Histograms of Oriented 3D Gradients for Fully Automated Fetal …downloads.hindawi.com/journals/bmri/2017/3956363.pdf · 2019-07-30 · tan−1(𝑦 𝑥) cos−1(𝑧 𝑟)). (1)

Research ArticleHistograms of Oriented 3D Gradients for Fully AutomatedFetal Brain Localization and Robust Motion Correction in 3 TMagnetic Resonance Images

Ahmed Serag1 Gillian Macnaught2 Fiona C Denison1 Rebecca M Reynolds1

Scott I Semple2 and James P Boardman1

1MRC Centre for Reproductive Health University of Edinburgh Edinburgh UK2Clinical Research Imaging Centre University of Edinburgh Edinburgh UK

Correspondence should be addressed to Ahmed Serag afseraggmailcom

Received 22 August 2016 Revised 13 December 2016 Accepted 26 December 2016 Published 30 January 2017

Academic Editor Yudong Cai

Copyright copy 2017 Ahmed Serag et al 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

Fetal brain magnetic resonance imaging (MRI) is a rapidly emerging diagnostic imaging tool However automated fetal brainlocalization is one of the biggest obstacles in expediting and fully automating large-scale fetalMRI processingWe propose amethodfor automatic localization of fetal brain in 3 T MRI when the images are acquired as a stack of 2D slices that are misaligned dueto fetal motion First the Histogram of Oriented Gradients (HOG) feature descriptor is extended from 2D to 3D images Then asliding window is used to assign a score to all possible windows in an image depending on the likelihood of it containing a brainand the window with the highest score is selected In our evaluation experiments using a leave-one-out cross-validation strategywe achieved 96 of complete brain localization using a database of 104 MRI scans at gestational ages between 34 and 38 weeks Wecarried out comparisons against templatematching and random forest based regressionmethods and the proposedmethod showedsuperior performance We also showed the application of the proposed method in the optimization of fetal motion correction andhow it is essential for the reconstruction process The method is robust and does not rely on any prior knowledge of fetal braindevelopment

1 Introduction

Recent successful application of magnetic resonance imaging(MRI) has provided us with an unprecedented opportunityto study in intricate detail the developing brain in the livingfetus or neonate [1ndash7] However quantitative fetal brain anal-ysis remains challenging due to the unique confrontationsassociated with fetalMRI such asmotion artifacts low signaland spatial resolution as shown in Figure 1

Advances in medical image processing techniques havefacilitated the reconstruction of motion-corrected high-resolution 3D fetal MR images [8ndash12] from stacks of 2Dintersecting images which in turn have laid the foundationfor modeling [13ndash15] and quantitative analysis [1 15 16] ofthe developing fetal brain Such reconstruction methods relyon initial localization and cropping of the brain region from

a standard wide field of view (FOV) MRI to assist the slice-to-volume registration process [17] by excluding surroundingmaternal tissues that can result in registration failure

Fetal brain localization as an automated processing step isone of the biggest obstacles in expediting and fully automat-ing large-scale fetal MRI processing and analysis To date theapproaches proposed to address the issue of automating fetalMRI brain localization can be classified into two categoriestemplate matching and machine learning approaches In theearlier category Anquez et al [18] proposed an approachthat starts by detecting the eyes using 3D template matchingfollowed by segmentation of the brain using a 2D graph-cut segmentation of the mid-sagittal slice rotated at severalangles The best matching segmentation is selected andused to initialize a 3D graph-cut segmentation Taimouriet al [19] proposed another template matching approach

HindawiBioMed Research InternationalVolume 2017 Article ID 3956363 8 pageshttpsdoiorg10115520173956363

2 BioMed Research International

Figure 1 An example of a fetal MR scan at 355 weeks of gestationalage (GA) The scan shows that the acquired image has an arbitraryfetal orientation motion artifacts and low spatial resolution

that registers a template to each slice of the fetal brain in3D

Examples of machine learning approaches include themethod proposed by Ison et al [20] which is based ona two-phase random forest (RF) classifier to suppress theinfluence of maternal tissues and provide likely positionsof tissue centroids inside the brain An approximation of ahigh-order Markov Random Field (MRF) finds the optimalselection of landmarks and these landmarks along with aconfidence weighted probability map provide an estimate ofthe center of a region of interest around the brain Keraudrenet al [21] described a different learning-based approach thatdecomposes the search space by performing a 2D detectionprocess based on Scale-Invariant-Feature-Transform (SIFT)features before accumulating the votes in 3D The approachrelies on providing the gestational age as a prior knowledgeto remove the scale component

Recently Kainz et al [22] proposed a localizationmethodwhere rotation invariant feature descriptors (based on Spher-ical Harmonics) for medical volumes of the uterus arecalculated in order to learn the appearance of the fetal brainin the feature space Alansary et al [23] proposed a differentlocalization method where superpixels are first extracted anda histogram of features for each superpixel is computed usingbag of words based on dense SIFT descriptors Then a graphof superpixels is constructed and a RF classifier is trained todistinguish between brain and nonbrain superpixels

