Ultrasound estimation of fetal weight in twins by artificial neural network

doi:10.4236/jbise.2011.41006

Paper Menu >>

Journal Menu >>

J. Biomedical Science and Engineering, 2011, 4, 46-50

doi:10.4236/jbise.2011.41006 Published Online January 2011 (http://www.SciRP.org/journal/jbise/ JBiSE

Published Online January 2011 in SciRes. http://www.scirp.org/journal/JBiSE

Ultrasound estimation of fetal weight in twins by artificial

neural network

Hanieh Mohammadi, Meshkat Nemati, Zohreh Allahmoradi, Hoda Forghani Raissi,

Somayeh Saraf Esmaili, Ali Sheikhani

Department Biomedical Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran.

Email: [email protected]

Received 8 November 2010; revised 12 November 2010; accepted 26 November 2010.

ABSTRACT

This study was undertaken to determine the accuracy

of using Ultrasound (US) estimation of twin fetuses

by use of Artificial Neural Network. At First, as the

training group, we performed US examinations on

186 healthy singleton fetuses within 3 days of delivery.

Three input variables were used to construct the

ANN model: abdominal circumference (AC), ab-

dominal diameter (AD), biparietal diameter (BPD).

Then, a total of 121 twin fetuses were assessed sub-

sequently as the validation group. In validation group,

the mean absolute error and the mean absolute per-

cent error between estimated fetal weight and actual

fetal weight was 261.77 g and 7.81%, respectively.

Results show that, twin estimation of birth weight by

ultrasound correlates fairly well with the actual

weights of twin fetuses.

Keywords: Ultrasound; Fetal Weight Estimation; Twin;

Artificial Neural Network

1. INTRODUCTION

Next to preterm labor, intrauterine growth retardation is

the most common risk factor leading to perinatal death

in twins [1-3]. Estimation of fetal weight by ultrasound

is an important investigation in the management of twin

pregnancies because of the high incidence of growth

retardation [4], the inaccurate estimation of fetal weight

by abdominal palpation, and the common excess of am-

niotic fluid [5]. Growth retardation occurs more fre-

quently in twin gestatio n than in singletons [6 ]. This fact

contributes heavily to the high frequency of perinatal

complications in twin gestation [7]. Weight discrepancy

in particular has been shown to be a most significant

factor in predicting perinatal mortality and morbidity [8].

Prediction of fetal weight by Ultrasonographic meas-

urements in twin gestation may be complicated by fetal

crowding, fetal position and oligohydramnios. Several

investigators have commented on the difficulty in ob-

taining reliable measurements of the fetal biparietal di-

ameter (BPD), and the abdominal circumference (AC) in

twins [3-5]. The most widely used weight formula for

both singletons and twins is the Shepard et al. revision

of the Warsoff et al. equation using the biparietal diame-

ter (BPD) and the abdominal circumference (AC) [1-9].

A slight imprvement of the method has been introduced

by incorporating femur length and/or measurements of

the head and abdominal areas [7,8]. It can, however, be

difficult to define the whole AC in twins, and thus di-

ameter measurements may result in more accurate fetal

weight estimations [10]. Various neural network archi-

tectures and learning methodologies have been used in

the literature for fetal weight estimation [8,10]. Although

different methods are available, a simple, quick and re-

liable method of assess in birth weight is still under de-

bate [10-13].

Artificial neural network (ANN), a computerized ana-

log of a biologic neural system, has been widely used in

many different professional fields [5-7]. The constructed

architecture of the ANN model would develop relation-

ship between the input and output data when training

proceeds. The way of training the ANN model simulates

how biologic neural connections are established and

rectified perpetually. After an appropriate training proc-

ess, the nonlinear neural network can afford a best fit

guess as a result. The architecture, principles, character-

istics and applications have been discussed in the litera-

ture [10-26]. Th e purpose of this study was to determine

the accuracy of using ANN model in predicting fetal

weight in Twins and probably develop a new way which

may be more relevant to assess accurate fetal weight

estimation and could develop nonlinear relationships

between input variables and output outcomes and reduce

the errors between estimated fetal weight and actual fetal

weight. In this paper we try to evaluate the accuracy of

H. Mohammadi et al. / J. Biomedical Science and Engineering 4 (2011) 46-50

using ANN model together with BPD, AC, and AD as

inputs for developing nonlinear relationships between

input variables and output outcomes and reduce the er-

rors between estimated fetal weight and actual fetal

weight in twins.

