Rafeek T, A Gunasundari / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.662-668662 | P a g eReliable Non invasive First Trimester Screening Test Using Imageprocessing and Artificial Neural NetworkRafeek T*, A Gunasundari**(*PG Scholar , ** Assistant Professor)(Department of Electrical &Electronics Engineering, PSG College of Technology, Coimbatore-641 004)ABSTRACTDown Syndrome is a chromosomalabnormality caused by the presence of all or partof a third copy of chromosome 21.Any womancan have a baby with Down Syndrome. Physicalgrowth of affected babies are delayed especiallythe facial characteristics. Recent study provesthat Down Syndrome can be detected in earlystage by identifying the absence of fetal nasalbone. During the first trimester of pregnancy,visual identification of nasal bone by examiningthe ultrsonogram is very difficult. Speckle noiseis also introducing errors in ultrasonic images.This work presents a new approach for thedetection of nasal bone by using different imageprocessing algorithms and Back propagationneural network(BPNN).A high performancehybrid Despeckling method is used in this systemwhich can dramatically increases the accuracy ofthe whole system. The features in the nasalregion are extracted in spatial domain as well astransform domain using Discrete CosineTransform (DCT) and wavelet transforms.Features extracted from images with nasal boneand images which don’t have nasal bone. Thenormalized data set is used to train BackPropagation Neural Network (BPNN).Thistrained artificial feed forward network is used toclassify different ultra sonogram. Experimentallyprove that the proposed method gives betterclassification rate than any other non invasivescreening method. This method combined withthe present detection methods can reduceoperator error and enhance overall detectionrate.Keywords: Back Propagation Neural Network(BPNN), DCT, Despeckling, First Trimester, NonInvasive Screening.1. IntroductionDown Syndrome is a genetic disorder thatwas named after John Langdon Down, who firstrecognized it as a distinct condition in 1866. Itaffects very small percent of world‟s population,approximately 1 of 800 live births. Individuals withDown syndrome tend to have a lower-than-averagecognitive ability, often ranging from mild tomoderate disabilities. It is difficult for a personaffected by Down Syndrome to lead a normalindependent life. Several studies show that maincause for this Syndrome is the presence of an extrachromosome in the 21st pair. Routine screening(Prenatal screening) for Down Syndrome is carriedout during pregnancy in order to identify womenwho are at high risk of giving birth to a child withDown Syndrome. Previously Down Syndrome wasdetected using invasive techniques, though they giveaccuracy of 90% or above, they carry a significantrisk of miscarriage. Recently Non invasivetechniques using Ultra sonogram are being are usedfor the early detection of Down Syndrome duringperiod of gestation. It is an established fact that,during first trimester period of gestation the absenceof nasal bone is reliable indicator for the presence ofDown syndrome. Prenatal ultrasound studies in11th-14th week fetus have shown that nasal bone isnot visible for the case of a Trisomy 21.In this work, a process for non invasive,accurate, and reliable method to enhance thedetection rate of Nasal bone from ultrasound imageshas been proposed. On an ultrasound scan imagenasal bone appears as a very small white patch. Thetexture of this white patch is significantly differentfrom that of surrounding tissues. These statisticalfeatures are extracted from image on both spatialdomain and Transform domain and are fed to neuralnetwork classifier. A three layer Back PropagationNeural Network is created and these extractedparameters are used to train the neural network.After the network has been trained sufficiently, itwill be able to distinguish images having nasal bone.2. Proposed MethodologyFig. 1.block diagramThe proposed work, presents a new method for theidentification of Down Syndrome through thedetection of nasal bone. The block diagram for theproposed system is shown in figure. We can classify
