Novel coarse-to-fine dual scale technique for tuberculosis cavity detection in chest radiographs

RESEARCH Open Access

Novel coarse-to-fine dual scale technique fortuberculosis cavity detection in chest radiographsTao Xu1, Irene Cheng2, Richard Long3 and Mrinal Mandal1*

Abstract

Although many lung disease diagnostic procedures can benefit from computer-aided detection (CAD), current CADsystems are mainly designed for lung nodule detection. In this article, we focus on tuberculosis (TB) cavitydetection because of its highly infectious nature. Infectious TB, such as adult-type pulmonary TB (APTB) andHIV-related TB, continues to be a public health problem of global proportion, especially in the developing countries.Cavities in the upper lung zone provide a useful cue to radiologists for potential infectious TB. However, thesuperimposed anatomical structures in the lung field hinder effective identification of these cavities. In order toaddress the deficiency of existing computer-aided TB cavity detection methods, we propose an efficientcoarse-to-fine dual scale technique for cavity detection in chest radiographs. Gaussian-based matching, local binarypattern, and gradient orientation features are applied at the coarse scale, while circularity, gradient inversecoefficient of variation and Kullback–Leibler divergence measures are applied at the fine scale. Experimental resultsdemonstrate that the proposed technique outperforms other existing techniques with respect to true cavitydetection rate and segmentation accuracy.

Keywords: Classification, Segmentation, Computer-aided detection (CAD), Tuberculosis (TB)

1. IntroductionChest radiographs or chest X-ray (CXR) images arewidely used to diagnose lung diseases such as lung can-cer, tuberculosis (TB), and pneumonia. Due to the super-imposed anatomical structures in the human chest, theCXR images are generally noisy and the diagnosisrequires careful examination by experienced radiologists.Computer-aided detection (CAD) systems in chest radi-ography have therefore been developed to reduce theworkload of radiologists. Ginneken et al. reviewed theCAD technological development in 2001 [1] and 2009[2]. Developing a single system that looks into all abnor-malities on a chest radiograph is practically impossibledue to the widely different characteristics of abnormal-ities, and specific focus of the image processing algo-rithms. Therefore, the current CAD systems often aimat a single aspect, e.g., detection of lung cancer nodules.This strategy has been proved to be successful, and

many effective algorithms have been developed for rou-tine diagnostic procedures [2].A general CAD system framework is shown in Figure 1.

There are four modules in the system. First, a CXRimage undergoes the preprocessing step, which gener-ally includes image enhancement, noise removal, andlung field segmentation. In the next step, candidatesthat may contain abnormalities are coarsely detectedusing pattern recognition techniques. In the thirdstep, features that can be used to identify abnormal-ities are identified from the candidates. Depending onthe radiographic manifestation of the abnormalities,these features could be geometric, photometric, ortextural. Finally, a classifier is applied to perform ahigh-level screening to reduce the false positive rate.An efficient CAD system relies on robust image pro-cessing, pattern recognition, and artificial intelligencetechniques. For instance, a recent CAD system [3]designed for identifying lung nodules uses an activeshape model for lung field segmentation, followed bya weighted multi-scale convergence-index filter fornodule candidates detection. To identify the nodulessuccessfully, an adaptive distance-based threshold

* Correspondence: [email protected] of Electrical and Computer Engineering, University of Alberta,Edmonton, AB T6G 2V4, CanadaFull list of author information is available at the end of the article

© 2013 Xu et al.; licensee Springer. This is an Open Access article distributed under the terms of the Creative CommonsAttribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproductionin any medium, provided the original work is properly cited.

