Computed tomography has improved precision for prediction of intramuscular fat percent in the M. longissimus thoracis et lumborum in cattle compared to manual grading

Accepted Manuscript Computed tomography has improved precision for prediction of intramuscular fat percent in the M. longissimus thoracis et lumborum in cattle compared to manual grading

F. Anderson, J. Cook, A. Williams, G.E. Gardner PII: DOI: Reference:

S0309-1740(18)30331-0 doi:10.1016/j.meatsci.2018.07.025 MESC 7639

To appear in:

Meat Science

Received date: Revised date: Accepted date:

22 March 2018 19 July 2018 19 July 2018

Please cite this article as: F. Anderson, J. Cook, A. Williams, G.E. Gardner , Computed tomography has improved precision for prediction of intramuscular fat percent in the M. longissimus thoracis et lumborum in cattle compared to manual grading. Mesc (2018), doi:10.1016/j.meatsci.2018.07.025

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT Computed tomography has improved precision for prediction of intramuscular fat percent in the M. longissimus thoracis et lumborum in cattle compared to manual grading. F. Andersona, J. Cookb, A. Williamsa, G.E. Gardnera

School of Veterinary and Life Sciences, Murdoch University, WA, Australia 6150

b

Scott Automation and Robotics Pty Ltd, Rydalmere, NSW, Australia, 2116

PT

a

AC

CE

PT E

D

MA

NU

SC

RI

*Corresponding author: Fiona Anderson. Email: [email protected]

ACCEPTED MANUSCRIPT Abstract This study assessed the ability of computed tomography (CT) to predict intramuscular fat (IMF) % in the M. longissimus thoracis et lumborum M. longissimus lumborum of 64 cattle undergoing Meat Standards Australia (MSA) marbling score. The M. longissimus lumborum was dissected from all carcasses, CT scanned from the 12th rib to the caudal aspect of the

PT

muscle, with IMF% determined in cranial and caudal regions. The striploin was dissected

RI

from carcasses at the 12th rib and CT scanned, with IMF% determined at the 12th rib of the M. longissimus thoracis (LT) and the caudal M. longissimus lumborum (LL). Prediction of IMF%

SC

using CT in the LT was moderate (R2 = 0.89, RMSE = 1.91) utilising CT pixel density and

NU

standard deviation, and better than prediction using MSA marbling score (R2 = 0.81, RMSE = 2.47). Prediction of IMF% in the caudal aspect of the LL was comparatively poor (R2 = 0.63,

MA

RMSE = 4.69), with IMF% 1.97% higher in this region (P<0.05). CT is a promising nondestructive method for determination of IMF% that was more accurate than MSA marbling

PT E

D

score.

Key words

AC

CE

MSA marbling score; fatness; beef; computed tomography; meat quality; Meat Standards Australia

ACCEPTED MANUSCRIPT 1.

Introduction The amount of intramuscular fat (IMF) % in meat is linked to eating quality assessments

for tenderness, juiciness and flavour (Pannier et al., 2014; Shorthose & Harris, 1991; Thompson, 2004), accounting for up to 15 % of the variation in palatability of beef (Dikeman, 1987). The use of Meat Standards Australia (MSA) carcass grading is important in the beef

PT

industry to predict eating quality (Polkinghorne, Thompson, Watson, Gee, & Porter, 2008;

RI

Watson, Polkinghorne, & Thompson, 2008) with MSA grade calculated using a number of

SC

observed or measured carcass traits. MSA Marbling Score is used as an estimate of IMF% in the prediction model and is visually determined by a carcass grader in the M. longissimus

NU

thoracis (LT) between the 5th and 13th rib (AUS-MEAT 2005, Polkinghorne et al. 2008). The accuracy and precision of visual marbling score for predicting IMF% is not well documented

MA

with early studies in Australian beef indicating it to be a poor predictor of IMF% (R2 = 0.32 at the 10/11th rib) (Taylor & Johnson, 1992), with other countries showing better results (R2 =

D

0.78 at 13th rib) (Savell, Cross, & Smith, 1986). Comparison of these studies is difficult with

PT E

only knowledge of the correlation coefficient, as the ranges in IMF% differed, as did the location of the assessment. Given that visual marbling scoring systems are captured by a

CE

human grader and at a single site of measurement, it is likely they are subject to imprecision and inaccuracy due to factors such as perception of numerosity (Bertamini, Zito, Scott-

AC

Samuel, & Hulleman, 2016) and individual grader skill. On this basis there is an opportunity to introduce a rapid and more accurate in-plant measurement of IMF%.