Although most previously reported methods achievedaccurate brain localization results all methods were imple-mented and tested using 15 T MR images and the general-ization of these approaches to 3 T MR images has not beenexamined Fetal MRI performed at 3 T is emerging as apromising modality for the evaluation of fetal anatomy as it

provides an increased signal-to-noise ratio (compared with15 T) This increase in the signal-to-noise ratio can allow fordecreased acquisition time increased spatial resolution or acombination of both with the overall goal of obtaining moredetailed imaging of fetal anatomy

In this article we focus on developing a novel methodfor fully automatic fetal brain localization from raw 3T MRimages with no need for preprocessing (such as intensityinhomogeneity correction) or prior knowledge about thedataset under study (such as gestational age (GA) or imagingplane) We treat the problem as a machine learning problemwhere a model is trained using a sample of positive andnegative examples From each sample discriminant featuresare extracted for both positive (brain) and negative (non-brain) examples based on a 3D extension of the success-ful Histogram of Oriented Gradients (HOG) [24] featuredescriptor A sliding window classifier based on SupportVector Regression (SVR) is used to assign a score to allpossible windows in an image and the window with thehighest score is selected

The remainder of this article is organized as follows Sec-tion 2 describes in detail the proposed method In Section 3the experimental results of applying the proposed methodto 3 T MR fetal images are presented We also compare theperformance of our method against a template matchingmethod and present its application in the optimization offetal motion correction Section 4 discusses our study andoutlines areas for prospectiveworkThis paper is an extensionof previous preliminary work [25]

2 Methods

21 Data The database for this study included a total of104 MR images from 32 singleton fetuses of healthy womenand women with diabetes between 3430 and 3760 weeks ofgestational age (mean and standard deviation of 3592plusmn083)Ethical approval was obtained from the National ResearchEthics Committee (South East Scotland Research EthicsCommittee) and written informed consent was obtained

22 MR Image Acquisition Images were acquired on aSiemens Magneton Verio 3 T MRI clinical scanner (SiemensHealthcare GmbH Erlangen Germany) T2-weighted half-Fourier acquisition single-shot turbo spin-echo images wereacquired of the fetal brain in sagittal coronal and transverseorientations where at least one stack is available in eachanatomical direction (HASTE TRTE = 180086ms FOV =400 times 400mm matrix = 192 (phase) times 256 (frequency) andvoxel size = 15 times 15 times 6mm)

23 The HOG Descriptor for 2D Images In the 2D imagedomain successful methods such as Harris-Corner detector[26] and the well-known SIFT descriptor [27] rely on aggre-gated gradients However the HOGdescriptor [24] organizesgradients into histograms As the first step the gradient imageis computed by convolving the input image with an appro-priate filter mask A grid of histograms is then constructedwhere each histogram organizes the respective gradients intobins according to their orientation To preserve locality a

BioMed Research International 3

histogram is computed for each cell in an evenly spaced gridAccordingly each cell contains the same number of gradients(depending on the cell size) and gets assigned exactly onehistogram The cells themselves are then organized in rect-angular blocks which may overlap The histogram values ofall cells within one block are concatenated to form a vectorThe vector of each block is then normalized and subsequentlythe concatenation of all those block-vectors yields the finalfeature vector Further details and an evaluation of the 2DHOG descriptor can be found in Dalal and Triggs [24]

24 Extending HOG to 3D Images Due to the nature ofthe analyzed data that is 3D volumetric images 3D featuredescriptors are ideal It also has been demonstrated that 3Ddescriptors aremuchmore descriptive than their correspond-ing 2D as richer information is encoded into the histograms[28]

Here our extension of the HOG approach to 3D medicalimages consists of two steps First we need to extractgradients from the images Second we need to organizethese three-dimensional gradients into bins using appropriatehistograms computed over uniformly spaced grid-blocksThis step is straightforward as we simply extend the grid andhistogram dimension each by oneWe then can convert eachgradient into spherical coordinates (1) and bin it according toits orientation (azimuth 120579 isin [0 2120587) and zenith 120601 isin [0 120587])

(119903120579120601) =(radic1199092 + 1199102 + 1199112tanminus1 (119910119909)cosminus1 (119911119903 ) ) (1)

The first step however does not generalize as easilyTo compute the image-gradient several approaches mightbe considered According to Dalal and Triggs [24] theconvolution of the image with a 1D [minus1 0 1] filter maskis most suitable This approximates the partial first-orderderivative according tonabla119891 (119909 119910) = (120575119891120575119909 120575119891120575119910)119879 (2)

asymp (119891 (119909 + 1 119910) minus 119891 (119909 minus 1 119910)119891 (119909 119910 + 1) minus 119891 (119909 119910 minus 1)) (3)

nabla119891 (119909 119910 119911) = (120575119891120575119909 120575119891120575119910 120575119891120575119911)119879 (4)

asymp (119891(119909 + 1 119910 119911) minus 119891 (119909 minus 1 119910 119911)119891 (119909 119910 + 1 119911) minus 119891 (119909 119910 minus 1 119911)119891 (119909 119910 119911 + 1) minus 119891 (119909 119910 119911 minus 1)) (5)