Rafeek T, A Gunasundari / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.662-668663 | P a g ethe proposed system into four different sections.First one is image acquisition. Set of sample imagesare collected from an ultrasound imaging system.Second one is image pre-processing section, applysome pre-processing techniques for the removal ofinherent noises, and certain characteristics of nasalbone can be easily detected. Next is to segment theregion of interest. Next section is feature extraction,nasal usually appears as very small white patch inUltra sonogram. The textures of white patches aresignificantly different from that of surroundingtissues. As the textures are different, statisticalparameters extracted from image segmentcontaining nasal bone will be different from regionwithout nasal bone. This statistical difference can beused for the identification of nasal bone. Extractedfeatures are sending to ANN network in both spatialdomain and converted to transform domain data.The last section is a neural network classifier whichis used to distinguish an image with nasal bone fromone which doesn‟t have nasal bone.2.1 Image acquisitionThe first stage of any vision system is theimage acquisition stage. Ultrasonography is able todetect many fetal structural and functionalabnormalities. Ultrasound works by using sound togenerate an image of the fetus. A special gel isapplied on the mother‟s abdomen and a transducer isused to transmit the sound waves into the abdomen,it directs small pulses of inaudible, high-frequencysound waves into the body. As the sound wavesbounce off of internal organs, fluids and tissues, thesensitive microphone in the transducer records tinychanges in the sound‟s pitch and direction. Thesesignature waves are instantly measured anddisplayed by a computer, which in turn creates areal-time picture on the monitor.2.1.1UltrasonographyUltra sonic imaging uses high frequencysound waves and their echoes to produce imagesthat can demonstrate organ movement in real time.Unlike electromagnetic waves, such as X-rays andgamma rays, ultrasound is non ionizing and as suchis considered safe at the intensities used in clinicalimaging systems. The transducer probe makes thesound waves and receives the echoes. It isconsidered as the mouth and ears of the ultrasoundmachine. In the probe, there is one or more quartzcrystals called piezoelectric crystals. When anelectric current is applied to these crystals, theychange shape rapidly. The rapid shape changes, orvibrations, of the crystals produce sound waves thattravel outward. Conversely, when sound or pressurewaves hit the crystals, they emit electrical currents.Therefore, the same crystals can be used to send andreceive sound waves. The probe also has a soundabsorbing substance to eliminate back reflectionsfrom the probe itself, and an acoustic lens to helpfocus the emitted sound waves.Ultrasonography hasemerged as a useful technique for imaging internalorgans and soft tissue structures in the human body.It is non invasive, portable, versatile, relatively lowcost. The fetal images are obtained from ultrasoundmachine (model HD-15 PHILIPS) .2 .2 Image pre-processing:The objective of image enhancement is toimprove the interpretability of the informationpresent in images for human viewers. Enhancementyields a better-quality image for the purpose ofsome particular application which can be done byeither suppressing the noise or increasing the imagecontrast. Images are often corrupted by impulsenoise due to noisy sensor or channel transmissionerror. This appears as discrete isolated pixelvariations that are not spatially correlated. The goalof removing impulse noise is to suppress the noisewhile preserving the integrity of the edges and detailinformation associated with the original image.Speckle reduction is usually used as a critical pre-processing step for clinical diagnosis by ultrasoundand ultrasound image processing. Image variance orspeckle noise is a granular noise that inherentlyexists and degrades the quality of the active images.The first step is to reduce the effect of speckle noise.2.2.1 Hybrid Despeckling FilterUltrasound imaging system is an importantimaging method in medical field. One of the majordrawbacks of ultrasound images is the poor imagequality due to speckle noise. Only skilledradiologists can make an effective diagnosis andhence limiting its use in a wide medical network.Speckle is a multiplicative noise which createsdifficulty for extracting fine details from an image.Speckle suppression by means of a digital imageprocessing is one of the techniques to improve theimage quality and possible the diagnostic potentialof medical ultrasound imaging. Despeckling is themethod to reduce speckle noise and to improve thevisual quality for better diagnosis. Many denoisingmethods have been proposed over the years, such aswavelet thresholding and bilateral filtering,anisotropic methods, median filtering etc. Amongthese, wavelet thresholding has been reported as ahighly successful method.Fig.2 .hybrid despeckling method-block diagram