Computed tomography (CT) has been used to determine the IMF% of muscle in beef (Navajas et al., 2009; Prieto et al., 2010), lamb (Clelland et al., 2014) and pork (Anderson, Pethick, & Gardner, 2015; Font-i-Furnols, Brun, Tous, & Gispert, 2013) with varying degrees of precision. Techniques with higher precision generally rely on additional information such

ACCEPTED MANUSCRIPT as carcass weight and measures of fatness (Anderson et al., 2015; Clelland et al., 2014), making independent and rapid analysis of IMF% in muscle difficult. CT scans are a threedimensional representation of an object, comprised of a number of individual images (slices), with each image consisting of a matrix of pixels. A voxel is the three-dimensional representation of the pixel, with voxel size influenced by slice thickness and the distance

PT

between slices. Larger voxels are created when there is increased distance between each

RI

image and this impacts on the density of a voxel through partial averaging of the tissue within

SC

the voxel (Dowsett, Kenny, & Johnston, 2006; Heuscher & Vembar, 1999). Differentiating small tissue components in 3-dimensions becomes difficult when voxel sizes are large and

NU

partial averaging is increased, therefore on this basis an increase in voxel size is likely to decrease the precision of prediction of IMF%. A key aim of this experiment was to determine

MA

the precision of prediction of IMF% using CT scanning alone and what aspects of image

D

acquisition can be altered to enable rapid scanning yet still achieve an adequate prediction.

PT E

There is evidence to suggest that IMF% varies along the length of the M. longissimus lumborum M. longissimus thoracis et lumborum (LTL) in pigs (Faucitano, Rivest, Daigle,

CE

Lévesque, & Gariepy, 2004), cattle (Cook, Bray, & Weckel, 1964; Zembayashi & Lunt, 1995) and lamb (Anderson et al unpublished). In all species IMF% was found to be highest at the

AC

cranial thoracic and lumbar regions of the M. longissimus lumborum M. longissimus thoracis (LT) and caudal M. longissimus lumborum (LL) compared to the 12th thoracic vertebrae, likely reflecting that the entire LTL extends across a range of anatomical regions of the animal. Given the variation in IMF% along the muscle, modalities such as CT which analyse images along its entire length may offer better accuracy and precision of prediction along the muscle rather than relying on a point measurement between the 5th or 13th thoracic vertebrae. where IMF % is currently predicted using visual marbling score. Additionally, CT may be a

ACCEPTED MANUSCRIPT better predictor of IMF% compared to manual marbling score at these sites and therefore improve eating quality assessment of the M. longissimus lumborum LTL. The precision of prediction of IMF% has been shown to vary between muscles in lamb (Anderson et al., 2015). However given the M. longissimus lumborum LTL is relatively homogenous in nature it would be expected for the precision of CT prediction along the length of this muscle to be

RI

PT

consistent along its length.

This study investigated the ability of MSA marbling score and CT to predict IMF% at the

SC

cranial aspect of the M. longissimus lumborum 12th rib of LT, testing the hypothesis that CT

NU

will predict IMF% with greater precision than MSA marbling score. Furthermore it was hypothesised that the LT M. longissimus lumborum will have lower IMF% than the caudal M.

MA

longissimus lumborum LL, but that CT will predict IMF% with similar precision in both locations of this muscle. Due to the partial averaging of densities it was hypothesised that as

2.

PT E

D

voxel size increases the prediction of IMF% using CT will decrease in precision.

Materials and methods

CE

Striploins Samples were obtained from 32 steers and 32 cows processed at a commercial abattoir. The cattle underwent standard assessment for MSA traits (AUS Meat, 2005) which

AC

are reported in table 1 and included: hot carcass weight (HCWT); MSA marbling scores (judged in tenths of grades from 100 to 1100 and in this experiment determined at the same site as the LT sample); and ossification score (maturity score for assessment of physiological age measured in increments of 10 from 100 to 500). Following slaughter, striploins were dissected from the LTL from the 12th vertebrae of the LT to the caudal end of the LL. Prior to CT scanning two 6 cm samples were excised from the LTL and trimmed of fat: a 6 cm sample at the 12th rib (LT) and the caudal 6 cm of this muscle (LT). A Siemens SOMATOM

ACCEPTED MANUSCRIPT Sensation 40/64 CT scanner was used with scan settings described in Table 2. Changes in scan settings will impact on spatial resolution, contrast resolution, noise and ultimately the quality of an image (Seeram, 2016). The scan settings in this experiment were chosen to reflect production of a high quality image (High Quality), or to reflect rapid processing time

PT

(High Speed), with these high speed images usually reflecting a lower image quality.

RI

Immediately following CT scanning, the 6 cm samples from the LT and LL were vacuum

SC

packed and frozen, before being sent to Murdoch University for IMF% determination. Samples were freeze dried using a ScanVac CoolSafeTM freeze drier (LabogeneTM,

NU

Vassinerod, Denmark) and IMF % of each muscle sample determined using a near infrared procedure (NIR). NIR measurements were taken using a Spectro Star 2400 and all samples

MA

were subsequently calibrated against chloroform solvent extraction as detailed by Perry,

PT E

2.1. Image analysis

D

Shorthose, Ferguson, and Thompson (2001) (R2 = 0.99, RMSE = 1.2).

Image J (version 1.37v, National Institutes of Health, Bethesda, MD, USA, used in

CE

conjunction with Microsoft Excel) was used to process the images with pixel information obtained from all CT images. CT images were captured consecutively along the length of the

AC

LTL with a voxel depth of 0.6 mm. The CT image information was used in a variety of ways to determine the impact of image analysis on IMF% prediction: firstly, the impact of voxel depth using mean and standard deviation of all pixels; secondly, the use of thresholding techniques; and thirdly impact of using surrounding pixel information at various weightings (nearest neighbour technique). These three methods were also used to determine the impact that using High Speed and High Quality settings had on IMF% prediction in both the LT sample and caudal LL sample.

