Food Chemistry 171 (2015) 258–265

Contents lists available at ScienceDirect

Food Chemistry journal homepage: www.elsevier.com/locate/foodchem

Analytical Methods

Suitability of hyperspectral imaging for rapid evaluation of thiobarbituric acid (TBA) value in grass carp (Ctenopharyngodon idella) fillet Jun-Hu Cheng a, Da-Wen Sun a,b,⇑, Hong-Bin Pu a, Qi-Jun Wang a, Yu-Nan Chen a a b

College of Light Industry and Food Sciences, South China University of Technology, Guangzhou 510641, China Food Refrigeration and Computerized Food Technology, Agriculture and Food Science Centre, University College Dublin, National University of Ireland, Belfield, Dublin 4, Ireland

a r t i c l e

i n f o

Article history: Received 6 June 2014 Received in revised form 25 August 2014 Accepted 30 August 2014 Available online 10 September 2014 Keywords: Hyperspectral imaging TBA value PLSR Wavelength selection Grass carp

a b s t r a c t The suitability of hyperspectral imaging technique (400–1000 nm) was investigated to determine the thiobarbituric acid (TBA) value for monitoring lipid oxidation in fish fillets during cold storage at 4 °C for 0, 2, 5, and 8 days. The PLSR calibration model was established with full spectral region between the spectral data extracted from the hyperspectral images and the reference TBA values and showed good performance for predicting TBA value with determination coefficients (R2P) of 0.8325 and root-meansquare errors of prediction (RMSEP) of 0.1172 mg MDA/kg flesh. Two simplified PLSR and MLR models were built and compared using the selected ten most important wavelengths. The optimised MLR model yielded satisfactory results with R2P of 0.8395 and RMSEP of 0.1147 mg MDA/kg flesh, which was used to visualise the TBA values distribution in fish fillets. The whole results confirmed that using hyperspectral imaging technique as a rapid and non-destructive tool is suitable for the determination of TBA values for monitoring lipid oxidation and evaluation of fish freshness. Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction It is well-known that fish is always considered as a significant source of high quality animal proteins for human nutrition and consumption. Particularly, fish is also an important and nutritional seafood due to the natural and high concentrations of polyunsaturated x-3 fatty acids (PUFA), such as eicosapentaenoic acid (EPA, C20:5 n-3) and docosahexaenoic acid (DHA, C22:6 n-3), which have been proved to have beneficial and special health effects to prevent cardiovascular disease, lower cholesterol levels and blood viscosity, and strengthen memory and thinking ability for humans (Karlsdottir et al., 2014; Iglesias & Medina, 2008). However, owing to the high content of PUFA, coupled with extremely active existing pro-oxidants, fish is very vulnerable to lipid oxidation. Degradation of PUFAs caused by self-acting or enzymatic oxidation during diverse storage conditions and processing operations can easily result in the formation of undesirable oxidation products such as peroxides, hydroperoxides, conjugated dienes/trienes, aldehydes, ketones, and others (Alishahi & Aïder, 2012; St. Angelo, Vercellotti, Jacks, & Legendre, 1996). They are able to modify fish ⇑ Corresponding author at: College of Light Industry and Food Sciences, South China University of Technology, Guangzhou 510641, China. Tel.: +353 1 7167342; fax: +353 1 7167493. E-mail address: [email protected] (D.-W. Sun). URLs: http://www.ucd.ie/refrig, http://www.ucd.ie/sun (D.-W. Sun). http://dx.doi.org/10.1016/j.foodchem.2014.08.124 0308-8146/Ó 2014 Elsevier Ltd. All rights reserved.

muscle components and to generate rancidity and hazardous substances (de Abreu, Losada, Maroto, & Cruz, 2011). For this reason, lipid oxidation has been acknowledged as a leading cause of freshness loss and quality deterioration in fish muscle. Therefore, measurement and evaluation of lipid oxidation of fish is an obviously important task. There are various analytical methods and techniques available for the determination and assessment of lipid oxidation in muscle foods. Traditionally, the methods currently used for monitoring lipid oxidation are based on chemical analysis to determine some important oxidative parameters that are able to provide useful and accurate information on reflecting lipid oxidation degree (Klaypradit, Kerdpiboon, & Singh, 2010; Karoui & Blecker, 2011). One of the most important products of lipid oxidation is malondialdehyde (MDA), which is considered to be a carcinogenic initiator and mutagen (Halliwell & Chirico, 1993). MDA has been used as an indicator of oxidative damage in biological samples and muscle foods (Guillen-Sans & Guzman-Chozas, 1998) and the pink fluorescent MDA-thiobarbituric acid (MDA-TBA) complex produced after reaction with 2-thiobarbituric acid (TBA) at low pH and high temperature is commonly measured (Fernández, Pérez-Álvarez, & Fernández-López, 1997). Although this TBA value evaluation method can provide relatively precise analysis and valuable measurement, it is usually time-consuming, work-intensive and requires the use of large amounts of chemical solvents and analytical

