A review of computed tomography and manual

Meat Science 123 (2017) 35–44

Contents lists available at ScienceDirect

Meat Science journal homepage: www.elsevier.com/locate/meatsci

Review

A review of computed tomography and manual dissection for calibration of devices for pig carcass classification - Evaluation of uncertainty Eli V. Olsen ⁎, Lars Bager Christensen, Dennis Brandborg Nielsen Danish Technological Institute, DMRI, Gregersensvej 9, DK-2630 Taastrup, Denmark

a r t i c l e

i n f o

Article history: Received 20 April 2016 Received in revised form 30 August 2016 Accepted 31 August 2016 Available online 01 September 2016 Keywords: Pig carcass classification Reference methods Uncertainty CT scanning Dissection Lean meat percentage, LMP

a b s t r a c t Online pig carcass classification methods require calibration against a reference standard. More than 30 years ago, the first reference standard in the EU was defined as the total amount of lean meat in the carcass obtained by manual dissection. Later, the definition was simplified to include only the most important parts of the carcass to obtain a better balance between accuracy and cost. Recently, computed tomography (CT) obtained using medical X-ray scanners has been proposed as a reference standard. The error sources of both traditional (manual) dissection methods and the new methods based on images from CT scanning of pig carcasses are discussed in this paper. The uncertainty resulting from the effect of various error sources is estimated. We conclude that, without standardisation, the uncertainty is considerable for all the methods. However, methods based on volume estimation using CT and image analysis might lead to higher accuracy if necessary precautions are taken with respect to measuring protocol and reference materials. © 2016 Published by Elsevier Ltd.

Contents 1.

Introduction . . . . . . . . . . . . . . . . . . . . . . . 1.1. Approved LMP reference methods . . . . . . . . . . 1.2. Computed tomography . . . . . . . . . . . . . . . 1.3. Purpose . . . . . . . . . . . . . . . . . . . . . . 1.4. The concept of uncertainty . . . . . . . . . . . . . 1.5. Metrological principles . . . . . . . . . . . . . . . 1.6. Type A and Type B estimates . . . . . . . . . . . . 2. Standard presentation and dissection . . . . . . . . . . . . 2.1. Evaluation of uncertainty of the standard presentation. 2.2. Evaluation of dissection uncertainty . . . . . . . . . 3. Reference methods based on computed tomography . . . . . 3.1. Spectral analysis . . . . . . . . . . . . . . . . . . 3.2. Segmentation . . . . . . . . . . . . . . . . . . . 3.3. Virtual LMP . . . . . . . . . . . . . . . . . . . . 3.4. Uncertainty in computed tomography . . . . . . . . 3.5. Traceability . . . . . . . . . . . . . . . . . . . . 3.6. Summary of uncertainty in computed tomography . . 4. Summary and general discussions . . . . . . . . . . . . . 5. Conclusions . . . . . . . . . . . . . . . . . . . . . . . Acknowledgement . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . .

⁎ Corresponding author. E-mail address: [email protected] (E.V. Olsen).

http://dx.doi.org/10.1016/j.meatsci.2016.08.013 0309-1740/© 2016 Published by Elsevier Ltd.

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

36 36 36 36 36 36 36 37 37 37 38 39 39 40 41 41 43 43 44 44 44

36

E.V. Olsen et al. / Meat Science 123 (2017) 35–44

1. Introduction Objective online classification of pig carcasses is obtained indirectly by measuring relevant characteristics at the slaughter line. Typically, these characteristics include the back fat thickness and frequently also the thickness of the loin. The characteristics must be measurable and highly correlated with the total lean meat percentage (LMP). It is necessary to use a reference method to determine the parameters of a model that converts the measurands of the online equipment to a predicted value of LMP. This process essentially constitutes a calibration of the online equipment. Ideally, the LMP should be obtained from measurements including the whole carcass, which is possible when using tomographic methods (X-ray or magnetic resonance imaging). However, the methods are still not available for online measurements. A commonly used online technology is a combined optical insertion probe and a ruler, which utilizes the different reflectivity from meat and fat obtained at near infrared wavelengths. Current online methods are non-invasive and based on, for instance, ultrasound, measuring the reflections from different layers of tissues and fascia. Market harmonisation and the need for transparency resulted in the adoption of EU regulations for objective classification in 1984/1985 (Council Regulation (EEC), 1984; Commission Regulation (EEC), 1985). Although the regulations do not limit the use of technology, they impose requirements on the predictive capability of the methods as evaluated by the quality of calibration. Statistical guidelines (Causeur et al. 2003) propose cost-effective solutions to practical problems related to calibration, including sampling.

actual facilities in 2015 appears in a report from the COST network FAIM (2011). Results and experiences from various investigations have been reported partly as international collaboration within the COST network. The uses of CT data are numerous, for example Clelland et al. (2014), Daumas & Monziols (2011), Font-i-Furnols et al. (2015), Petnehazy et al. (2012), Lambe et al. (2013), Vester-Christensen et al. (2009). This paper is restricted to a discussion of the application of CT as a reference method as an alternative to manual knife dissection. 1.3. Purpose The purpose of having harmonised rules for classification of pigs within the EU is to provide a reliable and common basis for evaluation of the carcass value expressed as LMP. It is expected that the common rules will ensure that all approved online methods will produce approximately the same results if they can be compared directly on the same carcasses. Approval of an online classification method is only granted if a satisfactory performance of the calibration for a specific pig population is documented. In this paper, we focus on the uncertainty of the approved reference methods. Lists of uncertainty contributions are drawn up based on available estimates from both published and unpublished experiments conducted during the past ten years, providing a budget of uncertainty of the reference methods. Our purpose is to draw attention to the metrological aspects and, in particular, to identify critical factors that should be standardised to improve the robustness and reliability of the European reference system for online classification of pig carcasses.