ACCEPTED MANUSCRIPT

2.1.1. Voxel depth To determine the impact of voxel depth on IMF% prediction, image sets had to be reconstructed as they were originally captured at 0.6 mm voxel depths. This was done by taking the mean of 5 slices (for a 3 mm reconstruction) or 10 slices (for a 6 mm

PT

reconstruction) matching voxels within adjacent images. This resulted in 3 separate image-

RI

sets for each muscle sample, one at 0.6 mm, one at 3 mm, and one at 6 mm voxel depths.

SC

Within each image-set the mean voxel density and standard deviation of voxel densities was then calculated using all images within that image-set. Thus mean voxel density and standard

NU

deviation of voxel densities was available for each muscle sample for a variety of different

MA

image acquisition methods, including high speed or high quality scanning.

The pixel density information and standard deviation of pixel density from the different

D

image acquisition methods was used to predict chemical IMF% in a general linear model

PT E

(SAS Version 9.1, SAS Institute, Cary, NC, USA). Within each model, the mean voxel density of all images for the sample and the average standard deviation of these voxel

CE

densities were included as covariates. Each model was also tested with HCWT as an additional covariate. Thus separate models were constructed for a variety of different image

AC

acquisition methods, including High Speed or High Quality scanning, with multiple consecutive images of voxel depths of 0.6, 3, or 6 mm.

2.1.2. Thresholding Thresholding analysis involved categorising pixels relative to a predetermined threshold of 970 HU with pixels above this value classified as lean, and below classified as fat. Given that scanned images were a two part mixture and captured only one muscle (i.e. no seams

ACCEPTED MANUSCRIPT between muscle groups), this implies that pixels categorised below the threshold of 970 HU can be classified as intramuscular fat. This threshold was identified as it represented the average pixel density of pure fat. Based on the number of fat and lean pixels determined by thresholding the following parameters were calculated: ratio of number of fat:lean pixels; average density of fat pixels; average density of lean pixels; percentage weight of fat in the

PT

sample (calculated by multiplying the pixels categorised as fat or lean by the density of each

SC

RI

tissue, summing and dividing by the calculated total tissue weight).

A correlation matrix (Table 3) was used to highlight which of these calculated measures

NU

were highly correlated with the average and standard deviation of all pixel densities and therefore likely to offer little extra precision when predicting IMF% beyond that already

MA

described. All threshold variables were included individually in a general linear model to determine their precision of prediction of IMF%, however in all cases these predictors

D

demonstrated less precision than when using individual mean and mean standard deviation of

PT E

pixel values for the entire image. Therefore these results not reported.

CE

2.1.3. Nearest neighbour technique This analysis used an “image smoothing” technique to adjust pixel densities within each

AC

image, after which the mean and standard deviation of pixel densities from every slice for each sample were used in general linear models to predict IMF%. This image smoothing technique was based upon adjusting pixel values for the densities of their nearest neighbour. For example if the pixel was weighted at 100%, then only the pixel information was used, compared to a weighting of 50% where the original pixel information was weighted at 50%, and the surrounding pixel information weighted at 50% to calculate a new pixel density. A range of weightings were used to calculate pixel densities including 75%, 50%, 25% and

ACCEPTED MANUSCRIPT 11%. This was repeated for both the High Speed and High Quality CT scan settings and similarly analysed as for voxel depth (2.11)

2.1.4. Comparison with industry standard A comparison was made between the current industry standard for IMF% prediction,

PT

MSA grading score, and the image analysis techniques. This was achieved using a general

RI

linear model containing MSA grading score with and without HCWT as a covariate to predict

Results

NU

3.

SC

IMF%.

A total of 2 samples (LT and LL sampling site) were collected from each of 64 beef cattle

MA

and analysed for IMF%, with the raw mean ± SD of IMF% of all samples was 8.4 ± 6.3. The raw mean ± SD (min, max) of IMF% and values in the LT samples was 7.32 ± 5.2 (1.8, 21.1)

D

and LL was 9.46 ± 7.1 (2.4, 31.0) samples were reported separately in Table 1. The mean ±

PT E

SD, minimum, and maximum for pixel average density and pixel standard deviation for the

CE

images are shown in Table 4.

3.1. IMF% positional analysis

AC

The IMF% at the caudal end of the LL M. longissimus lumborum was 1.97 IMF% greater (P<0.05) than the LT cranial sample. However, these two IMF% values were highly correlated with a correlation coefficient 0.96.

ACCEPTED MANUSCRIPT 3.2. Prediction of IMF% 3.2.1. Voxel depth There was a negative linear relationship (P<0.01) between IMF% and CT pixel density. The IMF% was initially predicted using all pixels from all slices and the standard deviation of these pixels. The ability of CT to predict IMF% was greater in the cranial LT section (Table

PT

5: Model 1, R2 = 0.82, RMSE = 2.46) compared to the caudal section LL (Table 5: Model 2,

RI

R2 = 0.63, RMSE = 4.69). This same result was reflected using the High Speed settings

SC

(Models 3 and 4, Table 5) with the difference in the precision in the LT cranial and LL caudal

NU

samples similar using both settings.