J.-H. Cheng et al. / Food Chemistry 171 (2015) 258–265

reagents that might be hazardous and harmful to analysts and the lab environment (Balasubramanian & Panigrahi, 2011; Iñón, Garrigues, Garrigues, Molina, & de la Guardia, 2003). Therefore, it is very difficult to be employed on-line in a rapid and non-destructive manner. Recently, hyperspectral imaging technique as an emerging and innovative tool has been increasingly used to non-destructively and rapidly determine and evaluate food quality and safety (Cheng et al., 2013; Cheng & Sun, in press; Barbin et al., 2012; ElMasry et al., 2011; ElMasry, Sun, & Allen, 2011; ElMasry, Sun, & Allen, 2012; Kamruzzaman, ElMasry, Sun, & Allen, 2011; Kamruzzaman, ElMasry, Sun, & Allen, 2012). The hyperspectral imaging system integrates the traditional spectroscopy and imaging or computer vision (Du & Sun, 2005; Jackman, Sun, Du, & Allen, 2008; Sun & Brosnan, 2003; Valous, Mendoza, Sun, & Allen, 2009) technique into one system and provides a hypercube I(x, y, k) including spatial (x and y) and spectral (k) information simultaneously, which presents a three-dimensional (3D) dataset that contains many images of the same object, and each of which is measured at a different wavelength (ElMasry, Kamruzzaman, Sun, & Allen, 2012; Sun, 2010). In addition, grass carp (Ctenopharyngodon idella) as one kind of popular freshwater fish is widely farm-cultured in China and has also been introduced and accepted in Europe and America due to its rapid growth rate, easy cultivation, high yield, and low price as well as high nutritional values (Cheng, Qu, Sun, & Zeng, 2014). Accordingly, some exploratory investigations about using hyperspectral imaging have been successfully conducted for quality evaluation of grass carp fillets based on some important parameters such as colour (Cheng, Sun, Pu, & Zeng, 2014), textural firmness (Cheng et al., 2014), and total volatile basic nitrogen (TVB-N) value (Cheng, Sun, Zeng, & Pu, in press). Based on these studies, it has been proved that hyperspectral imaging technique shows the potential for rapid and non-destructive assessment and analysis of fish quality and safety. However, to the best of our knowledge, no investigation on determination of TBA value in grass carp fillet using hyperspectral imaging technique has been reported until now. Therefore, the purpose of this study was first to investigate the suitability of using hyperspectral imaging (400–1000 nm) for nondestructive and rapid determination of TBA value for evaluation of lipid oxidation in farmed grass carp fillets and to generate the distribution map of TBA value in grass carp fillets with multivariate analysis.

2. Materials and methods 2.1. Fish samples preparation Eighteen farmed fresh grass carps with similar age of three months, approximately weight of 1.5 kg, and feeding environment from the same freshwater aquaculture ponds were purchased in a local market in Guangzhou, China, and directly transported to the laboratory alive in water using a big plastic bucket within 15 min. Upon arrival, the fishes were stunned by a sharp blow to the head with a wooden stick and then gill cutting. The internal organs were removed at the same time with bloodletting from the fish belly location. Then they were immediately beheaded, skinned, and filleted and then washed with cold water. In order to acquire more fish samples, the fresh fillets were instantly subsampled into a cuboid shape with analogous sizes of 3.0 cm  3.0 cm  1.0 cm (length  width  thickness). Consequently, a total of 180 subsamples of fish fillets were obtained from different locations of tested fish fillets. In order to acquire a practical and full range of TBA values indicating the fish flesh from fresh (fully acceptable) to badly oxidative (totally unacceptable) for further establishing better cal-

259

ibration and prediction models, all the subsamples were labelled and packaged into the sealed plastic bags and randomly divided into four groups subjected to cold storage for 0, 2, 5, and 8 days at 4 ± 1 °C in a refrigerator (Haier Company, Qingdao, China). Among these 180 subsamples, two thirds samples (n = 120) were used to create the calibration models and the remaining one third samples (n = 60) were used to establish the prediction model. Each group having 45 subsamples were first scanned by the hyperspectral imaging system and then these subsamples were immediately used to measure the reference TBA values using the traditional method described below. 2.2. Evaluation of TBA value In this study, lipid oxidation was monitored by the evaluation of thiobarbituric acid reactive substances (TBA-RS) according to the procedure of Salih, Smith, Price, and Dawson (1987) with some modifications. Five grams of grass carp fillet muscle was minced and then mixed with 25 mL of trichloroacetic acid (20%) and 20 mL of distilled water for centrifuging for 10 min with the revolving speed of 8000 rpm, and the filtrate was diluted with ultrapure water to 50 mL. The mixture of 10 mL of diluent and 10 mL of thiobarbituric acid solution was heated in a boiling water bath (95–100 °C) for 15 min to develop a pink colour, and then cooled with running tap water for 5 min. The absorbance of the cooled supernatant was measured at 532 nm by a spectrophotometer (UV-1800, Shimadzu, Instruments of Mfg. Co. Ltd., Suzhou, China). A standard curve was prepared using 1,1,3,3-tetrameth-oxypropane at a concentration ranging from 0 to 10 ppm, and the amounts of TBA-RS were expressed as mg of MDA/kg sample. 2.3. Hyperspectral imaging system A typical lab push-broom hyperspectral imaging system was used to acquire hyperspectral images of grass carp fillets in reflectance mode. This system is mainly composed of a line-scanning imaging spectrograph (Imspector V10E, Spectral Imaging Ltd., Oulu, Finland) covering the spectral range of 308–1105 nm, a high performance charge-coupled device (CCD) camera (DL-604M, Andor, Ireland) attached with the effective resolution of 1004  1002 pixels, a camera lens (OLE23, Schneider, German), an illumination unit consisting of two 150 W halogen lamps (2900-ER, Illumination Technologies Inc., New York, USA) equipped with a fiber optical line light situated at an angle of 45° to light the moving platform controlled by a stepping motor (IRCP0076-1COMB, Isuzu Optics Corp., Taiwan, China), and a computer control system with hyperspectral image data acquisition software (Spectral Image software, Isuzu Optics Corp., Taiwan, China). The software could regulate the exposure time, motor speed, combining mode, wavelength range, and image acquisition. In this study, the actual working spectral range of this hyperspectral imaging system was 308–1105 nm with a spectral increment of about 1.58 nm between the contiguous bands, thus generating a total of 501 bands. However, according to the visual inspection of spectral information of the acquired hyperspectral images, there is a low signal-to-noise at both ends of the spectral range that would influence the further reliability and prediction ability of models. Therefore, the spectral range of 308–399 nm and 1001– 1105 nm were removed and the effective spectral range of 400– 1000 nm with a total of 381 wavebands (variables) was considered and used for further analysis. 2.4. Image acquisition and calibration For each group fish samples, 45 cubed subsamples were placed on the moving platform and then conveyed to the field of the view

