Prediction of Carcass Composition and Meat and Fat Quality Using Sensing Technologies: A Review

Computed tomography

X-ray–based technologies for assessing carcass composition are based on differences in signal attenuation when X-rays pass through different tissues. Basically, body components with strong absorbance, such as bones, appear bright and light, and elements that do not absorb the X-rays appear dark (Narsaiah et al., 2019). Traditional X-ray radiography is in 2 dimensions (2D), and 3D scanning was made possible through CT, wherein either samples rotate on a translation stage or the equipment rotates around the sample (spiral CT).

The application of CT for measuring lean, fat, and bone has been largely described and reviewed (Olsen et al., 2017; López-Campos et al., 2019). Recently, a number of studies have reported the application of CT in live animals for the estimation of body tissue composition. In this context, Pan et al. (2021) used automated CT segmentation in live pigs together with deep neural network statistical analysis and reported accurate predictions for body fat, lean, and bone weights (coefficients of determination: R² = 0.94, 0.86, and 0.91, respectively). In turn, Geraldo et al. (2021) found high predictability for body tissue composition of the Santa Inés breed of sheep (R² of prediction: R²_p = 0.90, 0.81, and 0.81 for fat, lean and bone weights, respectively), using CT scans together with the Cavalieri method (Gundersen et al., 1988) of multiple linear regression (MLR) analysis.

Apart from traditional studies on estimation of carcass tissue composition, Font-i-Furnols et al. (2021) used CT technology in both live dairy calves and carcasses to estimate carcass fat, moisture, and protein content. These authors found that prediction for carcass fat content was moderate when scanning live calves and high when scanning carcasses (R²_p = 0.69 and 0.88, respectively), whereas predictability for moisture (R²_p = 0.36 and 0.59) and protein content (R²_p = 0.36 and 0.26) was low in both scanning modes. Font-i-Furnols et al. (2019) used CT in live pigs to estimate the intramuscular fat (IMF) content in different pig muscles; however, the predictions were not accurate (R² of cross-validation: R²_cv = 0.07 to 0.42).

Dual-energy X-ray absorptiometry

DXA is based on differential absorption of X-rays of varying energy levels by different tissues. In comparison to CT, DXA has some advantages in that it is less expensive in terms of instrument and installation costs and requires minimal radiation exposure.

In research abattoirs, recent studies have reported potential for DXA to be able to predict not only whole carcass but also primal and retail cut composition, using conventional manual cut-outs as the reference method. Using linear regression analysis, Kipper et al. (2019) reported high DXA predictability in pork for total side weight (R² = 0.97 to 0.99); total bone (R² = 0.76 to 0.95), fat (R² = 0.96 to 0.99), and lean (R² = 0.84 to 0.98) content; and tissue composition of primal cuts (R² = 0.76 to 0.99), although accuracy varied depending on factors such as the region of interest selected, carcass weight, and/or the selected prediction equation. These authors suggested that the regions of interest should be chosen based on the item to be analyzed in order to decrease the error of analysis. In beef, López-Campos et al. (2018) scanned steer primals using DXA and reported a high accuracy for predicting total lean (R²_p = 0.99), fat (R²_p = 0.98), and bone (R²_p = 0.94) in whole carcasses using partial least squares regression (PLSR), as well as the tissue composition from 12 out of 15 retail cuts (R²_p = 0.81 to 0.95) (Figure 1). Similarly, high predictability by DXA in cull cows was obtained by Segura et al. (2021) for total lean, fat, and bone content in the whole carcass (R²_p = 0.99, 0.99, and 0.92, respectively) and tissue composition of most individual cuts (R²_p = 0.74 to 0.99). However, lower predictability was found for bone estimation in the flank retail cut (R²_p = 0.31), probably because of the high variability in both DXA estimations and weight measurements, as a consequence of the low amount of bone included in this primal. Additionally, R²_p values of 0.86 and 0.81 were described for retail cut and lean meat yield percentages, respectively.

Figure 1.

Beef chuck primal cut scanned with an iDXA unit and X-ray image generated (Lopez-Campos et al. 2018). Source: Agriculture and Agri-Food Canada–Lacombe Research and Development Centre.

Some of the latest developments in DXA have focused on its online application. In this context, Gardner et al. (2018) reported satisfactory prediction models for total fat, lean, and bone in lamb carcasses (R²_p = 0.89, 0.69, and 0.68, respectively), using a unit installed at a commercial abattoir at line speed. Likewise, using online DXA at commercial abattoirs, high predictability was reported in lamb by Connaughton et al. (2020b) for carcass fat, lean, and bone across varying processing factors (R²_p = 0.81 to 0.95) and by Connaughton et al. (2020a) for carcass fat and lean content considering a range of phenotypic and genotypic variables (R²_p = 0.91 and 0.74, respectively). Additionally, Gardner et al. (2021), also using an online DXA unit installed at a commercial abattoir, reported accurate predictions (R²_p = 0.63 to 0.95) for several retail cut weights in lambs. In beef, Calnan et al. (2021) have recently developed a prototype of a rapid DXA in a shipping container and reported high predictability for lean, fat, and bone of entire carcass sides (R²_p = 0.85, 0.94, and 0.82, respectively) and forequarters (R²_p =0.83, 0.93, and 0.82, respectively) and moderate accuracy for hindquarters (R²_p = 0.68, 0.80, and 0.73, respectively). It is important to note that, in all studies, CT was used as the standard method for calibration purposes; however, DXA technology is also becoming a reference technology for research studies on growth performance and carcass composition evaluation (Sousa dos Santos et al., 2021). Commercial application of DXA has now also been extended to use with robotic arms (Scott Technology Ltd., Dunedin, New Zealand) for further carcass processing at high line speeds.

Beyond use to estimate carcass composition, a number of authors have either suggested the application of or successfully implemented DXA to allow for precision nutrition/feeding of ruminants (González et al., 2018), as a methodology for estimating growth performance and body protein mass in growing-finishing pigs (Remus et al., 2020) and to predict aspects of lamb eating quality using bone DXA values (Anderson et al., 2021).