1.1. Approved LMP reference methods 1.4. The concept of uncertainty The very first common LMP reference method was accepted in 1984. It was adopted from a German dissection method used as the reference method for tissue separation in an international experiment reported by Merkus (1979). The formal definition of LMP is the ratio between the total weight of lean meat in the carcass separated with a knife and the total weight of the carcass. It is a very time-consuming method, and quicker methods were accepted on the national scale until 1994, when a new common simplified reference standard was accepted. The agreement with the previous standard was obtained by introducing a correction factor of 1.3. The accuracy was estimated and reported by Nissen et al. (2006). As a consequence of the results, the method was revised once more in 2008, yielding a higher accuracy. Today, three reference methods are accepted in the EU (Commission regulation (EC), 2008). The first method is a total dissection, excluding the head, which is close to the original definition. The second method is a simplified method defined by the ratio of the weight of lean meat in four main cuts of the carcass plus the weight of the tenderloin and the total weight of the same four cuts and the tenderloin. The ratio is multiplied by a factor of 0.89 to account for the non-dissected parts compared to total dissection. In this paper, total dissection and simplified dissection are referred to as “knife dissection”. Recently, it has become possible to use computed tomography (CT), provided an acceptable correlation with knife dissection methods can be demonstrated. No formal requirements explaining what is meant by “acceptable correlation” are provided. 1.2. Computed tomography In the context of this paper, CT is an X-ray technique used for medical imaging. The use of CT for the study of farm animals is very encouraging. The method has been used for several years in breed selection programmes for lambs in the UK (Bünger et al. 2011) and pigs in Norway (Topigs Norsvin). The method's utility as an objective reference method for online classification was investigated in an inter-European project EUPIGCLASS (2000). Subsequently, several research institutions have acquired a CT scanner designed for human medical use. A list of the

Measurement imperfection can give rise to an error in the measurement result. The error can be viewed as having two components: a random component and a systematic component. The random effect arises from unpredictable variation, while the systematic effect arises from a recognised effect of an influence quantity. It might be possible to reduce the effect of influence quantities, although, typically, the degree of influence will vary from one situation to another. Consequently, influence quantities will be evaluated based on estimated contributions to the uncertainty together with random error. 1.5. Metrological principles The uncertainty of the accepted reference standard has been evaluated by a European research group and reported by Nissen et al. (2006). The work used the metrological principles described in the international standard ISO 5725 (2012) and was also inspired by the Guide to the Expression of Uncertainty in Measurements, (BIPM, 2008). The aim of the work was to document the complete chain of uncertainties included in the references. Traceability is a core concept in metrology. It is defined by the International Bureau of Weights and Measures, BIPM, as “the property of the result of a measurement or the value of a standard whereby it can be related to stated references, usually national or international standards, through an unbroken chain of comparisons, all having stated uncertainties.” The concept is illustrated in Fig. 1. The uncertainty of the standard references for calibration of online measurements will be discussed below in the framework of Fig. 1. 1.6. Type A and Type B estimates Uncertainty is estimated using two types of variance estimates (see BIPM, 2008). The type A uncertainty estimate is obtained from the experimental variance of observations, which are typically considered as outcomes from a Gaussian distribution. The type B uncertainty estimate is evaluated by scientific judgement based on available information. If


Fig. 1. Illustration of the concept of traceability to the International System of Units (SI).

the range of outcomes can be determined without any knowledge of the distribution, the uniform distribution is assumed together with the variance estimate a2/3, where “a” denotes the half range. 2. Standard presentation and dissection Using the terminology in Fig. 1, the international (European) primary standard for the content of lean meat in a pig carcass, LMP0, is defined as the ratio between the weight of “red striated muscles from all parts (except the head) of the carcass as far as separable by knife” and the weight of the carcass. LMP 0 ¼

Leanweight 100% Carcassweight

ð1Þ

The simplified method is defined by ∑ Leanweight i

LMP s ¼

i¼1−5

∑ Totalweight i

0:89 100%

ð2Þ

i¼1−5

where Total weighti, i = 1,…5, refers to the weight of the shoulder, loin, ham, belly and tenderloin, and Lean weighti, i = 1,…5, refers to the weight of lean meat in the shoulder, loin, ham and belly and the weight of the tenderloin. Transforming the definitions into practice gives rise to the following questions: • • • •

How should a carcass be defined? Can we assume a symmetrical carcass and only dissect half a carcass? Is “as far as separable by knife” an unambiguous instruction? Do tendons, glands and blood vessels belong to lean meat?

37

weight is approximately 1 kg, corresponding to approximately 0.5 LMP. Even though the effect is systematic, the true effect is not known. The effect depends on the slaughter process in question, and the contribution to the uncertainty is evaluated by a type B estimate, assuming a = 0.25 LMP. The effect of removing the genital organs, flare fat, etc. can be considered similar to that mentioned above, and, since no data are available, a type B estimate at the same level as the removing of bristles is assumed. The carcass is split into two halves using a saw or an axe. The row of processus spinosus describes the middle of the back, and ideally they should be split in the middle with half a bone to each side. This is the equivalent of the splitting of the sternum bone, which ideally also defines the line of symmetry of the carcass. Sometimes, the head is also split or is cut off, leaving the cheeks on the two sides. The splitting often introduces a degree of error irrespective of the method, be it manual or automatic, and investigations of the symmetry of carcasses are thus made difficult. However, it is possible to compensate for mis-splitting of the sternum bone and obtain the standard presentation. The error related to the splitting of the carcass is based on data available from a study reported by Hviid et al. (2011), and a type A estimate is possible. In Nissen et al. (2006), the right side LMP is estimated to be 0.6 LMP higher than the left side. An investigation by Hviid et al. (2011) supports this result, although the effect is only half the size, and furthermore the difference is caused primarily by a different amount of fat so that the effect is only through the denominator of the LMP ratio. As a standard, only the left side is dissected. The carcass head poses a specific problem. In Europe, the head represents very little value, with only the cheeks representing a slight value, and often the head is cut off, leaving the cheeks on the carcass sides. The cut must be carried out very carefully to minimise the uncertainty. The effect of the size of the cheek has been evaluated using data from a Scandinavian study that included two test samples using standard preparation and one test sample using the normal cutting procedure at the slaughter line. In Sweden, the head is cut off as part of the slaughter process, and the variation of the cheeks is significantly bigger than in Denmark and Norway, where the head is cut off in a controlled process. The uncertainty evaluation of type A is obtained from LMP estimated from CT scanning using methods explained in section 3.3 and by comparing calculations of LMP including and excluding the weight of the cheek. The data are shown in Table 1. A summary of all identified influence quantities related to carcass presentation is given in Table 2. In Table 2, it is assumed that the uncertainty contributions are independent. The total uncertainty is obtained from the sum of variances and is expressed as a standard deviation. The effect of preparation on the determination of LMP is estimated to be 0.77 LMP, with the major contribution from the splitting. However, if the standard procedure is followed, the uncertainty is reduced to 0.14 LMP. 2.2. Evaluation of dissection uncertainty

