NIR hyperspectral imaging for fat quantification in minced pork - ATB

NIR hyperspectral imaging for fat quantification in minced porkDouglas F. Barbin a ; Gamal Elmasry a ; Da-Wen Sun* a and Paul Allen ba Food Refrigeration and Computerised Food Technology (FRCFT), School of BiosystemsEngineering, University College Dublin, National University of Ireland, Agriculture & FoodScience Centre, Belfield, Dublin 4, Ireland.bAshtown Food Research Centre, Teagasc, Dublin 15, Ireland.*Corresponding author: dawen.sun@ucd.ieAbstractMinced meats are offered to the consumers with different levels of fat. In addition, they arethe main ingredients in a large assortment of processed products like hamburgers, pattiesand sausages. There is a need for a fast determination technique of different chemicalcomposition in minced meat to ensure that the right amount is presented. In this study, apushbroom hyperspectral imaging system in the reflectance mode was used as a fast andnon-destructive determination method of fat content in minced pork. Near-infrared (NIR)hyperspectral images (900-1700 nm) were acquired and the mean spectra from the mincedsamples were extracted by automatic segmentation of the region of interest (ROI). Fatcontent was determined by nuclear magnetic resonance (NMR) and related to the spectralinformation by partial least-squares regression (PLSR) models. The coefficient ofdetermination obtained by cross-validated PLSR models indicated that the NIR spectralrange has good ability to predict the amount of fat (R 2= 0.95) in pork. Feature-relatedwavelengths were selected to predict the fat amount using reduced spectral data. Theregression model obtained from selected wavelengths was applied back to the spectralimage to display the amount of fat in each pixel of the image. The proposed approachpermits the quantification and visualisation of the spatial distribution of fat within the sample.This technique represents a potential tool for high-speed assessment of fat content in meatproducts.Keywords: Hyperspectral imaging, pork, partial least squares regression, image processing.1. IntroductionConsidering the prominence of meat constitution regarding economical or health relatedaspects, there is a need to analyze the chemical composition of minced meat to ensure thatthe consumers and processors get the right quality from their suppliers. Analytical methodscurrently used suffer from some are sometimes healthy and environmentally harmful(Prevolnik et al., 2011). Other major drawbacks of such methods are the errors introducedby sampling procedures and lengthy preparation (Togersen et al., 2003). There is a need fora fast determination technique of fat content in minced meat to ensure that the right amountis presented.

Several investigations have demonstrated the ability of spectroscopy to predict the chemicalcomposition of organic and biological materials such as food products (Brondum et al.,2000a, b; Barlocco et al., 2006). However, spectroscopic systems usually have limitedspatial field of view, thus it can be easily affected by the selection of the region-of-interest(ROI) to be analyzed. By combining the chemical selectivity of spectroscopy with thepossibility of image visualisation, hyperspectral imaging allows to display hidden informationin the image that can quantitatively describe the properties of the tested samples, enabling amore complete description of component concentration in heterogeneous samples. In thisstudy, a pushbroom hyperspectral imaging system in the reflectance mode was used as afast and non-destructive method for determination of fat content in minced pork.2. Material and MethodsFresh pork samples (n = 120) from four different muscles including longissimus dorsi (LD),semimembranosus (SM), semitendinosus (ST) and biceps femoris (BF) were selected froman industrial supplier (Rosderra Irish Meats Group, Roscrea, Co. Tipperary, Ireland). Allvisible subcutaneous fat were trimmed and samples were minced using a food processor (R-201E Ultra, Robot-Coupe, France) until a homogeneous paste was obtained. Fat content ofpork samples was analysed using the Smart Trac (CEM Corporation, Matthews, NorthCarolina, USA) (AOAC Official Method 2008.06; Leefler et al., 2008). The samples weretransferred to a metallic can and imaged in a pushbroom NIR hyperspectral system.Spectral information was extracted from the hyperspectral images of the minced porksamples. Partial least square regression (PLSR) was applied to predict the amount of fatusing the NIR spectral information as predictors and the fat content measured by analyticalmethod as the response variable.The whole data set (120 samples) was divided into two groups, one for building thecalibration model consisting of 80 samples (training set), and another group used forvalidation consisting of 40 samples (testing set). The PLSR models were built with thetraining set under full cross validation by using leave-one-out cross-validation (LOOCV)method. The predictive ability of the regression model was evaluated by calculating thecoefficient of determination in calibration (R 2 C), standard error in calibration (SEC),coefficient of determination in cross-validation (R 2 CV) and standard error estimated by crossvalidation(SECV). The best model selected should have high coefficient of determination(R 2 C and R 2 CV) and low standard errors (SEC and SECV), in addition to smallest differencebetween SEC and SECV (ElMasry et al., 2011). Finally, the predictability of the establishedPLSR models were rather tested in the independent testing data set composed of theremaining 40 samples.