Based on the description above Algorithm 1 summarizesthe extraction of the 3DHOG feature vector from 3D fetalMRimages

It is worth mentioning that Histograms of Oriented 3DGradients have been previously reported in the literature for

Set the block size 119904 for image 119868Set the number of histogram bins 119861Set f3dhog to represent the 3DHOG feature vectorCompute the image gradient 119866foreach block do

foreach gradient g in 119866 doTransform g to spherical coordinatesInsert g into corresponding histogram bins

endNormalize block-vector bAppend normalized b to f3dhog

end

Algorithm 1 Histograms of Oriented 3D Gradients (3DHOG)

example in the areas of video sequence analysis [29 30] or 3Dobject retrieval [31] Those 3D extended HOGs are differentfrom ours as their data representation was 2D + time [29 30]or 3D mesh models [31] Therefore the previously proposedextensions of HOG cannot be directly (or simply) applied toour data as we deal with 3D volumetric medical images

25 Brain Localization Using Sliding Window A slidingwindow is used to move over all possible windows in animage and the window with the highest score is selectedGiven that one of our aims is to provide a method that doesnot require any prior knowledge we use a sliding windowof a fixed size that is large enough to hold the largest brainin our database To assign a score to each window we useSupport Vector Regression (SVR) For our experiments thedetector has the following default settings [minus1 0 1] gradientfilter with no smoothing 9 orientation bins in 0ndash180 degreesfor azimuth and zenith 3 times 3 times 3 voxel blocks with an overlapof half the block size L2-norm block normalization 40 times 40times 5 step size for sliding window SVR (119862 = 1 120598 = 01) with afirst-order polynomial kernel In the Results we perform anevaluation to measure the performance with respect to thevariation of these parameters and to find the best values forour learning task

26 Comparisons against Other Methods We implementeda basic method of template matching with the templaterepresenting an average fetal brain to compare our proposedmethod against We will call the template 119879(119909119905 119910119905 119911119905) where(119909119905 119910119905 119911119905) represent the coordinates of each voxel in thetemplate We will call the input or search image 119878(119909 119910 119911)where (119909 119910 119911) represent the coordinates of each voxel inthe search image We then simply move the center (or theorigin) of the template119879(119909119905 119910119905 119911119905) over each (119909 119910 119911) point inthe search image and calculate the sum of products betweenthe coefficients in 119878(119909 119910 119911) and 119879(119909119905 119910119905 119911119905) over the wholearea spanned by the template As all possible positions of thetemplate with respect to the search image are considered theposition with the best score is the best position Here the bestscore is defined as the lowest Sum of Absolute Differences(SAD)


SAD (119909 119910 119911)= 119873sum119894=0

119872sum119895=0

119871sum119896=0

Diff (119909 + 119894 119910 + 119895 119911 + 119896 119894 119895 119896) (6)

where119873119872 and 119871 denote the sizes of each dimension of thetemplate image We also experimented replacing the supportvector regressor with a random forest regressor [32] using thefollowing parameter settings number of trees in the forest =10 minimum number of samples required to be at a leaf node= 1 and unlimited number of leaf nodes

27 Validation Leave-one-out and 10-fold cross-validationprocedure were performed for the 104 MR images In caseof leave-one-out cross-validation each image in turn wasleft out as a test sample and the remaining 103 images wereused as the training dataset In 10-fold cross-validation theoriginal dataset was randomly partitioned into 10 equal sizedsubsamples Of the 10 subsamples a single subsample wasretained as the test data and the remaining 9 subsampleswereused as training data The cross-validation process was thenrepeated 10 times (the folds) with each of the 10 subsamplesused exactly once as the validation data A correct detectionor complete brain localization was defined as the enclosure ofall brain voxels within the detected box

We also measured the distance 119889 between the centerof the ground truth bounding box and the center of thedetected bounding box and median absolute error MdAEwas obtained using the following formula

MdAE = median119894=1119899

10038161003816100381610038161198891198941003816100381610038161003816 (7)

where 119899 is the total number of images used for validation

3 Results

Using leave-one-out cross-validation the proposed methodprovided 96 success rate in brain localization andMdAE of695mm When the 10-fold cross-validation was performedthe method provided comparable results with success rate of95 and MdAE of 773mm

