Research Article Backpropagation Neural Network ...downloads.hindawi.com/journals/jam/2013/453098.pdfResearch Article Backpropagation Neural Network Implementation for Medical Image

Hindawi Publishing CorporationJournal of Applied MathematicsVolume 2013 Article ID 453098 8 pageshttpdxdoiorg1011552013453098

Research ArticleBackpropagation Neural Network Implementation forMedical Image Compression

Kamil Dimililer

Electrical and Electronic Engineering Department Near East University Nicosia North Cyprus Mersin 10 Turkey

Correspondence should be addressed to Kamil Dimililer kdimililerneuedutr

Received 12 August 2013 Revised 27 November 2013 Accepted 28 November 2013

Academic Editor Ferenc Hartung

Copyright copy 2013 Kamil Dimililer This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Medical images require compression before transmission or storage due to constrained bandwidth and storage capacity An idealimage compression system must yield high-quality compressed image with high compression ratio In this paper Haar wavelettransform and discrete cosine transform are considered and a neural network is trained to relate the X-ray image contents to theirideal compression method and their optimum compression ratio

1 Introduction

Most images at some point will require compression beforetransmitting or storage due to constrained bandwidths andlimited storage capacity This is especially true of medicalimages X-ray images are widely used in medicine they arethe secondmost commonly used type of tests after laboratoryones

Teleradiology which is the term used for using technol-ogy to send radiographic images or X-rays across distancesfrom one location to another has become lately one ofthe most used clinical aspects of telemedicine Telemedicinerefers to the use of communication and information tech-nologies for the delivery of clinical care such as the transferradiological images from a site of image acquisition to aremote location for interpretation in hospitals such as tomog-raphy (CT) scans magnetic imaging (MRI) ultrasonography(US) and X-rays These radiological images are needed to becompressed before transmission to a distant location or dueto the bandwidth or storage limitations [1]

There has been a rapid development in compressionmethods to compress large data files such as images wheredata compression in various applications has become morevital [2] Efficient methods of compression to compress andstore or transfer image data files are needed while retaininghigh image quality and marginal reduction in size due to theimprovements of technology [3]

The discrete cosine transform (DCT) is one of the mostpopular transforms used in compression of medical imagesin standards like Joint Photographic Experts Group (JPEG)In DCT-based compression the image is split into smallerblocks for computational simplicity and the blocks areclassified on the basis of information content to maximizecompression ratio without sacrificing diagnostic information[1] DCT-based image compression has been popular dueto its simplicity and thus has been used in many medicalimage processing applications [4ndash7] The use of DCT andartificial neural networks has also been investigated in searchfor optimum compression methods In [8] DCT and neuralnetworks based medical image compression was proposedand applied toMRandCTmedical images In [9] a topology-preserving neural network was developed for medical imagecompression using a ldquoneural-gasrdquo network In [10] adaptivearchitecture neural networks were implemented for medicalimage compression In [11] a neural network classifier wasused with a combination of different image compressiontechniques for various applications In [12] a neural networkmodel called direct classification was used to compress imagedata In [13] a direct solution method applied to imagecompression using neural networks was suggested

More works using neural networks for image compres-sion applications emerged lately such as those in [14] aneural network quantizer was used to yield a high compres-sion ratio while maintaining high-quality images In [15]

2 Journal of Applied Mathematics

Preparationof targets

Training phase

Testing phase

CMandCR

CMandCR

1

2

4096

1

2

52

1

18

Training images64 times 64

8-bitgrayscale

Testing images64 times 64

8-bitgrayscale

1

2

4096

1

2

52

1

18

Figure 1 Block diagram of medical image compression system

a multiresolution neural network (MRNN) filter bank wasused as a transform for coding In [16] an image compres-sion algorithm based on image blocks complexity measuremethods and a neural network was proposed In [17] imagecompression using wavelet transform and a neural networkwas suggested In [18] a neural network basedDCT compres-sion system that finds the optimum compression ratios for avariety of images was also suggested