Rafeek T, A Gunasundari / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.662-668664 | P a g eIn wavelet thresholding, a signal is decomposed intoapproximation and detail sub bands and thecoefficients in the detail sub bands are processed viahard or soft thresholding. The hard thresholdingeliminates coefficients that are smaller than athreshold; the soft thresholding shrinks thecoefficients that are larger than the threshold aswell. The main task of wavelet thresholding is theselection of threshold value and the effect ofdenoising depends on the selected threshold. Abigger threshold value suppresses the usefulinformation and noise components, while a smallerthreshold cannot eliminate the noise effectively.The bilateral filter is an alternative towavelet thresholding. It applies spatially weightedaveraging without smoothing edges. This isachieved by combining two Gaussian filters. Onefilter works in spatial domain and the other in theintensity domain. Therefore, not only the spatialdistance but also the intensity distance is importantfor the determination of weights .Hence, these typesof filters can remove noise in an image whileretaining edges. The main objective of this work isto design a filter for effective Despeckling ofmedical ultrasound images without smoothingedges. The parameters of the bilateral filter used are,σ d =1.6, σ r = 0.5 and the window size is 3x3 withone level wavelet decomposition. The mainadvantage of hybrid despeckling technique is,retaining edges or preserves edge whendenoising.That means get a better performance ofstandard quality matrices likesignal to noise ratio (SNR), in the case of hybriddespeckling .Fig.3. original &despeckled image2.2.2 Selection of ROI:It is sometimes of interest to process asingle sub region of an image, leaving other regionsunchanged. This is commonly known as region-of-interest (ROI) processing. Image sub regions may bespecified by using Mathematics or Graphicsprimitives, such as Point, Line, Circle, Polygon, orsimply as a list of vertex positions. Region ofinterest (ROI) usually means the meaningful andimportant regions in the images. The use of ROI canavoid the processing of irreverent image points andaccelerate the processing. Extraction of regions ofinterest from images is an important topic in theimage processing area, especially in biomedicalimage processing area.2.3EdgeDetectionIn an image, between two regions, a set ofconnected pixels are seen. Such pixel group is calledas an edge. The detection of edge (boundary) is themost common method for detecting meaningfuldiscontinuities, especially in medical imaging field.The purpose of edge detection is to identify areas ofan image where a large change in intensity.Typically, edge detection is useful for segmentation,identification and registration of objects in a scene.Mathematically, first and second order derivativesare used for edge detection.Prewitt, Sobel andLaplacian operator methods are more common inedge detection. These operators work well forimages with sharp edges and low amount of noise.Edge detection algorithm should look for aneighborhood with strong signs of change. Most ofthe edge detectors work on measuring the intensitygradient at a point in the image .2.4 Watershed Segmentation AlgorithmThe watershed transformation is a powerfultool for image segmentation. In a watershedtransformation, the image is considered as atopographic surface. The gray level of the imagerepresents the altitudes. Marker controlledwatershed algorithm is the direct application ofwatershed algorithm generally leads to oversegmentation due to noise and other localirregularities of the gradient, i.e. large number ofsegmented regions. This can be a serious enough torender the result of algorithm virtually useless. Anapproach used to control over segmentation is basedon the concept of markers. A marker is a connectedcomponent belonging to an image. These markerscan be of two types: internal marker associated withobject of interest and external markers associatedwith the background. A procedure for markerselection typically consist of two principle steps Pre-processing and definition of a set of criteria that amarker must satisfy Marker selection can rangefrom simple procedure based on the gray levelvalues and connectivity, to more complexdescription involving size, shape, location, texturecontent and so on. For the proposed system externalmarkers shows effective partition of the nasal boneregion from its background. Watershed algorithmhas been applied to each individual region. In otherwords we simply take the gradient of thesmoothened image and then restrict the algorithm tooperate on a single watershed that contains themarker in that particular region. Using markersbring a priori knowledge to bear on thesegmentation problem .2.5 Image Feature ExtractionAn image feature is a distinguishingprimitive characteristic or attribute of an image.Some features are natural in the sense that suchfeatures are defined by the visual appearance of an