Computer vision systems

CVS—also known as machine vision systems, visual image systems, or just image systems—are based on sequential image acquisition, image processing (thresholding, binarization, etc.), image analysis (image segmentation, feature extraction, etc.), and data analysis (data normalization, model fitting, validation and tuning, prediction, etc.). The CVS have been applied in diverse areas to describe different characteristics through images by interpreting, reconstructing, and extracting properties (shapes, textures, densities, distances, etc.).

In livestock and meat science, CVS have mainly been applied to estimate carcass composition and quality for carcass classification and grading purposes. Lohumi et al. (2018) found predictions with moderate accuracy for lean meat yield (R²_p = 0.77) in commercial pork carcasses using images obtained from a fully automated VCS2000 camera (e+v^® Technology GmbH, Oranienburg, Germany) installed in a commercial slaughterhouse and MLR analyses. Additionally, in this study, the lean meat yield predictability was relatively high for main cuts such as ham, belly, and shoulder (R²_p = 0.80, 0.89, and 0.85, respectively) and moderate for loin and tenderloin (R²_p = 0.73 and 0.67, respectively). In beef, Segura et al. (2021) demonstrated the feasibility of using the variables from a combination of VBS2000 (whole-side carcass camera: e+v^® Technology GmbH, Oranienburg, Germany) and VBG2000 (rib-eye camera: e+v^® Technology GmbH, Oranienburg, Germany) cameras along with PLSR analysis to accurately predict carcass tissue composition (R²_p = 0.84 to 0.93) and tissue composition of most individual cuts (R²_p = 0.71 to 0.88), as well as lean meat and retail cut yield (R²_p = 0.90 and 0.86, respectively) (Figures 2 and 3). Regarding lamb, the ability of CVS to predict carcass commercial cut weight and yield of light lambs was evaluated by Batista et al. (2021). The results of this study confirmed previous reports on the ability of CVS technology to estimate the weights of different cuts (R²_cv = 0.96 to 0.99) and their lean meat yields (R²_cv = 0.96 to 0.99), whereas the accuracy was more limited for predicting cut weight (R²_cv < 0.43) and lean meat in percentage (R²_cv < 0.44).

Figure 2.

Carcass side image taken with a whole-side carcass camera (Segura et al., 2021). Source: Agriculture and Agri-Food Canada–Lacombe Research and Development Centre.

Figure 3.

Rib-eye image analysis performed with a rib-eye camera system at the 12^th-13^th rib (Segura et al., 2021). Source: Agriculture and Agri-Food Canada–Lacombe Research and Development.

Apart from carcass quality estimations, CVS have also been used to predict meat quality. In this context, Araújo et al. (2020) evaluated the relationships between carcass shape, carcass tissue characterization, and commercial cuts of hair sheep lambs to predict meat quality (R²_p = 0.82), suggesting the potential of the technology to establish categories for carcass classification in lambs. In fact, Stewart et al. (2021) observed accurate predictions for rib-eye area (R² = 0.83), Meat Standards Australia marbling (R² = 0.76), AUS-MEAT marbling (R² = 0.70), and chemical IMF (R² = 0.78) using a prototype vision system on a phenotypically diverse beef and lamb carcass population.

The industry’s need for automation and robotization, coupled with the concept of data objectivity without human intervention, have fostered interest in the development and use of 3D vision technologies (Vázquez-Arellano et al., 2016). In this context, Le Cozler et al. (2019b) demonstrated the potential of using laser cameras and image analysis to objectively predict body weight (R²_p = 0.93), total volume (R²_p = 0.99), and area (R²_p = 0.90) of dairy cows. In a second study, using the same laser cameras, Le Cozler et al. (2019a) also tested the feasibility of this technology for assessing animal morphology, namely heart girth, chest depth, and wither height as well as hip, backside, and ischial widths (R²_p = 0.38 to 0.79). A number of systems incorporating 3D CVS have been approved in Europe to estimate meat yield and grading in pork carcasses (Implementing Decisions of European Commission [2011] and European Commission [2019]). For instance, the Frontmatec company (Kolding, Denmark) has developed a 360° 3D scanning system—the Beef Classification Center BCC-3TM—which recreates a complete beef half-carcass. The system provides data to estimate conformation and fattening level according to EUROP standards and has estimated with relatively high accuracy the weight of beef commercial cuts in a commercial abattoir (R²_cv = 0.78, 0.85, and 0.81, for inside, knuckle, and rump, respectively; Esberg et al., 2019). Following a proof-of-concept using a 3D red, green, and blue camera, Alempijevic et al. (2021) reported an R²_cv value of 0.69 to predict lean meat yield in beef at 2 commercial abattoirs.

In cattle, Miller et al. (2019) predicted live weight, cold carcass weight, and saleable meat yield with moderate to high accuracy (R²_p = 0.70, 0.88, and 0.72, respectively), using a 3D time-of-flight camera and machine learning algorithms (artificial neural networks) to extract 60 predictor variables from images of live steers and heifers collected at both commercial and research farms. In addition, the models predicted EUROP fat and conformation grades with 54% and 55% accuracy (R²), respectively.

In pig carcasses, Masoumi et al. (2021) developed a method of digital imaging from full 3D models of half carcass sides and reported moderate predictability for the lean and fat content (R²_p = 0.77 and 0.73, respectively) (Figure 4). However, predictability varied for the weight of 12 different commercial cuts (R²_p = 0.41 to 0.80). In the case of lambs, the application of 3D-CVS to estimate live weight has also been described by Samperio et al. (2021), reporting weight estimates of the animal with an error of less than 6%.

Figure 4.

Assessment of pig half carcass composition using a novel 3D digital imaging approach (Masoumi et al., 2021). Source: Agriculture and Agri-Food Canada–Sherbrooke Research and Development Centre.