2. MATERIALS AND METHODS

A retrospective study of 186 consecutive twin pregnan-

cies delivered in the department of Obstetrics and Gy-

necology at Madaran medical faculty hospital, between

January 2010 and April 2010, was undertaken. In all

cases the estimated date of confinement had been estab-

lished by ultrasound scan at 20 weeks of gestation. Me-

dian maternal age was 26.7 years (range 15-44), median

number of pregnancies was 2 (1-5), number of previous

deliveries 0 (0-3), and median gestational age at delivery

was 36 weeks (14-41). The median birth weight of twin

A was 2390 g (160-2918 g), and of twin B 2265 g

(210-2868 g). We consecutively performed fetal biome-

try by US on every healthy twin fetuses which was ad-

mitted to the delivery room. The exclusion criteria were

fetal anomaly and fetuses not delivered within 3 days of

US examination [2]. All women had a normal pregnancy

with ultraso und documentation of the BPD, AC and AD

[8,10,14]. A total of 189 fetuses that met the above crite-

ria were included as the training group of the MNM

model. For further validate the established ANN model

in fetal weight estimation 81 fetuses that were delivered

within the subsequent 3 months from May 2010 to July

2010 and met the criteria described above were used as

the validation group. The fetal BPD was measured ac-

cording to Watmough et al. [10], from the outer to the

inner contour of the head. The fetal AD was calculated

as a mean of two diameters at right angles to one another,

measured on the outer contour on a transverse scan of

the fetal long axis, and at the level where the umbilical

vein enters the ductus venosus [2,11]. Also, the fetal

abdominal circumference (AC) was measured at the

level of the umbilical vein entry into the ductus venosus

[2]. The measurements were done 4 times and rounded

to the nearest millimeter. All the US measurements were

conducted by experienced sonographist according to the

methods previously described. We used the commer-

cially available 2-D US scanners (Aloka SSD-650) with

a 3.5-MHz transabdominal probe.

3. ANN MODEL

A neural network is a model that simulates the functions

of biologic neurons. The ability o f a single neuron could

be greatly improved by connecting multiple neuron s in a

layer. Artificial neural networks are powerful non-linear

models used vastly for classifying different types of data.

A neural network is composed of few layers of neurons.

Neurons in adjusting layers are connected with relative

quantitative weights. These weights are randomly cho-

sen, and they are changed through the training procedure,

so that the mean of the sum-of-squares error (MSE) is

minimized. The MSE is the squared difference between

the network outpu t and network target, averaged over all

of the cases [16,17]. Figure 1 shows the architecture of

the trained ANN model and Figure 2 is the neural net-

work development flow chart. Our trained ANN model

in our investigation was composed of four layers: 1) one

input layer with three inputs; 2) hidden layer with 6

neurons and 18 connections; 3) hidden layer with 2 neu-

rons and 12 connections 3. One output layer with one

outcome and 2 connections. The back propagation (BPN)

network algorithm is used as the learning algorithm to

train the artificial neural network [9,10].

Three inputs were BPD, AC, and AD. Furthermore, to

examine the performance of ANN model we calculate

mean absolute error (AE) and mean absolute percent

error (APE) between actual fetal weight and estimated

fetal weight.

4. RESULT

The median birth weight of twin A was 2390 g (150-

3200 g), and of twin B 2165 g (210-3250 g) in the vali-

dation group, respectiv ely. In validation group, the mean

absolute error (AE) and the mean absolute percent error

(APE) between estimated fetal weight and actual fetal

weight was 261.77 g and 7.81%. Also, the moderation

significant correlations between the actual fetal weight

and the estimated fetal weight of the validation group are

shown in Figure 3 (r = 0.9348, n = 121). Also, th e over-

all, high correlation between AC, AD, BPD and twin`s

EFW were 0.81, 0.87 and 0.84, respectively, which

shows the important effect of these parameters on twin`s

weight. It seems that the prediction error is known to

increase with increasing twin’s weight.