260

J.-H. Cheng et al. / Food Chemistry 171 (2015) 258–265

of the camera to be scanned line by line with the adjusted motor speed (1.5 mm/s) and exposure time (30 ms) to coordinate with the image acquisition. Thus, a total of 180 three dimensional hyperspectral images were created, recorded and stored in a raw format before being image processed. To reduce the effects of illumination and detector sensitivity as well as the differences in camera and physical configuration of the imaging system, the raw acquired hyperspectral images (R0) should be calibrated into the reflectance mode with two extra images for black (B) and standard white (W) reference images. The black image (0% reflectance) was obtained by recording a spectral image after fully covering the camera lens with its black cap. The white reference image was acquired using a uniform Teflon white calibration tile (100% reflectance). The calibrated image (Rc) was then calculated by the following equation:

Rc ¼

R0  B  100% W B

ð1Þ

2.5. ROI identification and spectra extraction After image acquisition and reflectance calibration, the region of interests (ROIs) with a rectangle shape within hyperspectral images were identified and selected based upon the important locations corresponding to areas where the reference subsamples had been collected. Then the pixel spectra within each region were extracted and averaged, which produced X-matrix of 180 spectra. All extraction routines were programmed using the software ENVI v4.8 (ITT Visual Information Solutions, Boulder, CO, USA). 2.6. Multivariable data analysis The numerous spectral data extracted from the hyperspectral images generally include amounts of effective and valuable information and inevitably there exists some redundant and interferential information affecting the calibration and validation models. In order to improve the predictive ability and reliability of models and reduce the variability between samples due to scattering and optical interference possibly caused by water movement in cold storage, the spectral pre-processing approach of multiplicative scatter correction (MSC) was applied to remove the undesirable scatter effect from the data matrix prior to data modelling (Jin et al., 2011). After spectral pre-processing, selecting a robust and reliable analytical method for processing and building a calibration model for quantitative analysis is of significance. Partial leastsquares regression (PLSR) has widely been proved to be a reliable calibration tool for spectral data modelling (Krishnan, Williams, McIntosh, & Abdi, 2011). The great advantage of this regression algorithm is to avoid colinearity problems and is very suitable in the situation when the number of variables is greater than the number of samples (Mehmood, Liland, Snipen, & Sæbø, 2012). In this study, the quantitative relationship was established between spectral data of samples and their reference measured TBA values based on PLSR analysis. PLSR can find a set of independent variables (wavelengths), the X-matrix (180  381), and the dependent variable (TBA values), the Y-matrix (180  1), where the X-matrix represents the average spectral data at the 381 wavelengths for the 180 subsamples, whereas, the Y-matrix shows the values of TBA in the subsamples. On the other hand, multiple linear regressions (MLR) is another method to establish the quantitative relationship between two or more explanatory independent variables and a dependent variable by fitting a linear equation to the observed data (Wu et al., 2012). However, the main disadvantage of this approach is to require the samples are more than variables and to be easily affected by the colinearity between variables (Guillén-Casla, Rosales-Conrado,

León-González, Pérez-Arribas, & Polo-Díez, 2011). In this study, the number of the wavelength variables was much greater than the number of samples (381 vs. 180). Therefore, after the optimal wavelength selection, running MLR algorithm would be helpful to establish a better calibration model. The implementations of PLSR and MLR were conducted by the chemometric software (Unscrambler version 9.7, CAMO, Trondheim, Norway). 2.7. Optimal wavelength selection The acquired hyperspectral images usually characterise the high dimensionality with redundancy and multicollinearity among contiguous wavelength bands, which can easily result in the consequent time-consuming calibration process and affect the speed of computation related to the processing of the hyperspectral images (Cheng et al., in press; Lorente, Aleixos, Gómez-Sanchis, Cubero, & Blasco, 2013). Thus, it is of interest to find the minimal number of wavelengths carrying the most valuable information, which may be equally or more efficient than the full wavelength range and provide satisfactory prediction results. In this study, having a great influence on lipid oxidation, the most sensitive wavelengths were selected by calculating weighted regression coefficients also called b-coefficients that were corresponding to the PLSR model with full range spectra. The wavelengths with the highest absolute values of weighted regression coefficients (regardless of the sign) were selected as the optimal wavelengths for TBA value prediction. On the basis of the selected characteristic wavelengths, new optimised PLSR and MLR models were developed and compared. The whole procedures for spectral analysis were carried out in the Unscrambler chemometric software (Unscrambler version 9.7, CAMO, Trondheim, Norway). 2.8. Model validation and evaluation There are a number of methods to validate multivariate analysis models. In this study, full cross-validation also called leave-oneout cross-validation was employed to validate the established PLSR model. The process of this method was conducted by removing one sample or a subset of samples from the calibration dataset and a new PLSR model was then constructed based on the remaining calibration samples (ElMasry & Wold, 2008). Finally, the obtained model was applied to predict the sample left out. The procedure was repeated for every subsample in the dataset, providing a more realistic measurement of the prediction errors of the model. In addition, the optimal number of latent variables (LV) based on the PLSR algorithm for building the calibration model was determined by using the minimum value of predicted residual error sum of squares (PRESS). As for the performance of the established models, it is necessary to look for effective methods to evaluate the predictive effectiveness, reliability and accuracy for practical applications. Commonly, the evaluation indicator systems are mainly related to the coefficients of determination (R2) and root mean square errors in calibration (R2C, RMSEC), cross-validation (R2CV, RMSECV) and prediction (R2P, RMSEP), respectively. Generally speaking, an admirable and comparable model should have higher values of R2C, R2CV, and R2P, and lower values of RMSEC, RMSECV and RMSEP as well as a small difference between them (Cheng et al., 2014). In details, R2 indicates the proportion of the variance in reference data that can be explained by the variance in the predicted data. The values of RMSEC, RMSECV and RMSEP are measurements of the root mean square errors in the analysis and assessment of the fitting degree of regression during calibration, cross-validation and prediction with lower values implying better predictive capacity (Hernández-Martínez et al., 2013). It is always anticipated to acquire RMSEs as close as zero and R2 as close as one. In fact, the value of R2 in the range of 0.82–0.90 usually