2.1. Evaluation of uncertainty of the standard presentation A number of factors such as those indicated above influence the presentation of a carcass, and, consequently, a standardisation has been established. The standard presentation of a carcass is defined as the body of a slaughtered pig, bled and eviscerated, without tongue, bristles, hooves, genital organs, flare fat, kidneys and diaphragm. An evaluation of the most important effects is included in this review to recall the importance of standardisation. Some factors are related to the equipment used in the slaughter process and the slaughterhouse workers carrying out the slaughtering and preparation work. For example, the bristles are removed with a combination of scalding with hot water or steam followed by mechanical scraping, a process which is known to differ from place to place. A highly effective scraping process can easily remove 1 mm of the skin surface compared with a less effective process, and, since the area of a carcass surface is approximately one square metre, the effect on the carcass

The definition of lean meat constitutes the next challenge, firstly in terms of the amount of meat “separable by knife”. Nissen et al. (2006) report the precision of the 1994 version of the simplified method of LMPs. An experiment was carried out to compare the results from dissections prepared by a team of European trained butchers. Differences of up to 2 LMP were observed. The detailed analysis showed that the main problem was related to the cutting of the carcass into the main

Table 1 The weight of the cheek and its effect on LMP. Combined results from studies in Denmark, Norway and Sweden. Variable

N

Mean

Std. dev.

Minimum

Maximum

Effect of cheek on LMP Amount of meat in cheek, kg Total weight of cheek, kg

632 632 632

0.550 0.309 0.824

0.241 0.143 0.363

0.053 0.006 0.052

1.091 0.560 1.518

38


Table 2 Summary of error sources related to preparation of the carcass and the estimated uncertainty contribution. Standard presentation of pig carcasses Error sources

Uncertainty estimate

Variance contribution

Removal of bristles

Type B: a =

0.02

Removal of tongue, hooves, genital organs, flare fat, kidneys and diaphragm Splitting into two halves

Cheeks

0.25 LMP Type B: a = 0.25 LMP Type A: Std. dev. = 0.71 Hviid et al. (2011) Type A: Std. dev. = 0.24 Based on Table 1

Total variance contribution assuming independency

0.02 The procedure is standardised 0.50 Only left side is used for reference purposes 0.06 The procedure is standardised 0.60 = 0.772 With standardisation: 0.02 = 0.142

cuts to be dissected. The dissection itself was carried out in a relatively uniform way. These results led to a revised version of the simplified LMPs (see definition (2)). The denominator was changed from being the weight of the whole carcass to being the weight of the pieces that were dissected together with a new correction factor. The change reduced the uncertainty significantly. The experiment was designed to estimate the repeatability standard deviation, which was calculated to be 0.5 LMP. It has been possible to recalculate the precision of the present version of LMPs (see definition (2)). The precision of LMP0 (see definition (1)) has not been estimated, but it is expected to be of the same size as (2). Visually, tendons, glands and blood vessels do not belong to any of the main tissue groups, i.e. fat, meat and bone. However, it is very time-consuming to separate these parts, and they are therefore included in the standard procedure. However, the decision to include or exclude some or all of them in the dissection is one of the differences between the working methods of various butchers. The extent of the problem is reported in Olsen et al. (2006). German and Danish experts evaluated the effect of including or excluding tendons and glands as part of meat tissue in a small study (9 carcasses). The average difference and standard deviation between LMP0 estimated with or without tendons and glands are estimated at −1.3 LMP and 0.18 LMP, respectively. Even though the standard procedure is to include tendons and glands, the butchers might perform the cutting in different ways, and the estimated standard deviation is considered as uncertainty contribution. In CT scanning, elements with approximately the same density cannot be separated. Consequently, since their protein content is high, tendons, glands, blood vessels and cartilage are considered as meat in CT scanning. Using prior knowledge of the anatomical structure in the image segmentation analysis might reduce the problem (see section 3.2). A

summary of identified influence quantities related to knife dissection is given in Table 3. Also in Table 3, the uncertainty contributions are assumed independent, giving an uncertainty estimate obtained from the sum of variances. The effect of the manual cutting and preparation on the determination of LMP is estimated to be 0.82 LMP. The estimate of minimum uncertainty (repeatability standard deviation), corresponding to an optimal knife dissection method with no systematic effect of the butchers, is reported by Nissen et al. (2006) and is estimated at 0.51 LMP. 3. Reference methods based on computed tomography X-rays are electromagnetic waves that can propagate through tissue with less attenuation than the electromagnetic waves of visible light. Xray detectors register the radiation transmitted through the measurement objects. Since various materials absorb X-rays with different degrees of efficiency depending on their density and atomic composition, an average attenuation projection is obtained corresponding to the different materials in the object penetrated by the ray. If this measurement is repeated all the way around an object, for example obtaining 360 projections, a cross section of the object can be reconstructed by means of a mathematical transformation of the projections. Different types of transformation exist, and the result is, for example, a 512 × 512 matrix (a tomogram), in which the matrix element values are the average attenuation in the corresponding volume element (ISO 15708-1, 2002), (ISO 15708-2, 2002). Typically, the cross-section values are reproduced as an image in grey tones reflecting different materials, preferably in a fixed scale format based on the absorption of air and water that defines two calibration points, often referred to as the Hounsfield Scale. Basically, the spatial resolution of the image depends on the resolution and size of the detector, the dimension of the focal spot size of the X-ray source, the slice thickness, the field of view and the number of projections, whereas the spectral resolution depends on the radiation dose. A selection of cross sections from a CT-scanned half pig carcass can be seen in Fig. 2. A close-up of one cross section can be seen in section 3.2. Each image has a resolution corresponding to 1 × 1 mm2 pixels, and each image represents a section of the object with a thickness of typically 1–10 mm. A pixel can be considered to represent a unit volume (V) of the measured object called a voxel, and the pixel value represents the average attenuation in the volume V. The CT scanning generates a three dimensional representation of the carcass in a large number of unit volumes. In addition to the three dimensional resolution, the final image quality depends on a number of factors such as the X-ray source voltage and current (typically in the range of 80–140 kV and 80–

Table 3 Summary of error sources related to knife dissection and the estimated uncertainty contribution. Knife dissection Error sources Butchers and working conditions

Uncertainty estimate

Type A: Std. dev. = 0.80 Data from Nissen et al. (2006) Tendons and glands Type A: Std. dev. = 0.17 Olsen et al. (2006) Total variance contribution assuming independency

Variance contribution 0.64

0.03 0.67 = 0.822

Fig. 2. Selection of cross sections from a CT-scanned half pig carcass.