Figure 1. Architecture of the trained ANN model, BPD = bipa-

rietal distance, AC = abdominal circumference, AD= abdomi-

al diameter. n

H. Mohammadi et al. / J. Biomedical Science and Engineering 4 (2011) 46-50

Figure 2. The neural network development flow chart.

5. CONCLUSIONS

A neural network is a model that simulates the functions

of biologic neurons. The ability of a single neuron could

be greatly improved by connecting multiple neurons in

layers. Artificial neural networks are powerful non-l i nea r

models used vastly for classifying different types of data.

A neural network is composed of few layers of neurons.

Neurons in adjusting layers are connected with relative

quantitative weights. These weights are randomly cho-

sen, and then are changed through the training procedure,

so that the mean of the sum-of-squares error (MSE) is

minimized [2,11,12]. The ANNs learning algorithms can

be divided into two main groups that are supervised (or

Associative learning) and unsupervised (Self-Organiza-

tion) learning [2,9,11,14]. One of the most commonly

used supervised ANN model is back propagation (BPN)

network that uses back propagation learning algorithm [9-

11]. Back propagation algorithm is one of the well-known

Figure 3. Scattergram of the estimated fetal body weight vs.

the actual fetal body weight the by the ANN model in the vali-

dation group.

JBiSE

H. Mohammadi et al. / J. Biomedical Science and Engineering 4 (2011) 46-50

algorithms in neural networks. The back propagation

neural network is essentially a network of simple proc-

essing elements working together to produce a complex

output [11]. These elements or nodes are arranged into

different layers: input, middle and output. The output

from a back propagation neural network is computed

using a procedure known as the forward pass [2,8-11,15]:

1) The input layer propagates a particular input vector’s

components to each node in the middle layer. 2) Middle

layer nodes compute output values, which become inputs

to the nodes of the output layer. 3) The output layer

nodes compute the network output for the particular in-

put vector. The forward pass produces an output vector

for a given input vector based on the current state of the

network weights. Since the network weights are initial-

ized to random values, it is u nlikely that reasonable out-

puts will result before training . The weights are adjusted

to reduce the error by propagating the output error

backward through the network [2,9,15]. In this study, we

used the BPN algorithm to develop the ANN and prove

our hypothesis that BPD, AC and AD wi t hin A NN model

could reduce errors between estimated fetal weight and

actual fetal weight. The subjects in our series were a

group of women with healthy singleton fetus with

documentation of US examination with, BPD, AC and

AD. Some may wonder at our choice of the three input

parameters, thinking that they are not well justified. The

three dimensional variables are reasonable because of

the previous literature [9]. Also, the overall, high corre-

lation between AC, AD, BPD and twin’s EFW were 0.81,

0.87 and 0.84, respectively, which shows the important

effect of these parameters on twin’s weight. In our study,

the definition of an anomaly was for any fetus with a

major structural anomaly that could be diagnosed prena-

tally, such as holoprosencephaly, omphalocele, cystic

hygroma, etc. These were excluded from the study. We

might include some fetuses with rare and nonstructural

anomalies that could only be diagnosed postnatally by

genetic screening or metabolic methods, in which the

prenatal ultrasonic examination cannot demonstrate any

structural abnormality. However, we believe that this

point makes only little impact on the stud y because these

nonstructural anomalies are too rare [9].

In our study, the mean absolute error (AE) and the

mean absolute percent error (APE) between estimated

fetal weight and actual fetal weight were 162.71 g and

7.81%, respectively. The fetuses in weight range of

(>2500 g) are the lowest accurate fetal weight estimation

in validation group (AE = 269 g, APE= 10.51%), we

think that, as the fetus grows are more quick at the last

trimester and we considered babies within 3 days of de-

livery, it might be one part of the error in this weight

range is related to fetus grows within this estimation of

fetal weight. In th is ANN model we have 4 layers; input

layer, two median layers and output layer. We have three

input variables AC, AD and BPD. In all cases the esti-

mated date of confinement had been established by ul-

trasound scan at 20 weeks of gestation. Median maternal

age was 26.7 years (range 15-44), median number of

pregnancies was 2 (1-5), number of previous deliveries 0

(0-3), and median gestational age at delivery was 36

weeks (14-41). The median birth weight of twin A was

2390 g (160-29 18 g) , an d of tw in B 2 265 g (210 -2868 g)

in the training group. Also, the median birth weight of

twin A was 2190 g (150-3200 g), and of twin B 2165 g

(210-2868 g) in the validation group.