31 Influence of Parameters Figure 2 summarizes the influ-ence of the various HOG parameters on overall detectionperformance Detector performance is sensitive to the wayin which gradients are computed but the simplest schemeprovided better accuracy As shown in Figure 2 increasingthe number of orientation bins did not greatly improveperformance up to 9 bins but beyond this performancestarted to degrade Gradient strengths vary over a widerange due to intensity inhomogeneity therefore effectivelocal contrast normalization between overlapping blocks isessential for good performance From our experiments itturns out that L1-norm is better than L2-norm When testedfor different block sizes the block size of 3 times 3 times 3 voxelsyielded the best performance Testing for different slidingwindow step sizes the step size of 40 times 40 times 5 had thebest performance Experiments also showed that SVR with

a second- and third-order polynomial kernels outperformedlinear SVR

32 Comparisons against OtherMethods To date nomethodhas been proposed for automatic localization of the fetal brainusing 3 T MR images making it difficult to compare ourproposed method against a benchmark method Howeverwe compared our results against a basic method of templatematching with the template representing an average fetalbrain The median localization error for the implementedtemplate-based matching approach was MdAE = 125mmThis shows a degradation of brain localization accuracycompared to proposedmethod usingHistograms of Oriented3D Gradients When we experimented replacing the supportvector regressor with a random forest regressor we obtainedresults with MdAE of 79mm

33 Application to Robust Motion Correction Due to thenature of the acquisition that is acquiring stacks of 2Dslices in real-time MRI in order to reduce the scan timewhile avoiding slice cross-talk artifacts slices are quite oftenmisaligned due to fetal motion and form an inconsistent3D volume (see Figure 1) Registration-based approachesfor reconstructing motion-corrected high-resolution fetalbrain volumes require a cropped box around the brain toexclude irrelevant tissues otherwise the registration and thereconstruction will likely fail Figure 3 shows an example ofthree different subjects that were successfully reconstructedusing the Baby ToolKit [17] It is clear that brain cropping isessential for the reconstruction process and the registrationfails when the whole FOV is used

4 Discussion

In this article Histogram of Oriented Gradients (HOG) isproposed to automatically localize the fetal brain in 3 T MRimages The main contribution of the article is the extensionof HOG to 3D for fetal brain MRI localization We choseto use HOG features partly because of its demonstratedsuperiority to other widely used features like SIFTSURF inmany applications [24 33 34] and partly because the useof HOG features in our task is rational as head boundariesare HOG rich relative to surrounding maternal tissues (seeFigure 4)

We also chose to use Support Vector Regression (SVR)onHOG features instead of Support VectorMachine (SVM)which was proposed by Dalal and Triggs [24] in the originalHOG feature descriptor Typically SVR is used with a slidingwindow approach to assign a score to all possible windowsin an image and the window with the highest score is thenselected Therefore by using SVR over SVM we avoid aninevitable problem that arises in overlapping sliding windowapproach which is the occurrence of multiple detectionwindows in the same neighborhood area [35]

We examined the influence of different HOG parameterson overall detection performance in terms ofmedian absoluteerror of a bounding box detected around the fetal brainand the ground truth (defined manually) In our resultswe achieved 96 for complete brain localization and the


MdA

E (m

m)

MdA

E (m

m)

Effect of normalization15

10

0

3 times 3 times 3

5 times 5 times 5

7 times 7 times 7

Effect of block size30

20

10

5

0

MdA

E (m

m)

Effect of SVR kernel10

5

0

1st order

2nd order

3rd order

MdA

E (m

m)

Effect of step size

10

0

30

20

40 times 40 times 5

45 times 45 times 5

50 times 50 times 5

Effect of gradient

MdA

E (m

m)

15

10

5

0

MdA

E (m

m)

Effect of bin size15

10

5

0

3

6

9

18

12

25 times 25 times 5

30 times 30 times 5

35times 35times 5

L1-norm

L2-norm [1 0 minus2 0 1]

[minus1 0 1]

Figure 2 The effect of the various HOG parameters on overall localization performance

automated detection process takes less than 5 seconds on anormal computer Our method is more general than Anquezet al [18] as it does not rely on localizing the eyes It also doesnot require any prior knowledge (such as gestational age) asin Keraudren et al [21]

We carried out comparisons against template match-ing and random forest based regression methods and theproposed method showed superior performance We alsoshowed the application of the proposed method in theoptimization of fetal motion correction and how it is essentialfor the motion correction process

Noteworthy is the fact that the original HOG featuredescriptor is not rotation invariant given that we relied onthe large variation of the training dataset to handle differentorientations Using a database of 104MRI scans aged between34 and 38 weeks of GA the rotation variance is representedwithin the training samples and hence themodel will learn toclassify it In addition the training samples were acquired indifferent planes (axial coronal and sagittal) which is anotherstrength of the proposed method

