Magnetic Resonance Im

Texture analysis of MR image for predicting the firmness of Huanghua

pears (Pyrus pyrifolia Nakai, cv. Huanghua) during storage using an

artificial neural network

Ran Zhoua, Yunfei Lib,4aInstitute of Refrigeration and Cryogenic Engineering, Shanghai Jiao Tong University, 201101 Shanghai, PR China

bDepartment of Food Science and Technology, Shanghai Jiao Tong University, 201101 Shanghai, PR China

Received 6 July 2006; accepted 27 September 2006

Abstract

Firmness, a main index of quality changes, is important for the quality evaluation of fruits. In the present study, texture analysis (TA) of

magnetic resonance images was applied to predict the firmness of Huanghua pears (Pyrus pyrifolia Nakai, cv. Huanghua) during storage

using an artificial neural network (ANN). Seven co-occurrence matrix-derived TA parameters and one run-length matrix TA parameter

significantly correlated with firmness were considered as inputs to the ANN. Several ANN models were evaluated when developing the

optimal topology. The optimal ANN model consisted of one hidden layer with 17 neurons in the hidden layer. This model was able to predict

the firmness of the pears with a mean absolute error (MAE) of 0.539 N and R=0.969. Our data showed the potential of TA parameters of MR

images combined with ANN for investigating the internal quality characteristics of fruits during storage.

D 2007 Elsevier Inc. All rights reserved.

Keywords: Magnetic resonance imaging; Texture analysis; Artificial neural network; Huanghua pears; Firmness

1. Introduction

Huanghua pears (Pyrus pyrifolia Nakai, cv. Huanghua)

are widely planted in the south of China. Because of their

acute physiological changes after harvest, Huanghua pears

can easily lose their firmness during storage [1]. In fact, the

softening of fruits during storage, which results from

changes in the structure of fruit tissue, is reflected in the

changes in the status of water in the tissue [2,3]. With the

development of the technique of magnetic resonance

imaging (MRI), interest in the application of MRI for

determining quality attributes like firmness and solubility of

solids (a factor related to water status in the MRI) of

agricultural products is increasing. Several kinds of fruits

and vegetables such as apples [4], pears [5], citrus [6],

potatoes [7] have been studied using this noninvasive and

nondestructive method.

MRI has the ability to provide highly resolved spatial

information about anatomical interior and water distribution

in plant tissues [8,9]. 1H MRI is the most straightforward,

0730-725X/$ – see front matter D 2007 Elsevier Inc. All rights reserved.

doi:10.1016/j.mri.2006.09.011

4 Corresponding author. Tel.: +86 21 64783085; fax: +86 21

64783085.

E-mail address: yfli@sjtu.edu.cn (Y. Li).

accessible and sensitive MRI probe for highly resolved

spatial information concerning the distribution and magnetic

environment of water in biologic tissues [3]. Due to its

sensitive nature, it can provide visualization of even minute

localized morphological changes in intact plant tissues [10].

Moreover, MRI has been used to quantitatively assess and

study physiological changes like firmness and solubility of

solids in fruit during maturation and storage as a function of

the relaxation times [3,4,11].

Compared to the traditional MRI method for predicting

the firmness of fruits during storage, in this contribution, the

application of two techniques is reported: texture analysis

(TA) of MR images, which is commonly used for the

quantification of medical images [12], and artificial neural

networks (ANNs) to relate the TA parameters to fruit

firmness. The TA parameters can be used to describe the

density of pixels and their variability in levels of grey tones

that are inaccessible to the human eye during observation

[13]. Also, some of the TA parameters were found to be

significantly correlated with apple firmness in the case of

maturation [2]. However, little is known of the application

of TA parameters and ANNs to the prediction of firmness of

fruits during storage.

aging 25 (2007) 727–732

Fig. 1. Equatorial transverse T2-weighted MR image of Huanghua pears.