Also estimation of fetal weight by ANN model at the

weight range of (<1500 g) are the most accurate result. It

seems that the prediction birth weight error is known to

increase with increasing weight of twins, week by week.

In conclusion, our study demonstrates that our single

multiplicative neuron model is a well-established model

and can be used to estimate fetal weight. However, more

accuracy of fetal weight estimation is in need of further

studies.

6. ACKNOWLEDGEMENTS

This study was supported in part by grants from university of science

and research branch Islamic Azad University, Tehran. The authors are

grateful to all doctors for the ultrasound measurements; Ms. Fatemeh

Nematollahi and Ms. Fatemeh Bani and their assistance; and Ali

Ghafari at the Department of Obstetrics and gynecology, Madaran

Medical Faculty for equipment supply.

REFERENCES

[1] Hendricks, C.H. (1966) Twinning in relation to birth

weight, mortality and congenital anomalies. Obstetrics

Gynecology, 27, 47-53.

[2] Powers, W.F. (1973) Twin pregnancy, complications and

treatment. Obstetrics Gynecology, 43, 795-808.

[3] Chang, F.M., Liang, R.I., Ko, H.C., Yao, B.L., Chang, C.H.

and Yu, C.H. (1997) Threedimensional ultrasound- as-

sessed fetal thigh volumetry in predictingbirt h weight, Ob-

stetrics Gynecology, 90, 331-339.

doi:10.1016/S0029-7844(97)00280-9

[4] Manlan, G. and Scott, K.E. (1978) Contribution of twin

pregnancy to perinatal mortality and fetal growth, retar-

dation: reversal of growth retardation after birth. Cana-

dian Medical Association Journal, 118, 365-368.

[5] Secher, N.J., Kaern, J., Hansen, P.K., et al. (1985) In-

tra-uterine growth in twin pregnancies: prediction of fetal

growth retardation. Obstetrics Gynecology, 66, 63-68.

[6] Shepard, M.J., Richards, V.A., Berkowitz, R.L. Warsof,

S.L. and Hobbins, J.C. (1982) An evaluation of two

equations for predicting fetal weight by ultrasound.

American Journal of Obstetrics Gynecology, 142, 47-54.

[7] Vintzileos, A.M., Campbell, W.A., Rodis, J.F., Bors-

Koefoed, R. and Nochimson, D.J. (1987) Fetal weight

estimation formulas with head, abdominal, femur, and

H. Mohammadi et al. / J. Biomedical Science and Engineering 4 (2011) 46-50

JBiSE

thigh circumference measurements. American Journal of

Obstetrics Gynecology, 157, 410-414.

[8] Campbell, S. and Wilkin, D. (1975) Ultrasonic measure-

ment of fetal abdomen circumference in the estimation of

fetal weight. British Journal of Obstetrics Gynecology, 82,

689-697. d oi: 10.1111/j .1471-0528.1975.tb00708.x

[9] Secher, N.J., Djursing, H., Kern Hansen, P., et al. (1987)

Estimation of fetal weight in the third trimester by ultra-

sound. European Journal of Obstetrics, Gynecology and

Reproductive Biology, 24, 1-11.

doi:10.1016/0028-2243(87)90031-1

[10] Sindberg Eriksen, P., Secher, N.J., Weis-Bentzon, M., et

al. (1985) Normal growth of the fetal biparietal diameter

and the abdominal diameter in a longitudinal study: An

evaluation of the two parameters in predicting fetal

weight. Acta Obstetricia et Gynecological Scandinavica,

64, 65-70. doi:10.3109/00016348509154690

[11] Jouannic, J.M., Grange, G., Goffinet, F., Benachi, A. and

Carbrol, D. (2001) Validity of sonographic formulas for

estimating fetal weight below 1, 250g: aseries of 119

cases. Fetal Diagnosis Therapy, 16, 254-258.

doi:10.1159/000053923

[12] Hendricks, C.H. (1966) Twinning in relation to birth

weight, mortality and congenital anomalies. Obstetrics

Gynecology, 7, 47-53.