Unlike the discrete cosine transform the wavelet trans-forms are not fourier based and therefore discontinuities inimage data can be handled with better results using wavelets[19] Wavelets are a mathematical tool for hierarchicallydecomposing functions There is a general preference touse wavelet transforms in image compression because thecompressed images will be obtained with higher compressionratios and higher PSNR values [20] Haar wavelet transformis one of the wavelet methods applied in compressing thedigital images Previousworks usingHaar image compressioninclude an application that is applied to adaptive data hidingfor the images dividing the original image into 8times 8 subblocksand reconstructing the images after compression with goodquality [21]

The use of compression methods in general and waveletcompression in particular with medical images has beenpreviously investigated in [1 22ndash25] More recently a neuralnetworkmodel was used to obtain the optimum compressionratio when using Haar wavelet image compression which wasapplied to a variety of images [19]

The aim of the work presented within this paper is todevelop an image compression system for medical X-rayimage compression using a neural network classifier andtwo popular compression methods namely Haar wavelettransform (HWT) and discrete cosine transform (DCT)Our novel method suggests that a trained neural networkcan learn the nonlinear relationship between the intensity(pixel values) of an X-ray image and its ideal compressionmethod and optimum compression ratio Once the ideal

compression method out of the two considered methods ischosen by the network and the highest compression ratiowhile maintaining good image quality using that chosenmethod is obtained then the result reduction in radiographimage size should make the storage and transmission ofradiographs more efficient

Figure 1 represents the block diagram for the suggestedmedical image compression system In training phase pre-determined compression methods with their compressionratios are prepared and trained using backpropagation neuralnetwork and in testing phase the images are given to theback propagation neural network system to achieve thecompression ratio and compression method

The paper is organized as follows Section 2 describes theX-ray image databasewhich is used for the implementation ofour proposed system Section 3 presents the X-rays compres-sion system describing image preprocessing and the neuralnetwork design and implementation Section 4 introduces themethod used to evaluate the results and provides an analysisof the system implementation Finally Section 5 concludesthe work that is presented within this paper and suggestsfurther work

2 The X-Ray Image Database

The development and implementation of the proposed sys-tem for medical X-rays compression uses 80 X-ray imagesfrom our medical image database which contains X-rayimages of fractured dislocated broken and healthy bonesin different parts of the body Both considered compressionmethods (DCT and HWT) were applied to X-ray imagesusing nine compression ratios (10 20 90) and anoriginal image with its compression using the ratios usingDCT method is shown in Figure 2 [26]

Contrast weighted entropy is a new term that combinesthe change of contrast values between the neighbor pixels

Journal of Applied Mathematics 3

Original

10 20 30

40 50 60

70 80 90

Figure 2 An original image its DCT compression at nine ratios


(a)

(b)

Figure 3 (a) Training Set (b) Testing Set

with the entropy of the image which includes the informationpresent within an image An image with high contrastweighted entropy should be compressed with a low amountof compression ratio in order not to make the informationwithin the pixels lost However an image with low contrastweighted entropy can be compressed with higher amountof compression ratio because there are not enough detailsthat can be lost during the compression process Equation (1)represents the contrast weighted entropy of an image

CWH = minussum119873

(119901119894minus 120583) times 119901

119894times log2(119901119894) times Tot (1)

where Tot represents the total number of pixels within theimage 119873 represents the number of intensity values of thepixels within the image 119901

119894represents the probability of

the intensity values and 120583 represents mean value of thefrequencies of the intensity values

The optimum HWT and DCT compression ratios forthe X-ray images were determined using the optimum com-pression criteria based on visual inspection and empiricalanalysis The visual inspection involved 10 experts in theirfield who were asked to observe the change in contrast withinthe pixels amount of data within an image and smoothnessand edge continuity of certain areas within the reconstructedimages [27] and consider the contrast within the adjacentpixels and their entropy within the certain areas of the imageswhich is the value of contrast weighted entropy inTable 1 [28]

Table 1 Original and compressed images with their CWH

Image 1 CWH Image 1 CWHOrg Image minus3441102 50 DCT minus483934710 DCT minus3821898 60 DCT minus464850220 DCT minus5087191 70 DCT minus484522130 DCT minus4950159 80 DCT minus748216240 DCT minus4902581 90 DCT minus1203251

(i) Training image set contains 40 images with knownoptimum compression ratios which are used fortraining the neural network within the compressionsystem Examples of training images are shown inFigure 3(a)

(ii) Testing image set it contains 40 images with knownoptimum compression ratios which are used to testand validate the efficiency of the trained neuralnetwork Examples of these testing images are shownin Figure 3(b)