R. Zhou, Y. Li / Magnetic Resonance Imaging 25 (2007) 727–732728

The ANN used in this study is a data-processing

technique inspired by studies in neuroscience. Many single

elements called neurons are connected to each other in

different ways, thus creating different kinds of ANNs. Each

neuron consists of a transfer function expressing internal

activation level. The most popular transfer function is log-

sigmoid for nonlinear relationship [14]. The rationale for

attempting to apply an ANN in relating TA parameters to

firmness of pears during storage is based on the notion that

an ANN has the capability to represent the linear and

nonlinear relationship by learning by experience when no

exact mathematical relationship is available [15,16]. The

objective of this study was to evaluate the potential of TA

of MR images and an ANN in predicting the firmness of

Huanghua pears during storage.

Table 1

Firmness values of Huanghua pears used as output for ANN

Time (days) 48C CA

9 31.96F1.65a 31.69F1.46a

27 29.92F1.45b 30.67F1.39ab

40 28.71F1.71b 29.97F1.96b

54 28.93F2.24b 30.72F2.67ab

Mean values (N) of firmness of Huanghua pears with S.D. for each testing

time were computed. Time indicates days of storage.

Values within a column followed by the same letter were not significantly

different ( P b.05).

2. Material and methods

2.1. Fruit material

Huanghua pears at commercial maturity, according to the

skin color of the fruit, were harvested from a commercial

orchard in Fengxian, Shanghai, China. The pears were

selected on the basis of uniform color and absence of bruises

and disease. All of the pears (about 120 kg) were

transported to the laboratory within 2 h. According to the

commonly used storage method for Huanghua pears, the

fruits were divided into two groups: the first was stored

under 48C and the second was maintained at controlled

atmospheric (CA) condition (6% O2 and 2% CO2 at 48C) for2 months. MRI techniques were performed on the 9th, 27th,

40th and 54th day after storage. Five fruits were selected

randomly from each group and a number was affixed to

their stem. Then, the selected pears were transported in a

refrigerated box to the MRI laboratory about 2 h prior

to measurements.

2.2. Magnetic resonance imaging

The experiments were performed on a whole-body 1.5-T

MRI scanner (General Electric, Waukesha, WI, USA)

using the conventional head coil in the First Hospital of

Shanghai Jiao Tong University. A cross was first drawn on

the equatorial region of each fruit. Then, two pears were

tested together to save examination time; their equatorial

regions were about on the same plane and fixed by

adhesive tape. Each of the pears was packaged by a plastic

net to prevent the tape from sticking to the fruit’s skin. The

following parameters of a T2-weighted sequence were used

[2]: TR=3500 ms; TE=62.7 ms; number of acquisi-

tions=4; FOV=20 cm; slice thickness=1 mm; 256�256 pixels of the imaging matrix (pixel size=0.78 mm);

and three image slices with a 3-mm gap in between. The

middle image slice passing through the equatorial regions

of the pears in the transverse plane at each imaging session

was selected for further studying (see Fig. 1). Then, the

equatorial images were converted into bitmap format on

the MRI spectrometer and analyzed with MaZda software

ver. 3.20 [17].

2.3. Firmness

When the MRI experiment was completed, the 10 pears

were transported back into the refrigerated box. Then,

firmness was measured using a TA-XT2i texture analyzer

(Stable Micro Systems, Surrey, UK) with a 5-kg load cell

and a 2-mm-diameter cylinder probe. The test was

performed with a pretest and test speed of 5 mm/s, a post-

test speed of 10 mm/s, and auto-25 g trigger force. Firmness

was measured on three sides at intervals of 1208 of each

fruit at the equatorial region. So there were 5�3 examina-

tions per group. The firmness values of the pears during

storage are shown in Table 1.

2.4. Texture analysis and statistics

Three regions of interests (ROIs) on the equatorial MR

image of each pear were carefully located on three sides at

intervals of 1208 within the area of the flesh of the fruit.