Recently, CVS has been applied to predict meat quality variables. For instance, Tomasević et al. (2019) reported that CVS colors were more accurate and precise in classifying bi-colored and non-uniformly colored beef and pork products (frequency of similarity ranged from 92.9% to 100%) and better mirrored their real color as perceived by panelists than a traditional colorimeter. In turn, Uttaro et al. (2021) developed an image analysis approach for identifying marbling on intact pork loin and reported a moderate estimation for IMF content, similar to both RAW (R² = 0.72) and JPEG (R² = 0.72) images. However, lower IMF predictability was reported at very high marbling levels.

Ultrasound

US is a nondestructive, cost-effective, easy-to-operate and reliable technology that provides in-depth measurements useful for evaluating carcass composition (Xiong et al., 2017). However, US measurements can be affected by different factors such as repeatability and reproducibility of the measures as consequences of operator performance (Vargas et al., 2021), position (transverse vs. longitudinal dorsal) (Cisneros et al., 1996), and pressure of the probe on live animals (Ripoll et al., 2009). To minimize these sources of error, many associations that assess the genetic potential of cattle in North America have certification policies through the Ultrasound Guidelines Council (Tait, 2016). There is extensive literature on the use of US in animal husbandry and evaluation of carcass composition in different species (Pathak et al., 2011; Xiong et al., 2017; López-Campos et al., 2019). Nevertheless, recent studies have reported limitations of current commercial US devices, and new applications of high-intensity US (HIUS) to livestock end-products have emerged.

In many cases, the US devices provide an image that must be processed to extract useful information such as marbling (Fukuda et al., 2017) or fat and muscle thickness (Theriault et al., 2009). Nevertheless, owing to software advances, current commercial US devices can directly provide those measurements. Lucas et al. (2017) found that fat and muscle thickness obtained with a commercial US Piglog 105 device (Carometec A/S, Herlev, Denmark) in live pigs were highly and moderately correlated with the reference method (correlation coefficients: r = 0.97 and 0.65, respectively). Additionally, these authors successfully predicted carcass lean yield (R²_p = 0.86). Using the same US equipment, Szyndler-Nędza et al. (2016) estimated with moderate to high accuracy pig carcass lean yield (R² = 0.51 to 0.84) and lean weight (R² = 0.40 to 0.74) in different breeds. When all breeds were combined in the analysis, lower US predictability was reported for carcass lean yield and weight (r = 0.63 and 0.58, respectively), although lean yield estimation improved (r = 0.71) when breed was considered in the model. Using other automatic US systems installed on the slaughter line, Font-i-Furnols and Gispert (2009) found moderate predictions for pig carcass lean yield using 42 and 2 variables from AutoFOM (Frontmatec, Kloding, Denmark) and Ultrafom (SFK Technology A/S, Herlev, Denmark) devices in equation models (R² = 0.78 and 0.64, respectively). Higher predictability (R² = 0.93, 0.82, 0.81) was observed by Janiszewski et al. (2019) for loin, ham, and belly lean yield, respectively, using 31 variables from AutoFOM, although lower US predictability was observed for other major cuts such as shoulder (R² = 0.68) and neck (R² = 0.42).

AutoFOM has also been used to estimate commercial cut weights. Choi et al. (2018) reported more accurate AutoFOM predictions for larger (R²_cv = 0.77 to 0.86) compared with smaller commercial cut weights (R²_cv = 0.34 to 0.57), which might be due to the lack of anatomical reference points and the lack of a strong correlation between the scanned area of the carcass and the smaller commercial cut weight. Nevertheless, the predictability obtained for heavier primal cuts shows US as an advantageous management tool enabling slaughterhouses to optimize product sorting prior to carcasses entering the cold room. Recently, Kress et al. (2020) estimated primal cut weights in different sex groups (gilts, boars, immunocastrates, and barrows) using an AutoFOM US device, and they concluded that the ability of the AutoFOM III estimation formulas to provide precise data for boars and immunocastrates is unclear. Particularly, boars might be misevaluated because of an increased proportion of lean meat and a different carcass composition due to sexual dimorphism. Hence, these authors suggested that the AutoFOM estimation formulas must be further examined in a detailed dissection trial and validation study in order to ensure fair pricing conditions for all sex groups and to avoid market distortions.

As biological tissue characteristics are sensitive to ultrasonic frequency (Park et al., 1994), HIUS has been applied recently to modify the physical and chemical properties of meat, in order to improve meat quality and safety. According to Jayasooriya et al. (2004), acoustic parameters such as frequency, intensity, duration of treatment, and temperature can influence the result from the ultrasonication. In this context, Diaz-Almanza et al. (2019) applied HIUS (37 kHz, 90 W cm⁻²) at different exposure times (0, 10, 20, and 40 min) in beef longissimus thoracis et lumborum (LTL) and observed the highest (P < 0.05) tenderness at 40 min; however, microbial reduction was higher at 10 min. Peña-González et al. (2017) observed that lipid oxidation increased (P < 0.0089), shear force decreased (P < 0.0001), and meat was perceived as more tender and juicy in HIUS-treated beef longissimus dorsi (40 kHz, 11 W cm⁻², 60 min) after 14 d of storage. Fallavena et al. (2020) applied HIUS at different temperatures (10°C to 35°C) and intensities (22 to 84 W cm⁻²) on beef biceps femoris and observed that high temperature and US intensity resulted in less tenderness whereas intermediate US intensity and low temperatures improved meat tenderness (P < 0.05). Overall, HIUS reduced water-holding capacity (WHC) and had a negative effect on lipid oxidation but did not influence pH or color.

Bioelectrical impedance analysis

The basic principle of BIA is to introduce an alternating current (1–500 kHz) through a body (fat and muscle tissues) using 2 transmission electrodes, while the electrical resistance produced is measured by 2 detector electrodes at the opposite side of the body (Swantek et al., 1992) (Figure 5). The BIA is a nondestructive technology used for characterizing different biological tissues that has decisive benefits compared with other characterization methods, because it is more efficient owing to the relatively reduced time, hardware, and software required (Kanoun, 2018). Although BIA technology was evaluated several years ago for estimating lamb carcass composition (Berg and Marchello, 1994), interest in using and exploring the potential of this technology is ongoing.