In the LT cranial samples, using either the High Quality or High Speed settings, when

MA

slices were reconstructed at 3 mm voxels the ability to predict IMF% was lower (Table 5: Model 5 and Model 7) compared to 0.6 mm voxel widths (Table 5: Model 1 and Model 3).

D

Precision was further reduced when the slice widths were 6 mm (Table 5: Model 9 and Model

PT E

11). . Alternatively, increasing the voxel width in the caudal samples from 0.6 mm (Table 5: Model 2, and Model 4) to 6 mm (Table 5: Model 10, and Model 12) made little difference to

CE

the precision of IMF% prediction for either the High Quality or High Speed images.

AC

3.2.2. Use of surrounding pixels

When pixel density was calculated using the nearest neighbour technique, the precision of prediction of IMF% varied depending on the weighting of the central pixel. As the weighting of the central pixel was reduced, the prediction of IMF% improved. This maximised when the weighting of the central pixel was 11% (Table 5: Models 13-16), compared to 100% weighting of the central pixel information (Table 5, Models 1-4). The best prediction of IMF% were those achieved in the cranial LT sample from the M. longissimus lumborum,

ACCEPTED MANUSCRIPT using an 11% weighting of the central pixel, with similar precision evident in the High Quality and High Speed scan settings (Table 5: Model 13, R2=0.88, RMSE=2.02 and Model 15, R2=0.89, RMSE=1.89). This precision is also demonstrated in Figure 1 with the prediction of IMF% shown in relation to ideal prediction.

PT

3.2.3. Thresholding techniques

RI

None of the variable tested in the thresholding technique improved the precision of

SC

IMF% prediction. The parameter with lowest correlation with average pixel density and standard deviation was average lean pixel density (correlation coefficient -0.54 and 0.28),

NU

however it added no further improvement to the precision of the models. Independently, lean

MA

pixel density had very poor precision for predicting IMF% (R2= 0.14, RMSE =2.66).

3.3. Use of additional carcass information and comparison with MSA grading

D

A comparison of MSA marbling score to predict IMF% was only made with the IMF% in

PT E

the LT cranial sample from the M. longissimus lumborum given the visual marbling score was based on examination of the muscle at this site most cranial surface of this muscle after

CE

carcass splitting. The MSA marbling score had good precision for predicting IMF% (R2 = 0.81, RMSE = 2.47), although this was less than that of CT (Table 5: Model 15, R2 = 0.89,

AC

RMSE = 1.91). The use of HCWT alone offers moderate prediction of IMF% in the LT M. longissimus lumborum, (R2 = 0.39, RMSE = 4.44). There was minimal improvement in the precision of IMF% prediction when HCWT was used in conjunction with MSA marbling score or the CT scanning information. The inclusion of other parameters such as ossification score and eye muscle area did not improve the ability to predict IMF% in either the High Quality or High Speed Scans.

ACCEPTED MANUSCRIPT 4.

Discussion

Scan settings and image analysis In support of our hypothesis when images were reconstructed into larger voxel widths of 3 mm and 6 mm, the prediction of IMF% using average pixel density and standard deviation was reduced. This indicates the importance of utilising the appropriate scan settings to

RI

PT

maximise precision of IMF% prediction.

SC

Further improvements to precision of prediction of IMF% in both the LT and LL within the LTL utilised the M. longissimus lumborum average pixel density and standard deviation

NU

of all pixels at 0.6 mm slice widths after image smoothing using the nearest neighbour technique (11% weighting of central pixel). Therefore, the nearest neighbour method has a

MA

significant advantage over using only the raw pixel values and their standard deviations.

D

The lack of significant difference in the ability to predict IMF% using the High Quality v

PT E

High Speed settings suggests that future scanning protocols can make use of high speed settings to further reduce scanning times. This will require further investigation under

CE

scenarios where greater voxel depths are employed, however this may imply that similar accuracies can be achieved by further lowering the exposure time enabling greater product

AC

through-put in future commercial prototypes. One possible consideration for commercial deployment is the large processing time taken when multiple images are scanned. However, carefully written improved code and future increases in computer processing speeds are likely to make this point redundant.

This study has utilised least squares regression to predict IMF% using CT pixel density with good precision. Partial least squares regression not shown to be advantageous in a

ACCEPTED MANUSCRIPT number of other studies (Font-i-Furnols et al., 2013; Kongsro & Gjerlaug-Enger, 2013; Lambe et al., 2017). In contrast, Prieto et al. (2010) used partial least squares regression to predict of IMF% to achieve superior precision of prediction (R2 = 0.76, RMSEC of 471.99 mg FA 100g-1 muscle) in beef.

PT

Differences in IMF% and precision of prediction along the M. longissimus thoracis et

RI

lumborum.