261

J.-H. Cheng et al. / Food Chemistry 171 (2015) 258–265

Fresh fish

o

Storing at 4 C for 0, 2, 5, 8 days

Hyperspectral imaging system

ROI selection

Fish fillets

Raw hyperspectral images

Spectral extration

Fish fillet subsamples

Calibrated hyperspectral images

Spectral data matrix

Reference measured TBA values

Segmented images

PLS regression

Images at optimal wavelengths

Wavelength selection

Distribution maps

Spectral data at optimal wavelengths

PLSR MLR

Best prediction model Fig. 1. Main steps of determination of TBA value in grass carp fillet by hyperspectral imaging.

indicates good performance of a model, while the value of R2 lower than 0.82 reveals inaccurate and relatively poor performance, and the value of R2 higher than 0.90 shows excellent performance (Williams, 2001). The values of PRESS, RMSEC, RMSECV and R2 were calculated with the following equations on the basis of the reference measured and predicted TBA values:

PRESS ¼

n X ðycal  yact Þ2

ð2Þ

i¼1

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn 2 i¼1 ðycal  yact Þ RMSEC ¼ n sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn 2 i¼1 ðypred  yact Þ RMSECV ¼ n Pn 2 ðy  y act Þ R2 ¼ 1  Pni¼1 cal 2 i¼1 ðycal  ymean Þ

ð3Þ ð4Þ ð5Þ

where n is the number of samples; yact is the actual value; ycal is the calibrated value; ypred is the predicted value; ymean is the mean of the reference measured value.

images into the chemical images for prediction of TBA value distribution of the examined fish fillets. The visualisation process was conducted by calculating the dot product between spectrum of each image pixel and the PLSR regression coefficients (Elmasry et al., 2012). The resulting multiplication deduced TBA value in all spots of the sample, which assisted to obtain the observation of the lipid oxidation degree within one subsample as well as from fillets to fillets. The finally obtained chemical images or visualised distribution maps were shown in a linear colour bar with different colours (the pixels with low values were shown in blue and those with high values were shown in red), representing corresponding value of the predicted TBA in the whole fish fillet subsample, which is helpful to recognise and interpret the lipid oxidation degree by inspection of the different colour distribution. All the procedures of visualisation were programmed in the software Matlab 2010a (The Mathworks Inc., Mass, USA). Fig. 1 illustrates the main steps of determination of TBA values in grass carp fillets by hyperspectral imaging technique. 3. Results and discussion 3.1. Spectral feature analysis

2.9. Visualisation of lipid oxidation The visualisation of spatial distribution of TBA values is very significant to observe the changes of lipid oxidation from sample to sample even from position to position within the same sample. The great advantage of the hyperspectral imaging technique is able to generate the distribution map of different TBA values during cold storage on the basis of spatial position of every pixel and the corresponding colour values. In this study, the new selected optimised calibration model was developed using only the spectral data at those characteristic wavelengths and then the new model was used to transfer and visualise every pixel of the hyperspectral

The average reflectance spectra extracted from the pixels within the ROIs of the tested grass carp fillets in the spectral range (400– 1000 nm) during different cold storage days at 4 °C are presented in Fig. 2. It can be observed that there was a similar general trend throughout the examined wavelength region at different storage days. However, some differences still existed in the magnitude of spectral reflectance value, suggesting that lipid oxidation in fish muscle occurred during post-mortem handling to some extent. For example, the fresh (0 d) samples showing the lowest reflectance values that was corresponding the smallest TBA values indicated great absorptions occurred and the minimum lipid oxidation

262

J.-H. Cheng et al. / Food Chemistry 171 (2015) 258–265

0.45

2d

0.35

5d

Predicted TBA value (mg/kg)

0.4

8d

0.3 Reflectance

1.2

0d

0.25 0.2 0.15 0.1 0.05

R² P = 0.8325 RMSEP = 0.1172 1

0.8

0.6

0.4

0.2 0.2

0 400

500

600 700 800 Wavelengths/nm

900

0.4

1000

Fig. 2. Average spectral features of the tested grass carp fillets during cold storage.