Rafeek T, A Gunasundari / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.662-668665 | P a g eimage, while other, artificial features result fromspecific manipulations of an image. Natural featuresinclude the luminance of a region of pixels and grayscale textural regions. Image amplitude histogramsand spatial frequency spectra are examples ofartificial features. Image features are of majorimportance in the isolation of regions of commonproperty within an image (image segmentation) andsubsequent identification or labelling of suchregions (image classification). There are twoquantitative approaches to the evaluation of imagefeatures: prototype performance and figure of merit.In the prototype performance approach for imageclassification, a prototype image with regions(segments) that have been independentlycategorized is classified by a classificationprocedure using various image features to beevaluated. The classification error is then measuredfor each feature set. The best set of features is, ofcourse, that which results in the least classificationerror. The figure-of-merit approach to featureevaluation involves the establishment of somefunctional distance measurements between sets ofimage features such that a large distance implies alow classification error. The most basic of all imagefeatures is some measure of image amplitude interms of luminance, spectral value or other units.There are many degrees of freedom in establishingimage amplitude features. Image variables such asluminance values may be utilized directly, oralternatively, some linear, nonlinear, or perhapsnon-invertible transformation can be performed togenerate variables in a new amplitude space. On anultra sonogram nasal bone appears as a very smallwhite patch. The texture of this white patch issignificantly different from that of surroundingtissues. This characteristic of the image is extractedfor the detection purpose. Statistical parameters suchas mean, variance, skewness, kurtosis, fifth ordermoment and sixth order moment are taken here fordetection purposes.2.5.1MeanThe simple mathematical average of set of two ormore numbers. In a data set, the arithmetic mean isequal to the sum of the values divided by thenumber of values. The arithmetic mean of a set ofnumbers x11, x12... xmn is typically denoted by ,The mean of the values:Mean x =1M∗Nni=1 (𝑥𝑖𝑗 )mj=1 (1)Mean indicates the tendency to cluster aroundsome particular value.2.5.2VarianceIn probability theory and statistics,the variance is a measure of how far a set ofnumbers is spread out. It is one of severaldescriptors of a probability distribution, describinghow far the numbers lie from the mean (expectedvalue). In particular, the variance is one ofthe moments of a distribution. In that context, itforms part of a systematic approach todistinguishing between probability distributions.While other such approaches have been developed,those based on moments are advantageous in termsof mathematical and computational simplicity. Thevalue, which characterizes its “width” or“variability” around the mean value, is the variance:Variance =1M−1 ∗(N−1)ni=1 (xij − x)mj=1 (2)2.5.3 SkewnessThe skewness of a random variable „x‟ isthe third standardized moment and it is representedas:Skewness = 1/(M ∗ N) [xi,j−xσ]³ (3)Where M, N represents the image size, x is themean and σ is the standard deviation. In probabilitytheory and statistics, skewness is a measure of theasymmetry of the probability distribution of a real-valued random variable. The skewness value can bepositive or negative, or even undefined.2.5.4 KurtosisIn probability theory and statistics, thekurtosis is also a non-dimensional quantity. Itmeasures the relative flatness of a distribution to anormal distribution. The conventional definition ofthe kurtosis is:Kurtosis= [1/(M ∗ N) [xij − x/σ]⁴ −3 (4)Where the -3 term makes the value zero for anormal distribution, σ is the standard deviation.2.5.5 Fifth order features:[1/(M ∗ N) [xi, j − x/σ] 5(5)2.5.6 Sixth order features:𝑆𝑖𝑥𝑡ℎ 𝑂𝑟𝑑𝑒𝑟 = [1/(𝑀 ∗ 𝑁) [𝑥𝑖, 𝑗 − 𝑥^/𝜎] 6(6)2.6 Transform domain AnalysisIt represents the analysis and representationof image in Transform domain. By using transforms,signals can be represented with less number ofcoefficients with minimum distortion. Thecommonly used transform domain analysis isDiscrete Cosine Transform, Sine transform, Wavelettransform etc. In this work, multiresolutiontransform and DCT has been used for the analysis.2.6.1 Discrete Cosine TransformA discrete cosine transform (DCT)expresses a sequence of finitely many data points interms of a sum of cosine functions oscillating atdifferent frequencies. Like other transforms, theDCT attempts to decorrelate the image data.Discrete-Cosine Transform (DCT) has foundpopularity due to its comparative concentration ofinformation in a small number of coefficients, andincreased tolerance to variation of illumination TheDCT has been proved successful at de-correlating