In support of our hypothesis the cranial 6 cm LT sample from the 12th rib of the excised

SC

striploin M. longissimus lumborum (12th rib) was significantly lower in IMF% (7.32 ± 5.2)

NU

compared to the caudal 6 cm LL sample (9.5 ± 7.1) from of this muscle. This is similar to work in pigs (Faucitano et al., 2004), lamb (Anderson et al. unpublished) and cattle (Cook et

MA

al., 1964; Zembayashi & Lunt, 1995). Given that this experiment sampled only two regions of the LTL muscle we cannot determine whether this change occurs in a linear fashion from the

D

lumbar to sacral region of the muscle M. longissimus lumborum, or the magnitude and

PT E

direction of change in IMF% in the cranial the thoracic region of the muscle. Future work is poised to explore this relationship. The lower IMF% in the LT portion of the LTL cranial

CE

region of the M. longissimus lumborum is in contrast to the predicted eating quality scores using the MSA model, which are consistently higher for the cranial end of the striploin (LT)

AC

than the caudal region (LL) (Watson et al., 2008). However Alternatively, IMF% is just one of multiple components that contribute to the prediction of eating quality within this muscle (Muchenje et al., 2009).

In contrast to our hypothesis, CT scanning provided a superior prediction of IMF% in the LT compared to the caudal LL region of the striploin for all analysis techniques investigated. The variation in IMF% or the marbling distribution along the muscle may be responsible for

ACCEPTED MANUSCRIPT the change in precision and highlights the importance of developing predictions in individual locations of a muscle and between different muscles. The study of Prieto et al. (2010) states that prediction in Angus cattle was more precise than in Limousin cattle due to their higher IMF%. However, in the current study the caudal LL had higher IMF% and a greater standard deviation in values but lower precision of prediction than the LT sample, which indicates

PT

IMF% alone is unlikely to be the reason for variation in precision of IMF% prediction.

RI

Anderson et al. (2015) observed variation in the precision of prediction of IMF% between

SC

muscles irrespective of IMF% which further supports calibration of CT prediction in different carcass locations. The structure of the IMF within the muscle and size of marbling are likely

NU

to have an impact on prediction of IMF%. Smaller flecks of IMF are likely to be recorded as higher density due to the partial averaging of density within a voxel which also supports the

MA

findings of decreased precision with increasing voxel size observed in this study.

D

Comparison with MSA grading and inclusion of other carcass data

PT E

The precision of prediction of IMF% was greater using CT (R2 = 0.88, RMSE = 2.02) than MSA grading (R2 = 0.81, RMSE = 2.47) in this data set. These results supports the

CE

potential use of CT technology for the prediction of IMF% and therefore eating quality in the MSA grading system. CT will provide a more repeatable and accurate result in a commercial

AC

setting than the visual MSA marble score which is likely to vary between graders and within grader between days due to human error and factors such as perception of numerosity (Bertamini, Zito, Scott-Samuel, & Hulleman, 2016). Thus in a commercial setting where MSA marbling score is assessed by a range of graders of varied skill levels and at greater speed, the accuracy of IMF% prediction using MSA marbling score may diminish. A further advantage of technologies such as CT is its potential for use in “Hot” carcasses, and therefore employed sooner after processing to enable decisions regarding boning to better utilise

ACCEPTED MANUSCRIPT carcasses and maximise carcass value. The cost of this technology is a consideration for implementation and its use for IMF% prediction alone is unlikely, however these results indicate a useful application in the event of CT technology being incorporated into processing plants for other reasons such as yield and cut prediction.

PT

When HCWT was included in the models, the accuracy of IMF% prediction was

RI

improved only marginally, similar to results in beef of Prieto et al (2010). However, ideally

SC

the IMF% method should rely on direct information, rather than phenotypically correlated information from other traits such as carcass weight. This is because future breeding values

NU

need to control these traits independently, thus predicting IMF% using carcass weight limits the capacity to select for a high carcass weight and high IMF% independently. None-the-less,

MA

at the level of paying producers for the phenotype delivered, HCWT is likely to be useful for

D

further improving the precision of this feedback.

PT E

Comparisons with other species such as lamb are difficult to make as some utilise pixel information from the cross section of the carcass, which therefore includes other fat depots

CE

(Clelland et al., 2014). Another study in lamb, IMF% was predicted with moderate precision (R2 = 0.36, RMSE = 0.62) (Lambe et al., 2017), however this prediction model utilised

AC

information from carcass weight, and loin weight and thus is not an independent predictor of IMF%either. It is ideal for CT prediction methods to utilise pixel information only from within the muscle rather than including other information such as muscle weight, subcutaneous or intermuscular fat so that IMF% is predicted independently from correlates with other fat measures.

ACCEPTED MANUSCRIPT 5.