200 mA), the field of view, the matrix size (usually 512 × 512), and the mathematical methods used to reconstruct the images. Ideally, these factors should be described and fixed in a measuring protocol. 3.1. Spectral analysis Spectral analysis utilizes only absorption data and not the spatial information. The tissue absorption, μx, is described on the so-called scale of Hounsfield units, HU: HU ¼ 1000

μ x −μ water μ water

ð3Þ

where the value 0 HU corresponds to water and −1000 HU corresponds to air. Meat and bone have a higher absorption than water and correspond to positive values, approximately + 60 HU and above + 250 HU respectively, while fat corresponds to negative values, approximately −80 HU. Fig. 3a shows the distribution of voxel values from a piece of scanned muscle (blue curve). Nearly all values are positive. As each voxel corresponds to a volume (V) the total volume of the piece of muscle is simply calculated as the area of the histogram multiplied by the size of V. The weight is obtained by multiplying the muscle volume by the muscle density. The green curve shows a similar histogram of pure fat, where most values are negative. The threshold value of 0 HU ideally separates pure

Fig. 3. a. Distribution of Hounsfield units (HU) in pure muscle, pure fat and dissected lean meat. b. Distribution of Hounsfield units (HU) in three carcasses at three levels of LMP. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

39

muscle and fat. Fig. 3a also shows the histogram of a piece of “lean meat”, i.e. meat including a small amount of fat. Observe that the three histograms overlap. Even though the histograms (spectra) do not separate for lean meat and fat exactly, it is usual to take advantage of the underlying assumption that meat and fat occupy separate regions of the Hounsfield scale and analyse the CT data as histograms instead of images. Examples of histograms of three scanned carcasses at three levels of LMP are shown in Fig. 3b. The method is used in Dobrowolski et al. (2004) and Font‐i‐Furnols et al. (2009), where the histogram is used as a multivariate regressor, and the LMP obtained with knife dissection is the dependent variate in a linear model. The parameters of the model are estimated using partial least squares regression (PLSR). Since the estimation of the parameters is dependent on data from a knife dissection, the solution is not a proper alternative to knife dissection and will not be discussed further in this paper.

3.2. Segmentation Segmentation analysis utilizes both absorption data and the spatial information. The overlap between the histograms shown in Fig. 3a is a consequence of ambiguities like those illustrated by the cross-section image in Fig. 4a. The white area illustrates bones, light grey depicts the meat, dark grey the fat and finally black the surrounding air. In the lower left part of the cross section, there is no clear distinction between fat and meat, resulting in intermediate pixel values close to 0, the same as water. The skin 1 is light grey, which is similar to meat due to its high protein content, and the marrow 2, which has an absorption like fat, is dark grey. By means of image analysis, it is possible to classify the pixels to obtain a better differentiation between fat and meat. Furthermore, the marrow can be identified and included in bone, and the skin

Fig. 4. a Cross-section image. [1] indicates the skin and [2] the marrow. b Cross-section image after segmentation in meat (red), fat (white) and bone (grey). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

40


included in fat despite having a density similar to that of meat, in both cases to simulate the manual dissection performed by a butcher (Fig. 4b). Vester-Christensen et al. (2009) and Lyckegaard et al. (2006) apply a multivariate Bayesian 2D contextual classification scheme to each slice as described by Larsen (2000), the so-called Owen-Hjort-Mohn classification (OHM). The intensity of a pixel as well as certain combinations of neighbouring pixels are taken into account and modelled in a Bayesian scheme with priors obtained from thresholds in the ideal histogram. In this way, the context, i.e. neighbours of each pixel, also influence the final classification of the specific pixel, making the method less susceptible to random noise and artefacts and thereby more robust. It could be relevant to use solutions based on similar smoothing methods, but in 3D, to make maximum use of information. OHM classification is considered robust to the parameters that form the measuring protocol. This is supported by the results from a small experiment (N = 5), in which the middle part of a pig carcass was scanned five times using varying radiation doses, giving varying spectrometric resolutions (10 mm slice thickness, 140 kV, soft reconstruction, 470 mm reconstruction diameter (0.918 mm pixel size)). LMPCT, formulas (4) and (5) in section 3.3, was estimated using three different algorithms: ‘OHM’, ‘THRH’ (based on the fixed threshold between meat and fat spectrum) and ‘Full width at half maximum’, FWHM (based on a threshold method, in which the threshold is obtained as the intersection between two Gaussian distributions which are assumed to fit the fat and meat histogram). The values relative to the value at 80 mA are calculated and shown in Fig. 5. The OHM algorithm appears to be less sensitive to the random noise originating from the reduced dose. The algorithm performance for higher doses will be published elsewhere. Furthermore, the uncertainty arising from using simple thresholds to classify the tissue types in the segmentation compared to the more robust OHM algorithm has been investigated in a realistic grading experiment in Finland, where 129 carcasses were scanned twice using two different scanning methods. First, step-and-shoot (similar to Axial in other scanner models) (8 mm steps/slice thickness) was used, and this was followed by a helical (10 mm slice thickness) scan. All other parameters were kept fixed (135 keV, 200 mA). Subsequently, the two segmentation methods were used, resulting in four sets of images. These data form the basis for evaluating the sensitivity of LMPCT, formulas (4) and (5) in section 3.3, with respect to the scanning

Table 4 The effect of scanning method (helical and step-and-shoot) on the estimation of LMPCT using either OHM or THRH segmentation.