Figure 5.

Bioelectrical impedance analysis applied to a beef carcass. Source: Agriculture and Agri-Food Canada–Sherbrooke Research and Development Centre.

Using BIA measurements collected in vivo from lambs, Moro et al. (2019) predicted carcass weight and yield at slaughter with high (R²_cv = 0.96) and moderate (R²_cv = 0.57) accuracy, respectively. Moro et al. (2020) successfully predicted both lean (R²_cv = 0.98 and 0.99) and fat (kilograms) (R²_cv = 0.92 and 0.91) content in lamb carcasses using BIA measurements collected at 50 kHz on hot and cold carcasses, respectively. An et al. (2021) reported high (P < 0.01) correlations between fat percentage obtained with a BIA device on live pigs and lean meat percentage (r = −0.97) or backfat thickness (millimeters) (r = 0.93) measured with US equipment, indicating that BIA technology might be useful for predicting body compositions of live finishing pigs to facilitate swine feed management.

Beyond the application of BIA to evaluate carcass composition, BIA measurements have been recently used to predict meat quality. In this context, Huh et al. (2021) recently demonstrated that multifrequency impedance measurements were significantly correlated with moisture, fat (r = −0.83 and 0.86 at 128 kHz, respectively), and protein (r = −0.86 at 80 kHz) content of beef. However, crude ash content was not associated with impedance values (r < −0.43). Afonso et al. (2020) found that BIA predictability was relatively high for IMF content (R²_cv = 0.79) and low to moderate for collagen content and shear force value (R²_cv = 0.51 and 0.60, respectively) in beef LTL muscle. They also reported low predictions for cooking loss, pH, and color values (R²_cv ≤ 0.36).

BIA spectroscopy is increasingly used for real-time measurements in portable and embedded solutions owing to favorable developments in the field of microelectronics. Nevertheless, despite the potential of BIA technology to evaluate the quality of carcasses and meat, some studies have emphasized the need to refine the use of BIA for acceptance and transfer to industrial environments (slaughterhouses, food processing cutting rooms, etc.) (Zhao et al., 2017; Afonso et al., 2020).

Near infrared spectroscopy

In NIRS (12,500 to 4,000 cm⁻¹ or 800 to 2,500 nm), there is a change in dipole moment during vibrational transitions of a molecule, resulting in absorption of an infrared photon by the molecule. The absorbed photon has the same amount of energy as the energy difference between the 2 vibrational states of the molecule (Agarwal and Atalla, 2010). The vibrations involve C–H, O–H, and N–H chemical bonds (Prieto et al., 2017). There are many advantages of NIRS over conventional methods for meat quality analysis. NIRS is a sensitive, fast, nondestructive, cost-effective, and environmentally friendly technology (Dixit et al., 2017). Additionally, NIRS has a higher penetration depth than other spectroscopic techniques (Hassoun et al., 2020). One disadvantage of NIRS, however, is that meat is composed of approximately 75% water, which is strongly absorbed in the infrared region and can influence NIRS predictions (Andueza et al., 2019).

Recently, NIRS has been used to predict the content of chemical components in intact meat samples. Dixit et al. (2020b) successfully predicted IMF content of intact beef LTL at 24 h postmortem using offline and online a portable visible (VIS)-NIR spectrometer (350 to 2,500 nm) in a commercial pilot plant and PLSR (R²_p = 0.88 and 0.89, respectively). Patel et al. (2021) found moderate IMF (R²_p = 0.46, 0.66, and 0.62) and moisture (R²_p = 0.51, 0.63, and 0.70) content estimations for 7-d aged beef using a portable Vis-NIRS (350 to 1,830 nm), a portable NIRS (950 to 1,650 nm), and a handheld micro-NIRS (905 to 1,649 nm), respectively, in a research lab. Unexpectedly, the smallest instrument (Micro-NIRS) was the most precise for moisture and—together with portable NIRS—for IMF. According to Patel et al. (2021), this result could be due to data redundancy problems of the spectrometers with wider and more defined spectra. In pork, Wang et al. (2018a) and Wang et al. (2020b) found high correlations between NIR spectra and IMF, protein, and moisture content (r_p ≥ 0.86) using portable dual-band Vis-NIR spectrometers in research labs and PLSR. In contrast, several studies reported low or unreliable NIRS predictability for IMF content in lambs at 24 to 28 h postmortem (R²_cv = 0.52 to 0.55 or R²_p = 0.29 to 0.54; Knight et al., 2019; Fowler et al., 2021b; Hitchman et al., 2021a; Lambe et al., 2021) and beef longissimus thoracis aged for 2 and 7 d (R²_cv = 0.08 and 0.12, respectively; de Nadai Bonin et al., 2020), using portable Vis-NIR spectrometers (350 to 2,500 or 400 to 1,395 nm) and PLSR at commercial abattoirs. The lower NIRS predictability in these studies could be partly due to the narrower range of wavelengths and/or IMF content compared with those used in Dixit et al. (2020b). The latter could be due to differences among species, as lambs tend to have a narrower range of IMF than beef (Hoehne et al., 2012; Stewart et al., 2021; Hitchman et al., 2021b). Low NIRS predictability was also observed by Patel et al. (2021) for protein content in beef (coefficient of determination of calibration: R²_c = 0.61, 0.51, and 0.62; R²_p = 0.46, 0.23, and 0.43) using a portable VIS-NIR, a portable NIR, and a handheld micro-NIR spectrometer, respectively. These authors attributed the low R²_c to the low variability of this chemical component in meat samples, whereas the low R²_p were due to a greater proportion of farm/date variance, which was not reflected in the meat spectra.