Conclusions and future work CT offers a rapid and moderately precise method for prediction of IMF%. The use of CT

with the settings described in this report predict IMF% in the LTL M. longissimus lumborum with moderate to high precision, with 6 mm slice width and the ‘nearest neighbour technique’ using the High Quality scan settings offering the best prediction. The precision to which CT

PT

predicts IMF% in this data set is better than the current industry standard, which utilises MSA

RI

marbling score and is similar to the precision of other CT techniques published for beef. This

SC

study also highlights the change in IMF% along the length of the striploin (LTL) M. longissimus lumborum and the difference in precision of prediction along the length of at the

NU

cranial and caudal aspects of the muscle. Extension of CT imaging technology to whole carcasses or beef primals would allow better prediction of IMF% and eating quality of cuts

MA

and has the potential to improve carcass utilisation. Furthermore, three-dimensional image analysis such as CT offers advantages over image analysis or visual scoring at a cut surface,

D

which must rely on prediction of IMF% at a site distant to the site of image acquisition. The

PT E

use of CT to determine IMF% in a commercial setting is a realistic proposition for the future,

CE

especially given the advances in technology which make rapid CT assessment a possibility.

Future investigations may involve alternative methods of data analysis such as a 3

AC

dimensional nearest neighbour calculation or the use of alternative scan settings to improve precision and speed of scanning. There is scope for extending the range of IMF% and MSA marbling score of samples to determine the precision of IMF% prediction across this extended range in a commercial setting for both CT and MSA manual grading. The methods used to predict IMF% needs validation in hot samples, and validation across a number of larger independent data.

ACCEPTED MANUSCRIPT Acknowledgments The authors would like to thank Meat and Livestock Australia for funding of this experiment. The authors gratefully acknowledge the contributions of staff employed at Murdoch University and Scott Automation and Robotics Pty Ltd for funding and data collection

PT

support.

AC

CE

PT E

D

MA

NU

SC

RI

Anderson, F., Pethick, D., & Gardner, G. (2015). The correlation of intramuscular fat content between muscles of the lamb carcass and the use of computed tomography to predict intramuscular fat percentage in lambs. animal, 9(7), 1239-1249. AUS-MEAT (2005). Handbook of Australian Meat 7th Edition (International Red Meat Manual). AUSMEAT, South Brisbane, QLD, Australia. Bertamini, M., Zito, M., Scott-Samuel, N. E., & Hulleman, J. (2016). Spatial clustering and its effect on perceived clustering, numerosity, and dispersion. Attention, Perception, & Psychophysics, 78(5), 1460-1471. Clelland, N., Bunger, L., McLean, K. A., Conington, J., Maltin, C., Knott, S., & Lambe, N. R. (2014). Prediction of intramuscular fat levels in Texel lamb loins using X-ray computed tomography scanning. Meat Science, 98(2), 263-271. doi: http://dx.doi.org/10.1016/j.meatsci.2014.06.004 Cook, C., Bray, R., & Weckel, K. (1964). Variations in the quantity and distribution of lipid in the bovine longissimus dorsi. Journal of Animal Science, 23(2), 329-331. Dikeman, M. (1987). Fat reduction in animals and the effects on palatability and consumer acceptance of meat products. Proceedings- 40th Annual Reciprocal Meat Conference of the American Meat Science Association, June 28th - July 1st 1987, Chicago, USA, 93-103. Dowsett, D. J., Kenny, P. A., & Johnston, R. E. (2006). The Physics of Diagnostic Imaging Second Edition. London, UK: Hodder Arnold. Faucitano, L., Rivest, J., Daigle, J. P., Lévesque, J., & Gariepy, C. (2004). Distribution of intramuscular fat content and marbling within the longissimus muscle of pigs. Canadian Journal of Animal Science, 84(1), 57-61. doi: 10.4141/A03-064 Font-i-Furnols, M., Brun, A., Tous, N., & Gispert, M. (2013). Use of linear regression and partial least square regression to predict intramuscular fat of pig loin computed tomography images. Chemometrics and Intelligent Laboratory Systems, 122, 58-64. doi: http://dx.doi.org/10.1016/j.chemolab.2013.01.005 Heuscher, D., & Vembar, M. (1999). Reduced partial volume artifacts using spiral computed tomography and an integrating interpolator. Medical physics, 26(2), 276-286. Kongsro, J., & Gjerlaug-Enger, E. (2013). In vivo prediction of intramuscular fat in pigs using computed tomography. Open Journal of Animal Sciences, 3, 321. Lambe, N. R., McLean, K. A., Gordon, J., Evans, D., Clelland, N., & Bunger, L. (2017). Prediction of intramuscular fat content using CT scanning of packaged lamb cuts and relationships with meat eating quality. Meat Science, 123(Supplement C), 112-119. doi: https://doi.org/10.1016/j.meatsci.2016.09.008 Muchenje, V., Dzama, K., Chimonyo, M., Strydom, P. E., Hugo, A., & Raats, J. G. (2009). Some biochemical aspects pertaining to beef eating quality and consumer health: A review. Food Chemistry, 112(2), 279-289. doi: https://doi.org/10.1016/j.foodchem.2008.05.103 Navajas, E., Richardson, R., Glasbey, C., Prieto, N., Ross, D., Hyslop, J., Simm, G., & Roehe, R. (2009). Associations between beef density by X-ray computed tomography, intramuscular fat and

