non-invasive ‘through-package’ assessment of the microstructural

Upload: claire-s-malibiran

Post on 07-Apr-2018

221 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/6/2019 Non-invasive through-package assessment of the microstructural

    1/7

    LWT 40 (2007) 737743

    Research Note

    Non-invasive through-package assessment of the microstructural

    quality of a model food emulsion by the NMR MOUSE

    Adrian M. Haiduca, Elena E. Trezzaa, Dagmar van Dusschotenb, Aleksander A. Reszkac,John P.M. van Duynhovena,

    aUnilever Food and Health Research Institute, Unilever R&D, P.O. Box 120, 3133 AT Vlaardingen, The NetherlandsbDSM Research, P.O. Box 18, 6160 MD Geleen, The Netherlands

    cSCC Global Technology Centre, Unilever R & D, P.O. Box 120, 3133 AT Vlaardingen, The Netherlands

    Received 4 August 2005; received in revised form 11 February 2006; accepted 16 February 2006

    Abstract

    Recently, NMR has been demonstrated in a truly non-invasive through-package sensor mode, also denoted as the MObile Universal

    Surface Explorer (MOUSE). In this feasibility study, we present a first example where we use the MOUSE sensor for assessment of the

    microstructural quality of a food material. We have taken model systems consisting of protein-stabilized oil-in-water emulsions, where an

    important microstructural quality parameter is water exudation (WE). In order to establish a sound relation between MOUSE signals

    and WE, it was necessary to deploy multivariate calibration techniques. It was found that the performance of the MOUSE is comparable

    to that of conventional benchtop NMR. Thus it was demonstrated that the NMR MOUSE presents a good option for non-invasive

    assessment of microstructural quality parameters, e.g. in manufacturing and in the supply chain.

    r 2006 Swiss Society of Food Science and Technology. Published by Elsevier Ltd. All rights reserved.

    Keywords: NMR MOUSE sensor emulsions

    1. Introduction

    Within the foods industry we are witnessing a transition

    from the classical process and quality control taking place

    in an off-line laboratory, towards non-invasive, on-line

    and real-time measurement of product quality parameters.

    This has resulted in the tight integration of process and

    quality control, and is yielding significant economical

    benefits. Many on-line measurement technologies have

    become available that can be integrated with process and

    quality control in a versatile manner (Senorans, Ibanez, &Cifuentes, 2003). Most of these measurement tools are

    based on spectrophotometry (Williams & Norris, 2003)

    and provide only feedback on the chemical composition of

    the food system of interest, however. Although food

    quality is often perceived as compositional, i.e. the presence

    of molecules that can be tasted or smelled, the dominating

    factor is often the microstructure of the product. The

    microstructure of a product is particularly important for

    semi-solid foods, ranging from sauces and dressings to

    yoghurt, spreads and ice cream. Sofar primarily ultrasound

    technology has matured enough to allow the on-line

    measurement of solid-to-liquid ratios (Prakash & Ramana,

    2003) and (for specific systems) rheological properties

    (Bamberger & Greenwood, 2004; Ross, Pyrak-Nolte, &

    Campanella, 2004). Despite these successes, one has to

    conclude that these techniques have their limitations withrespect to versatility when it comes to assessment of a

    broad range of microstructural parameters. NMR is a

    good candidate to fulfil such a role, and within the foods

    industry it is applied to assess a range of microstructural

    features on (relatively) low-cost benchtop spectrometers

    (Todt, Burk, Guthausen, Kamlowski, & Schmalbein,

    2001). However, since these applications still require that

    a sample is taken and placed in the benchtop NMR

    spectrometer, this technology still belongs to domain of the

    classical process and quality control laboratory.

    ARTICLE IN PRESS

    www.elsevier.com/locate/lwt

    0023-6438/$30.00 r 2006 Swiss Society of Food Science and Technology. Published by Elsevier Ltd. All rights reserved.

    doi:10.1016/j.lwt.2006.02.026

    Corresponding author. Tel.: 3110 4605534; fax: 31 10 4605310.