3 The X-Rays Image Compression System

The X-ray image compression system uses a supervisedneural network based on the back propagation learn-ing algorithm due to its implementation simplicity and


1

Original image

ICMand

OCR

2

1

2

52

1

18Reduced size

pixelsCompressed image

4096(256 times 256) pixels

image (64 times 64)

Figure 4 X-ray image compression system ICM ideal compression method OCR optimum compression ratio

the availability of sufficient ldquoinputtargetrdquo database for train-ing this supervised learner The neural network relates theX-ray image intensity (pixel values) to the image ideal com-pression method and its optimum compression ratio havingbeen trained using images with predetermined compressionmethods and ratios Once trained the neural network wouldchoose the ideal method and the optimum compression ratiofor an X-ray image upon presenting it to the neural networkby using its intensity values

Sample image resizing was used to resize the originalimages of size (256times 256) pixels into (64times 64) pixels deletingthe next three pixels while keeping the first pixel valueFurther reduction to the size of the images was attemptedin order to reduce the number of input layer neurons andconsequently the training time however meaningful neuralnetwork training could not be achieved and thus the use ofwhole images of the reduced size of 64 times 64 pixels

In a previous research [19] 32 times 32 = 1024 nodes of inputwas applied to the back propagation neural network In thispaper 64 times 64 = 4096 nodes of input has been preferred togive adequate information to the suggested image compres-sion system Although 1024 nodes of input takes less trainingtime the 4096 nodes of input were preferred because therecognition ratio for the 4096 nodes of input is higher thanthe recognition ratio of the 1024 nodes of input The testingtime for both networks was equal As a result 64 times 64 =4096 nodes of input has been preferred taking into accountthe recognition ratio rather than the training time

The size of the input X-ray images affects the choice of thenumber of neurons in the neural networkrsquos input layer whichhas three layers input hidden and output layers

Figure 3 represents sample images from the training setand testing set of the back propagation neural network

Using one-pixel-per-neuron approach the neural net-workrsquos input layer has 4096 neurons according to the topol-ogy of the neural network the hidden layer has been decidedto be used as 52 neurons after several experiments whichassures meaningful training while keeping the time cost toa minimum and its output layer has 18 neurons representingthe output classification of the ideal compressionmethod andthe optimumcompression ratio as follows output neurons 1ndash9 represent DCT compression at ratios (10ndash90) whereas

014

012

01

008

006

004

002

0

Mea

n sq

uare

erro

r

0 500 1000 1500 2000 2500 3000 3500 4000

Iteration

Error graph

Figure 5 Neural networkrsquos learning curve

output neurons 10ndash18 represent HWT compression at ratios(10ndash90)

During the learning phase the learning coefficient andthe momentum rate were adjusted during various exper-iments in order to achieve the required minimum errorvalue of 0002 which was considered as sufficient for thisapplication Figure 4 shows the topology of this neuralnetwork within the image compression system

4 Results and Discussion

The evaluation of the training and testing results was per-formed using two measurements the recognition rate andthe accuracy rate Once the choice of the ideal compressionmethod is made by the neural network the recognition rateis defined as follows

RROC = (119868OC119868119879

) lowast 100 (2)

where RROC is the recognition rate for the neural networkwithin the X-ray compression system 119868OC is the number


Original image

Original image

Original image

DCT 20

HWT 70

DCT 30

Figure 6 Examples of testing set image compression using the developed system

of optimally compressed X-ray images and 119868119879is the total

number of X-ray images in the database setThe accuracy rate RAOC for the neural network output

results is defined as follows

RAOC = 100 + (1 minus((

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816) lowast 10) + 119878

119879

119878119879

) (3)

where 119878119875represents the pre-determined (expected) optimum

compression ratio in percentage 119878119894represents the optimum

compression ratio as determined by the trained neuralnetwork in percentage and 119878

119879represents the total number

of compression ratios used for both of the compressionmethods

The optimum compression deviation (OCD) is anotherterm that is used in our evaluation OCD is the differencebetween the pre-determined or expected optimum com-pression ratio 119878

119875and the optimum compression ratio 119878

119894as

determined by the trained neural network and is defined asfollows

OCD =(

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816)

10

(4)