5 Conclusion

The main motivation behind this work was to automate thefetal brain localization step which is one of the obstaclespreventing rapid deployment and full automation of large-scale fetal MRI postprocessing Due to the nature of theacquisition that is acquiring stacks of 2D slices in real-time MRI in order to reduce the scan time while avoidingslice cross-talk artifacts slices are quite often misaligneddue to fetal motion and form an inconsistent 3D volumeRegistration-based approaches for reconstructing motion-corrected high-resolution fetal brain volumes require acropped box around the brain to exclude irrelevant tissuesotherwise the registration and the reconstruction will likelyfail Future work will include applying and evaluating thebrain localization framework to a larger GA range (iesecond trimester) and different MRI modalities of the fetalbrain such as T1-weighted Echo Planar Imaging (EPI) andDiffusion Tensor Imaging (DTI) This will allow for moreaccurate assessment and analysis of the developing fetal


(a)

(b)

Figure 3 Three MRI scans reconstructed using raw MRI without brain localization and cropping (a) Inside the green colored frames (b)the same images were motion-corrected after automatic brain localization and cropping using the proposed method

(a) (b)

Figure 4 An example of an axial fetal MR image used from the training database (a) and the resulting computed HOG descriptor overlaidon the example image (b)

brain across several stages of development using differentimaging modalities The method is available to the researchcommunity at httpbrainsquareorg

Competing Interests

The authors declare that the research was conducted in theabsence of any commercial or financial relationships thatcould be construed as a potential conflict of interests

Authorsrsquo Contributions

Ahmed Serag designed and performed the experimentsanalyzed output data and wrote the manuscript James PBoardman Fiona C Denison and Rebecca M Reynoldsprepared the grant application Gillian Macnaught and ScottI Semple acquired imaging data All authors approved thefinal submitted version and agree to be accountable for itscontent


Acknowledgments

This work was supported by the Theirworld (httpwwwtheirworldorg) NHS Research Scotland and NHS LothianResearch andDevelopmentThis work was undertaken in theMRC Centre for Reproductive Health which is funded by theMRC Centre Grant MRN0225561 The authors are gratefulto the families who consented to take part in the study and tothe midwives for recruiting

References

[1] J A Scott P A Habas K Kim et al ldquoGrowth trajectories ofthe human fetal brain tissues estimated from 3D reconstructedin utero MRIrdquo International Journal of Developmental Neuro-science vol 29 no 5 pp 529ndash536 2011

[2] P S Huppi SWarfield R Kikinis et al ldquoQuantitativemagneticresonance imaging of brain development in premature andmature newbornsrdquo Annals of Neurology vol 43 no 2 pp 224ndash235 1998

[3] A Serag I S Gousias A Makropoulos et al ldquoUnsupervisedlearning of shape complexity application to brain develop-mentrdquo in Spatio-Temporal Image Analysis for Longitudinal andTime-Series Image Data Lecture Notes in Computer ScienceSpringer 2012

[4] P Batchelor A Castellano Smith D Hill D Hawkes T Coxand A Dean ldquoMeasures of folding applied to the developmentof the human fetal brainrdquo IEEE Transactions on Medical Imag-ing vol 21 no 8 pp 953ndash965 2002

[5] L R Ment D Hirtz and P S Huppi ldquoImaging biomarkersof outcome in the developing preterm brainrdquo The LancetNeurology vol 8 no 11 pp 1042ndash1055 2009

[6] A Serag M Blesa E J Moore et al ldquoAccurate Learning withFew Atlases (ALFA) an algorithm for MRI neonatal brainextraction and comparison with 11 publicly available methodsrdquoScientific Reports vol 6 Article ID 23470 2016

[7] N N Andescavage A du Plessis R McCarter et al ldquoComplextrajectories of brain development in the healthy human fetusrdquoCerebral Cortex 2016

[8] F Rousseau O A Glenn B Iordanova et al ldquoRegistration-based approach for reconstruction of high resolution in uteroMR brain imagesrdquo Academic Radiology vol 13 no 9 pp 1072ndash1081 2006

[9] A Gholipour J A Estroff and S K Warfield ldquoRobustsuper-resolution volume reconstruction from slice acquisitionsapplication to fetal brain MRIrdquo IEEE Transactions on MedicalImaging vol 29 no 10 pp 1739ndash1758 2010

[10] K Kim P A Habas F Rousseau O A Glenn A J Barkovichand C Studholme ldquoIntersection based motion correction ofmultislice MRI for 3-D in utero fetal brain image formationrdquoIEEE Transactions on Medical Imaging vol 29 no 1 pp 146ndash158 2010

[11] M Kuklisova-Murgasova G Quaghebeur M A Rutherford JVHajnal and J A Schnabel ldquoReconstruction of fetal brainMRIwith intensity matching and complete outlier removalrdquoMedicalImage Analysis vol 16 no 8 pp 1550ndash1564 2012

[12] B Kainz M Steinberger W Wein et al ldquoFast volume recon-struction from motion corrupted stacks of 2D slicesrdquo IEEETransactions on Medical Imaging vol 34 no 9 pp 1901ndash19132015