Regarding NIRS ability to predict fatty acid (FA) composition, Prieto et al. (2018a) successfully estimated the content of FA groups and iodine value (R²_p = 0.81 to 0.94), although lower predictability was reported for individual FA with low variability such as palmitic, stearic, and oleic FA (R²_p = 0.60 to 0.70) in pork subcutaneous fat, using a portable NIRS device (350 to 2,500 nm) in a research abattoir and PLSR (Figure 6). Piao et al. (2018) found moderate NIRS predictability for monounsaturated FA (MUFA), oleic, and saturated FA (SFA) in beef intermuscular fat (R²_p = 0.69, 0.64, and 0.67, respectively), using a 700 to 1,050 nm portable NIR spectrometer at meat markets and PLSR. When spectra were collected on intact lamb meat samples, accurate NIRS predictions were found for SFA and polyunsaturated FA (R²_cv = 0.84 and 0.94, respectively), and moderately accurate predictions were found for MUFA content (R²_cv = 0.57), using a 350 to 2,500 nm portable spectrometer in a research abattoir and PLSR (Juárez et al., 2017). However, low NIRS predictability for individual and groups of FA (R²_cv = 0.11 to 0.56) were found in intact lamb meat samples by Fiorentini et al. (2017), using a 350 to 2,500 nm portable instrument and PLSR. Apart from low variability, another limiting factor for FA prediction by NIRS is the similar absorption patterns of different FA due to similarities in chemical structure and functional groups (Prieto et al., 2017; Tao and Ngadi, 2018). This condition could generate complex relationships between the spectra and the response variables that are not capable of being predicted under a PLSR model. In fact, Barragán-Hernández et al. (2020) considerably improved NIRS predictability of individual and groups of FA in ground beef samples when support vector machine-learning (SVM) regression was applied (R²_p = 0.74 to 0.99) compared with PLSR (R²_p = 0.06 to 0.58). Therefore, regression based on SVM could be a novel alternative that manages to identify several vectors in hyperspace capable of constructing a generalizable prediction model for estimating FA profiles in beef with higher accuracy than conventional methods.

Figure 6.

Collection of near-infrared spectra from the inner layer of pig carcass subcutaneous fat using a portable NIRS device (Prieto et al., 2018a). Source: Agriculture and Agri-Food Canada–Lacombe Research and Development Centre.

Portable and handheld NIR spectrometers have also been used recently to predict pH of intact meat samples. Dixit et al. (2020b) and Wang et al. (2018a) reported high NIRS predictability and correlation for pH in beef at a commercial pilot plant (R²_p ≥ 0.84) and pork at a research laboratory (r_p ≥ 0.81), respectively. However, several studies reported low predictive ability (R²_cv ≤ 0.29 or R²_p ≤ 0.30) in beef, pork, and lamb meat at research laboratories or commercial abattoirs (Fiorentini et al., 2017; Andersen et al., 2018; Patel et al., 2021; Savoia et al., 2021); this is likely due to narrower ranges in pH values compared with the 2 above-mentioned studies.

Furthermore, portable and handheld spectrometers have been used to predict color values (L*, a*, and b*) of intact meat samples at laboratories and research or commercial abattoirs. Savoia et al. (2021) and Juárez et al. (2017) reported high L* predictability in beef (R²_p ≥ 0.80) and lamb (R²_cv = 0.93), respectively, whereas Wang et al. (2018a) reported high correlations between NIR spectra and L* (r_p = 0.89 to 0.90), a* (r_p = 0.88 to 0.94), and b* (r_p = 0.90 to 0.95) in pork. In contrast, other studies reported low to moderate NIRS predictability for beef L* (R²_cv = 0.33 to 0.49 or R²_p = 0.42 to 0.52), a* (R²_cv = 0.07 to 0.32 or R²_p = 0.52 to 0.71), and b* (R²_cv = 0.19 to 0.39 or R²_p = 0.35 to 0.63; Sahar et al., 2019; Patel et al., 2021; Savoia et al., 2021) as well as for pork L* (R²_cv = 0.63), a* (R²_cv = 0.72), and b* color values (R²_cv = 0.65; Furtado et al., 2019). These differences among studies for NIRS predictability of color values could be partly due to instrument variations.

In the last few years, WHC of intact meat has also been estimated using portable and handheld NIR spectrometers. NIRS poorly predicted purge loss (R²_p ≤ 0.31; Savoia et al., 2021) and cook loss in beef (R²_cv ≤ 0.45 or R²_p ≤ 0.25; Sahar et al., 2019; Patel et al., 2021; Savoia et al., 2021) and drip loss in pork (R²_cv ≤ 0.12; Andersen et al., 2018; Savoia et al., 2021). Nevertheless, a relatively high correlation between NIR spectra and cook loss in pork (r_p = 0.77) was found by Wang et al. (2020b), probably due to a wider variability (higher standard deviation) in cook loss from this study compared with that from the above-mentioned studies.

Recently, many studies have tested the prediction performance of portable and handheld Vis-NIR spectrometers at research or commercial abattoirs to predict Warner-Bratzler shear force (WBSF) values in meat from different species, as tenderness is a driver of consumer satisfaction. However, low NIRS predictability for WBSF values has been reported for both beef and lamb intact meat samples (R²_cv < 0.41 or R²_p < 0.20; Juárez et al., 2017; Knight et al., 2019; Cafferky et al., 2020; Savoia et al., 2021). Apart from the moderate to low repeatability and reproducibility reported for meat quality physical attributes such as shear force (Patel et al., 2021), the poor NIRS predictability for WBSF might be just because tenderness of cooked meat is not explained by compositional differences or vibrational properties of fresh meat provided by the NIR spectra. Similarly, Cafferky et al. (2020) reported unreliable predictions for taste panel tenderness scores in beef at 1 and 2 d postmortem using a portable NIR spectrometer at a commercial chilling room (R²_cv = 0.06) and a laboratory (R²_cv = 0.13), respectively, and for other sensory attributes such as juiciness, crumbliness, and beef flavor (R²_cv = 0.04 to 0.41).