Each ROI had 700 pixels (each pear area had 8000–9000

pixels depending on the size). Texture analyses were

performed using the MaZda software. Two kinds of TA

Table 2

TA parameters used as inputs for ANNs

Time 1 2 3 4 5 6 7 8

S (0,1)

contrast

S (0,1)

correlation

S (0,1) difference

variance

S (0,1) difference

entropy

S (1,1)

correlation

S (0,3)

correlation

S (5,�5)correlation

458fraction

9 48C 10.58F2.86 0.42F0.08 4.29F1.21 0.86F0.04 0.25F0.09 0.17F0.07 �0.01F0.08 0.89F0.01

CA 10.89F3.11 0.45F0.07 4.48F1.31 0.86F0.05 0.29F0.12 0.17F0.10 �0.07F0.07 0.89F0.01

27 48C 10.81F6.01 0.66F0.06 4.64F2.82 0.84F0.10 0.45F0.08 0.25F0.12 0.08F0.14 0.89F0.02

CA 4.40F1.50 0.70F0.10 1.76F0.59 0.68F0.07 0.48F0.17 0.34F0.15 0.08F0.13 0.85F0.03

40 48C 6.23F1.74 0.67F0.06 2.82F0.89 0.75F0.05 0.43F0.10 0.27F0.08 0.06F0.11 0.86F0.02

CA 2.41F0.56 0.66F0.04 1.03F0.22 0.58F0.04 0.37F0.09 0.36F0.07 0.06F0.11 0.81F0.02

54 48C 9.99F1.77 0.61F0.07 4.13F0.94 0.84F0.03 0.52F0.06 0.23F0.07 0.13F0.12 0.88F0.01

CA 7.84F2.64 0.62F0.10 3.21F1.15 0.79F0.06 0.35F0.15 0.23F0.17 0.08F0.09 0.89F0.02

Correlation coefficients 0.21 �0.37 0.20 0.19 �0.31 �0.26 �0.21 0.18

Significance level 4 44 4 4 44 44 4 4

Mean values of TA parameters with S.D. for each testing time were computed.

Correlation coefficients indicate Pearson correlation coefficients between the TA parameter and firmness of pears during storage.

Parameters 1–7 are co-occurrence matrix-derived parameters. Parameter 8 is a run-length matrix parameter. Full definitions of these parameters were given by

Lerski et al. [13].

4 Significance level: P b.05.

44 Significance level: P b.01.

R. Zhou, Y. Li / Magnetic Resonance Imaging 25 (2007) 727–732 729

parameters were used: first-order and second-order param-

eters [13]. The texture parameters are listed below:

Histogram parameters (mean, variance, skewness, kurtosis)

Absolute gradient parameters (absolute gradient mean,

variance, skewness, kurtosis, percentage of pixels with

nonzero gradient)

Run-length matrix parameters (horizontal, vertical, 458and 1358 run-length and grey level nonuniformity, long-

run and short-run emphasis, fraction of image in runs)

Co-occurrence matrix parameters (distances d=1, 3 and

5 pixels with angles=08, 458, 908 and 1358 were

considered — a total of 132 parameters)

The total number of TA parameters of one ROI was 161.

The matrix of TA parameters and their corresponding

firmness was obtained with 161 columns and 120 rows

(15 replications for the two groups from four periods of

Fig. 2. Basic structure of back-propagation feed-forward ANN models used in thi

connection. The dot in the hidden layer represents the omitted neurons. Eight TA p

firmness information. The number of neurons in the hidden layer varied from 2 t

measurement). Pearson’s correlation coefficients were com-

puted between the TA parameters and firmness of the pears

using SAS 8.0. The results of correlation analysis are shown

in Table 2. Then, eight TA parameters that were signifi-

cantly correlated with firmness of the pears were used as

input for the ANN.