[13] E Dittrich T Riklin Raviv G Kasprian et al ldquoA spatio-temporal latent atlas for semi-supervised learning of fetal brainsegmentations and morphological age estimationrdquo MedicalImage Analysis vol 18 no 1 pp 9ndash21 2014

[14] R Wright D Vatansever V Kyriakopoulou et al ldquoAge depen-dent fetal MR segmentation using manual and automatedapproachesrdquo MICCAI Workshop on Perinatal and PaediatricImaging pp 97ndash104 2012

[15] A Serag V Kyriakopoulou M A Rutherford et al ldquoA multi-channel 4D probabilistic atlas of the developing brain applica-tion to fetuses and neonatesrdquoAnnals of the BMVA vol 2012 no3 pp 1ndash14 2012

[16] A Gholipour A Akhondi-Asl J A Estroff and S K WarfieldldquoMulti-atlas multi-shape segmentation of fetal brain MRI forvolumetric and morphometric analysis of ventriculomegalyrdquoNeuroImage vol 60 no 3 pp 1819ndash1831 2012

[17] F Rousseau E Oubel J Pontabry et al ldquoBTK an open-sourcetoolkit for fetal brainMR image processingrdquoComputerMethodsand Programs in Biomedicine vol 109 no 1 pp 65ndash73 2013

[18] J Anquez E D Angelini and I Bloch ldquoAutomatic segmenta-tion of head structures on fetal MRIrdquo in Proceedings of the IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo09) pp 109ndash112 IEEE Boston Mass USA July2009

[19] V Taimouri A Gholipour C Velasco-Annis J A Estroff andS KWarfield ldquoA template-to-slice blockmatching approach forautomatic localization of brain in fetal MRIrdquo in Proceedings ofthe 12th IEEE International Symposium on Biomedical Imaging(ISBI rsquo15) Brooklyn NY USA April 2015

[20] M Ison R Donner E Weigl and G Langs ldquoFully automatedbrain extraction and orientation in raw fetal MRIrdquo in Proceed-ings of the MICCAI WorkshopmdashPAPI pp 17ndash24 October 2012

[21] K Keraudren V Kyriakopoulou M Rutherford J V Hajnaland D Rueckert ldquoLocalisation of the brain in fetal MRIusing bundled SIFT featuresrdquo Medical Image Computing andComputer-Assisted Intervention vol 16 part 1 pp 582ndash589 2013

[22] B Kainz K Keraudren V Kyriakopoulou M Rutherford J VHajnal and D Rueckert ldquoFast fully automatic brain detectionin fetal MRI using dense rotation invariant image descriptorsrdquoin Proceedings of the IEEE 11th International Symposium onBiomedical Imaging (ISBI rsquo14) pp 1230ndash1233 Beijing ChinaMay 2014

[23] A AlansaryM Lee K Keraudren et al ldquoAutomatic brain local-ization in fetal MRI using superpixel graphsrdquo in Proceedingsof the 1st International Workshop on Machine Learning MeetsMedical Imaging ICMLWorkshop (MLMMI rsquo15) vol 9487 LilleFrance July 2015

[24] N Dalal and B Triggs ldquoHistograms of oriented gradients forhuman detectionrdquo in Proceedings of the IEEE Computer SocietyConference on Computer Vision and Pattern Recognition (CVPRrsquo05) vol 1 June 2005

[25] A Serag J P Boardman R M Reynolds F C Denison GMacnaught and S I Semple ldquoAutomatic fetal brain localizationin 3T MR images using Histograms of Oriented 3D Gradientsrdquoin Proceedings of the 24th Signal Processing and CommunicationApplication Conference (SIU rsquo16) pp 2273ndash2276 ZonguldakTurkey May 2016

[26] C Harris and M Stephens ldquoA combined corner and edgedetectorrdquo in Proceedings of the Alvey Vision Conference C JTaylor Ed vol 15 pp 231ndash236 September 1988


[27] D G Lowe ldquoDistinctive image features from scale-invariantkeypointsrdquo International Journal of Computer Vision vol 60 no2 pp 91ndash110 2004

[28] P Scovanner S Ali andM Shah ldquoA 3-dimensional sift descrip-tor and its application to action recognitionrdquo in Proceedingsof the 15th ACM International Conference on Multimedia (MMrsquo07) pp 357ndash360 Augsburg Germany September 2007

[29] A Klaser M Marszałek and C Schmid ldquoA spatio-temporaldescriptor based on 3D-gradientsrdquo in Proceedings of the 19thBritish Machine Vision Conference (BMVC rsquo08) BMVA PressSeptember 2008

[30] N Buch J Orwell and S A Velastin ldquo3D extended histogramof oriented gradients (3DHOG) for classification of road usersin urban scenesrdquo in Proceedings of the 20th British MachineVision Conference (BMVC rsquo09) September 2009