Rafeek T, A Gunasundari / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.662-668666 | P a g eand concentrating the energy of image data. DCTsare important to numerous applications in scienceand engineering. The DCT packs energy in the lowfrequency regions. Therefore, some of the highfrequency components can be discarded withoutsignificant quality degradation. Note that thetransform image has zeros or low level intensitiesexcept at the top left corner where the intensities arevery high. These low frequencies, high intensitycoefficients, are therefore, the most importantcoefficients in the frequency matrix and carry mostof the information about the original image. Thediscrete cosine transform of an NxN image(x, y) isdefined by:𝐶 u, v = α u α v N−1x=0 f x, y cosπ 2x+1 u2Ncosπ 2x+1 v2NN−1y=0(7)For u, v = 0, 1, 2 ……...N -1and α u , α(v) =1Nfor u = 02Nfor u ≠ 0The proposed technique calculates the 2D-DCT for each cropped region. A subset of thesecoefficient values is taken to construct the featurevector. Empirically, the upper left corner of the 2D-DCT matrix contains the most important valuesbecause they correspond to low-frequencycomponents within the processed image block. Theextracted coefficients/features are then used fortraining purpose. There is no redundant informationbecause the wavelet functions are orthogonal. Thecomputation is efficient due to the existence of apyramidal algorithm based on convolutions withquadrature mirror filters. The original signal can bereconstructed from the wavelet decomposition witha similar algorithm.2.6.2 Wavelet TransformA wavelet is a wave-like oscillation withamplitude that starts out at zero, increases, and thendecreases back to zero. It can typically be visualizedas a "brief oscillation" like one might see recordedby a seismograph or heart monitor. Generally,wavelets are purposefully crafted to have specificproperties that make them useful for signalprocessing. Wavelet theory is applicable to severalsubjects. All wavelet transforms may be consideredforms of time-frequency representation forcontinuous-time (analog) signals and so are relatedto harmonic analysis. Wavelets are mathematicalfunction which decomposes data into differentfrequency components, each component with aresolution matched to its state. It has moreadvantage when we analyzing physical situationwith discontinuities and sharp edges. The wavelettransform is identical to hierarchical subbandfiltering systemAlmost all practically useful discretewavelet transforms use discrete-time filter banks.These filter banks are called the wavelet and scalingcoefficients in wavelets nomenclature. The wavelettransform (WT) is a relatively new type oftransform. One of the main strength thatcharacterizes this transform is its ability to provideinformation about the time-frequency representationof the signal. For most practical applications thereare two kinds of wavelets available which are thecontinuous wavelet transform (CWT) and thediscrete wavelet transform (DWT). Waveletcoefficient at every scale generating a huge amountof data. Due to the huge amount of data generatedthrough CWT, training classifiers based on itscoefficients at different scales can often becomedifficult. The family of Daubechies wavelets waschosen as the basis functions for the decomposition.Daubechies wavelets are classified according to thenumber of vanishing moments, N. The smoothnessof the wavelets increases with the number ofvanishing moments. For the case when N = 1, theDaubechies scaling function and wavelet functionresembles that of the Haar and are discontinuous. Itis desirable to have smooth wavelets and thereforeN is increased. Although the Daubechies2 wavelet iscontinuous, its derivative is discontinuous. For Ngreater than 2, the wavelet and its derivative areboth continuous.By applying DWT, the data is actually dividedi.e. decomposed into four subbands correspondingto different resolution levels. The Transformfunction used for the analysis is Daubechies D4wavelet transform; it has four wavelet and scalingfunction coefficients. The scaling functioncoefficients are:h0 =1+√34√2h1 =3+√34√2h2 =3−√34√2h3=1−√34√2The wavelet function coefficient values are:A1 = h3H1 = -h2V1 = h1D1 = -h0The sublevels labeled H1, V1 and D1 represent thefinest scale wavelet coefficients i.e. image details,while A1 represents coarse level coefficients(approximation of image). The image is analyzed inthree resolutions. The mean, variance, skewnesskurtosis, fifth order and sixth order are taken for allthe three levels. These constitute 6 parameters in asingle resolution and hence 72 coefficientparameters (3 resolutions) for each image. Thecoefficients were stored in separate files in the formof a vector and later these were used for training ofartificial neural network .