The OCD is used to indicate the accuracy of the systemand depending on its value the recognition rates varyTable 2 shows the three considered values of OCD andtheir corresponding accuracy rates and recognition rates


Table 2 Accuracy and recognition rates according to OCD

OCD Accuracy rate (RAOC) Recognition rate (RROC)0 100 2140 (525)1 9444 3640 (900)2 8889 3740 (925)3 Below 8889 4040 (100)

Table 3 Neural network final training parameters

Input nodes 4096Hidden nodes 52Output nodes 18Learning rate 00005Momentum rate 03Error 0002Iterations 3524Training time (seconds) 2015Run time (seconds) 0015

The evaluation of the system implementation results uses(OCD = 1) as it provides a minimum accuracy rate of 875which is considered sufficient for this application

The neural network learnt and converged after 3524iterations or epochs and within 33 minutes and 12 secondswhereas the running time for the generalized neural networksafter training and using one forward pass was 001 secondsThese results were obtained using a 28GHz PC with 8GB ofRAM Windows 8 64-bit OS and MATLAB 2013a software

Table 3 lists the final parameters of the successfullytrained neural network whereas Figure 5 shows the errorminimization curve of the neural network during learning

The trained neural network recognized correctly the idealmethod and optimum compression ratios for all 40 trainingimages as would be expected thus yielding 100 recognitionof the training set Testing the trained neural network usingthe 40 images from testing set that were not presented to thenetwork before yielded 90 recognition rate where 36 outof the 40 images with known ideal method and optimumcompression ratios were correctly classified

The results of this application are demonstrated inFigure 6 which shows examples of the optimally compressedX-ray images as determined by the trained neural network

5 Conclusions

This paper describes a novelmethod for compressingmedicalX-Ray images using a neural network with Discrete CosineTransform and Haar Wavelet Transform The method usesHWT and DCT based compression with nine compressionratios each and a supervised neural network that learnsto associate the gray X-ray image intensity (pixel values)with the ideal compression method and a single optimumcompression ratio to be used with the selected method

The implementation of the proposed system choosesHWT and DCT based image compression methods as theyhave been popularly and efficiently used in previousworks for

this kind of applicationThe aim of an optimum compressionratio is to combine high compression with good-qualitycompressed X-ray images thus making the storage andtransmission of these medical images more efficient

The proposed system was developed and implementedusing 80 X-ray images of fractured dislocated broken andhealthy bones in different parts of the body The neural net-workwithin the compression system learnt to associate the 40training images with their predetermined ideal compressionmethod and optimum compression ratios within 3524 sec-onds Once trained the neural network could recognize theideal method and optimum ratio for an X-ray image within0015 seconds

In this work a minimum accuracy level of 9444 wasconsidered as acceptable Using this accuracy level the neuralnetwork yielded 900 correct recognition rate of idealmethod and optimum ratios

Future work will include the implementation of thissystem using multilayer perceptron neural network structureto be used in the decision of optimum compression ratio withoptimum compression method using Daubechies waveletand biorthogonal wavelet transform based image compres-sion

References

[1] S Singh V Kumar and H K Verma ldquoAdaptive threshold-based block classification in medical image compression forteleradiologyrdquo Computers in Biology and Medicine vol 37 no6 pp 811ndash819 2007

[2] M J Nadenau J Reichel and M Kunt ldquoWavelet-based colorimage compression exploiting the contrast sensitivity func-tionrdquo IEEE Transactions on Image Processing vol 12 no 1 pp58ndash70 2003

[3] K Ratakonda and N Ahuja ldquoLossless image compressionwith multiscale segmentationrdquo IEEE Transactions on ImageProcessing vol 11 no 11 pp 1228ndash1237 2002

[4] D Chikouche R Benzid andM Bentoumi ldquoApplication of theDCT and arithmetic coding to medical image compressionrdquo inProceedings of the 3rd International Conference on Informationand Communication Technologies From Theory to Applications(ICTTA rsquo08) pp 1ndash5 Damascus Syria April 2008

[5] V N P Raj and T Venkateswarlu ldquoA novel approach to medicalimage compression using sequential 3D DCTrdquo in Proceedingsof the International Conference on Computational Intelligenceand Multimedia Applications (ICCIMA rsquo07) vol 3 pp 146ndash152Sivakasi India December 2007