appeared. As to the differences between spectral reflectance values, they were mainly attributed to the variations of chemical compositions of fish muscle induced possibly by microbial spoilage, enzyme activity and cold storage time and temperature. In details, the TBA value of grass carp fillet during refrigerated storage tended to increase throughout the storage (data not shown). The increase of TBA values were caused by the formation of secondary lipid oxidation products, including n-alkanals, trans-2-alkenals, 4-hydroxytrans-2-alkenals, and malonaldehyde, especially aldehydes (Thanonkaew, Benjakul, Visessanguan, & Decker, 2006). These lipid oxidation products to some extent could result in the crosslinking of myofibrillar proteins, and the structural and functional changes of these proteins (de Abreu et al., 2011; Yanishlieva & Marinova, 2001). It can also be noticed that the reflectance values (8 d) were different from the other days obtained in the wavelength range of 400–600 nm, which mainly located in the day 2 and day 5. The possible reason was contributed to the slight decrease in TBA value at day 8 induced by the decomposition of aldehydes or the interaction of aldehydes with fish muscle proteins (Rawdkuen, Jongjareonrak, Benjakul, & Chaijan, 2008). On the other hand, the overtone and combination vibrations of the molecular chemical bonds (O–H, C–H, C–O, N–H and S–H) are commonly used to explain the variations of the spectra. In the visible range (400–780 nm), there were two obvious absorption bands located near 420 nm and 560 nm, respectively, which was possibly due to the absorption of pigments such as hemoglobin and myoglobin reported by Zhu, Zhang, He, Liu, and Sun (2013) for halibut. In the wavelength region of 780–1000 nm, there was a weak absorption sited at about 780 nm relating to the third overtone O–H stretching (water) or generated by lipid oxidation (Iqbal, Sun, & Allen, 2013). Lipid has an absorbance peak at 930 nm corresponding to the third overtone C–H stretching in the methylene group of fat (Klaypradit et al., 2010), which was a little difficult to discern in Fig. 2. Another significant absorption peak was observed at 970 nm that was usually ascribed to the second overtone O–H stretching in water (Garini, Young, & McNamara, 2006).

0.6

0.8

1

1.2

Measured TBA value (mg MDA/kg flesh) Fig. 3. Predicted and measured TBA values for PLSR model using the whole spectral wavelengths range.

3.2. Prediction of TBA value using full spectral range In order to estimate the performance of the calibration and prediction models, the PLSR model was developed using full spectral range after arranging the spectral data extracted from all fish subsamples in the matrix (X) and their corresponding TBA values in the column vector (Y). Specifically, the PLSR calibration model was obtained based on the reflectance spectra of the tested samples at all 381 wavelengths as X-variables and the reference measured TBA values as Y-variable. Table 1 shows the calibration and cross-validation analytical parameters of PLSR model for predicting TBA values of fish samples. As indicated in Table 1, it can be seen that the PLSR model exhibited a good performance to predict the TBA values with the values of determination coefficients (R2C = 0.8522 and R2P = 0.8244) and root-mean-square errors (RMSEC = 0.1105 mg MDA/kg flesh and RMSEP = 0.1201 mg MDA/kg flesh). To further improve the ability of prediction model, MSC as a widely used pre-processing technique was applied to eliminate the undesirable scattering effect from the data matrix. The resulting performance of the developed MSC-PLSR model is illustrated in Table 1. Obviously, application of this pre-processing method was useful for improving the predictive model performance compared with the previous results with the improvement of R2 (0.0245 and 0.0081) and diminution of RMSEs (0.0100 mg MDA/kg flesh and 0.0029 mg MDA/kg flesh). Moreover, the selection of the ideal number of latent variables (LVs) is an important step to establish a robust PLSR model with the best calibration results. In this study, the ideal number of ten LVs of the PLSR model was determined and obtained by the lowest value of PRESS. According to the requirement, ten latent variables were obtained and comparable prediction ability was shown with all the number of wavelengths (10 vs. 381). Fig. 3 shows the efficiency and robustness of the PLSR model (R2P = 0.8325 and RMSEP = 0.1172 mg MDA/kg flesh) for predicting the TBA values of grass carp fillets. The obtained model confirms the suitability of this hyperspectral imaging technique for TBA values prediction in a non-destructive and rapid manner. To validate and further improve the performance of prediction model for

Table 1 Performances of models for prediction of TBA values of grass carp fillet by hyperspectral imaging technique. Model

PLSR MSC-PLSR RC-PLSR RC-MLR

Variable number

381 381 10 10

LVs

10 10 8 /

Calibration

Cross-validation

Prediction

R2C

RMSEC

R2CV

RMSECV

R2P

RMSEP

0.8522 0.8767 0.8465 0.8524

0.1105 0.1005 0.1121 0.1098

0.8322 0.8510 0.8450 0.8479

0.1182 0.1107 0.1127 0.1116

0.8244 0.8325 0.8324 0.8395

0.1201 0.1172 0.1175 0.1147

J.-H. Cheng et al. / Food Chemistry 171 (2015) 258–265

practical applications, more samples should be considered and included in the calibration set to account for more variability of chemical composition and biological variations in fish muscle. 3.3. Prediction of TBA values using selected optimal wavelengths Using the selected optimal wavelengths carrying the most characteristic information may be equally or more efficient than using full wavelengths for prediction of fish muscle attributes. This great advantage is very important to optimise the calibration models, to eliminate the high dimensionality of the hyperspectral images, to reduce the computation time and further to satisfy the real-time inspection (Liu, Sun, & Zeng, 2014). In this study, the regression coefficients (RC) from PLSR model analysis were employed to allocate the optimal wavelengths for simplifying the original calibration models. As a result, Fig. 4 shows the ten individual wavelengths obtained (444, 475, 553, 577, 590, 623, 710, 795, 847 and 937 nm), which were recognised as the effective wavelengths to replace the full spectral region for further predicting TBA values in fish fillet muscle. In order to find a better reduced calibration model using the selected ten most informative wavelengths, two classical linear regression algorithms of PLSR and MLR were used to correlate reduced spectral data with reference measured TBA values. The performances of simplified models of RC-PLSR and RC-MLR models for monitoring the lipid oxidation in grass carp fillet muscle are presented in Table 1. It can be noticed that the prediction ability of the optimised models (regardless of RC-PLSR or RC-MLR) possessed comparable or better performance than the models developed using the whole spectral range. It can thus be proved that the hyperspectral imaging technique with the only selected ten important wavelengths is also suitable to predict and monitor lipid oxidation in fish fillet. As illustrated in

Fig. 4. The selection of optimal wavelengths by weighted regression coefficients.