NIRS has also been used recently to estimate pork belly firmness. Lam et al. (2020) and Soladoye et al. (2018) successfully predicted belly firmness (subjective scoring: R²_p = 0.66 to 0.84; belly flop angle: R²_p = 0.71 to 0.91) when subcutaneous fat at the shoulder and lean and subcutaneous fat belly layers were scanned, respectively, using portable VIS-NIR spectrometers (350 to 2,500 nm) in cold rooms of a research abattoir and PLSR.

In recent years, NIRS technology along with machine learning techniques have become popular tools for authenticating meat products based on species, geographic origins, animal diet, and freshness in research laboratories or abattoirs. Yang et al. (2018) classified beef, pork, and mutton samples with 100% accuracy using SVM. Weng et al. (2020) used a portable NIR spectrometer and deep convolutional neural network (DCNN) to successfully identify beef adulterated with pork (R²_p = 0.94). Revilla et al. (2020) correctly classified 84% of dry-cured beef samples of a Protected Geographic Indication quality label and 100% of non–Protected Geographic Indication, using a benchtop NIR spectrometer and artificial neural network. Barragán et al. (2021) correctly classified 75% to 100% of subcutaneous fat and intact meat samples from cattle fed barley or corn using a portable Vis-NIRS instrument and applying linear SVM. These authors confirmed the successful application of the SVM technique in studies with relatively small data sets while allowing an external validation. Finally, Moon et al. (2020) classified “fresh,” “likely spoiled,” and “spoiled” sirloin beef samples with 86% to 96% accuracy, using a portable Vis-NIR spectrometer and a convolutional neural network–based machine learning algorithm.

Hyperspectral imaging

By integrating spectroscopy and imaging techniques, imaging spectrometry, also known as HSI, is used to acquire 20 or more equally distributed contiguous spectral bands with ranges throughout both the visible and near infrared regions, resulting in a highly informative characterization of reflectance and emittance spectrum (Goetz, 2009). This advantage has led to the application of HSI to identify components and characterize their spatial distribution in a product, substance, tissue, or environment. The use of HSI for predicting chemical composition, technological attributes, and adulteration in meat has been previously reviewed (Xiong et al., 2014; Siche et al., 2016; Prieto et al., 2018b). Since the latter reviews, multiple studies using HSI to predict these attributes have emerged in beef, lamb, and pork.

Recent studies successfully predicted pH values in red meat using various statistical approaches. Using moveable benchtop HSI scanners, Craigie et al. (2017) and Dixit et al. (2020a) reported a relatively high HSI predictability for pH in lamb (R²_p = 0.71) and in a combination of beef, lamb, and venison (R²_p = 0.75–0.86), respectively, using PLSR. Nevertheless, Dixit et al. (2020a) and Yao et al. (2019) found higher predictability for pH using DCNN in beef/lamb/venison (R²_p = 0.89) and support vector regression (SVR) in pork using a portable scanner (R²_p = 0.93), respectively. The higher performance observed with SVR compared with PLSR models agrees with Thissen et al. (2004), who concluded that, for spectral applications, SVR is more robust and able to account for nonlinear effects from temperature changes. In addition, Craigie et al. (2017) included outliers in the model in an attempt to replicate industry application results, which could result in the lower performance compared with the other studies. These authors suggested continued calibration and validation to improve the robustness of the model. When comparing HSI line (550 to 1,700 nm) and snapshot (670 to 950 nm) scan systems using a robotic platform inline in a meat processing pilot plant to predict muscle pH in beef, the line scanner showed superior performance (R²_p = 0.89) to the HSI snapshot scanner (R²_p = 0.77; Dixit et al., 2020b). Although HSI snapshot scanners exhibit faster image acquisition, HSI line scanning systems are able to provide better sample representation, explaining this difference in performance.

Similar to pH studies discussed earlier, DCNN models performed slightly better (R²_p = 0.89) than PLSR (R²_p = 0.84 to 0.88) for IMF proportion prediction, when performing a combined analysis of beef, lamb, and venison using a benchtop HSI scanner with a moving stage (Dixit et al., 2020a). Likewise, HSI line scanning (R²_p = 0.90) outperformed snapshot scanning systems (R²_p = 0.72) when predicting beef IMF content inline in a meat processing pilot plant (Dixit et al., 2020b). In pork, several studies revealed high HSI prediction performance for IMF content in loin (R²_p = 0.86 to 0.96), using benchtop and pushbroom lab HSI systems and several prediction models (step-wise regression, PLSR, SVM, and back-propagation neural networks) (Huang et al., 2014; Ma et al., 2018; Kucha et al., 2021a). Aheto et al. (2020) used a pushbroom HSI system in the laboratory to estimate IMF content in pork bellies and reported relatively high predictability when a median spectral features model (R²_p = 0.81) and a mean image feature model (R²_p = 0.73) were used. When using HSI to predict IMF proportion in longissimus lumborum from lambs, moderate accuracy (R²_p = 0.67; Craigie et al., 2017) and high accuracy (R²_p = 0.92) were observed using PLSR and progressive PLSR (which cumulatively included data chronologically and recalculated the model) models, respectively. The latter used a line scanning system with a moving stage in a lab setting and suggested that, for inline HSI systems, updating the models over time (particularly where lighting and environmental conditions cannot be carefully controlled) might be a useful strategy to improve future predictions. Velásquez et al. (2017) reported pushbroom lab HSI as a successful technology to classify beef based on the degree of marbling using the decision tree method (99.92%).

The HSI has also been applied to predict FA concentration. Craigie et al. (2017) predicted groups (SFA, MUFA and polyunsaturated FA) and individual FA content with moderate (R²_p = 0.53 to 0.70) and low accuracy (R²_p = 0.00 to 0.48), respectively, in lamb meat using PLSR. The low HSI performance was likely due to the inclusion of outliers in an attempt to replicate an industry application scenario. In contrast, Wang et al. (2020a) found high HSI predictability using a lab benchtop moving stage HSI system for palmitic (R²_p = 0.91) and oleic FA content (R²_p = 0.88) in lamb meat, when HSI spectra were combined with texture data (using gray-level co-occurrence matrix to extract textural features) and variable combination population analysis–iteratively retaining informative variables was applied to an SVM regression (for palmitic acid prediction) and a PLSR model (for oleic acid prediction). For specific FA in pork, using full wavelength spectral data and mean spectral features, a recent study showed prediction performance for myristic, palmitic, palmitoleic, stearic, oleic, and linoleic FA ranging from R²_p 0.69 to R² 0.84 (Kucha et al., 2021b).