    E-mail address: [email protected]

    (J.P.M. van Duynhoven).

    http://www.elsevier.com/locate/lwthttp://dx.doi.org/10.1016/j.lwt.2006.02.026mailto:[email protected]:[email protected]:[email protected]:[email protected]://dx.doi.org/10.1016/j.lwt.2006.02.026http://www.elsevier.com/locate/lwt
  • 8/6/2019 Non-invasive through-package assessment of the microstructural

    2/7

    Recently, NMR has also been presented in a truly non-

    invasive mode, by deploying one-sided magnets with built-

    in measurement coils (Blu mich et al., 1998), also denoted as

    the MObile Universal Surface Explorer (MOUSE). The

    MOUSE consists of two separate magnets attached

    antiparallel wise to an iron yoke. Sandwiched between

    these magnets is a simple RF coil, which is used fortransmitting RF pulses and receiving the NMR signal

    (Fig. 1). The first applications of the NMR MOUSE were

    in non-invasive assessment of polymer quality (Guthausen

    et al., 2000), but also the first applications in food

    technology have appeared. These food applications of the

    MOUSE focussed on the non-invasive assessment of

    chemical compositional parameters (Guthausen et al.,

    2004; Pedersen, Ablett, Martin, Mallet, & Engelsen,

    2003). However, the aforementioned sensitivity of NMR

    to distribution and dynamic state of water implies that

    the MOUSE should also be a versatile sensor for the

    microstructural quality of foods (Martinez et al., 2003). In

    this study, we will present a first example where we use the

    MOUSE sensor in this respect.

    For this feasibility study we have taken model systems

    consisting of oil-in-water emulsion gels, stabilized by

    proteins. These systems form complex structures (Haiduc

    & van Duynhoven, 2005; Haiduc, van Duynhoven,

    Heussen, Reszka, & Reiffers-Magnani, submitted; Pudney,

    Hancewicz, Cunningham, & Brown, 2004), built from fat

    droplets and a protein aggregate network in which water is

    dispersed in pores with a range of sizes (schematically

    depicted in Fig. 2). One important quality parameter of

    these materials is their water exudation (WE). In the

    classical approach, WE is determined by measuring theamount of water that is lost from the material when it is

    subjected to gravitational forces. As long as these materials

    are contained, e.g. in a tub, water cannot leave the

    structure and water is distributed homogeneously. How-

    ever, when the material is put in geometry where water can

    leave the structure, exudation takes place, which can take

    between 1 and 8 h. Recently, also a NMR method was

    described that could measure loss of water from a food

    microstructure (Hinrichs, Go tz, & Weisser, 2003), but also

    this method is intrinsically slow. When a more rapidalternative methodology would become available, in a truly

    non-invasive, through-package mode, this would open

    opportunities for process and quality control in manufac-

    turing and the supply chain.

    In our study, we first identified the underlying micro-

    structural basis of WE by conventional benchtop NMR.

    Subsequently we assessed the microstructural quality

    of the model systems by non-invasive, through-package

    MOUSE measurements. We will also outline the require-

    ment of multivariate calibration techniques for exploiting

    experimental data generated by the MOUSE sensor.

    2. Materials and methods

    2.1. Acquisition

    A Bruker Minispec MQ 20 NMR spectrometer (Bruker

    BioSpin, Rheinstetten, Germany), equipped with a variable

    temperature relaxation probe (dead time 7 ms) was used for

    the conventional benchtop NMR experiments. Samples

    were measured in a 10-mm diameter tube, filled to a height

    of 1 cm. Transverse relation decays, consisting of a free

    induction decay (FID) and a subsequent CarrPurcell

    MeiboomGill (CPMG) pulse train (Meiboom & Gill,1958) were recorded, using an in-house pulse programme

    written in the Bruker Exspel programming language. Using

    this combined FID-CPMG method, a large range of

    transverse relaxation times can be assessed in a single

    experiment. The 901801 pulse spacing in the CPMG

    sequence was 0.3 ms. The CPMG decay was sampled by

    averaging 20 points at the top of each echo. The FID part

    of the sequence was sampled with a time interval of 1 ms.

    The decays were measured at 25 1C. All MOUSE NMR

    measurements were performed on a Bruker MQ MiniSpec

    spectrometer that was connected to a separate radio

    frequency (RF) unit and a portable magnet with a built-

    in RF coil (RWTH, Aachen, Germany). In order to get

    meaningful data, a CPMG train of very short RF pulses is

    required. In between these pulses, the NMR signal arises in

    the form of very sharp echoes that are each fitted to a

    polynomial to determine the amplitude. The phase of the

    first five echoes can be determined and the whole echo train

    is then subjected to a so-called zero-order phase correction,

    using this phase information. Now, the imaginary part of

    the complex signal can be disregarded. This improves the

    signal to noise ratio and significantly reduces the baseline.