ACCEPTED MANUSCRIPT

AC

CE

PT E

D

MA

NU

SC

RI

PT

fatty acid composition: Preliminary results. Proceedings of the British Society of Animal Science, Southport, UK, 117. Pannier, L., Gardner, G. E., Pearce, K. L., McDonagh, M., Ball, A. J., Jacob, R. H., & Pethick, D. W. (2014). Associations of sire estimated breeding values and objective meat quality measurements with sensory scores in Australian lamb. Meat Science, 96(2, Part B), 10761087. doi: http://dx.doi.org/10.1016/j.meatsci.2013.07.037 Perry, D., Shorthose, W. R., Ferguson, D. M., & Thompson, J. M. (2001). Methods used in the CRC program for the determination of carcass yield and beef quality. Australian Journal of Experimental Agriculture, 41(7), 953-957. doi: http://dx.doi.org/10.1071/EA00092 Polkinghorne, R., Thompson, J., Watson, R., Gee, A., & Porter, M. (2008). Evolution of the Meat Standards Australia (MSA) beef grading system. Australian Journal of Experimental Agriculture, 48(11), 1351-1359. Prieto, N., Navajas, E., Richardson, R., Ross, D., Hyslop, J., Simm, G., & Roehe, R. (2010). Predicting beef cuts composition, fatty acids and meat quality characteristics by spiral computed tomography. Meat Science, 86(3), 770-779. Savell, J., Cross, H., & Smith, G. (1986). Percentage ether extractable fat and moisture content of beef longissimus muscle as related to USDA marbling score. Journal of Food Science, 51(3), 838839. Seeram, E. (2016). Computed Tomography, Physical Principles, Clinical Applications, and Quality Control 4th Edition. St Louis, Missouri, USA: Elsevier Health Sciences. Shorthose, W., & Harris, P. (1991). Effects of growth and composition on meat quality. Meat Science, 7, 515-549. Taylor, D., & Johnson, E. (1992). Visual marbling score and chemical fat content of M. longissimus in beef carcasses. Proceedings of the Australian Society of Animal Production, Melbourne, Victoria, Australia, 71-73. Thompson, J. (2004). The effects of marbling on flavour and juiciness scores of cooked beef, after adjusting to a constant tenderness. Animal Production Science, 44(7), 645-652. Watson, R., Polkinghorne, R., & Thompson, J. M. (2008). Development of the Meat Standards Australia (MSA) prediction model for beef palatability. Australian Journal of Experimental Agriculture, 48(11), 1368-1379. doi: https://doi.org/10.1071/EA07184 Zembayashi, M., & Lunt, D. (1995). Distribution of intramuscular lipid throughout M. longissimus thoracis et lumborum in Japanese Black, Japanese Shorthorn, Holstein and Japanese Black crossbreds. Meat science, 40(2), 211-216.

ACCEPTED MANUSCRIPT Figure 1. The prediction of intramuscular fat % in the M. longissimus thoracis at the 12th vertebrae using computed tomography and the nearest neighbour technique (11% weighting of central pixel) at High Quality scan settings. Solid

AC

CE

PT E

D

MA

NU

SC

RI

PT

line represents perfect prediction, with dots representing residuals to this relationship.

ACCEPTED MANUSCRIPT

Table 1. Carcass data including mean ± standard deviation, minimum and maximum values for M. longissimus thoracis et lumborum (LTL) length (mm), hot standard carcass weight (kg), hump height (mm), eye muscle area (cm2), ossification score, Meat Standards Australia (MSA) marbling score and intramuscular fat % of cranial and caudal samples.

390, 555

318.6 ± 56.3 224.0, 443.0

50.2 ± 14.4

73.5 ± 10.1

20.0, 85.0

40.0, 94.0

CE

PT E

D

MA

NU

SC

LTL: M. longissimus thoracis et lumborum

AC

Max

Min,

452 ± 31.9

Eye muscle area (cm2)

Ossificatio MSA n score Marbling Score (MSA range: (MSA Range: 100-500) 100-1190) 259.1 ± 425.5 ± 172.2 253.7 130.0, 253.7, 590.0 1120.0

PT

Mean ± SD

Hump height (mm)

RI

Hot LTL length standard carcass (mm) weight (kg)

ACCEPTED MANUSCRIPT Table 2. Computed tomography scan settings used on the Siemens SOMATOM Sensation 40/64 scanner for High Quality and High Speed protocols for prediction of intramuscular fat in the M. longissimus thoracic et lumborum.

mm) Scan Speed (mm/s)

AC

CE

PT E

D

MA

NU

Pitch

PT

512 x 512 140 500 500 500 471 157 65 0.977 54 1.5

RI

X-ray Tube Current Exposure (mAs) Generator Power Pixel Spacing (x and y;

512 x 512140 500 500 500 383 383 53 0.977 17.5 0.5

High Speed scan settings 0.6

SC

Slice Thickness (mm) Matrix size (pixels) kVp Data Collection Diameter (mm)Reconstruction Diameter (mm)Rotation Time (ms)

High Quality scan settings 0.6

ACCEPTED MANUSCRIPT Table 3. Simple correlation coefficients of the computer tomography derived parameters used to determine IMF% in the M. longissimus thoracic et lumborum of beef: mean and standard deviation of all pixel densities, ratio of fat:lean pixels, average density of fat pixels, average density of lean pixels and percentage weight of fat. Mean of all pixel densities