[13] Eik-Nes, S.H., Groettum, P. and Anderson, N.J. (1982)

Estimation of fetal weight by ultrasound measurement.

Acta Obstetricia et Gynecological Scandinavica, 61,

307-312.

[14] Bjerkedal, T. and Skjaerven, R. (1980) Percentiles of

birth weight and crown-heel length in relation to gesta-

tion period of single livebirths. Tidsskrift for den Norske

Laegeforening, 100 , 88-91.

[15] Chuang, L., Hwang, J.Y., Chang, C.H., Yu, C.H. and

Chang, F.M. (2002) U-utrasound estimation of fetal

weight with the use of computerized artificial neural

network model. Ultrasound in Medicine Biology, 28,

991-996. doi:10.1016/S0301-5629(02)00554-9

[16] Merz, E., Lieser, H., Schicketanz, K.H. and Harle, J.

(1988) Intrauterine fetal weight assessment using ultra-

sound: A comparison of several weight assessment

methods and development of a new formula for the de-

termination of fetal weight. Ultraschall in der Medizin, 9,

15-24. doi:10.1055/s-2007-1011588

[17] Schild, R.L., Fell, K., Fimmers, R, Gembruch, U. and

Hansmann, M. (1988) A new formula for calculating

weight in the fetus of < or = 1600g. Ultrasound Obstet-

rics Gynecology, 24, 775-780. doi:10.1002/uog.1741

[18] Farmer, R.M., Medearis, A.L., Hirata, G.I. and Platt, L.D.

(1992) The use of a neural net work for the ultrasound

estimation of fetal weight in the macrosomic fetus.

American Journal of Obstetrics Gynecology, 166, 467-

1472.

[19] Secher, N.J., Toettrup, A., Djursing, H. and Weber, T.

(1989) Fetal weight estimation in twin pregnancies.

European Journal of Obstetrics, Gynecology and Repro-

ductive Biology, 32, 181-186.

doi:10.1016/0028-2243(89)90033-6

[20] Eriksen, P.S., Secher, N.J. and Weis-Bentzon, M. (1985)

Normal growth of the fetal biparietal diameter and the

abdominal diameter in a longitudinal study. Acta Obstet-

ricia et Gynecological Scandinavica, 64, 65-70.

doi:10.3109/00016348509154690

[21] Halaska, M.G., Vlk, R., Feldmar, P., et al. (2006) Pre-

dicting term birth weight using ultrasound and maternal

characteristics. European Journal of Obstetrics, Gyne-

cology and Reproductive Biology, 128, 231-235.

doi:10.1016/j.ejogrb.2006.01.020

[22] Mielke, G., Pietsch-Breitfeld, B., Salinas, R., Risse, T.

and Marzusch, K. (1995) A new formula for prenatal ul-

trasonographic weight estimation inextreme lypreterm

fetuses. Gynecological Obstetric Investigation, 40, 84-88.

doi:10.1159/000292311

[23] Yadav, R.N., Kalra, P.K. and John, J. (2007) Time series

prediction with single multiplicative neuron model. Ap-

plied Soft Computing, 7, 1157-1163.

doi:10.1016/j.asoc.2006.01.003

[24] Houlton, M.C.C., Marivate, M. and Philpott, R.H. (1981)

The prediction of fetal growth retardation in twin preg-

nancy. British Journal of Obstetrics Gynecology, 88,

264-273. d oi: 10.1111/j .1471-0528.1981.tb00980.x

[25] Blickstein, I. and Lancet, M. (1988) The growth discor-

dant twin. Obstet Gynecol Surv, 43, 509-515.

doi:10.1097/00006254-198809000-00002

[26] Pielet, B.W., Sabbagba, E. , Scott, N., et al. (1987) Ultra-

sonic prediction of birth weight in preterm fetuses:

Which formula is best? American Journal of Obstetrics

Gynecology, 157, 1411-1414.