Rafeek T, A Gunasundari / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.662-668667 | P a g e3. Artificial Neural NetworkThis section focuses on the neural networkapproach for the detection of nasal bone. Neuralnetworks are introduced with an emphasis onmultilayer Back propagation feed forward neuralnetworks. An artificial neural network is a systembased on the operation of biological neural networksand it is called as an emulation of biological neuralsystem. Artificial neural networks are adaptivemodels. They can learn from the data andgeneralized things which they have learned. Theycan find good solution, were traditional models havefailed. Artificial neural network learns by updatingits network parameters according to a predefinedrule called learning rules. Neural network aretrained, so that a particular input leads to specifictarget output. The network is trained in such a waythat it minimizes the deviation between output andthe specified target. Neural network has been usedin pattern recognition, identification, classification,and computer vision and control systems3.1 Training using Back Propagation NeuralNetworkA tool for the detection of Nasal bone wasdesigned and was implemented in MATLAB. Theimage is analyzed in the transform domain. Thevarious parameters, which are used for the analysis,are mean, variance, skewness, kurtosis, fifth orderand sixth order. These parameters are extracted fromthe transform domain. This analysis helps in findingout some statistical difference that can be used forthe detection of the presence of nasal bone. Topredict whether a new image contain DownSyndrome or not, need to have a training set fromwhich the details of the new image can be predicted.Use a neural network, which will take in the inputsand train the network according to the imagecharacteristics and creates two well-definedboundaries for images with and without Nasal bone.Images are converted to gray scale using utilitiesavailable in MATLAB before they are fed to thetool designed. Because of large fetal movementduring the scanning process it is necessary to defineregion of interest which can compensate for changesin the fetal head position. The image is analyzed inthree resolutions. The mean, variance, skewness,kurtosis, fifth order and sixth order are taken for allthe three levels. These constitute 72 parameters for athree level decomposition for each image .Thecoefficient parameters are stored in separate filesand are normalized separately. These 72 parametersare used in the training phase. Parameters areextracted from a large number of images and arethen normalized. The normalized patterns are usedin the training of Back propagation neural network.Training is continued till the error converges to areasonably minimum value. During the trainingprocess the numeral “0”is used torepresent images with Down Syndrome and “1”isused to represent images having normal anatomicalfeatures.4. Results and ConclusionUltrasonography is the most popular technique usedin the field of imaging of soft tissue structures in thehuman body especially fetal images. In this work,offered a new approach to the use of computer aidedexamination to enhance the detection rate of DownSyndrome. Initially selected the nasal bone portionwith the help of ROI, then applied different imageprocessing algorithms including a new type ofdespeckling method, then used a feed forward ANNnetwork for training the data base and detected theexistence of nasal bone accurately. Here, used alatest denoising method, that is hybrid despecklefilter instead of common median filter. Performanceof this filter is superior as compared to medianbecause of the use of bilateral as well as waveletdenoising methods. Results show that the proposedsystem can easily detected the presence of nasalbone.Table1: Detection using DWTTable 2:Despeckled image parametersSINOMETHOD MEDIAN BILATERAL1. MSE 1.0442 1.01782. PSNR 17.9431 21.40113. STRUCTURALCONTENT0.9222 0.99044 NORMALIZEDABSOLUTEERROR0.1696 0.16555. NORMALIZEDCROSS-CORRELATION0.9840 0.9737Images No ofimagestestedTruedetectionFalseDetectionWith NB 78 70 8WithoutNB25 21 4