In addition to fat variables, other meat chemical components have been estimated using HSI. Water content was successfully predicted using a lab-based benchtop HSI and MLR in lambs (R²_p = 0.92; Pu et al., 2014). Although still relatively high, lower HSI predictability was reported for protein content in lambs using MLR or PLSR (R²_p = 0.80 to 0.85; Pu et al., 2014) and in pork using back-propagation neural networks (R²_cv = 0.78 to 0.83; Ma et al., 2019). The overall lower performance for predicting protein content could be due to the low range/variability of this component (Prieto et al., 2018b).

Regarding color in red meat, high predictability with benchtop lab HSI systems was reported for L*, a*, and b* values in combined beef, lamb, and pork samples using MLR (R²_p = 0.97, 0.84, and 0.82, respectively; Kamruzzaman et al., 2016b) and in beef using PLSR (R² = 0.98, 0.92, and 0.95, respectively; Yu et al., 2020). Metmyoglobin content is known to influence meat color stability, and many studies recently found high HSI predictability for different metmyoglobin content variables in lambs using benchtop lab HSI systems and different prediction approaches including PLSR and least-squares support vector machines (LSSVM) (R²_p = 0.85 and 0.91, respectively; Cheng et al., 2020), competitive adaptive reweighted sampling–LSSVM (R²_p = 0.81 to 0.91; Yu et al., 2020), NIR-HSI data combined with generalized 2D correlation spectroscopy method (R²_p = 0.85; Cheng et al., 2021), and competitive adaptive reweighted sampling–PLSR (R²_p = 0.77; Yuan et al., 2020). Few studies have recently applied HSI to predict WHC in meat. Kamruzzaman et al. (2016a) predicted with high accuracy drip loss in beef, lamb, and pork together using a benchtop HSI in a lab environment and PLSR (R²_p = 0.92) and SVM (R²_p = 0.94).

Regarding the prediction of instrumental tenderness (WBSF) using a benchtop HSI system in a lab setting, Balage et al. (2018) reported low predictability in beef using a specific region of interest (R²_p = 0.06 to 0.27) and PLSR. These authors concluded that HSI applied to a specific region of the sample does not accurately represent the tenderness of the whole sample. Nubiato et al. (2018) used a lab-based benchtop HSI to classify beef samples based on WBSF values and reported that 89.9% and 84.8% of samples were correctly classified using partial (928 to 1,413 nm) and full (928 to 2,524 nm) wavelength, respectively. When HSI systems were recently used inline in a meat processing pilot plant, Dixit et al. (2020b) reported a poor prediction performance using an HSI line (R²_p = 0.17 to 0.36) and a snapshot scanner (R²_p = 0.13 to 0.18) for WBSF value in beef aged for different aging periods. These authors suggested that the low performance could either be due to a lack of direct correlation between WBSF and spectral features or be due to the fact that all the muscles were not in rigor at the scanning time. Considering that the spectroscopic measurements were taken at 24 h postmortem and WBSF measurements were performed in aged meat, this might have limited the prediction performance of WBSF.

The detection of meat adulteration (i.e., with spoiled meat or other animal meats) has been successful using HSI. Recent studies have shown high HSI prediction performance of detecting proportions of adulterated meat (by mixing spoiled meat) in beef (R²_p = 0.94; Zhao et al., 2019) and mutton (R²_p = 0.93; Zhu et al., 2021) using benchtop systems in a lab setting and SVM. Minced lamb meat and beef adulteration with duck meat have been accurately predicted using a lab HSI system with moving stage and PLSR and principal component regression modeling methods (R²_p > 0.95; Jiang et al., 2019; Zheng et al., 2019). Rady and Adedeji (2020) used a pushbroom HSI system in a laboratory and PLSR model to detect adulteration in minced meat for beef and pork, reporting a classification rate of 75%–100% and 100% for pure and adulterated samples, respectively.

Raman spectroscopy

Raman spectroscopy is a type of vibrational spectroscopy that uses light spectra between 200 and 1,800 nm (50,000 to 5,556 cm⁻¹) and is based on light scattering. When photons collide with molecules, they sometimes exchange energy, which is referred to as Raman scattering (Smith and Dent, 2005). Biomolecules such as amino acids, collagen, elastin, carotenoids, FA, and cholesterols contribute to Raman scattering in meat and fish, and Raman spectroscopy is used to obtain information regarding the concentration, structure, and interactions among these biomolecules (Damez and Clerjon, 2008; Yang and Ying, 2011; Czamara et al., 2015). There are several advantages of Raman spectroscopy over traditional techniques that have been used to assess meat quality traits. Raman spectroscopy is a nondestructive, low-cost, and highly sensitive technology that requires a small sample and short analysis time. Additionally, there is no solvent use or toxic waste involved, and the spectra of water do not interfere with those of other meat components (Motoyama, 2017; Pallone et al., 2018; Chen et al., 2020). However, one disadvantage of Raman spectroscopy is fluorescent background interference, especially in lean meats that contain fluorescing compounds, but algorithms have been recently developed to subtract fluorescent background (Wang et al., 2018b).

Raman spectroscopy has been widely used to predict chemical composition of meat (Yang and Ying, 2011; Troy et al., 2016). When scanning intact meat using lab benchtop instruments, Andersen et al. (2018) and Cama-Moncunill et al. (2020) found moderate predictability for IMF content from 4- to 5-d aged pork (R²_cv = 0.73) and 2-d aged beef LTL (R²_cv = 0.64), respectively, using PLSR. The latter indicated that contribution from non–readily identified bands probably not related to the variation of IMF might explain the modest spectral variance explained by the model. To date, there are not many studies applying handheld Raman spectrometers to predict chemical composition of intact meat. In this context, low Raman predictability was reported by Fowler et al. (2015a) for IMF and FA group content in lamb meat at a research laboratory (R²_cv = 0.01 to 0.21). These authors hypothesized that the process of separating protein and lipid spectra might have reduced the ability to predict major FA composition, because removing all spectra that contained mixed lipid and protein signals might also have removed information on the phospholipids bound in membranes.