[6] M J Zukoski T Boult and T Iyriboz ldquoA novel approach tomedical image compressionrdquo International Journal of Bioinfor-matics Research and Applications vol 2 no 1 pp 89ndash103 2006

[7] F Y Shih and Y TWu ldquoRobust watermarking and compressionfor medical images based on genetic algorithmsrdquo InformationSciences vol 175 no 3 pp 200ndash216 2005

[8] Z Dokur ldquoA unified framework for image compression andsegmentation by using an incremental neural networkrdquo ExpertSystems with Applications vol 34 no 1 pp 611ndash619 2008

[9] A Meyer-Base K Jancke A Wismuller S Foo and T Mar-tinetz ldquoMedical image compression using topology-preservingneural networksrdquo Engineering Applications of Artificial Intelli-gence vol 18 no 4 pp 383ndash392 2005


[10] R Ashraf and M Akbar ldquoAdaptive architecture neural netsfor medical image compressionrdquo in Proceedings of the IEEEInternational Conference on Engineering of Intelligent Systems(ICEIS rsquo06) pp 1ndash4 Islamabad Pakistan 2007

[11] L Ma and K Khashayar ldquoAdaptive constructive neural net-works using hermite polynomials for image compressionrdquo inAdvances in Neural Networks vol 3497 of Lecture Notes inComputer Science pp 713ndash722 Springer Berlin Germany2005

[12] H S Soliman and M Omari ldquoA neural networks approach toimage data compressionrdquo Applied Soft Computing vol 6 no 3pp 258ndash271 2006

[13] I Vilovic ldquoAn experience in image compression using neuralnetworksrdquo in Proceedings of the 48th International SymposiumELMAR-2006 Focused on Multimedia Signal Processing andCommunications pp 95ndash98 IEEE Press Zadar Croatia June2006

[14] R Ashraf and M Akbar ldquoAbsolutely lossless compression ofmedical imagesrdquo in Proceedings of the IEEE-EMBS 27th AnnualInternational Conference of the Engineering in Medicine andBiology Society pp 4006ndash4009 Shanghai China January 2006

[15] B Northan and R D Dony ldquoImage compression with amultiresolution neural networkrdquo Canadian Journal of Electricaland Computer Engineering vol 31 no 1 pp 49ndash58 2006

[16] H Veisi and M Jamzad ldquoImage compression with neuralnetworks using complexity level of imagesrdquo in Proceedings of the5th International Symposium on Image and Signal Processing andAnalysis (ISPA rsquo07) pp 282ndash287 Istanbul Turkey September2007

[17] S Osowski RWaszczuk and P Bojarczak ldquoImage compressionusing feed forward neural networksmdashhierarchical approachrdquo inNatural to Artificial Neural Computation vol 3497 of LectureNotes in Computer Science pp 1009ndash1015 Springer BerlinGermany 2006

[18] A Khashman and K Dimililer ldquoNeural networks arbitrationfor optimum DCT image compressionrdquo in Proceedings of theInternational Conference on ldquoComputer as a Toolrdquo (EUROCONrsquo07) pp 151ndash156 Warsaw Poland September 2007

[19] A Khashman and K Dimililer ldquoImage compression usingneural networks and Haar waveletrdquo WSEAS Transactions onSignal Processing vol 4 no 5 pp 330ndash339 2008

[20] K H Talukder and K Harada ldquoHaar wavelet based approachfor image compression and quality assessment of compressedimagerdquo International Journal of Applied Mathematics vol 36no 1 pp 1ndash8 2007

[21] B L Lai and L W Chang ldquoAdaptive data hiding for imagesbased on haar discrete wavelet transformrdquo inAdvances in Imageand Video Technology vol 4319 of Lecture Notes in ComputerScience pp 1085ndash1093 Springer Heidelberg Germany 2006

[22] W Li X Wu Y Zhang and M Yu ldquoWavelet-based far infraredmedical image compression with histogram shiftingrdquo in Pro-ceedings of the 1st International Conference on Bioinformaticsand Biomedical Engineering (ICBBE rsquo07) pp 1025ndash1027WuhanChina July 2007