[31] M Scherer M Walter and T Schreck ldquoHistograms of orientedgradients for 3D object retrievalrdquo in Proceedings of the Inter-national Conferences in Central Europe on Computer GraphicsVisualization and Computer Vision 2010

[32] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001

[33] C Yi X Yang and Y Tian ldquoFeature representations for scenetext character recognition a comparative studyrdquo in Proceedingsof the 12th International Conference on Document Analysis andRecognition (ICDAR rsquo13) pp 907ndash911 IEEE Washington DCUSA August 2013

[34] A Kembhavi D Harwood and L S Davis ldquoVehicle detectionusing partial least squaresrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 33 no 6 pp 1250ndash12652011

[35] C Premebida and U Nunes ldquoFusing LIDAR camera andsemantic information a context-based approach for pedestriandetectionrdquo International Journal of Robotics Research vol 32 no3 pp 371ndash384 2013

Submit your manuscripts athttpswwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


Figure 1 An example of a fetal MR scan at 355 weeks of gestationalage (GA) The scan shows that the acquired image has an arbitraryfetal orientation motion artifacts and low spatial resolution

that registers a template to each slice of the fetal brain in3D

Examples of machine learning approaches include themethod proposed by Ison et al [20] which is based ona two-phase random forest (RF) classifier to suppress theinfluence of maternal tissues and provide likely positionsof tissue centroids inside the brain An approximation of ahigh-order Markov Random Field (MRF) finds the optimalselection of landmarks and these landmarks along with aconfidence weighted probability map provide an estimate ofthe center of a region of interest around the brain Keraudrenet al [21] described a different learning-based approach thatdecomposes the search space by performing a 2D detectionprocess based on Scale-Invariant-Feature-Transform (SIFT)features before accumulating the votes in 3D The approachrelies on providing the gestational age as a prior knowledgeto remove the scale component

Recently Kainz et al [22] proposed a localizationmethodwhere rotation invariant feature descriptors (based on Spher-ical Harmonics) for medical volumes of the uterus arecalculated in order to learn the appearance of the fetal brainin the feature space Alansary et al [23] proposed a differentlocalization method where superpixels are first extracted anda histogram of features for each superpixel is computed usingbag of words based on dense SIFT descriptors Then a graphof superpixels is constructed and a RF classifier is trained todistinguish between brain and nonbrain superpixels

Although most previously reported methods achievedaccurate brain localization results all methods were imple-mented and tested using 15 T MR images and the general-ization of these approaches to 3 T MR images has not beenexamined Fetal MRI performed at 3 T is emerging as apromising modality for the evaluation of fetal anatomy as it

provides an increased signal-to-noise ratio (compared with15 T) This increase in the signal-to-noise ratio can allow fordecreased acquisition time increased spatial resolution or acombination of both with the overall goal of obtaining moredetailed imaging of fetal anatomy

In this article we focus on developing a novel methodfor fully automatic fetal brain localization from raw 3T MRimages with no need for preprocessing (such as intensityinhomogeneity correction) or prior knowledge about thedataset under study (such as gestational age (GA) or imagingplane) We treat the problem as a machine learning problemwhere a model is trained using a sample of positive andnegative examples From each sample discriminant featuresare extracted for both positive (brain) and negative (non-brain) examples based on a 3D extension of the success-ful Histogram of Oriented Gradients (HOG) [24] featuredescriptor A sliding window classifier based on SupportVector Regression (SVR) is used to assign a score to allpossible windows in an image and the window with thehighest score is selected

The remainder of this article is organized as follows Sec-tion 2 describes in detail the proposed method In Section 3the experimental results of applying the proposed methodto 3 T MR fetal images are presented We also compare theperformance of our method against a template matchingmethod and present its application in the optimization offetal motion correction Section 4 discusses our study andoutlines areas for prospectiveworkThis paper is an extensionof previous preliminary work [25]

2 Methods

21 Data The database for this study included a total of104 MR images from 32 singleton fetuses of healthy womenand women with diabetes between 3430 and 3760 weeks ofgestational age (mean and standard deviation of 3592plusmn083)Ethical approval was obtained from the National ResearchEthics Committee (South East Scotland Research EthicsCommittee) and written informed consent was obtained

22 MR Image Acquisition Images were acquired on aSiemens Magneton Verio 3 T MRI clinical scanner (SiemensHealthcare GmbH Erlangen Germany) T2-weighted half-Fourier acquisition single-shot turbo spin-echo images wereacquired of the fetal brain in sagittal coronal and transverseorientations where at least one stack is available in eachanatomical direction (HASTE TRTE = 180086ms FOV =400 times 400mm matrix = 192 (phase) times 256 (frequency) andvoxel size = 15 times 15 times 6mm)