    These improvements were implemented directly into the

    Bruker control software. Decay signals were recorded

    with the RF pulse spacing varying between 50 and 250 ms.

    ARTICLE IN PRESS

    N

    tub

    S N

    S

    Fig. 1. Schematic presentation of the NMR MOUSE.

    A

    B

    C

    Fig. 2. Schematic representation of oil-in-water emulsion gel stabilized by

    protein: (A) continuous protein gel phase, (B) oil droplets and (C) pores.

    A.M. Haiduc et al. / LWT 40 (2007) 737743738

  • 8/6/2019 Non-invasive through-package assessment of the microstructural

    3/7

    The number of echoes we acquired was always set to 512.

    The number of averages was set to 128. Due to slow T1-

    relaxation, the delay between the individual averages was

    set to 4 s, which yields an experimental time of 8 min for

    one MOUSE measurement.

    2.2. Samples

    A series of 9 protein-stabilized oil-in-water model

    emulsions was prepared, containing vegetable fat, pre-

    dominantly native whey protein concentrate (Nutrilac

    QU7560, ex Arla foods), skimmed milk powder, potassium

    sorbate as a preservative, demineralized water and gelatine.

    An emulsion premix was prepared at 60 1C, subsequently

    heated to 851C and then homogenized. Heating and

    holding steps took 20 min. Homogenization was performed

    using an APV Lab1000 homogenizer, single stage at 300

    bar by default, and is always preceded by turraxing the

    mixture at 8000 rpm for 2 min (Kiokias, Reiffers-Magnani,

    & Bot, 2004). Samples were acidified to a pH in the

    range 4.54.9 using a 50% citric acid solution in

    demineralized water, and homogenized again at 200 bar.

    Emulsions were filled in tubs, sealed, and stored at 5 1C

    until further analysis. The WE of these samples was

    characterized by means of a simple leakage test, and is

    expressed as percentage water lost. The repeatability and

    reproducibility of the leakage test is 1.6% and 2.8%,

    respectively.

    2.3. NMR data analysis

    Partial least squares (PLS) and multilinear regression(MLR) are well known techniques and their theoretical

    background will not be discussed here (Massart, Vande-

    ginste, Buydens, de Jong, & Smeyers-Verbeke, 1997). In

    this work PLS is used directly on the NMR decays. For

    MLR, the decays are first transformed from the continuous

    domain to the discrete domain of T2 distributions and

    amplitudes using the Nonnegative Least Squares (NNLS)

    algorithm (Bro & De Jong, 1997). The decays are averaged,

    and the resultant data is fitted without an initial guess of

    the number of components or T2 values. The basis matrix

    used in the NNLS regression is constructed from the NMR

    CPMG magnetization function:Mt M0 expt=T2 (1)

    where M0 is the amplitude at time t 0, T2 is the

    corresponding relaxation time constant and t is the

    acquisition time axis. M0 is set to 1 and the T2 value is

    varied across the interval 0.23000 ms with an equidistant

    step of 30 ms. In matrix notation, the equation solved with

    NNLS is

    S AK, (2)

    subject to AX0, where S is the experimental NMR decay

    matrix, and A are the unknown amplitudes of the

    exponential basis function K. The CPMG part follows a

    sum of exponential decays fitted with NNLS:

    St X

    Mn expt=T2n; n 1 to N, (3)

    where Mn

    is the amplitude of the nth exponential, T2n is the

    corresponding relaxation time constant, t is the acquisition

    time axis and N is the number of exponential functions or

    components in the sample. The sum SMn is smaller thanthe normalized value, the difference is due to components

    with very short T2s for which the magnetization decays to

    zero in the FID part.

    For MLR the following calibration equation is used:

    P B0CX

    BnMn, (4)

    where P is the calibrated parameter, B0 is the MLR

    constant and the unit constant C the sum of all

    magnetizations. This equation can be rewritten as

    P B0M0 X

    Mn X

    BnMn B0M0

    X

    B0 BnMn. 5In this work, the term B0 Bn is denoted as scaled

    coefficients.