The weighted regression coefficients yielded from the best PLSR models under full crossvalidationcan be used successfully for the appropriate selection of the feature-relatedwavelengths (Garrido Fernich, 1995). A new PLSR was carried out using only the featurerelatedwavelengths as predictor variables. The regression coefficients resulting from thePLSR models with selected wavelengths were applied in a pixel-wise manner to obtain theconcentration maps that display the distribution of fat within the sample.3. Results and DiscussionThe overall measured fat content varied from 0.30% to 8.96%, with an average value of2.43%. Fat composition was basically related to marbling since the external subcutaneousfat layer of the LD was trimmed out before analysis.Figure 1. Average spectra for different pork muscles (ST: semitendinosus, LD: longissimusdorsi, SM: semimembranosus, BF: biceps femoris).The main NIR spectral patterns of the pork samples originated from different muscles areshown in Figure 1. Spectral profiles of the four examined muscles exhibited similar patternwith differences in the magnitude of reflectance. Variations observed in spectral reflectanceamong pork muscles could be related to differences in the sample properties. Among themost prominent features that influenced the near-infrared spectra in the pork samples is theC-H stretching overtone associated to fat (Brondum et al., 2000a, b; Barlocco et al., 2006).The developed PLSR models for fat composition of minced pork samples under crossvalidation had a reasonable accuracy when applied to an independent test set, withcoefficients of prediction (R 2 P) of 0.95 using the full spectra (Table 1).

Table 1. Calibration statistics for predicting chemical composition with spectra extracted fromintact (I) and minced (M) pork using the whole spectra and selected wavelengths from thePLSR models.Model bands LV R 2 C R 2 CV R 2 P SEC SECV SEP RER RPDFull spectra 237 8 0.96 0.95 0.95 0.30 0.37 0.37 23.4 4.2Selected bands 9 6 0.95 0.94 0.93 0.34 0.39 0.42 22.2 4.0The weighted regression coefficients resulting from the best PLSR were considered as anindication of the possible feature-related wavelengths that could explain most of thevariation. By means of this approach, nine wavelengths were identified (927, 937, 990, 1047,1134, 1211, 1275, 1382, 1645 nm). New PLSR were then developed using these particularwavelengths. The performance of the new PLSR model is shown also in Table 1. Fat contentin the minced pork could be predicted accurately using the selected wavelengths with acoefficient of determination for (R 2 P) of 0.93 with small difference between SEC and SECV.(a)(b)Figure 2. Predicted versus measured values of tested pork samples using PLSR model with:(a) full spectra, and (b) selected wavelengths.Figure 2 shows the efficiency of PLSR model for predicting the fat content of pork samplesusing both full spectral range (Figure 2a) and selected wavelengths (Figure 2b) for anindependent set of samples (40 samples). As illustrated in Figure 2b, the PLSR regressionmodel with few selected wavelengths has a good performance similar to the predictionmodel with full spectra.The results obtained for the PLSR models were all based on using a single spectrum fromthe selected ROI from each image. The main advantage of the hyperspectral image is that itcontains abundant spectral information in every pixel which could be as a prediction data setby itself. Figure 3 shows the results of applying the PLSR model in a pixel-wise manner to