Xu et al. EURASIP Journal on Image and Video Processing 2013, 2013:3http://jivp.eurasipjournals.com/content/2013/1/3

mailto:[email protected]

http://creativecommons.org/licenses/by/2.0

technique is applied to segment the contour of eachcandidate. The geometric, intensity, and gradient fea-tures are then extracted from the segmentation results.After the first level screening, a Fisher linear discrimin-ant classifier is used on a subset of these features toperform the final detection.Nodule detection has been the main focus in current

CXR CAD systems. However, as Ginneken et al. pointedout [2], there are other diseases, e.g. TB, that rely heavilyon chest radiograph examination can benefit from theCAD systems. Infectious TB is still a public health prob-lem in many countries [4]. Therefore, our research focusis on developing a CAD system for the diagnosis of in-fectious TB. The TB can be identified based on differentradiographic patterns, such as cavity, airspace consolida-tion, and interstitial opacities [5]. A few existing CADsystems use texture analysis to detect interstitial changes[2]. However, the interstitial pattern is not a reliableradiographic cue for infectious TB. According to a re-cent research article on TB [6], cavitation in the upperlung zone (ULZ) is a typical radiographic feature ofAPTB. So far, insufficient research has been done for ef-ficient detection of TB cavities. Shen et al. [7] recentlyproposed a hybrid knowledge-guided (HKG) frameworkfor TB cavity detection, which contains three majorsteps. In Step 1, the cavity candidates are detected usingadaptive thresholding on the mean-shift clustered CXRs.In Step 2, a segmentation technique is applied to thecandidates to generate contours of important objects

present in the CXR image. In Step 3, the contour-basedcircularity and gradient inverse coefficient of variation(GICOV) features are extracted for the final cavity clas-sification using a Bayesian classifier. Although, this tech-nique provides a good performance, it has severallimitations. First, due to cavity size variation and the oc-clusion from neighboring superimposed anatomicalstructures, the mean shift cluster result is sensitive tothe parameter values used. Second, the adaptive thresh-old, which is a quadratic polynomial of GICOV score,does not perform well when the cavity boundary is weak.These two limitations lead to a high missing rate (MR)of true cavities. To overcome these problems, wepropose a dual scale feature classification strategy for TBcavity detection in chest radiographs. First, a coarse fea-ture classification step is performed to detect the cavitycandidates by capturing the geometric, textural, and gra-dient features in the lung field. Second, a Hessianmatrix-based technique is applied to enhance the cavitycandidates, which leads to a more accurate contour seg-mentation. Finally, fine features based on the shape,edge, and region are extracted from the segmented con-tours for the final cavity classification. Experimentalresults show that the performance of the proposed can-didates detection, segmentation, and cavity classificationmodules is superior compared to the results obtainedusing other related CAD systems.The rest of this article is organized as follows. Section 2

explains the cavity pattern in CXRs. Section 3 describesour proposed method in detail. Section 4 reports andanalyzes the performance of the proposed technique.Conclusion and future work are presented in Section 5.

2. Manifestation of cavity in chest radiographsIn chest radiography, a cavity is typically defined as aparenchymal cyst greater than 1 cm in diameter, con-taining either air or fluid or both [5]. Since the cavitiesare created by tissue necrosis within nodules or masses,their radiographic features are usually demonstrated asannular rings with variable wall thickness. Figure 2ashows a CXR image with a typical cavity (inside the

CXR

Image

Detected

Abnormality

PreprocessingCandidates

Detection

Feature

Extraction

Feature

Classification

Figure 1 The processing steps of a CAD system in chestradiology.

(c)(b)(a)

Figure 2 Occlusion of cavities in chest radiographs (in the red rectangle).

Xu et al. EURASIP Journal on Image and Video Processing 2013, 2013:3 of 18http://jivp.eurasipjournals.com/content/2013/1/3

rectangle region), which manifests as a focal lucent areaon the image and appears as a “hole” in the patient’s leftupper lung zone. However, these holes might be blurreddue to the overlapping projection of anatomical struc-tures or some other abnormalities in the neighborhood,which makes the identification of cavities a difficult taskfor radiologists. Figure 2b is another example of a TBcavity obscured by the left clavicle. Figure 2c shows anexample where the cavity is overlapped with interstitialopacities.