Rafeek T, A Gunasundari / International Journal of Engineering Research and Applications(IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.662-668668 | P a g eFig.4. Selection of ROIFig.5. Sobel and Prewitt edge detected ResultsFig.6. Watershed resultsREFERENCES Nicolaides K. H, Azar G, Byrne D, MansurC, Marks K. "Fetal nuchal translucency:ultrasound screening for chromosomaldefects in first trimester of pregnancy".BMJ 1992; 304: pp 867-889.. R.Vanithamani,G.Umamaheswari,A.AjayKrishnan, C.Ilaiyarasan, K.Iswariya and C.G.Kritika, “A HybridDespeckling Model for Medical UltrasoundImages”, Journal of computing, Volume 3,Issue 9, ISSN2151-9617, September 2011.. Anjit, T.A., Rishidas, S., “Identification ofNasal Bone for the Early Detection ofDown Syndrome using Back PropagationNeural Network ,” IEEE Conference onCommunications and Signal Processing(ICCSP),2011.. Nirmala.S and Palanisamy.V,“Measurement of nuchal translucencythickness in first trimester ultrasound fetalimages for detection of chromosomalabnormalities", International conference oncontrol, automation, communication andenergy conservation-2009.. B.Priestley shan and M.Madheswaran,“Revised estimates of ultrasonographicmarkers for gestational age assessment ofsingleton pregnancies among Indianpopulation.” International journal ofadvanced science and technology vol.17,april,2010.. Rafael C.Gonzalez, “Digital imageprocessing,” Prentice Hall , 2005.. C. Larose, P. Massoc, Y. Hillion, J. P.Bernard, Y. Ville, "Comparison of fetalnasal bone assessment by ultrasound at 11-14 weeks and by postmortem X- ray intrisomy 21: a prospective observationalstudy", Ultrasound in Obstetrics andGynecology, John Wiley & Sons, Ltd,Volume 22, Pages 27-30, 2003.. Stephane G. Mallat "A Theory forMultiresolution Signal Decomposition:TheWavelet Representation" ,IEEETransaction Pattern on Analysis andmachine Intelligence. Vol. II, No. 7. Joly1989.. K. O. kagan, S.Cicero, I.Staboulidou,D.Wright, K.H.Nicolaides, Fetal nasalbone in screening for trisomies 21, 18 and13 and turner syndrome at 11- 13 weeks ofgestation", Ultrasound in Obstetrics andGynecology, John Wiley & Sons, Ltd,Volume 33, Pages 259-264, 2004.. R.Vanithamani, G.Umamaheswari,“Performance Analysis of Filters forSpeckle Reduction in Medical UltrasoundImages”, International Journal ofComputer Application (0975-8887),volume 12-No.6, December 2010.. Geoff Dougherty “Digital Imageprocessing for medical applications”California State University, ChannelIslands.. User manual 4535 613 83101 Rev A “HD-15 ultrasound System” Philips ElectronicsN.V, May 2009.. B.Yegnanarayana “Artificial NeuralNetworks” Prentice Hall of India ,NewDelhi-2001.. S N Sivanandam,S Sumathi,S N Deepa“Introduction to Neural Networks usingMatlab 6.0”Tata McGraw Hill EducationPrivate Limited, new Delhi,2009.. S Jayaraman,S Esakkirajan,TVeerakumar”Digital ImageProcessing”,Tata McGraw Hill EducationPrivate Limited,New Delhi,2010.. Carol M Rumack, Stephanie R. Wilson, J.William Charboneau, and Deborah Levine,“Diagnostic Ultrasound”. ES gopi, “Digital Image Processing usingMATLAB” Scitech Publications Pvt.Ltd,2009