Segmentation method

Mean difference of LMPCT (Helical – Step-and-shoot)

Std. dev. of difference

OHM THRH

−0.08 −0.62

0.146 0.201

protocol and segmentation method presented in 2015 to the COST network FAIM (2011). The results are included in Tables 4 and 5. The effect of different scanning and segmentation methods is estimated as the average variance of deviation including bias, i.e. average (std. dev.2 + bias2) (see Table 7). The benefit of using image analysis to segment the data based on OHM classification or similar methods is the avoidance of including a dissection, thus minimising the influence of the human factor, which often forms the major part of the uncertainty. Even though the image analysis mimics the dissection performed by a butcher, the agreement is not perfect, and calibration against the standard reference, LMP0, is necessary to obtain traceability to the present reference (see section 3.5). 3.3. Virtual LMP Assume a sample of N scanned carcasses and the corresponding CT scans (see Fig. 2). The total number of voxels classified in each tissue type (see section 3.2) multiplied by the voxel volume determines the tissue volume. The tissue volume multiplied by the tissue density determines the tissue weight. Since the three tissue groups do not consist of pure fat, meat and bone, respectively, the average tissue densities are estimated by means of the ordinary least squares regression (4). W i ¼ βfat V fat;i þ βmeat V meat;i þ βbone V bone;i þ εi ; i ¼ 1; …:; N ð4Þ where Wi is the total (scale) weight of the ith carcass. Vfat , i , Vmeat, i , Vbone , i the volume of fat, meat and bone, respectively, obtained with image analysis (section 3.2) from the ith CT scan. βfat , βmeat ,βbone the average density parameters of the three types of tissue.

Fig. 5. LMPCT estimated for the middle part of a pig carcass (average of five examples) at five different radiation doses using three segmentation algorithms: ‘Owen-Hjort-Mohn’ (OHM), ‘Threshold’ (THRH) and ‘Full width at half maximum’ (FWHM). The LMP values relative to the value at 80 mA are shown.

E.V. Olsen et al. / Meat Science 123 (2017) 35–44 Table 5 The effect of segmentation method (OHM and THRH) on the estimation of LMPCT using either helical or step-and-shoot scanning.

Scanning method

Mean Difference of LMPCT (OHM-THRH)

Std. dev. of difference

Helical Step-and-shoot

0.48 −0.06

0.244 0.224

εi Normal distributed error N(0,σ2). From (4), the lean meat percentage obtained with CT (LMPCT) is defined by (5) LMP CT ¼

βmeat V meat 100% W

ð5Þ

Three data sets are available from Denmark (N = 299), Norway (N = 232) and Sweden (N = 140). From these data sets, three sets of average densities defined in (4) are estimated (Table 6). The effect of density estimates are obtained from re-estimation of LMPCT using the three sets of data and densities. The results are shown in Table 7. LMTCT (OHM method) is used to investigate the effect of the cheek (see section 2.1). Data are obtained from three studies in Denmark, Norway and Sweden (N = 632). Before CT scanning of the carcass, the cheek was cut off and placed on a pad on top of the carcass during the scanning, and subsequently the voxels corresponding to the cheeks were isolated. The weight is estimated using the estimated density parameters (Table 6). The effect on LMP is evaluated using formula (5) with and without the cheek included (see Table 1). 3.4. Uncertainty in computed tomography Carmignato et al. (2009) list possible influencing factors in micro tomography such as the environment, the detector, the X-ray source and the data processing. Similar influencing factors are present when using medical scanners for geometric (quantitative) measurements. The reconstruction algorithms in medical scanners are optimised for the type of diagnosis, i.e. depending on whether you are looking for fractured bones or defects in soft tissues (e.g. brain or muscle). The choice of data filtering and threshold values are influencing factors together with the calibration against the Hounsfield scale, with water as “0” and air as “−1000” as the two-point calibration anchor points. A large number of parameters need to be chosen, and each one can only be set within certain tolerance limits. For instance, the Hounsfield calibration is typically set within ±5 HU. The size of pig carcasses in a national/regional pig population typically varies within a weight range of 30 kg, but the average weight depends on regional traditions. Ideally, a reference method should be valid for all types of pig carcasses used in meat production. The combined population of pigs in Europe varies from 50 kg to 180 kg, and it will hardly be possible to define an international standard measuring protocol that is valid for all types of carcasses. The field of view, which determines the carcass section included in the measurement, might be a limiting factor preventing the scanning of the entire carcass of very heavy pigs. Alternatively, the carcass must be divided into pieces to be scanned separately. However, the pieces will have a larger surface area than the entire carcass, resulting in a higher number of voxels being a mixture of air and fat, meat or bone. A higher resolution will reduce the problem but will result in a larger amount of data to be recorded and analysed. The attenuation is higher in heavy Table 6 Average densities (standard error) for fat, meat and bone tissue types in three cases.

Denmark (N = 299) Norway (N = 232) Sweden (N = 140)

Fat

Meat

Bone

0.997 (0.005) 0.976 (0.009) 0.990 (0.007)

1.117 (0.006) 1.105 (0.008) 1.120 (0.009)

1.433 (0.065) 1.434 (0.086) 1.419 (0.085)