    2.4. Monte-Carlo analysis

    The value of the amplitudes obtained by a NNLS is

    influenced by the noise on the experimental decays. This is

    a well-known problem in numerical inversion and the usual

    work-around is by the use of a regularization parameter

    (Butler, Reeds, & Dawson, 1981; Provencher, 1982).

    However, a regularized fit leads to a continuous distribu-

    tion of T2s over a wide range, which would impede the useMLR for performing a calibration. Furthermore, in

    regularized fitting approaches small disturbances in the

    decays are known to introduce relatively large artefacts in

    the T2 distributions. Hence, in this work, it was chosen not

    to deploy regularization. As an alternative, the influence of

    the noise on the NNLS result was tested using a Monte-

    Carlo approach. For each recorded decay, the difference

    between the NNLS fit and the experimental data (residuals)

    is assumed to be due to normally distributed noise.

    Simulated noise is constructed using a random number

    generator with the same variance as the residuals and then

    added to the NNLS fit, thus simulating a new experiment

    on the same samples. The resultant new decays are then

    fitted again. The process is repeated 100 times, resulting in

    a distribution of amplitude values which can be used to

    derive statistical parameters.

    2.5. Software

    The commercial software package SIMCA-P (Umetrics,

    Umea, Sweden) was used for (PLS) model development.

    Orthogonal Signal Correction (Wold, Antti, Lindgren, &

    Ohman, 1998) was applied in order to yield a PLS

    model with a single principal component. Data pre-

    processing was performed in EXCEL and MATLAB

    ARTICLE IN PRESS

    A.M. Haiduc et al. / LWT 40 (2007) 737743 739

  • 8/6/2019 Non-invasive through-package assessment of the microstructural

    4/7

    (The MathWorks, Natick, MS, USA), using in house

    developed routines. Transformation of the NMR decays to

    discrete distributions of transverse relaxation times (T2) by

    means of NNLS and MLR of the resulting amplitudes

    versus functional material parameters was achieved using

    Matlab routines developed in house.

    3. Results and discussion

    3.1. Inspection of conventional and MOUSE NMR decays

    Raw NMR decays obtained by conventional benchtop

    NMR and the MOUSE are presented in Fig. 3. Whereas

    both types of signal were recorded within comparable

    measurement times (ca. 8 min), they differ significantly with

    respect to signal-to-noise ratio. Furthermore, the MOUSE

    signals decay much more rapidly. In order to extract

    common T2 populations in the NMR decays, a NNLS

    transformation was carried out (Bro & De Jong, 1997). The

    NNLS procedure was performed on decays recorded by

    conventional (benchtop) means, and by the MOUSE.

    The resulting T2 distributions are presented in Fig. 4.

    In the NMR decays that were measured by conventional

    means, three T2 components can be discerned, which

    can be assigned (Haiduc & van Duynhoven, 2005) to (I)

    oil (E170 ms), (II) oil and protein-associated water

    (E200 ms) and (III) water in pores (4500 ms). For the

    NMR MOUSE, two T2 population distributions are

    obtained, centred around 10 and 70 ms. The T2 populations

    as measured by the MOUSE cannot be related to the ones

    assessed by regular benchtop NMR in a straightforward

    manner. The strong magnetic field inhomogeneities that are

    inherent to the one-sided MOUSE magnets lead to strong

    diffusional contributions to the transverse relaxation time.

    This is particularly effective for rapidly diffusing watercomponents and less for slowly diffusion oil. From this, we

    may assume that the component at 10 and 70ms can

    (roughly) be assigned to water and oil, respectively.

    3.2. Prediction of WE from NMR decays recorded by

    conventional NMR and the MOUSE

    Since conventional benchtop NMR decays allow for

    physical interpretation, these were first used for explorative

    multivariate modelling. The most straightforward method

    to build such a model is by means of PLS. PLS has the

    advantage that it can predict WE directly from the

    continuous NMR exponential decays. As is shown in

    Table 1A, this results in a good predicting model for WE.

    The PLS model also generates a so-called loading which

    represents the signal in the NMR decays that is responsible

    for the relation between decays and WE. This signal is

    presented in Fig. 5A, but is difficult to be interpreted in a

    physical manner. This is due to use of centred and scaled

    data in the PLS model. This procedure has the advantage

    that it avoids extracting a very strong component which is

    just an average NMR decay and has no relation with WE.