3. Proposed techniqueComputer-aided feature identification in CXR imagesis comparatively more challenging than feature identi-fication in medical images of other body parts becauseof the rib cage and other superimposed anatomicalstructures in the lung field as illustrated in Figure 2.After examining the geometric, textural, and photo-metric characteristics of TB cavities, we propose acoarse-to-fine feature classification technique for cav-ity detection. Figure 3 shows a schematic of the pro-posed technique. It is observed that there are threemajor steps: (i) coarse feature classification, (ii) con-tour segmentation, and (iii) fine feature classification.A CXR image is first divided into patches. In the firststep, a coarse feature classification is performed oneach image patch to identify candidates which aresuspected to contain cavities. Two modules are usedto capture the coarse features: Gaussian-model-basedtemplate matching (GTM), and local binary pattern(LBP) and histogram of oriented gradient (HOG)basedfeature classification (LHFC). In the second step, con-tours of the chosen candidates are segmented usingtwo modules: Hessian-matrix-based image enhance-ment (HIE) and active contour-based segmentation(ACS). The HIE is used to boost the cavity edges. Theedge-based ACS is then applied to segment theenhanced images. In the third step, a contour-based

feature classification (CFC) module is applied. Fine fea-tures including shape, edge, and region are extractedfrom the contours. Cavity classification is then per-formed based on these features. A detailed descriptionof these five modules is presented in the followingsections.

3.1. GTMThe template matching (TM) is a widely used tech-nique in pattern recognition, where the presence of apattern in an image is detected by comparing differentparts of an image with a reference pattern known astemplate. In many TM techniques, instead of compar-ing a given template directly, a transformation of thetemplate is matched with similar transformation of acandidate region using a similarity measure. Normalizedcross correlation is often used to measure similarity be-cause of its fast implementation using the fast Fouriertransform. Since traditional TM is sensitive to rotationand scale, rotation and scale invariant transform suchas Fourier–Mellin transform [8], or ring-projectiontransform [9] can be incorporated into TM. However,these transforms provide good results only when a cav-ity shape/size deviates very little from the templateshape/size. To avoid missing true cavities, a solution isto use a large set of templates covering different cavitysizes and rotation angles.Using a large set of templates can be computationally

expensive but still cannot guarantee to detect all cavities.Therefore, the proposed technique makes use of priorknowledge given by TB experts to generate a customizedtemplate database specific for TB cavities. Observe thatin the “hole” like cavity shown in Figure 2a, line-cut in-tensity profiles in various directions of the cavity regionappear similar. Figure 4a shows the magnified region ofa cavity, and Figure 4b–e shows plots of the four inten-sity lines passing through the image center at 0°, 45°,90°, 135°. Each line’s intensity profile appears as a bi-modal Gaussian function. Based on the similarity ofthese intensity profiles, it is reasonable to mimic the cav-ity pattern using rotationally symmetric pattern such as2D circular or elliptical Gaussian ring distribution (asshown in Figure 4f ). Note that if a line-cut intensity pro-file of Figure 4f is calculated, a bimodal Gaussian distri-bution is obtained where the two major peakscorrespond to the two sides of the ring.A generic 2D Gaussian ring is defined as follows

I x; yð Þ ¼ e�1�wð Þ2 x2þy2ð Þ

2ξ2 ð1Þ

where w ¼ abffiffiffiffiffiffiffiffiffiffiffiffiffiffiffia2y2þb2x2

p , a and b are the two radii (distance

between the origin and the peaks on x,y axes), I(x,y) isthe image intensity function in the 2D domain, and ξ is

Contour Segmentation

Coarse Feature Classification

Fine Feature Classification

CXRImage

GTM Cavity

Candidates

DetectedCavities

LHFC

CFC ACS

HIE

Contours ofCavity

Candidates

Figure 3 Schematic of the proposed CAD framework. It containsthree major steps, which are built upon five modules: GTM, LHFC,HIE, ACS, and the CFC.