Furthermore, portable and handheld Raman spectrometers have been used in research laboratories to predict pH and color values of meat. Moderate predictions for ultimate pH were reported for intact pork (R²_cv = 0.52 to 0.72; Andersen et al., 2018, 2021) and lamb meat (R²_cv = 0.59; Fowler et al., 2015b), but predictions were lower for beef (R² = 0.42; Fowler et al., 2018). Fowler et al. (2018) indicated that a high loin purge (7%, probably due to a rapid pH decline) could have caused less Raman information present in spectra, resulting in poorer predictions of pH in meat. Additionally, Yang et al. (2020) reported high Raman predictability for pH and L* at 0-d aged beef steaks (R²_cv = 0.99 and 0.78, respectively), but these values declined steadily over a 21-d aging period, reaching a minimum R²_cv of 0.04 for both attributes. Attempts to predict a* and b* were unsuccessful. Similarly, poor predictions were reported for L* in lamb at 1 d and 5 d postmortem (R²_cv = 0.32 and 0.22, respectively; Fowler et al., 2015b).

Raman spectroscopy has also been applied to estimate WHC. When scanning intact meat in a laboratory, moderate Raman predictability was found by Andersen et al. (2021) for drip loss (R²_cv = 0.56 to 0.75). However, low predictability was reported for drip loss in pork (R²_cv = 0.49; Andersen et al., 2018) and drip loss and cook loss in beef (R²_cv = 0.59 and 0.49, respectively; Cama-Moncunill et al., 2020), using lab benchtop instruments. Cama-Moncunill et al. (2020) indicated the contribution from non–readily identified bands in the predictions that might not be related to the variation of physicochemical characteristics of the samples and the experimental design (drip and cook loss were measured on 2-d aged fresh and 14-d aged thawed muscle, respectively, whereas Raman spectra were collected on 2-d aged thawed muscle) as factors that could have jeopardized the accuracy of the predictions. Fowler et al. (2018) predicted with low accuracy purge loss in beef (R² = 0.46) using a handheld Raman spectrometer at the laboratory.

Discrepancies among recent studies have been found predicting instrumental tenderness by Raman spectroscopy in intact meat from different species. Using a lab benchtop spectrometer to predict WBSF values, Cama-Moncunill et al. (2020) found low accuracy in beef (R²_cv = 0.36). These authors suggested that the low predictability could be a result of including signals unrelated to physical/chemical changes in their model because some bands in their plots were uninterpretable and only small spectral changes were observed between tender and tough meat. Chen et al. (2020) found high predictive ability for shear force in beef, as measured with a Meullenet–Owens razor shear probe, using a portable Raman spectrometer in a laboratory (R²_p = 0.81). When handheld Raman spectrometers were used at the laboratory, Fowler et al. (2018) and Fowler et al. (2014) found unreliable Raman predictability for WBSF in beef (R² = 0.11) and lamb (R²_cv = 0.06), respectively. Fowler et al. (2014) attributed the low performance to Raman measurements being taken on the thinnest side of the longissimus lumborum samples because, owing to the close proximity of 3 muscle surfaces, diffuse scattering might have increased, thereby preventing signal discrimination in deeper areas of the muscle. Additionally, these authors hypothesized that increasing the integration time from 3 s or increasing the total accumulation by including repetitions would improve the accuracy of prediction by reducing the signal to noise ratio. When Raman spectroscopy was used for classification purposes, Santos et al. (2018) correctly classified >67% and >85% of pork samples based on instrumental tenderness at 1 d postmortem in a commercial abattoir and at 15 d postmortem in a laboratory, respectively, using a portable Raman spectrometer and SVM.

When tenderness was evaluated by untrained consumers, high accuracy predictions were reported using a handheld Raman spectrometer in intact lamb meat (R² = 0.99; Fowler et al., 2021a). Additionally, Santos et al. (2018) correctly classified pork samples based on tenderness assessed by trained panelists at 1 d postmortem in a commercial abattoir (>69%) and at 15 d postmortem in a laboratory (>93%), using a portable Raman spectrometer and SVM. Overall, these results indicate that Raman spectroscopy might be a better predictor of sensory tenderness than shear force. This is likely due to sensory tenderness being associated with more than one biochemical characteristic, whereas the measurement of shear force does not take into account the contribution of water and fat content to the sensory perception of juiciness and the impact this has on the perception of tenderness (Perry et al., 2001).

Recent cases of meat adulteration, such as the report of horse DNA in 33% of frozen beef burgers (Laurence, 2013), have increased the use of Raman spectroscopy to discriminate meat based on species, feeding regimes, or cuts. Robert et al. (2021) discriminated among beef, venison, and lamb with over 80% accuracy using partial least square discriminant analysis (PLS-DA) and a benchtop spectrometer. Additionally, Logan et al. (2021) correctly classify beef fed long-term grain, short-term grain, grass, and a supplemented grass diet (96%, 85%, 83%, and 83%, respectively), using a handheld Raman spectrometer directly on the carcass subcutaneous fat and PLS-DA. Martin-Gomez et al. (2021) used a portable Raman spectrometer in a laboratory and a K-nearest neighbours algorithm to correctly classify 83.3% of dry-cured hams based on pork diet and 77.8% to 100% of dry-cured hams based on breed. Furthermore, Ostovar Pour et al. (2020) distinguished beef cuts (rump, Scotch fillet, round, chuck, tenderloin, and T-bone) with 84.4% accuracy, using a portable Raman spectrometer in a laboratory and principal component discriminant function analysis.