    The disadvantage, however, is that when dealing with

    continuous NMR decays, a loading is obtained with both

    positive and negative amplitudes. This complicates the

    ARTICLE IN PRESS

    0

    0.2

    0.4

    0.6

    0.8

    1.0

    0.1 1 10 100 1000

    Time [ms]

    0

    0.2

    0.4

    0.6

    0.8

    1

    0.1 1 10 100 1000

    Time [ms]

    (B)

    (A)

    Fig. 3. Raw NMR decays obtained by (A) conventional benchtop time-

    domain NMR and (B) the MOUSE. Decays were normalized with respect

    to intensity at 0.1 ms and time scale (horizontal) is plotted logarithmically.

    0 150 300 450 600 750 900 1050 1200 1350 1500

    time [ms]

    Amplitude

    Amplitude

    I

    II

    III

    (A)

    (B)

    Fig. 4. T2 distribution obtained by NNLS fitting of CPMG data obtained

    by conventional benchtop NMR (A) and the NMR-MOUSE (B). The

    assignments of T2 populations IIII in (A) are outlined in the text.

    A.M. Haiduc et al. / LWT 40 (2007) 737743740

  • 8/6/2019 Non-invasive through-package assessment of the microstructural

    5/7

    straightforward explanation of WE in terms of micro-

    structural features, i.e. T2 populations.

    In order to obtain calibration models that allow for

    physical interpretation, another approach was explored

    where MLR is applied to NNLS signal amplitudes

    extracted from decays measured by conventional means.

    The MLR model is built using the NNLS amplitudes asresponses, and the functional parameters as predictors. The

    quality of the NNLSMLR models is illustrated by the plot

    in Fig. 6A. MLR is performing similar, or better, as PLS as

    is illustrated in Table 1A. The MLR model has the

    additional advantage that these can be interpreted in a

    physical manner (Haiduc & van Duynhoven, 2005). The

    coefficients of the MLR establish a direct linear relation

    between the amplitudes of the three T2 populations and the

    calibrated parameter. The relaxation times T2 for thesethree populations relate to different microstructural

    features of the emulsion, thus enabling physical interpreta-

    tion. The scaled MLR coefficients in Fig. 5B represent the

    physical relation between microstructural features (in the

    form of T2 populations) and the calibrating (functional)

    parameters P. In Fig. 5B one can observe that WE is

    dependent on population I and (mostly) on pore popula-

    tion III. The large T2 value associated with Population III

    point towards water confined in relatively large pores in the

    continuous phase of the emulsion. The water in these pores

    is more prone to drain from the emulsion than other

    populations.

    The results of both PLS and MLR demonstrate that a

    microstructure related quality property can be predicted

    from decays recorded on benchtop NMR equipment.

    The MLR model is not simply a multivariate black box,

    it also allows for building understanding of WE in

    ARTICLE IN PRESS

    -0.1

    -0.05

    0

    0.05

    0.1

    0 500 1000 1500 2000 2500 3000

    Time [s]

    Am

    plitude

    -50

    0

    50

    100

    150

    200

    250

    0 150 300 450 600 750 900 1050 1200 1350 1500

    Time [ms]

    Amp

    litude

    (A) (B)

    Fig. 5. (A) PLS loading (coefficients) for calibration of conventional benchtop NMR decays with Water Exudation. (B) Scaled coefficients for an MLR

    model that predicts Water Exudation based on conventional benchtop decays.

    Table 1

    Goodness of fit (R2) and root mean square errors (RMSE) of calibration

    (Cal) and cross validation (xVal) for PLS and MLR models constructed

    from (A) conventional benchtop and (B) MOUSE decays

    A B

    Conventional MOUSE

    PLS MLR PLS MLR

    R2 0.93 0.96 0.78 0.97

    RMSE-Cal 1.17 1.13 3.05 0.99

    RMSE-xVal 1.57 1.71 3.42 1.99

    Values in bold indicate strong models.

    0

    5

    10

    15

    20

    25

    0 5 10 15 20 25

    Predicted Water Exudation [%]

    MeasuredWaterExudatio

    n[%]

    0

    5

    10

    15

    20

    25

    0 5 10 15 20 25

    Predicted Water Exudation [%]

    MeasuredWaterExudatio

    n[%]

    (A) (B)

    Fig. 6. Goodness of fits of NNLS-MLR models based on (A) conventional benchtop NMR decays and (B) MOUSE decays.