2.5. Artificial neural network model

Our study obtained a total data set of 120 cases. The data

set was divided into two independent groups: one group

consisting of 90 cases (including results from two storage

groups (48C and CA) examined three times) was considered

as the training data set; and another consisting of 30 cases

as the validation data set (the remaining results from two

storage groups (48C and CA) during storage). Each of the

cases consisted of two parts: input (TA parameters) and

output (firmness). The numerical values of the input and

s study. Each circle represents a neuron, and each line represents a synaptic

arameters were used to input data and one neuron in the output layer to yield

o 20.

Table 3

Mean absolute error and training iteration of different ANN models with

one hidden layer

Neurons of

hidden layer

MAE Number of iteration

12 0.121 7.66�10513 0.126 5.99�10514 0.133 5.14�10515 0.123 3.63�10516 0.127 6.31�10517 0.120 3.29�10518 0.137 6.36�10519 0.137 4.18�10520 0.130 3.92�105

Fig. 3. Correlation between predicted and experimental values of firmness

of pears using the optimal network during training. The optimal network

included one hidden layer with 17 neurons. The training data set included

90 cases.

R. Zhou, Y. Li / Magnetic Resonance Imaging 25 (2007) 727–732730

output variables used by the ANN were normalized to the

range of 0 and 1. Input vectors were normalized to avoid

numerical overflows due to very large or very low weights;

output vectors were normalized to make them lie in the range

of the output transfer function used by the ANN [18]. The

normalization variable xnorm is represented as follows:

xnorm ¼x� xmin

xmax � xmin

ð1Þ

where x is an input or output variable, and xmax and

xmin are the maximum and minimum value of the vari-

able, respectively.The MATLAB toolbox and environment were used in

Windows XP system.MATLAB software (MATLAB version

7.01; the MathWorks, Inc., Natick, MA, USA) and its Neural

Network Toolbox were used. A feed-forward neural network

with back-propagation algorithm was designed. The training

data set was used to train the weights and biases of ANNs

with log-sigmoid transfer functions. It used eight neurons in

an input layer, corresponding to the eight TA parameters

listed in Table 2, and an output layer having one neuron

representing the corresponding firmness value. The basic

structure of the ANN model is shown in Fig. 2.

The number of hidden layers was influenced by the

complexity of the problem and size of the data. It had been

reported that one hidden layer was sufficient to approximate

any continuous nonlinear function [16,19], even thoughmore

complex networks could be used in special applications [16].

However, more hidden layers could result in over-fitting,

since the network focused excessively on the characteristics

of individual samples [18,20]. Additionally, there is no theory

yet that can tell how many neurons in a hidden layer are

needed to approximate any given function. In this study, a

neural network model with one hidden layer was used. The

number of neurons in the hidden layer varied from2 to 20with

increments of 1. This led to a total of 19 networks.

In the training process, the sum square error (SSE) was

used as an index of the learning efficiency of all the networks.

The training algorithm minimized the SSE between the

desired and actual network output following an iterative

gradient search until the SSE converged at the level of the

error goal [21]. In this study, the error goal was defined to be

equal to 0.02. The selection of a learning rate was important

for the rate of change of connection weights during training.

In this study, the learning rate was fixed at 0.1. A low learning

rate was selected because high fluctuations in errors during

the training process took place when a higher learning rate

was used.

2.6. Selection of optimal ANN configuration

The performances of the various ANN models were

compared using mean absolute error (MAE) [22].

The error function was defined as

MAE ¼ 1

n

Xn

i¼1jYP � YDj ð2Þ

where YP represents the ANN output for a given input and

YD is the desired output for the same input. n is the number

of data points.