[23] S S Devi and K Vidhya ldquoDevelopment of medical image com-pression techniquesrdquo in Proceedings of the International Confer-ence on Computational Intelligence andMultimedia Applications(ICCIMA rsquo07) pp 97ndash101 Sivakasi India December 2007

[24] Y T Chen and D C Tseng ldquoWavelet-based medical imagecompression with adaptive predictionrdquo Computerized MedicalImaging and Graphics vol 31 no 1 pp 1ndash8 2007

[25] B Karlik ldquoMedical image compression by using Vector Quan-tization Neural Network (VQNN)rdquo Neural Network World vol16 no 4 pp 341ndash348 2006

[26] Famagusta General Hospital and Radiology Department ldquoX-ray radiograph databaserdquo Paralimni Cyprus 2013

[27] A Khashman and K Dimililer ldquoComparison criteria for opti-mum image compressionrdquo in Proceeding of the InternationalConference on ldquoComputer as a Toolrdquo (EUROCON rsquo05) pp 935ndash935 Belgrade Serbia amp Montenegro November 2005

[28] K Dimililer ldquoNeural network implementation for image com-pression of X-raysrdquo Electronics World vol 118 no 1911 pp 26ndash29 2012

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MathematicsJournal of


Mathematical Problems in Engineering

Hindawi Publishing Corporationhttpwwwhindawicom

Differential EquationsInternational Journal of

Volume 2014

Applied MathematicsJournal of


Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of


Mathematical PhysicsAdvances in

Complex AnalysisJournal of


OptimizationJournal of


CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of


Operations ResearchAdvances in

Journal of


Function Spaces

Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of Mathematics and Mathematical Sciences


The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Algebra

Discrete Dynamics in Nature and Society



Decision SciencesAdvances in

Discrete MathematicsJournal of


Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Stochastic AnalysisInternational Journal of


Preparationof targets

Training phase

Testing phase

CMandCR

CMandCR

1

2

4096

1

2

52

1

18

Training images64 times 64

8-bitgrayscale

Testing images64 times 64

8-bitgrayscale

1

2

4096

1

2

52

1

18

Figure 1 Block diagram of medical image compression system

a multiresolution neural network (MRNN) filter bank wasused as a transform for coding In [16] an image compres-sion algorithm based on image blocks complexity measuremethods and a neural network was proposed In [17] imagecompression using wavelet transform and a neural networkwas suggested In [18] a neural network basedDCT compres-sion system that finds the optimum compression ratios for avariety of images was also suggested

Unlike the discrete cosine transform the wavelet trans-forms are not fourier based and therefore discontinuities inimage data can be handled with better results using wavelets[19] Wavelets are a mathematical tool for hierarchicallydecomposing functions There is a general preference touse wavelet transforms in image compression because thecompressed images will be obtained with higher compressionratios and higher PSNR values [20] Haar wavelet transformis one of the wavelet methods applied in compressing thedigital images Previousworks usingHaar image compressioninclude an application that is applied to adaptive data hidingfor the images dividing the original image into 8times 8 subblocksand reconstructing the images after compression with goodquality [21]

The use of compression methods in general and waveletcompression in particular with medical images has beenpreviously investigated in [1 22ndash25] More recently a neuralnetworkmodel was used to obtain the optimum compressionratio when using Haar wavelet image compression which wasapplied to a variety of images [19]

The aim of the work presented within this paper is todevelop an image compression system for medical X-rayimage compression using a neural network classifier andtwo popular compression methods namely Haar wavelettransform (HWT) and discrete cosine transform (DCT)Our novel method suggests that a trained neural networkcan learn the nonlinear relationship between the intensity(pixel values) of an X-ray image and its ideal compressionmethod and optimum compression ratio Once the ideal

compression method out of the two considered methods ischosen by the network and the highest compression ratiowhile maintaining good image quality using that chosenmethod is obtained then the result reduction in radiographimage size should make the storage and transmission ofradiographs more efficient

Figure 1 represents the block diagram for the suggestedmedical image compression system In training phase pre-determined compression methods with their compressionratios are prepared and trained using backpropagation neuralnetwork and in testing phase the images are given to theback propagation neural network system to achieve thecompression ratio and compression method