263

Table 1, compared with the RC-PLSR calibration model, the simplified RC-MLR model showed better effectiveness and predictability in predicting TBA values with R2P of 0.8325 and RMSEP of 0.1147 mg MDA/kg flesh, confirming that MLR is very advantageous in the aspects of the variables much less than samples and less colinearity than full spectral range. On the basis of the better RC-MLR model for the measurements of TVB values in grass carp fillets, the quantitative regression equation for monitoring the lipid oxidation in fish fillet was acquired and is described as follows:

Y TBA ¼ 0:196  0:446X 444nm  1:710X 475nm þ 5:6286X 553nm  36:415X 577nm þ 50:756X 590nm  29:523X 623nm þ 4:190X 710nm þ 61:857X 795nm  56:386X 847nm þ 2:105X 937nm

ð6Þ

where Xi nm is the reflectance spectral value at the wavelength of i nm and YTBA is the predicted TBA value. Therefore, the subsequent distribution map of TBA values during cold storage was visualised based on the RC-MLR model. 3.4. Distribution map of TBA value The great advantage of hyperspectral imaging technology is its ability to provide spatial information that is very important for revealing the detailed detection of fish and visualising chemical images of fish muscle components to observe the specific differences of lipid oxidation in fish fillet. For that reason, the final obtained RC-MLR model established with the help of the characteristic wavelengths was used to transfer each pixel of the image to predict TBA values in the whole fish fillet sample. The distribution map of lipid oxidation was produced after multiplying the MLR regression coefficients by the spectrum of each pixel in the image. In the obtained distribution map, pixels showing the similar spectral information provided the same predicted values of TBA, which were then visualised and described using similar colour in the image. Different colours in the distribution map indicate different prediction values in the image in proportion to the spectral feature differences of the corresponding pixels (Elmasry et al., 2012). Fig. 5 shows examples of visualisation of lipid oxidation distribution map of some examined grass carp fillets at different TBA values. It was easy to observe the differences in TBA values using the pseudo colour images within the samples. It was also noticed that the distribution density of TBA value of the grass carp fillet sample was non-uniform and asymmetric, which could demonstrate the lipid oxidation degree and variations of different locations of the same sample. In particular, the distribution of lipid oxidation illustrated in Fig. 5b (TBA = 0.4602 mg MDA/kg flesh) were fairly non-uniform along with different locations of fish fillet samples. It was possibly due to the oxygen penetration and ruptured cells at the cut surface that to a great extent could accelerate fish muscle degradation and

Fig. 5. Examples of distribution maps of TBA value in fish fillets at four different TBA values. (a) TBA = 0.2355 mg/kg; (b) TBA = 0.4602 mg/kg; (c) TBA = 0.9524 mg/kg; (d) TBA = 1.1213 mg/kg.

264

J.-H. Cheng et al. / Food Chemistry 171 (2015) 258–265

lipid oxidation or may be related to the effects of microorganism activities and occurrence of autolysis. When the TBA value was increased, the lipid oxidation was synchronously intensified. For example, Fig. 5c (TBA = 0.9524 mg MDA/kg flesh) and Fig. 5d (TBA = 1.1213 mg MDA/kg flesh) indicates the homogenous distribution of lipid oxidation using the same red colour, implying that a great level of degradation of chemical compounds occurred and consequently caused severe loss of freshness of fish fillets. In details, the density of red colour in Fig. 5d was much higher than that of in Fig. 5c, which means that the fish fillet suffered a greater degree of lipid oxidation. In addition, this visualisation of lipid oxidation was significant for further understanding of the dynamic changes of freshness loss and for direct determination of the purchase decision of consumers. Therefore, using hyperspectral imaging technique for successful prediction of TBA value confirmed that the technique is appropriate and has great potential for its use in monitoring and evaluation of lipid oxidation in fish muscle. 4. Conclusions The suitability of hyperspectral imaging technique in the wavelength range of 400–1100 nm for rapid and non-destructive evaluation of lipid oxidation was investigated based on the determination of TBA values in grass carp fillets. The PLSR calibration model established using full wavelengths generated good and satisfactory prediction results for TBA value with R2C and R2P of 0.8767 and 0.8325, and RMSEC and RMSEP of 0.1005 and 0.1172 mg MDA/kg flesh. The simplified PLSR and MLR calibration models using the ten optimal wavelengths (444, 475, 553, 577, 590, 623, 710, 795, 847 and 937 nm) selected by weighted regression coefficients both also showed good performance in TBA value prediction. The better developed RC-MLR model resulted in R2P of 0.8395 with RMSEP of 0.1147 mg MDA/kg flesh. The distribution maps visualised and generated by transferring the RC-MLR model to each pixel of the image were used to interpret the dynamic changes of lipid oxidation and fish freshness loss during cold storage. The current feasibility study showed the suitability of using hyperspectral imaging technique to determine the TBA value for evaluation of lipid oxidation in fish fillet muscle. Acknowledgments The authors were grateful to the Guangdong Province Government (China) for its support through the program of ‘‘Leading Talent of Guangdong Province (Da-Wen Sun)’’. This research was also supported by the National Key Technologies R&D Program (2014BAD08B09). References Alishahi, A., & Aïder, M. (2012). Applications of chitosan in the seafood industry and aquaculture: a review. Food and Bioprocess Technology, 5(3), 817–830. Barbin, Douglas F., ElMasry, G., Sun, D.-W., et al. (2012). Predicting quality and sensory attributes of pork using near-infrared hyperspectral imaging. Analytica Chimica Acta, 719, 30–42. Balasubramanian, S., & Panigrahi, S. (2011). Solid-phase microextraction (SPME) techniques for quality characterization of food products: a review. Food and Bioprocess Technology, 4(1), 1–26. Cheng, J.-H., & Sun, D.-W. (in press). Hyperspectral imaging as an effective tool for quality analysis and control of fish and other seafoods: Current research and potential applications. Trends in Food Science & Technology. http://dx.doi.org/ 10.1016/j.tifs.2014.03.006. Cheng, J.-H., Dai, Q., Sun, D.-W., Zeng, X.-A., Liu, D., & Pu, H.-B. (2013). Applications of non-destructive spectroscopic techniques for fish quality and safety evaluation and inspection. Trends in Food Science & Technology, 34(1), 18–31. Cheng, J.-H., Sun, D.-W., Zeng, X.-A., & Pu, H.-B. (in press). Non-destructive and rapid determination of TVB-N content for freshness evaluation of grass carp (Ctenopharyngodon idella) by hyperspectral imaging. Innovative Food Science & Emerging Technologies. http://dx.doi.org/10.1016/j.ifset.2013.10.013.