the standard deviation of the Gaussian distributionwhich determines the wall thickness of the ring. Notingthat when a = b = r, Equation (1) represents a 2D circu-lar Gaussian ring, where r is the inner radius. Rotatedpatterns can be generated by incorporating a rotationangle θ into the following coordinate transformation:

x ¼ x0 cosθ þ y0 sinθy ¼ y0 cosθ � x0 sinθ

�ð2Þ

where x’, y’ are the pixel’s location before rotation. UsingEquations (1) and (2), the template database can be builtwith various sizes, wall thicknesses, and rotation anglesby changing the value of parameters a, b, ξ, and θ. Forexample, given a 512 × 512 CXR image with a pixel spa-cing [0.8 mm, 0.8 mm], the physical size represented bythe image is 40.96 × 40.96 cm2. Since the diameter ofthe largest cavity is usually less than 6 cm, we define thetemplate size as 75 × 75. While the wall thickness is

within the range of [4 mm, 16 mm], parameter ξ is var-ied from 5 to 20 pixels. Figure 5 shows a set of tem-plates, with various radii, rotation angle, and wallthickness, used in this article.

3.2. LHFCAlthough the proposed GTM module works well forcavities of typical shape and intensity, it is difficult to de-tect cavities obscured by anatomical structures or someother abnormalities in the lung field. To address thisissue, we combine the LBP and HOG features, whichhave been shown to be useful in human detection inhandling partial occlusion [10]. The LBP [11] is a hybridtexture feature widely used in image processing. It com-bines the traditionally divergent statistical and structuralmodels of texture analysis. The LBP feature has somekey advantages, such as its invariance to monotonic graylevel changes and computational efficiency. The HOGfeature [12], similar to Lowe’s scale-invariant feature

Figure 5 An example of cavity templates. a/b < 1.6, wall thickness σ within [6,20], and θ = 0°, 45°, 90°, 135°.

(a) (b) (c) (d) (e) (f)

Figure 4 Line-cut intensity profile analysis of ‘hole’ like cavity region. (a) A cavity region; (b–d) line-cut intensity profile in four directions;(e) customized template for mimicking the cavity pattern.


transform feature, is regarded as an excellent descriptorto capture the edge or local shape information. It has agreat advantage of being robust to changes in illumin-ation or shadowing. These two features are expected tocomplement well the GTM technique, especially inblurred regions containing cavities, to detect TB cavitycandidates.In the LHFC module, a feature vector, which combines

the LBP and HOG features, is calculated for each candi-date window. The feature vector is then fed to a classi-fier, which is trained offline using ground-truth (cavityand non-cavity) training data. The classifier will assessthe windows as cavity candidates (positive samples) ornot (negative samples). The candidate windows are gen-erated using a sliding-window paradigm where an imageis scanned from the top left to the bottom right withoverlapping rectangular sliding windows. The windowsare scanned row wise. The window size is consistentwith the template size in GTM, i.e., each window has asize of 75 × 75. The overlap between two consecutivewindows is 2/3 of the window size.The computation of these two features and the classifi-

cation using support vector machine (SVM) [13] areexplained in the following sections.

3.2.1. Computation of the LBP featureIn this article, the LBP feature vector for a window iscalculated in three steps. In Step 1, explained in Figure 6,the LBP values are calculated by applying the LBP label-ing on each pixel. Here, each pixel in the window iscompared to each of its eight neighbors. The LBP valuefor the pixel is then calculated as follows

LBPP;R ¼XP�1

p¼0

u gp � gc� �

2p Note : u xð Þ ¼ 1 if x≥00 otherwise

� ��

ð3Þwhere gp, gc are gray levels of the neighborhood pixelsand center pixel, respectively, and u(·) is the unit-stepfunction. For a window of 75 × 75, there will be 5,625LBP values, with dynamic range between 0 and 255. InStep 2, an LBP-histogram, with 256 bins, is generatedfor the window from the 5,625 computed LBP values.