    A.M. Haiduc et al. / LWT 40 (2007) 737743 741

  • 8/6/2019 Non-invasive through-package assessment of the microstructural

    6/7

    microstructural terms. The aforementioned PLS and MLR

    approaches were also applied to the NMR decays recorded

    with the MOUSE. There is a small difference between the

    calibration and cross-validation errors of the MOUSE PLS

    model (Table 1B), indicating that the model is not

    overfitted. Compared to the model based on conventional

    benchtop NMR (Table 1A), the MOUSE PLS model is less

    favourable. This is because PLS needed only one statisti-

    cally relevant component in the MOUSE model, whereas

    three components were needed to describe the conventional

    benchtop NMR data. Much more promising are the MLR

    results obtained after the NMR MOUSE decays were

    transformed with NNLS into sets of amplitudes and

    observed T2* values (Fig. 4B). The quality of the MOUSE

    MLR model for WE (Fig. 6B) is comparable to the results

    from the conventional benchtop NMR experiments

    (Fig. 6A). We note that only two NNLS amplitudes

    were used in the MLR model, hence there is a minimal

    risk of overfitting the data. This is also illustrated by

    the relative small difference between calibration and cross-

    validation errors. MLR model errors for the MOUSE

    data are larger than for conventional benchtop NMR

    (Table 1B), which can be attributed to the poorer

    signal to noise ratio of the MOUSE data. This will havean impact on the value of the amplitudes obtained by a

    NNLS fit to the exponential decays. The influence of the

    noise on the amplitudes was tested using a Monte-Carlo

    approach. The result for the WE model shows that this

    parameter is independent of the noise (Table 2), which on

    average has a variance (square of standard deviation) of

    7105 for the NMR-MOUSE and 2106 for a

    conventional benchtop NMR experiment. Simulated

    data showed that the NNLS fit becomes unstable when

    the noise variance reaches 5 104, almost an order of

    magnitude higher than in our experiments. Hence, we

    conclude that MOUSE decays can be used to make reliable

    predictions of WE.

    This work demonstrates that low field NMR relaxome-

    try can be deployed to assess the microstructural quality of

    food products, by both conventional benchtop NMR as

    well as the NMR MOUSE. The models constructed from

    conventional benchtop NMR data, have the advantage

    that they allow for the understanding of WE in micro-

    structural terms. The NMR MOUSE data are more

    difficult to be interpreted, but allow for assessment of

    WE in a truly non-invasive (through package) mode. This

    may offer promising options for non-invasive assessment

    of other microstructural quality parameters, i.e. in manu-

    facturing and in the supply chain.

    Acknowledgements

    We acknowledge the European Union (Marie Curie

    Fellowship, HPMI-CT-2002-00167, Benchtop NMR tech-

    niques for the structural characterization of foods) for the

    funding of A.H. and E.T.

    References

    Bamberger, J. A., & Greenwood, M. S. (2004). Non-invasive character-

    isation of fluid foodstuffs based on ultrasonic measurements. Food

    Research International, 37(6), 621625.

    Blu mich, B., Blu mler, P., Eidmann, G., Guthausen, A., Haken, R.,

    Schmitz, U., et al. (1998). The NMR-MOUSE: Construction,

    excitation and applications. Magnetic Resonance Imaging, 16(5/6),

    479484.

    Bro, R., & De Jong, S. (1997). A fast non-negativity-constrained least

    squares algorithm. Journal of Chemometrics, 11, 392401.

    Butler, J. P., Reeds, J. A., & Dawson, S. V. (1981). Estimating solutions

    of first kind integral equations with nonnegative constraints and

    optimal smoothing. SIAM Journal of Numerical Analysis, 18(3),381397.

    Guthausen, A., Guthausen, G., Kamlowski, A., Todt, H., Burk, W., &

    Schmalbein, D. (2004). Measurement of fat content with single

    sided NMR. Journal of the American Oil Chemists Society, 81,

    727731.

    Guthausen, G., Guthausen, A., Balibanu, F., Eymael, R., Hailu, K.,

    Schmitz, U., et al. (2000). Soft matter analysis by the NMR-MOUSE.