Cheng, J.-H., Qu, J.-H., Sun, D.-W., & Zeng, X.-A. (2014). Visible/near-infrared hyperspectral imaging prediction of textural firmness of grass carp (Ctenopharyngodon idella) as affected by frozen storage. Food Research International, 56, 190–198. Cheng, J.-H., Sun, D.-W., Pu, H., & Zeng, X.-A. (2014). Comparison of visible and lLong-wave near-infrared hyperspectral imaging for colour measurement of Grass Carp (Ctenopharyngodon idella). Food and Bioprocess Technology. http:// dx.doi.org/10.1007/s11947-014-1325-7. de Abreu, D. A. P., Losada, P. P., Maroto, J., & Cruz, J. M. (2011). Lipid damage during frozen storage of Atlantic halibut (Hippoglossus hippoglossus) in active packaging film containing antioxidants. Food Chemistry, 126(1), 315–320. Du, C. J., & Sun, D.-W. (2005). Comparison of three methods for classification of pizza topping using different colour space transformations. Journal of Food Engineering, 68(3), 277–287. ElMasry, G., Iqbal, A., Sun, D.-W., et al. (2011). Quality classification of cooked, sliced turkey hams using NIR hyperspectral imaging system. Journal of Food Engineering, 103(3), 333–344. ElMasry, G., Kamruzzaman, M., Sun, D.-W., & Allen, P. (2012). Principles and applications of hyperspectral imaging in quality evaluation of agro-food products: A review. Critical Reviews in Food Science and Nutrition, 52(11), 999–1023. ElMasry, G., Sun, D.-W., & Allen, P. (2011). Non-destructive determination of waterholding capacity in fresh beef by using NIR hyperspectral imaging. Food Research International, 44(9), 2624–2633. ElMasry, G., Sun, D.-W., & Allen, P. (2012). Near-infrared hyperspectral imaging for predicting colour, pH and tenderness of fresh beef. Journal of Food Engineering, 110(1), 127–140. ElMasry, G., & Wold, J. P. (2008). High-speed assessment of fat and water content distribution in fish fillets using online imaging spectroscopy. Journal of Agricultural and Food Chemistry, 56(17), 7672–7677. Fernández, J., Pérez-Álvarez, J. A., & Fernández-López, J. A. (1997). Thiobarbituric acid test for monitoring lipid oxidation in meat. Food Chemistry, 59(3), 345–353. Garini, Y., Young, I. T., & McNamara, G. (2006). Spectral imaging: Principles and applications. Cytometry Part A, 69(8), 735–747. Guillén-Casla, V., Rosales-Conrado, N., León-González, M. E., Pérez-Arribas, L. V., & Polo-Díez, L. M. (2011). Principal component analysis (PCA) and multiple linear regression (MLR) statistical tools to evaluate the effect of E-beam irradiation on ready-to-eat food. Journal of Food Composition and Analysis, 24(3), 456– 464. Guillen-Sans, R., & Guzman-Chozas, M. (1998). The thiobarbituric acid (TBA) reaction in foods: A review. Critical Reviews in Food Science and Nutrition, 38(4), 315–350. Halliwell, B., & Chirico, S. (1993). Lipid peroxidation: Its mechanism, measurement, and significance. The American Journal of Clinical Nutrition, 57(5), 715S–724S. Hernández-Martínez, M., Gallardo-Velázquez, T., Osorio-Revilla, G., AlmarazAbarca, N., Ponce-Mendoza, A., & Vásquez-Murrieta, M. S. (2013). Prediction of total fat, fatty acid composition and nutritional parameters in fish fillets using MID-FTIR spectroscopy and chemometrics. LWT-Food Science and Technology, 52(1), 12–20. Iglesias, J., & Medina, I. (2008). Solid-phase microextraction method for the determination of volatile compounds associated to oxidation of fish muscle. Journal of Chromatography A, 1192(1), 9–16. Iñón, F. A., Garrigues, J. M., Garrigues, S., Molina, A., & de la Guardia, M. (2003). Selection of calibration set samples in determination of olive oil acidity by partial least squares–attenuated total reflectance–Fourier transform infrared spectroscopy. Analytica Chimica Acta, 489(1), 59–75. Iqbal, A., Sun, D.-W., & Allen, P. (2013). Prediction of moisture, color and pH in cooked, pre-sliced turkey hams by NIR hyperspectral imaging system. Journal of Food Engineering, 117(1), 42–51. Jackman, P., Sun, D.-W., Du, C.-J., & Allen, P. (2008). Prediction of beef eating quality from colour, marbling and wavelet texture features. Meat Science, 80(4), 1273–1281. Jin, J.-W., Chen, Z.-P., Li, L.-M., Steponavicius, R., Thennadil, S. N., Yang, J., et al. (2011). Quantitative spectroscopic analysis of heterogeneous mixtures: The correction of multiplicative effects caused by variations in physical properties of samples. Analytical Chemistry, 84(1), 320–326. Kamruzzaman, M., ElMasry, G., Sun, D.-W., & Allen, P. (2011). Application of NIR hyperspectral imaging for discrimination of lamb muscles. Journal of Food Engineering, 104(3), 332–340. Kamruzzaman, M., ElMasry, G., Sun, D.-W., & Allen, P. (2012). Prediction of some quality attributes of lamb meat using near-infrared hyperspectral imaging and multivariate analysis. Analytica Chimica Acta, 714, 57–67. Karlsdottir, M. G., Sveinsdottir, K., Kristinsson, H. G., Villot, D., Craft, B. D., & Arason, S. (2014). Effects of temperature during frozen storage on lipid deterioration of saithe (Pollachius virens) and hoki (Macruronus novaezelandiae) muscles. Food Chemistry, 156, 234–242. Karoui, R., & Blecker, C. (2011). Fluorescence spectroscopy measurement for quality assessment of food systems—a review. Food and Bioprocess technology, 4(3), 364–386. Klaypradit, W., Kerdpiboon, S., & Singh, R. K. (2010). Application of Artificial neural networks to predict the oxidation of menhaden fish oil obtained from Fourier transforms infrared spectroscopy method. Food and Bioprocess Technology, 4(3), 475–480. Krishnan, A., Williams, L. J., McIntosh, A. R., & Abdi, H. (2011). Partial Least Squares (PLS) methods for neuroimaging: A tutorial and review. Neuroimage, 56(2), 455–475.