23 The HOG Descriptor for 2D Images In the 2D imagedomain successful methods such as Harris-Corner detector[26] and the well-known SIFT descriptor [27] rely on aggre-gated gradients However the HOGdescriptor [24] organizesgradients into histograms As the first step the gradient imageis computed by convolving the input image with an appro-priate filter mask A grid of histograms is then constructedwhere each histogram organizes the respective gradients intobins according to their orientation To preserve locality a








nabla119891 (119909 119910 119911) = (120575119891120575119909 120575119891120575119910 120575119891120575119911)119879 (4)







end






SAD (119909 119910 119911)= 119873sum119894=0

119872sum119895=0

119871sum119896=0

Diff (119909 + 119894 119910 + 119895 119911 + 119896 119894 119895 119896) (6)




MdAE = median119894=1119899

10038161003816100381610038161198891198941003816100381610038161003816 (7)


3 Results






4 Discussion





MdA

E (m

m)

MdA

E (m

m)


10

0

3 times 3 times 3

5 times 5 times 5

7 times 7 times 7


20

10

5

0

MdA

E (m

m)


5

0

1st order

2nd order

3rd order

MdA

E (m

m)

Effect of step size

10

0

30

20

40 times 40 times 5

45 times 45 times 5

50 times 50 times 5

Effect of gradient

MdA

E (m

m)

15

10

5

0

MdA

E (m

m)


10

5

0

3

6

9

18

12

25 times 25 times 5

30 times 30 times 5

35times 35times 5

L1-norm


[minus1 0 1]





5 Conclusion



(a)

(b)


(a) (b)



Competing Interests





Acknowledgments


References










































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of























nabla119891 (119909 119910 119911) = (120575119891120575119909 120575119891120575119910 120575119891120575119911)119879 (4)







end






SAD (119909 119910 119911)= 119873sum119894=0

119872sum119895=0

119871sum119896=0

Diff (119909 + 119894 119910 + 119895 119911 + 119896 119894 119895 119896) (6)




MdAE = median119894=1119899

10038161003816100381610038161198891198941003816100381610038161003816 (7)


3 Results






4 Discussion





MdA

E (m

m)

MdA

E (m

m)


10

0

3 times 3 times 3

5 times 5 times 5

7 times 7 times 7


20

10

5

0

MdA

E (m

m)


5

0

1st order

2nd order

3rd order

MdA

E (m

m)

Effect of step size

10

0

30

20

40 times 40 times 5

45 times 45 times 5

50 times 50 times 5

Effect of gradient

MdA

E (m

m)

15

10

5

0

MdA

E (m

m)


10

5

0

3

6

9

18

12

25 times 25 times 5

30 times 30 times 5

35times 35times 5

L1-norm


[minus1 0 1]





5 Conclusion



(a)

(b)


(a) (b)



Competing Interests





Acknowledgments


References










































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















SAD (119909 119910 119911)= 119873sum119894=0

119872sum119895=0

119871sum119896=0

Diff (119909 + 119894 119910 + 119895 119911 + 119896 119894 119895 119896) (6)




MdAE = median119894=1119899

10038161003816100381610038161198891198941003816100381610038161003816 (7)


3 Results






4 Discussion





MdA

E (m

m)

MdA

E (m

m)


10

0

3 times 3 times 3

5 times 5 times 5

7 times 7 times 7


20

10

5

0

MdA

E (m

m)


5

0

1st order

2nd order

3rd order

MdA

E (m

m)

Effect of step size

10

0

30

20

40 times 40 times 5

45 times 45 times 5

50 times 50 times 5

Effect of gradient

MdA

E (m

m)

15

10

5

0

MdA

E (m

m)


10

5

0

3

6

9

18

12

25 times 25 times 5

30 times 30 times 5

35times 35times 5

L1-norm


[minus1 0 1]





5 Conclusion



(a)

(b)


(a) (b)



Competing Interests





Acknowledgments


References










































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















MdA

E (m

m)

MdA

E (m

m)


10

0

3 times 3 times 3

5 times 5 times 5

7 times 7 times 7


20

10

5

0

MdA

E (m

m)


5

0

1st order

2nd order

3rd order

MdA

E (m

m)

Effect of step size

10

0

30

20

40 times 40 times 5

45 times 45 times 5

50 times 50 times 5

Effect of gradient

MdA

E (m

m)

15

10

5

0

MdA

E (m

m)


10

5

0

3

6

9

18

12

25 times 25 times 5

30 times 30 times 5

35times 35times 5

L1-norm


[minus1 0 1]





5 Conclusion



(a)

(b)


(a) (b)



Competing Interests





Acknowledgments


References










































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a)

(b)


(a) (b)



Competing Interests





Acknowledgments


References










































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Acknowledgments


References










































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of





















of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















Documents

Histograms of Oriented 3D Gradients for Fully Automated Fetal …downloads.hindawi.com/journals/bmri/2017/3956363.pdf · 2019-07-30 · tan−1(𝑦 𝑥) cos−1(𝑧 𝑟)). (1)