the minced pork samples. Variations in concentration of fat were assigned with a linearcolour scale. The negative value in the colour scale was used to make the background morecontrasted to the sample. Although it was impossible to identify the fat composition ofminced pork by the naked eye, the spatial variation of these components within the samplecould be visualized by the NIR hyperspectral imaging system.Figure 3. Concentration maps for minced pork samples with predicted composition (%): (a)pseudo-colour image composed by three selected wavelengths (1081nm, 1275nm,1329nm), (b) concentration map for fat.The predicted composition represents an average of thousands of pixels within the ROI,compared to the measurements in few particular points of the sample in traditional methods.By including all the pixels within the selected ROI, the presented approach has theadvantage to display more detailed and accurate information. As shown in the concentrationmaps, values for fat content may vary within the same sample. Some fat filaments can beseen as a cluster of highly predicted values of fat. Consequently, small differences betweenthe values from predicted and measured components reported could be influenced by thedifferent location of samples used in the chemical analysis. The proposed approach permitsthe quantification and visualisation of the spatial distribution of fat within the sample.4. ConclusionConcentration maps obtained by this method justified to a great extent the effectiveness ofthis technique to predict and visualize fat content in minced pork. By combining both spatialand spectral features in one single system, NIR hyperspectral imaging represents a potentialtool for high-speed assessment of fat content in meat products.Acknowledgements

The authors gratefully acknowledge the financial support from the Food InstitutionalResearch Measure (FIRM) strategic research initiative administered by the Irish Departmentof Agriculture, Fisheries and Food.ReferencesAOAC Official Method 2008.06. Moisture and Fat in Meats by Microwave and NuclearMagnetic Resonance Analysis. June 23, 2008.Barlocco, N.; Vadell, A.; Ballesteros, F.; Galietta, G.; & Cozzolino, D. (2006). Predictingintramuscular fat, moisture and Warner-Bratzler shear force in pork muscle usingnear infrared reflectance spectroscopy. Animal Science, 82 (1), 111-116.Brøndum, J., Byrne, D.V., Bak, L.S., Bertelsen, G.; Engelsen, S.B. (2000b). Warmed-overflavour in porcine meat - a combined spectroscopic, sensory and chemometric study.Meat Science, 54, 83-95.Brøndum, J.; Munck, L.; Henckel, P., Karlsson, A.; Tornberg, E.; Engelsen, S. B. (2000a).Prediction of water-holding capacity and composition of porcine meat by comparativespectroscopy. Meat Science, 55, 177-185.ElMasry, G., Sun, D.-W., Allen, P. (2011). Non-destructive determination of water-holdingcapacity in fresh beef by using NIR hyperspectral imaging, Food ResearchInternational, doi:10.1016/j.foodres.2011.05.001Garrido Frenich, A.; Jouan-Rimbaud, D.; Massart, D. L.; Kuttatharmmakul, S.; MartinezGalera, M.; Martinez Vidal, J.L. (1995). Wavelength Selection Method forMulticomponent Spectrophotometric Determinations Using Partial Least Squares.Analyst 120, 2787-2792.Leefler T. P., Moser C. R., McManus B. J., Urh J. J., Keeton J. T., Claflin A. (2008).Determination of moisture and fat in meats by microwave and nuclear magneticresonance analysis: Collaborative study. Journal of AOAC International 91, 802–810.Prevolnik, M. Škrlep, M., Janeš, L., Velikonja-Bolta, Š., Škorjanc, D., Čandek-Potokar, M.(2011). Accuracy of near infrared spectroscopy for prediction of chemicalcomposition, salt content and free amino acids in dry-cured ham. Meat Science 88.299–304.Tøgersen, G.; Arnesen; J. F.; Nilsen; B.N.; Hildrum, K.I. (2003). On-line prediction ofchemical composition of semi-frozen ground beef by non-invasive NIR spectroscopy.Meat Science 63, 515–523.