a more dramatic effect on MD and consequently its association with fractures. Another study by Wong et al demonstrated that lower pqct-derived leg MD at 66% of the tibial length was associated with increased odds for fragility fractures in women 50 yr of age (OR = 7.16 [2.17, 23.42]) (9). The significant association between fragility fractures and MD after physically correcting for motion but not by statistical adjustment for MS-CSA suggests that it is not primarily motion streak size that may be introducing measurement error, but likely motion streak s contribution to variance in density that prevented the observation of a more precise relationship between MD and fractures. Although the present crosssectional study did not focus specifically on lower-limb fractures and did not examine incident fractures, the relationship between mid-thigh MD and incident hip fractures (RR = 1.58 [1.10, 1.99]) has been reported by Lang et al in 2941 men and women of the Health, Aging, and Body Composition study (age: years) (14).

10 pqct Motion Artifact on Muscle Density and Fractures 139 Study Limitations Although the precision of the motion streak quantification methods have been described here, the motioncorrected MD measures have not been truly validated against scans of the same participant without motion. Future studies will further validate these two techniques by comparing MD in a grade 1 scan and MD C in scans of higher motion grades within the same individual. Semi-quantitative motion streak assessment was completed on SliceOmatic software package, which is separate from the pqct manufacturer s Stratec software and would require additional costs. However, watershed-guided segmentation with manual editing can be developed in custom analysis packages such as Matlab. The sample size used in this study was sufficiently large to generate precise effect sizes for ORs involving all-cause fractures (N = 103), but the number of participants with prior fragility fractures was lower (N = 32). Therefore, associations concerning this outcome may appear smaller with wider confidence intervals. This study may also have lacked statistical power when modeling associations that excluded images with motion grades 1 and 2. The majority of the analyzed scans were comprised of these images (144/223 = 64.6%), which may have diluted the strength of the reported associations. That being said, all models showed adequate fit according to Hosmer-Lemeshow tests. This study was also a cross-sectional analysis of prevalent fractures, which makes it difficult to establish the direction of causality between fractures and muscle properties. It is possible that lower leg MD and greater motion are consequences of a previous fracture, rather than a risk factor for future fractures. Prospective studies that capture incident fractures as a primary outcome are needed to further explain the relationship between muscle density, motion, and fractures. Furthermore, this older adult cohort only included women. It remains unknown whether there are significant sex-related differences for the association between MD and fractures, and how the presence and severity of motion streaks may impact this relationship. Finally, it would be beneficial to study how motion severity may have an effect on measurements at anatomical sites where motion is more common, such as the mid- or distal radius. In conclusion, the presence of motion streaks within pqct lower leg scans leads to an overestimation of muscle density, particularly in those with more severe motion grades ( 3). Motion streak properties themselves were not direct risk factors for fragility or all-cause fractures but may be a factor limiting the observation of significant associations between lower muscle density and fragility or allcause fractures. Future studies should account for the amount of motion within these scans, and repeated scanning is recommended in those with motion grades above 3. For more meticulous assessment of motion, the fully automated threshold-based method using Stratec software may be useful for lower motion grades (<3), but it may underestimate the amount of motion in motion grades 3 or above. Alternatively, semi-automated segmentation using the watershed algorithm can capture motion streaks more completely, but lacked reanalysis precision. In either case, muscle density values generated by correcting for motion streaks using this method were successful in recovering significant relationships between muscle density and fragility fractures. References 1. Engelke K, Adams JE, Armbrecht G, et al Clinical use of quantitative computed tomography and peripheral quantitative computed tomography in the management of osteoporosis in adults: the 2007 ISCD Official Positions. J Clin Densitom 11(1): Frank-Wilson AW, Farthing JP, Chilibeck PD, et al Lower leg muscle density is independently associated with fall status in community-dwelling older adults. Osteoporos Int 27(7): Moore CD, Craven BC, Thabane L, et al Lowerextremity muscle atrophy and fat infiltration after chronic spinal cord injury. J Musculoskelet Neuronal Interact 15(1): Frost HM Bone s mechanostat: a 2003 update. Anat Rec A Discov Mol Cell Evol Biol 275(2): Furrer R, van Schoor NM, de Haan A, et al Genderspecific associations between physical functioning, bone quality, and fracture risk in older people. Calcif Tissue Int 94(5): Blew RM, Lee VR, Farr JN, et al Standardizing evaluation of pqct image quality in the presence of subject movement: qualitative versus quantitative assessment. Calcif Tissue Int 94(2): Wong AK, Hummel K, Moore C, et al Improving reliability of pqct-derived muscle area and density measures using a watershed algorithm for muscle and fat segmentation. J Clin Densitom 18(1): Tenenhouse A, Joseph L, Kreiger N, et al Estimation of the prevalence of low bone density in Canadian women and men using a population-specific DXA reference standard: the Canadian Multicentre Osteoporosis Study (CaMos). Osteoporos Int 11(10): Wong AK, Beattie KA, Min KK, et al Peripheral quantitative computed tomography-derived muscle density and peripheral magnetic resonance imaging-derived muscle adiposity: precision and associations with fragility fractures in women. J Musculoskelet Neuronal Interact 14(4): Schiferl DJ Fat vs. Lean Analysis Manual V7.3x. Fort Atkinson, WI: Bone Diagnostic, Inc, Gluer CC, Blake G, Lu Y, et al Accurate assessment of precision errors: how to measure the reproducibility of bone densitometry techniques. Osteoporos Int 5(4): Pialat JB, Burghardt AJ, Sode M, et al Visual grading of motion induced image degradation in high resolution peripheral computed tomography: impact of image quality on measures of bone density and micro-architecture. Bone 50(1): Pauchard Y, Liphardt AM, Macdonald HM, et al Quality control for bone quality parameters affected by subject motion in high-resolution peripheral quantitative computed tomography. Bone 50(6): Lang T, Cauley JA, Tylavsky F, et al Computed tomographic measurements of thigh muscle cross-sectional area and attenuation coefficient predict hip fracture: the health, aging, and body composition study. J Bone Miner Res 25(3):