3. Result

3.1. Performances of the developed ANN models

Each of the feed-forward ANNs was trained using

90 cases. The number of neurons in hidden layers of the

ANNs varied from 2 to 20. The optimization process of

ANNs lay in selecting an adequate number of neurons in the

hidden layer. The reason is that, with too few nodes, the

network did not have enough power to finish the learning task

[16]. In this study, the SSEs of ANNs with 2 to 11 neurons in

hidden layers did not converge at the error goal after 1�106training epochs. Since the main goal was to find the optimal

network having the best performance, these ANNs were

directly rejected. Another nine ANNs with 12–20 hidden

neurons had been considered to select the optimal topology

and are shown in Table 3. The criterion consisted of selecting

the optimal ANN that gave a minimum final error in a

Fig. 4. Correlation between predicted and experimental values of firmness

of pears using the optimal network during validation. The optimal network

included one hidden layer with 17 neurons. The verification data set

included 30 cases. MAE=0.539 N.

able 5

tatistical variables of experimental and predicted values (validation step)

tatistical values Experimental values Predicted values

ean 29.83 30.09

edian 29.43 29.87

tandard error 0.49 0.45

tandard deviation 2.67 2.47

oefficient variation 8.94 8.21

kewness 0.58 0.42

urtosis �0.52 �0.70

R. Zhou, Y. Li / Magnetic Resonance Imaging 25 (2007) 727–732 731

minimum number of iterations during the training process.

According to the table, the topology 8, 17, 1 (8 nodes in the

input layer, 17 neurons in the hidden layer and 1 neuron in the

output layer) yielded the minimum MAE at the lowest

iteration and was thus selected as the optimal ANN model.

The object of the validation step was to evaluate the

capability of the developed network. In this step, 30

independent cases were used for the validation process.

The performance of the optimal topology 8, 17, 1 during

training and validating process is shown in Figs. 3 and 4.

The correlation coefficient for the training process was

0.998 and for the validating process 0.969. The MAE

between the predicted data set and the validation data set

was equal to 0.539 N during the validation process (Fig. 4).

Moreover, since the validation data included 15 cases for the

pears from the first group and 15 cases for the pears from

the second group, the two subgroups were still respectively

analyzed. The correlation coefficients were 0.942 for the

pears from the first group (MAE=0.715 N) and 0.986 for the

pears from the second group (MAE=0.362 N). These results

could be considered as accurate enough for the ANN model

to predict the firmness of Huanghua pears during storage.

3.2. Statistical analysis

Tables 4 and 5 show some statistical properties of the

experimental firmness values of Huanghua pears and the

Table 4

Statistical variables of experimental and predicted values (training step)

Statistical values Experimental values Predicted values

Mean 30.49 30.61

Median 30.39 30.52

Standard error 0.21 0.21

Standard deviation 1.97 1.96

Coefficient variation 6.46 6.40

Skewness �0.09 �0.09Kurtosis �0.04 �0.13

T

S

S

M

M

S

S

C

S

K

corresponding predicted values in the training step and

validation step of the selected topology (8, 17, 1).

According to Table 4, the values and the distribution of

experimental and predicted firmness were similar during the

training process. Table 5 shows that the statistical param-

eters of ANN predictions were closer to those of the

experimental values for the validation set.

Two kinds of statistical analyses were still performed to

determine whether there were significant differences be-

tween firmness values provided by the experimental step

and those predicted by the ANN during the validation step.

These were based on the difference in the mean and S.D.

between the experimental and predicted values. The null

hypothesis assumed that the two parameters (mean, S.D.) of

both groups were equal. Otherwise, an alternative hypoth-

esis was defined. The threshold of the P value for each

hypothesis was 0.05. A Student t-test was used to determine

whether the mean values of both groups were significantly

different. The obtained P value was 0.35, which was much

higher than the threshold value, 0.05. Moreover, the two

subgroups of the validation cases were still respectively

analyzed. The P values were 0.30 for the pears from the first

group and 0.47 for the pears from the second group, which

were still higher than 0.05. Thus, the null hypothesis cannot

be rejected. The mean values of the experimental and

predicted values were not significantly different.