    Macromolecular Material Engineering, 276, 2537.

    Haiduc, A. M., & van Duynhoven, J. P. M. (2005). Correlation of

    porous and functional properties of food materials by NMR

    relaxometry and multivariate analysis. Magnetic Resonance Imaging,

    23(2), 343345.

    Haiduc, A.M., van Duynhoven, J.P.M., Heussen, P.C.H., Reszka, A.A.,

    Reiffers-Magnani, C., Multivariate modelling of the microstructural

    quality of food emulsions based on NMR, Submitted to FoodResearch International.

    Hinrichs, R., Go tz, J., & Weisser, H. (2003). Water-holding capacity and

    structure of hydrocolloid-gels, WPC-gels and Yoghurt characterised

    by means of NMR. Food Chemistry, 82(1), 155160.

    Kiokias, S., Reiffers-Magnani, C. K., & Bot, A. (2004). Stability of whey

    protein stabilized oil-in-water emulsions during chilled storage and

    temperature cycling. Journal of Agriculture and Food Chemistry, 52,

    38233830.

    Martinez, I., Aursand, M., Erikson, U., Singstand, T. E., Veliyulin, E., &

    van der Zwaag, C. (2003). Destructive and non-destructive analytical

    techniques for authentication and composition analysis of foodstuffs.

    Trends in Food Science Technology, 14, 489498.

    Massart, D. L., Vandeginste, B. G. M., Buydens, L. M. C., Jong, S. de, &

    Smeyers-Verbeke, J. (1997). Handbook of chemometrics and quali-

    metrics: Part A, B. Book Series: Data Handling in science andtechnology, vol. 20B. Amsterdam: Elsevier.

    Meiboom, S., & Gill, D. (1958). Modified spin-echo method for measuring

    nuclear relaxation times. Reviews of Scientific Instrumentation, 29,

    688691.

    Pedersen, H. T., Ablett, S., Martin, D. R., Mallet, M. J. D., & Engelsen, S.

    B. (2003). Application of the NMR MOUSE to food emulsions.

    Journal of Magnetic Resonance, 165, 4958.

    Prakash, M. N. K., & Ramana, K. V. R. (2003). Ultrasound and its

    application in the food industry. Journal of Food Science and

    Technology, 40(6), 563570.

    Provencher, S. (1982). CONTIN: A general purpose constrained

    regularization program for inverting noisy linear algebraic and integral

    equations. Computer Physics Communications, 27, 213227.

    Pudney, P. D. A., Hancewicz, T. M., Cunningham, D. G., & Brown, M.

    C. (2004). Quantifying the microstructures of soft materials

    ARTICLE IN PRESS

    Table 2

    Monte-Carlo simulation results on the MLR regression coefficient (R2)

    for water exudation

    R2 original 0.970

    R2 average from MC simulation 0.971

    R2 95% confidence 0.9680.973

    A.M. Haiduc et al. / LWT 40 (2007) 737743742

  • 8/6/2019 Non-invasive through-package assessment of the microstructural

    7/7

    by confocal Raman spectroscopy. Vibrational Spectroscopy, 34,

    123135.

    Ross, K. A., Pyrak-Nolte, L. J., & Campanella, O. H. (2004). The use of

    ultrasound and shear oscillatory tests to characterize the effect of

    mixing time on the rheological properties of dough. Food Research

    International, 37(6), 567577.

    Senorans, F. J., Ibanez, E., & Cifuentes, A. (2003). New trends in food

    processing. Critical Reviews in Food Science and Nutrition, 43(5),507526.

    Todt, H., Burk, W., Guthausen, G., Kamlowski, A., & Schmalbein, D.

    (2001). European Journal of Lipid Science and Technology, 103(12),

    835840.

    Williams, P., & Norris, K. (Eds.) (2003). Near-infrared technology in the

    agricultural and food industries. Am. Assoc. Cereal Chem., St. Paul,

    MN, USA.

    Wold, S., Antti, H., Lindgren, F., & Ohman, J. (1998). Orthogonal signal

    correction of near-infrared spectra. Chemometrics and IntelligentLaboratory Systems, 44(12), 175185.

    ARTICLE IN PRESS

    A.M. Haiduc et al. / LWT 40 (2007) 737743 743