AC

CE

PT E

PT

D

MA

NU

SC

Average density of fat pixels Average density of lean pixelsPercentage weight of fat

1 -

RI

Mean pixel density Standard deviation of pixel density Ratio of fat:lean pixels

Stand Ratio Avera Perce Avera ard of number ge density ntage ge density deviation fat:lean of lean weight of of fat pixels of all pixel pixels pixels fat densities -0.87 0.945 0.93 -0.54 0.93 1 -0.89 -0.64 0.28 -0.87 1 0.79 -0.66 0.99 1 -0.56 0.79 1 -0.68 1

ACCEPTED MANUSCRIPT Table 4. Mean ± SD, minimum and maximum of sample average pixel density and their standard deviations for CT images using all slices (0.6 mm voxels), all slices with 3 mm reconstructed voxels, and all slices 6 mm reconstructed voxels. Pixel density

PT

SD 8.9 6.1 6.3

AC

CE

PT E

D

MA

NU

SC

RI

All slices: 0.6 mm voxels All slices: 3 mm reconstructed voxelsAll slices: 6 mm reconstructed voxels

Mean ± (min, max) SD 1039.6±1 (1006.8, 1.9 1034.2 1062.7) (1014.7, ±10.61021.7 1055.7) (998.4, ±11.0 1046.7)

Standard deviation of pixels Mean ± (min, 103.6 ± max)(82.8, 113.2 ± 130.6)(97.8, 128.0 ± 128.5)(108.3, 146.5)

ACCEPTED MANUSCRIPT Table 5. F-values, coefficient, intercept, coefficient of determination (R-square), and root mean square error (RMSE) for models predicting intramuscular fat % in beef using the High Quality and High Speed settings, average computed tomography pixel density, standard deviation, with 100% weighting of central pixel, 3 mm and 6 mm voxel sizes and 11% weighting of the central pixel in the M. longissimus thoracis (LT) and M. longissimus lumborum (LL) caudal end of the M. longissimus thoracis et lumborum. Mod M el odel 1 2

Scan type Location

Mod el 3

Mo del 4

Mo del 6

Mo del 7

Mo del 8

100% weighing of original pixel (0.6 mm slice width)

All slices: 3 mm reconstructed voxels

High Quality

High Quality

High Speed

LT

LL

LT

LL

Average pixel density Standard Deviation

178. 39* 234. 98*

88. 32* 50. 1*

217. 07* 273. 82*

80. 26* 53. 77*

Average pixel density Standard Deviation

-0.71

Intercept

Mod el 5

-

-0.74

0.71 -1.03 856. 6

R2

0.82

RMSE

2.46

0.72 81 7.6

* P<0.01 M. longissimus thoracis: LT

E C

-1.05 884. 6

C A

0.6 3 4.6 9

PT

0.69

0.84 2.31

0.65 78 8.7

0.6 1 4.5 8

LL

M

N A 87. 72* 40. 52*

-

I R

High Quality

0.6 1 4.5 7

High Speed

LT

LL

LT

LL

54.9 * 132. 5*

89. 95* 57. 53*

52.9 2* 115. 37*

71. 94* 43. 27*

-0.48

0.69 0.85 814 .2

T P

Mod M el odel 12 11

All slices: 6 mm reconstructed voxels

C S U

High Speed

LT LL LT F values 86.86 90. 103. * 14* 56* 170.5 51. 172. 8* 97* 56* Coefficients and intercept -0.57 -0.6 0.69 -0.85 0.83 0.97 697.4 81 733. 3.4 3 Precision estimates 0.6 0.74 0.76 4 4.6 2.89 2.8 3

D E

Mod M el odel 10 9

-

-0.51

0.66 -0.66 587. 7

0.68 3.26

0.77 77 4.1

0.6 2 4.7 2

-0.67 614

0.65 3.4

M Mod M M odel el odel 14 odel 16 13 15 11% weighing of original pixel (0.6 mm slice width) High High Speed Quality LT LL LT LL

18 11 3.37* 8.65* 43 14 9.16* 1.73*

-

-

-

0.6

0.57

0.62

0.66 69 7.7

0.95 70 0

1.18 78 3.4

0.5 7

0.8 8

4.7 8

0.7 6

2.0 2

3.7 5

221. 94. 7* 5* 498. 11 45* 2.43*

-0.59

0.59

-0.94 720. 9

0.89 1.91

1.01 73 6

0.7 1 3.9 1

ACCEPTED MANUSCRIPT M. longissimus lumborum: LL

T P

I R

C S U

N A

D E

T P E

A

C C

M

ACCEPTED MANUSCRIPT

Highlights 

Computed tomography is more precise than MSA marbling score when predicting IMF %



CT can predict IMF% in the M. longissimus lumborum of beef with moderate

CE

PT E

D

MA

NU

SC

RI

IMF% prediction is more precise at the 12th rib than caudal M. longissimus lumborum

AC



PT

precision

Figure 1