The difference in S.D. of both groups was analyzed using

the F test. The P value was equal to 0.34 for the whole

validation cases. Also, the P values of the two subgroups

were computed: 0.46 for the first group and 0.38 for the

second group. None of these P values was lower than 0.05.

Hence, the S.D. of the experimental and predicted values did

not differ significantly.

4. Discussion and conclusions

For the detection of fruit firmness by MRI, previous

studies mainly focused on studying the correlations between

relaxation measurements and those derived from physical

methods of firmness testing [3]. Texture analysis of MR

images provided a new notion about characterizing this kind

of physiological stage of Huanghua pears. Moreover,

according to the results, the successful application of ANN

combined with TA parameters in determining the firmness of

pears provides MRI more insight into the development of

physiological changes in fruit and vegetables.

R. Zhou, Y. Li / Magnetic Resonance Imaging 25 (2007) 727–732732

In fact, degradation of the cell wall and decomposition of

the polysaccharides generally lead to changes in firmness

during storage in climacteric fruits, such as Huanghua pears

[1,23]. MR images are complicated by such subcompart-

ments of plant cells as the cell wall, the cytoplasm, the

vacuole and the exchange rates between these compartments

[24], which reflect the water status of plant tissues. TA

parameters of T2-weighted MR images are important in

representing the physical status of water molecules related

to the firmness of fruits, even though only a few of the TA

parameters take effect [2]. Therefore, to apply TA param-

eters in predicting pear firmness, it is necessary to find out

the parameters that are significantly correlated with changes

in fruit firmness during storage.

In this contribution, numerical analysis was used to

extract information on the changes in macroscopic structure

from MR images of the Huanghua pears. The relative order

of importance of various TA parameters in determining the

firmness of pears was determined through Pearson’s

correlation analysis. According to Table 2, there were seven

co-occurrence matrix-derived parameters and one run-length

matrix parameter which were significantly correlated with

the firmness values of pears during storage. Use of these

parameters produced a more reliable input set selection for

the ANN design by rejecting uncorrected TA parameters.

According to the results, the optimal model, whose hidden

layer consisted of 17 neurons, was able to predict the

firmness values of pears stored in two kinds of conditions

(48C and CA) with enough accuracy. The validity of the

ANN had been tested with very good results for unused

data. Additionally, the statistical results confirmed that the

predicted firmness values of the ANN model were similar to

the experimental results. Therefore, the ANN model in this

study became an effective tool by using TA parameters in

the prediction of firmness changes of pears during storage.

In summary, the main innovations of this work lay in the

use of appropriate TA parameters in predicting the firmness

of pears by ANN. The optimal ANN model, which

consisted of one hidden layer with 17 neurons, was able

to predict firmness values with an MAE of 0.539 N. Taken

together, this work lent support to the idea that firmness of

pear during storage can be detected by TA of MR images.

Acknowledgments

This work is the main part of the project bResearch and

Development of Fresh Produce Modern Logistics Technol-

ogy and Trading DemonstrationQ (2004BA527B) financed bytheMinistry of Science andTechnology of China. The authors

thank Yan LP for assistance in performing the experiments.

References

[1] Lin HT, Xi YF, Chen SJ. Postharvest softening physiological

mechanism of Huanghua pear fruit. Sci Agri Sinic 2003;36(3):

349–52.

[2] Letal J, Irak D, Suderlova L, Hajek M. MRI dtextureT analysis of MR

images of apples during ripening and storage. Lebensm-Wiss

U-Technol 2003;36:719–27.

[3] Clark CJ, Hockings PD, Joyce DC, Mazucco RA. Application of

magnetic resonance imaging to pre- and post-harvest studies of fruits

and vegetables. Postharvest Biol Technol 1997;11:1–21.

[4] Barreiro P, Ruiz-Cabello J, Fernandez-Valle ME, Ortiz C, Ruiz-

Altisent M. Mealiness assessment in apples using MRI techniques.

Magn Reson Imaging 1999;17(2):275–81.