41

carcasses than in light carcasses, and a higher dose is probably necessary, leading to faster wear of the X-ray source. The effect of different scanners and measuring protocols has been investigated in a small study in Denmark. Three medical scanners situated at different locations, but all within a short travelling distance of each other, were compared using hams as objects, which could easily be moved from place to place in a refrigerated van. Nine hams were scanned in the three scanners within a short interval of time (one day). The weight of each ham was estimated using the density estimates shown in the first row (Denmark) in Table 6. The estimated weights were compared to the scale weight. The average differences were −25 g, 34 g and 107 g for fat, meat and bone, respectively. The maximum effect on LMP is estimated at 1 LMP, resulting in a Type B variance estimate (a = 0.5) equal to 0.08 LMP (see Table 7). The robustness with respect to the measuring protocol has also been investigated by Christensen et al. (2006). The studies include the following factors: voltage/current (100 kV/110 mA vs. 140 kV/80 mA), voxel size (1 × 1 × 10 mm3 vs. 1 × 1 × 1 mm3) and the carcass position on the scanner table (rind side up vs. rind side down). The factors' effect on LMP was analysed using an analysis of variance model, and the contributions to the uncertainty are shown in Table 7. To overcome the difference between scanners (type and brand) and measuring protocols depending on the carcasses to be scanned, some robust and solid material that can detect potentially serious differences should be established. Carmignato et al. (2009) suggest reference standards that have the appearance of a regular array of inner and outer cylindrical features to obtain geometrical accuracy at micro level. Similar standards called phantoms have been proposed for CT scanners used for scanning pig carcasses. The analysis of the very first experiences from a round-robin test, based on phantoms that mimic the middle part of a pig carcass, is reported in Angel & De Chiffre (2014). The potential is promising, but the method needs to be optimised. An attempt to provide an overview of the stability over time has been made in the Danish scanner. Bars consisting of different materials are included in the scanner table, which is used as a support for the carcass during the measurement. An example is shown in Fig. 6a, with a close-up in Fig. 6b. Due to attenuation, the average value from each bar is influenced by the object being scanned. Consequently, only cross sections from the empty table can be used. In Fig. 7, the values obtained on each experimental day during four time periods are plotted, covering approximately two and a half years. Each point reflects the value of the cross sections of bars consisting of acrylic (approx. 120 HU), water in a tube (0 HU) and PVC (approx. 900 HU). The scanner was serviced three times, and a small shift in the level can be seen in the middle of the chart. This is probably an indication of the reproducibility of the maintenance, including the Hounsfield calibration, which is typically set within ±5 HU in water (0 HU) and air (−1000 HU). The same change must be expected in the cross sections of the carcass together with a small shift in the histograms in Fig. 3b. In this example, the consequence is probably small and of no importance for diagnostic purposes. However, geometrical measurements are, in principle, sensitive to errors of this type because of fixed thresholds, and it might be valuable to monitor the measurements using control charts like the one in Fig. 7. This could lead to new performance tests and possibly a Hounsfield calibration using four points instead of two. A complete account estimate of all the influencing instrumental factors is outside the scope of this paper. Hopefully, the example above illustrates some of the problems to consider to minimise the uncertainty and at same time obtain a balance between accuracy and cost. 3.5. Traceability The original definition of the lean meat percentage (LMP0) defined in (1) is based on the kilogram SI unit. The challenge is to transform

42


Fig. 6. a. Cross section of table and carcass placed on a pad filled with air (not visible). The three circles are cross sections of three bars consisting of acrylic (approx. 150 HU), water in a tube (0 HU) and PVC (approx. 900 HU). b. Close-up of the cross sections of bars of acrylic (approx. 120 HU), water in a tube (0 HU) and PVC (approx. 900 HU).

the reference method, including the simplified dissection reference, to national working standards, while ensuring a high degree of accuracy. Often, laboratories are inter-calibrated using samples of test materials, with the mean or median value being the consensus value if no accredited reference materials exist. A similar method cannot be established due to the lack of identical pigs. Alternatively, an

international team of butchers could be established to perform national reference dissections on samples of pigs representing the national pig populations. This idea has not been realised, and a pragmatic solution has been accepted for the time being. National reference methods for calibration of online measurement equipment are carried out by local butchers performing either total or simplified dissection.

Fig. 7. Response/time (experimental days) plot for average bar values in the scanner table.


43

Recently, CT has also become a legal reference. The method is traceable to the metre SI unit because of an unbroken chain of calibrations from metre to voxels, tissue volumes and LMPCT defined in (5). The chain can be extended in order to ensure comparable online results by adding a calibration of LMPCT to obtain the same level and scale of LMP as that obtained by knife dissection:

The total uncertainty obtained from Table 7 is 1.30 LMP, which might be a worst case scenario. A repeatability study resulted in a repeatability standard deviation estimated at 0.22 LMP, which means that computed tomography is potentially a promising reference method. However, the challenge is to reduce the effect of the identified error sources by standardisation.

LMP0 ¼ b LMPCT þ ε

4. Summary and general discussions

ð6Þ

where b is a scale parameter, and ε is normal distributed error N(0,σ2). The calibration parameter, b, is estimated using ordinary least squares regression. Data from Denmark (48 scanned carcasses), Norway (85 scanned carcasses), and Sweden (30 scanned carcasses) have been used to evaluate the influence quantity on uncertainty. LMPCT defined in (5) is estimated using segmentations carried out by the OHM algorithm. The scale transformation, b, is estimated for the three data sets together with a common estimate. A common estimate of b is 0.91 (std.err. 0.001). However, a bias is present in each data set (−0.24 (0.06), 0.41 (0.08), −0.81 (0.15), and the contribution to uncertainty is estimated by the std.err. (b) + average (bias2) (see Table 7). Sampling and experimental differences have been identified as main sources of differences between the data sets.

0.23

The uncertainty of legal reference standards that can be used for approval of online classification equipment in the EU has been estimated. The test material, i.e. the pig carcasses and especially the handling methods, represents a specific challenge for all methods. The splitting of the entire carcass into two halves is the main contribution to the uncertainty related to carcass presentation. However, standardisation reduces the uncertainty significantly. The effect of the butchers' working routines in simplified dissection has been minimised using the revised definition (2) compared to the one reported in Nissen et al. (2006). Tendons and glands might be treated differently and might also contribute to uncertainty. Computed tomography has still not been standardised. Based on the available estimation of uncertainty estimates, the measuring protocol seems to contribute the most. The estimated uncertainties are summarised in Table 8. Looking at knife dissection only, simplified dissection is the most commonly used reference standard. The standard is related to total dissection through a correction factor, and total dissection is defined by the kilogram SI unit. In principle, an unbroken chain of relations exists, which ensures that a measurement obtained online can be related to a precise reference defined by an SI unit and, consequently, form a traceable system. The only problem is that uncertainty of the relation between total dissection and simplified dissection has not been estimated. However, a European data set exists as a basis for the first investigations of CT as a reference method (EUPIGCLASS, 2000), in which both total and simplified dissection can be estimated for 120 pigs that are representative of the Hungarian pig population in the beginning of the century, excluding the Mangalica pigs, which are a specific Hungarian lard-type breed. Based on these data, a bias equal to −0.76 LMP between total (1) and simplified (2) dissection was estimated, and the standard error of the difference was also found to be 0.76 LMP. The mean squared error of the agreement between the two methods can be evaluated by Variance (mean difference) + bias2, which is a frequently used method to evaluate biased results or methods. In this case, the uncertainty between total and simplified dissection is estimated by ((0.762/120) + (−0.76)2)1/2 i.e. 0.76 LMP. The system of reference standards based on knife dissection is clearly not ideal regarding uncertainty. Currently, the only realistic alternatives are methods based on CT. LMPCT estimated in (5) can be considered traceable to the metre SI unit, even though the uncertainty at all levels has not been identified in detail. In principle, the relation ensuring agreement between LMPCT and LMP obtained with knife dissection introduces uncertainty from one method to the other. However, this uncertainty is less important because the correlation between the methods is high. In Daumas & Monziols (2011), the definition of LMPCT in parallel to (5) is claimed to be at the same level and scale as LMP obtained with a knife, and no correction is required. The method is not based on the entire carcass, but on the main cuts that introduce uncertainty related to the butchers. The potential exists for making a

0.18

Table 8 Summary of uncertainty related to LMP working standards.