The paper is organized as follows Section 2 describes theX-ray image databasewhich is used for the implementation ofour proposed system Section 3 presents the X-rays compres-sion system describing image preprocessing and the neuralnetwork design and implementation Section 4 introduces themethod used to evaluate the results and provides an analysisof the system implementation Finally Section 5 concludesthe work that is presented within this paper and suggestsfurther work

2 The X-Ray Image Database

The development and implementation of the proposed sys-tem for medical X-rays compression uses 80 X-ray imagesfrom our medical image database which contains X-rayimages of fractured dislocated broken and healthy bonesin different parts of the body Both considered compressionmethods (DCT and HWT) were applied to X-ray imagesusing nine compression ratios (10 20 90) and anoriginal image with its compression using the ratios usingDCT method is shown in Figure 2 [26]

Contrast weighted entropy is a new term that combinesthe change of contrast values between the neighbor pixels


Original

10 20 30

40 50 60

70 80 90



(a)

(b)




(119901119894minus 120583) times 119901













1

Original image

ICMand

OCR

2

1

2

52

1

18Reduced size



image (64 times 64)








014

012

01

008

006

004

002

0

Mea

n sq

uare

erro

r

0 500 1000 1500 2000 2500 3000 3500 4000

Iteration

Error graph






RROC = (119868OC119868119879

) lowast 100 (2)



Original image

Original image

Original image

DCT 20

HWT 70

DCT 30





RAOC = 100 + (1 minus((

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816) lowast 10) + 119878

119879

119878119879

) (3)








119894as


OCD =(

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816)

10

(4)












5 Conclusions







References





































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










Original

10 20 30

40 50 60

70 80 90



(a)

(b)




(119901119894minus 120583) times 119901













1

Original image

ICMand

OCR

2

1

2

52

1

18Reduced size



image (64 times 64)








014

012

01

008

006

004

002

0

Mea

n sq

uare

erro

r

0 500 1000 1500 2000 2500 3000 3500 4000

Iteration

Error graph






RROC = (119868OC119868119879

) lowast 100 (2)



Original image

Original image

Original image

DCT 20

HWT 70

DCT 30





RAOC = 100 + (1 minus((

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816) lowast 10) + 119878

119879

119878119879

) (3)








119894as


OCD =(

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816)

10

(4)












5 Conclusions







References





































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










(a)

(b)




(119901119894minus 120583) times 119901













1

Original image

ICMand

OCR

2

1

2

52

1

18Reduced size



image (64 times 64)








014

012

01

008

006

004

002

0

Mea

n sq

uare

erro

r

0 500 1000 1500 2000 2500 3000 3500 4000

Iteration

Error graph






RROC = (119868OC119868119879

) lowast 100 (2)



Original image

Original image

Original image

DCT 20

HWT 70

DCT 30





RAOC = 100 + (1 minus((

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816) lowast 10) + 119878

119879

119878119879

) (3)








119894as


OCD =(

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816)

10

(4)












5 Conclusions







References





































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










1

Original image

ICMand

OCR

2

1

2

52

1

18Reduced size



image (64 times 64)








014

012

01

008

006

004

002

0

Mea

n sq

uare

erro

r

0 500 1000 1500 2000 2500 3000 3500 4000

Iteration

Error graph






RROC = (119868OC119868119879

) lowast 100 (2)



Original image

Original image

Original image

DCT 20

HWT 70

DCT 30





RAOC = 100 + (1 minus((

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816) lowast 10) + 119878

119879

119878119879

) (3)








119894as


OCD =(

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816)

10

(4)












5 Conclusions







References





































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










Original image

Original image

Original image

DCT 20

HWT 70

DCT 30





RAOC = 100 + (1 minus((

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816) lowast 10) + 119878

119879

119878119879

) (3)








119894as


OCD =(

10038161003816100381610038161003816119878119901minus 119878119894

10038161003816100381610038161003816)

10

(4)












5 Conclusions







References





































Volume 2014




Journal of











Journal of


Function Spaces






Algebra



















5 Conclusions







References





































Volume 2014




Journal of











Journal of


Function Spaces






Algebra




































Volume 2014




Journal of











Journal of


Function Spaces






Algebra
















Volume 2014




Journal of











Journal of


Function Spaces






Algebra









Documents

Research Article Backpropagation Neural Network ...downloads.hindawi.com/journals/jam/2013/453098.pdfResearch Article Backpropagation Neural Network Implementation for Medical Image