J.-H. Cheng et al. / Food Chemistry 171 (2015) 258–265 Liu, D., Sun, D.-W., & Zeng, X. A. (2014). Recent advances in wavelength selection techniques for hyperspectral image processing in the food industry. Food and Bioprocess Technology, 7(2), 307–323. Lorente, D., Aleixos, N., Gómez-Sanchis, J., Cubero, S., & Blasco, J. (2013). Selection of optimal wavelength features for decay detection in citrus fruit using the ROC curve and neural networks. Food and Bioprocess Technology, 6(2), 530–541. Mehmood, T., Liland, K. H., Snipen, L., & Sæbø, S. (2012). A review of variable selection methods in partial least squares regression. Chemometrics and Intelligent Laboratory Systems, 118, 62–69. Rawdkuen, S., Jongjareonrak, A., Benjakul, S., & Chaijan, M. (2008). Discoloration and lipid deterioration of farmed giant catfish (Pangasianodon gigas) muscle during refrigerated storage. Journal of Food Science, 73(3), C179–C184. Salih, A., Smith, D., Price, J., & Dawson, L. (1987). Modified extraction 2thiobarbituric acid method for measuring lipid oxidation in poultry. Poultry Science, 66(9), 1483–1488. St. Angelo, A. J., Vercellotti, J., Jacks, T., & Legendre, M. (1996). Lipid oxidation in foods. Critical Reviews in Food Science & Nutrition, 36(3), 175–224. Sun, D.-W. (2010). Hyperspectral imaging for food quality analysis and control. San Diego, California, USA: Academic Press/Elsevier. 496, ISBN: 978-0-12-374753-2.

265

Sun, D.-W., & Brosnan, T. (2003). Pizza quality evaluation using computer vision part 1 - Pizza base and sauce spread. Journal of Food Engineering, 57(1), 81–89. Thanonkaew, A., Benjakul, S., Visessanguan, W., & Decker, E. A. (2006). The effect of metal ions on lipid oxidation, colour and physicochemical properties of cuttlefish (Sepia pharaonis) subjected to multiple freeze–thaw cycles. Food Chemistry, 95(4), 591–599. Valous, N. A., Mendoza, F., Sun, D.-W., & Allen, P. (2009). Colour calibration of a laboratory computer vision system for quality evaluation of pre-sliced hams. Meat Science, 81(1), 132–141. Williams, P. C. (2001). Implementation of near-infrared technology. Near-infrared technology in the agricultural and food industries, 2, 143. Wu, D., Shi, H., Wang, S., He, Y., Bao, Y., & Liu, K. (2012). Rapid prediction of moisture content of dehydrated prawns using online hyperspectral imaging system. Analytica Chimica Acta, 726, 57–66. Yanishlieva, N. V., & Marinova, E. M. (2001). Stabilisation of edible oils with natural antioxidants. European Journal of Lipid Science and Technology, 103(11), 752–767. Zhu, F., Zhang, D., He, Y., Liu, F., & Sun, D.-W. (2013). Application of visible and near infrared hyperspectral imaging to differentiate between fresh and frozen– thawed fish fillets. Food and Bioprocess Technology, 6(10), 2931–2937.