3.6. Summary of uncertainty in computed tomography A summary of the investigated factors is shown in Table 7. There are still a number of factors that have not been investigated, such as the effect of the number and size of mixed voxels along the edges. Environmental factors are standardised in the studies available for this investigation, although they might differ between other studies. Regular inspections performed by authorised technicians ensure that the detector, X-ray and physical conditions are kept stable. The effect of scanner adjustments within tolerance limits together with wear and tear has not been investigated thoroughly yet.

Table 7 Summary of error sources related to LMP obtained with computed tomography and the estimated error contribution. Computed tomography – segmentation methods Error sources Scanner variation Measuring protocol • Voxel size: 1 × 1 × 10 mm3 vs. 1 × 1 × 1 mm3 Measuring protocol • Volt/Amp: 100 kV/110 mA vs. 140 kV/80 mA Measuring protocol • Rind side up vs. rind side down Type of scanning method

Uncertainty estimate Type B: a = 0.5 Type A: Std. dev. =

0.08 0.11

0.33 Christensen et al. (2006) Type A: Std. dev. =

0.46

0.68 Christensen et al. (2006) Type A: Std. dev. = 0.32 Christensen et al. (2006) Type A: Std. dev. =

Traceability

0.48 Based on Table 4 Type A Based on Table 5 Type A: Std. dev. =

Density estimates

0.54 Type A: Std. dev. =

Type of segmentation

Variance contribution

0.10

0.24

0.49 Based on Table 6 Total variance contribution assuming independency

Repeatability std. Uncertainty dev.

0.29

1.69 = 1.302

Carcass presentation • With standardisation Knife dissection, EU reference standards Computed tomography, without standardisation

0.77 LMP 0.14 LMP 0.82 LMP 1.30 LMP

Unknown 0.51 LMP 0.22 LMP

44


precise reference standard based on CT. However, more experience is needed. In particular, it must be determined how differences between scanners (brands and types (multi or single slice) and measuring protocols) can be managed. Effective and reliable reference materials (phantoms) might solve the problem. 5. Conclusions The present reference standards for calibration of online classification devices based on knife dissections can only be obtained with considerable global uncertainty, although the local uncertainty might be acceptable. The error sources and the effects are relatively well known. Reducing the uncertainty is not only a question of costs and skills; the experimental conditions and working methods also have a considerable impact on the results and might be difficult and costly to standardise. Replacing reference standards based on knife dissections with reference methods based on CT will generally improve the uncertainty, but only if precautions are taken with respect to robustness and stability of the individual CT scanner. The complete set of influencing factors must be identified and quantified. It is known that both X-ray source voltage and current influence the results. However, different brands of CT scanners do not have the same ranges of settings, and it will therefore only be possible to standardise these parameters within certain tolerance limits. However, the software used to reconstruct the images is even more difficult to standardise, because the software is treated as a “black box” by the producer, especially the mathematics used to reconstruct cross sections from helical multi-slice scanning. The main focus needs to be on developing phantoms that mimic the items we want to measure. The aim is to adjust the settings dependent on the outcome from scanned phantoms, i.e. carry out an instrumental calibration to ensure comparable measurements on the primary items (carcasses). A preliminary test shows that it is possible to simulate a middle part of a pig carcass. However, it is not evident how the test should be carried out and, in particular, how the data can be used to calibrate the scanner in question. Once these problems have been solved, it might be possible to define a true instrumental reference method based on CT scanning, i.e. a method that is not secondary to knife dissection, but traceable to an SI value and parallel to knife dissection. Acknowledgement The results reported in this article have mainly been obtained from research and development financed by the Danish Pig Levy Fund (year-toyear grants, in 2016 Grant No. 10) and the Danish Government (Grant No. 3414-05-01358 and 3414-08-02284). The authors are very grateful for the collaboration with the Danish Technological University, without which it would not have been possible to perform some of the basic research. The authors also thank the technicians who carried out the practical work, often at inconvenient times of the day. The authors are also very grateful for collaboration with Scandinavian colleagues, who agreed to provide data for some of the results. Finally, the authors thank Holger Dirac for his critical and constructive reading of the manuscript. References Angel, J., & De Chiffre, L. (2014). Comparison on computed tomography using industrial items. CIRP Annals - Manufacturing Technology, 63(1), 473–476 (http://doi.org/10. 1016/j.cirp.2014.03.034). BIPM (2008). Evaluation of measurement data — Guide to the expression of uncertainty in measurement. International Organization for Standardization Geneva ISBN, 50(September), 134 (http://doi.org/10.1373/clinchem.2003.030528). Bünger, L., Macfarlane, J. M., Lambe, N. R., Conington, J., McLean, K. A., & Moore, K. (2011). Use of X-ray computed tomography (CT) in UK sheep production and breeding. In K. Subburaj (Ed.), CT Scanning-Techniques and Applications, INTECH (pp. 329–348). InTech Retrieved from http://www.intechopen.com/books/ct-scanning-techniques-andapplications/use-of-x-ray-computed-tomography-ct-in-uk-sheep-production-andbreeding