[5] Lammertyn J, Dresselaers T, Van Hecke P, Jancsok P, Wevers M, Nicolai

BM. MRI and X-ray CT study of spatial distribution of core breakdown

in dConferenceT pears. Magn Reson Imaging 2003;21:805–15.

[6] Galed G, Fernandez-Valle ME, Martınez A, Heras A. Application of

MRI to monitor the process of ripening and decay in citrus treated

with chitosan solutions. Magn Reson Imaging 2004;22:127–37.

[7] Thybo AK, Jespersen SN, L&rke PE, Stbdkilde-Jbrgensen HJ. Nonde-

structive detection of internal bruise and spraing disease symptoms in

potatoes using magnetic resonance imaging. Magn Reson Imaging

2004;22:1311–7.

[8] Chen P, McCarthy MJ, Kauten R. NMR for internal quality evaluation

of fruits and vegetables. T ASAE 1989;32:1747–53.

[9] Donker HCW, Van As H. Cell water balance of white button

mushrooms (Agaricus bisporus) during its post-harvest lifetime

studied by quantitative magnetic resonance imaging. Biochim Bio-

phys Acta 1999;1427:287–97.

[10] Zemah H, Bendel P, Rabinowitch HD, Kamenetsky R. Visualization

of morphological structure and water status during storage of Allium

aflatunense bulbs by NMR imaging. Plant Sci 1999;147:65–73.

[11] Andaur JE, Guesalaga AR, Agosin EE, Guarini MW, Irarrazaval P.

Magnetic resonance imaging for nondestructive analysis of wine

grapes. J Agric Food Chem 2004;52:165–70.

[12] Lerski RA, Schad LR. The use of reticulated foam in texture test

objects for magnetic resonance imaging. Magn Reson Imaging

1998;16(9):1139–44.

[13] Lerski RA, Straughan K, Schad LR, Boyce D, Blqml S, Zuna I. MR

image texture analysis — an approach to tissue characterization. Magn

Reson Imaging 1993;11:873–87.

[14] Therdthai N, Zhou WB. Artificial neural network modeling of the

electrical conductivity property of recombined milk. J Food Eng

2001;50:107–11.

[15] Izadifar M. Neural network modeling of trans isomer formation and

unsaturated fatty acid changes during vegetable oil hydrogenation.

J Food Eng 2005;66:227–32.

[16] Rai P, Majumdar GC, DasGupta S, De S. Modeling the performance

of batch ultrafiltration of synthetic fruit juice and mosambi juice using

artificial neural network. J Food Eng 2005;71:273–81.

[17] Szczypinski PM. MaZda 3.20 (1998–2002). MRI analysis software.

Technical University of Lodz, Lodz, Poland.

[18] Torrecilla JS, Otero L, Sanz PD. A neural network approach for

thermal/pressure food processing. J Food Eng 2004;62:89–95.

[19] Hornik K. Some new results on neural network approximation. Neural

Netw 1993;6:1069–72.

[20] Ruan R, Almaer S, Zhang J. Prediction of dough rheological

properties using neural networks. Cereal Chem 1995;72(3):308–11.

[21] Abdolmaleki P, Mihara F, Masuda K, Buadu LD. Neural networks

analysis of astrocytic gliomas from MRI appearances. Cancer Lett

1997;118:69–78.

[22] Sablani SS, Shayya WH. Computerization of Stumbo’s method of

thermal process calculations using neural networks. J Food Eng

2001;47:233–40.

[23] Lohani S, Trivedi PK, Nath P. Changes in activities of cell wall

hydrolases during ethylene-induced ripening in banana: effect of

1-MCP, ABA and IAA. Postharvest Biol Technol 2004;31:119–26.

[24] Bendel P, Zemah H, Kamenetsky R, Vergeldt F, Van As H.

Magnetization transfer and double-quantum filtered imaging as probes

for motional restricted water in tulip bulbs. Magn Reson Imaging

2001;19:857–65.