Carmignato, S., Dreossi, D., Mancini, L., Marinello, F., Tromba, G., & Savio, E. (2009). Testing of x-ray microtomography systems using a traceable geometrical standard. Measurement Science and Technology, 20(8), 084021 (http://doi.org/10.1088/0957-0233/20/8/ 084021). Causeur, D., Daumas, G., Dhorne, T., Engel, B., Font, I., Furnols, M., & Højsgaard, S. (2003). Statistical handbook for assessing pig classification methods. Retrieved from http://ec. europa.eu/agriculture/pigmeat/policy-instruments/statistical-handbook-for-assessingpig-classification-methods_en.pdf Christensen, L. B., Borggaard, C., & Olsen, E. V. (2006). Robustness of weight and meat content in pigs determined by CT. In ICoMST, 1–4. 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), 1–9 (http://doi.org/10.1016/j.meatsci.2014.06. 004). Commission regulation (EC) (2008). Commission regulation (EC) No 1249/2008 of 10 December 2008 laying down detailed rules on the implementation of the Community scales for the classification of beef, pig and sheep carcases and the reporting of prices thereof. Retrieved February 10, 2016, from http://eur-lex.europa.eu/legal-content/ EN/TXT/PDF/?uri=CELEX:32008R1249&rid=1 Commission Regulation (EEC) (1985). Commission Regulation (EEC) No 2967/85 of 24 October 1985 laying down detailed rules for the application of the Community scale for grading pig carcases. Retrieved from http://eur-lex.europa.eu/legal-content/EN/TXT/? qid=1453975333482&uri=CELEX:31985R2967 Council regulation (EEC) (1984). Council regulation (EEC) No 3220/84 of 13 November 1984 determining the community scale for grading pig carcases. Retrieved from http://eurlex.europa.eu/legal-content/EN/TXT/?qid=1453975333482&uri=CELEX:31984R3220 Daumas, G., & Monziols, M. (2011). An accurate and simple computed tomography approach for measuring the lean meat percentage of pig cuts, (1249), 2009–2012. Dobrowolski, A., Romvári, R., Allen, P., Branscheid, W., & Horn, P. (2004). Schlachtkörperwertbestimmung beim Schwein. Fleischwirtschaft, 3, 109–112. EUPIGCLASS (2000). Standardisation of pig carcass classification in the eu through improved statistical procedures and new technological developments (EUPIGCLASS). Retrieved from http://www.cordis.europa.eu/project/rcn/51397_en.html FAIM (2011). Action FA1102: Optimising and standardising non-destructive imaging and spectroscopic methods to improve the determination of body composition and meat quality in farm animals (FAIM). Retrieved from http://w3.cost.eu/index.php?id= 181&action_number=FA1102 Font-i-Furnols, M., Carabús, A., Pomar, C., & Gispert, M. (2015). Estimation of carcass composition and cut composition from computed tomography images of live growing pigs of different genotypes. Animal: An International Journal of Animal Bioscience, 9(1), 166–178 (http://doi.org/10.1017/S1751731114002237). Font-i-Furnols, M., Teran, M. F., & Gispert, M. (2009). Estimation of lean meat content in pig carcasses using X-ray computed tomography and PLS regression. Chemometrics and Intelligent Laboratory Systems, 98(1), 31–37 (http://doi.org/10.1016/j.chemolab. 2009.04.009). Hviid, M., Erbou, S. G. H., & Olsen, E. V. (2011). The content of meat – Is there a difference between left and right sides of pig carcasses? ICoMST(August), 10–13. ISO 15708-1 (2002). ISO 15708-1:2002 - Non-destructive testing - Radiation methods Computed tomography - Part 1: Principles. Retrieved February 3, 2016, from http:// www.iso.org/iso/iso_catalogue/catalogue_tc/catalogue_detail.htm?csnumber=28106 ISO 15708-2 (2002). ISO 15708-2:2002 - Non-destructive testing - Radiation methods - Computed tomography - Part 2: Examination practices. (Retrieved February 3, 2016, from http://www.iso.org/iso/iso_catalogue/catalogue_tc/catalogue_detail.htm?csnumber= 29722). ISO 5725 (2012). Accuracy (trueness and precision) of measurement methods and results Part 1: General principles and definitions. Retrieved from http://www.iso.org/iso/iso_ catalogue/catalogue_tc/catalogue_detail.htm?csnumber=11833 Lambe, N. R., Wood, J. D., McLean, K. A., Walling, G. A., Whitney, H., Jagger, S., ... Bünger, L. (2013). Effects of low protein diets on pigs with a lean genotype 2. Compositional traits measured with computed tomography (CT). Meat Science, 95(1), 129–136 (http://doi.org/10.1016/j.meatsci.2013.04.038). Larsen, R. (2000). 3-D contextual Bayesian classifiers. IEEE Transactions on Image Processing: A Publication of the IEEE Signal Processing Society. 9(3), 518–524 (Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/18255423). Lyckegaard, A., Larsen, B., Christensen, L. B., Vester-Christensen, M., & Olsen, E. V. (2006). Contextual analysis of CT scanned pig carcasses. In ICoMST. Merkus, G. S. (1979). Development of uniform methods for pig carcass classification in the EC (Vol. 70). (Retrieved from http://bookshop.europa.eu/en/development-of-uniformmethods-for-pig-carcass-classification-in-the-ec-pbCBNA79070/downloads/CB-NA-79070-EN-C/CBNA79070ENC_001.pdf;pgid=Iq1Ekni0. 1lSR0OOK4MycO9B0000dM3f3igD;sid=S8v7fdOrJ1v7foUZWn9l2rGO6QiHLXFoN5s=? FileName=C). Nissen, P. M., Busk, H., Oksama, M., Seynaeve, M., Gispert, M., Walstra, P., ... Olsen, E. (2006). The estimated accuracy of the EU reference dissection method for pig carcass classification. Meat Science, 73(1), 22–28 (http://doi.org/10.1016/j.meatsci.2005.10.009). Olsen, E. V., Christensen, L. B., Hansen, M. F., Judas, M., & Höreth, R. (2006). Challenges developing an instrumental reference for classification of pigs. In ICoMST, 1–2. Petnehazy, O., Benczik, J., Takacs, I., Petrasi, Z., Süto, Z., Horn, P., & Repa, I. (2012). Computed tomographical (CT) anatomy of the thoracoabdominal cavity of the male turkey (Meleagris gallopavo). Journal of Veterinary Medicine Series C: Anatomia Histologia Embryologia, 41(1), 12–20 (http://doi.org/10.1111/j.1439-0264.2011.01099.x). Topigs Norsvin. (n.d.). Topigs Norsvin. (Retrieved from http://topigsnorsvin.com/) Vester-Christensen, M., Erbou, S. G. H., Hansen, M. F., Olsen, E. V., Christensen, L. B., Hviid, M., ... Larsen, R. (2009). Virtual dissection of pig carcasses. Meat Science, 81(4), 699–704 (http://doi.org/10.1016/j.meatsci.2008.11.015).