Finally, in Step 3, to reduce the dimensional numbers offeatures, we adopt a popular approach used in textureanalysis, e.g., [14], by calculating the six statistical fea-tures (mean, standard deviation, smoothness, skewness,uniformity, and entropy) based on the LBP histogram.Figure 7b shows the six LBP features calculated from theimage window shown in Figure 7a.

3.2.2. Computation of HOG featureFor computational convenience, we first resize each 75 ×75 image window into a 64 × 64 window using bicubicinterpolation. The HOG feature for each resized windowis then calculated as follows.

Step 1.Gradient computation: The gradient of eachpixel in the window is calculated using two fil-ter kernels: [−1, 0, 1] and [−1, 0, 1]T. Let themagnitude and orientation of the gradient ofthe ith pixel (1 ≤ i ≤ 4096) be denoted by mi

and ϕi, respectively.Step 2.Orientation histogram: Each window is first

divided into non-overlapping cells of equal di-mension, e.g., a rectangular cell of 8 × 8. Theorientation histogram is then generated byquantizing ϕi into one of the nine major orien-tations: 2k�1ð Þπ

9 � π9 , 1 ≤ k ≤ 9. The vote of the

pixel is weighted by its gradient magnitude mi.Thus, a cell orientation histogram Hc is a vec-tor with dimension of 1 × 9.

Step 3. Block normalization: In order to account forchanges in illumination and contrast, the cellhistogram must locally be normalized, whichrequires grouping the cells together into larger,spatially connected blocks. The block size weuse is 2 × 2 cells (i.e., 16 × 16 pixels), and theoverlap between two neighboring blocks is 1/2of the block size. Therefore, a whole windowcontains 49 blocks. The block divisions for awindow image are shown in Figure 8. The fea-ture vector of one block Hb is concatenated byfour cell histograms: Hb = [Hc1 Hc2 Hc3 Hc4].Note that the orientation histogram of a blockHb is a vector with a dimension of 1 × 36. The

Figure 6 An example of calculating LBP values in an eight-neighbor cell.


normalized HOG vector is then calculated asfollows [12].

H b ¼ Hb

Hbk k ð4Þ

where ‖.‖ represents the L2 norm.

The HOG feature vector of an image window (with 49blocks) is a concatenated vector of all 49 normalizedblock orientation histogram (H b), and will have a dimen-sion of 1 × 1764 in our case. Figure 7c shows the plot ofthe HOG feature vector of the image window shown inFigure 7a.Combining the LBP and HOG features, a feature vector

of size 1 × 1770 is obtained for each image window. Thesefeatures vectors are fed to the SVM classifier, explained inthe following section, for cavity candidates detection.

3.2.3. Classification using SVMAlthough SVM can perform both linear and nonlinearclassifications, the basic SVM is a non-probabilistic binarylinear classifier [13]. It is commonly used in machinelearning as a supervised learning technique for recogniz-ing patterns. Our goal is to use a pattern’s feature vectorsto identify which class it belongs to. The classification de-cision is based on the value of a linear combination ofthese feature vectors. Researchers use SVM classifiers inapplications because of its efficiency in handling both lin-ear and nonlinear classification problems. Once the separ-ating hyperplane is obtained after the training step andthe classification accuracy is satisfied, the given task (data)could linearly be separated in a high-dimensional featurespace using this hyperplane.For two-class classification, the optimal separating

hyperplane in SVM to separate two sets of data in afeature vector space is defined by w→:x→ þ b ¼ 0, wherex→ is the feature vector space, w→ is the normal vector tothe hyperplane, and b is the offset of the hyperplanefrom the origin. Given M training feature vectorsx→k ; 1≤k≤Mgf , and the corresponding ground-truth

classification result {yk ∈ [1, −1], 1 ≤ k ≤ M}, the opti-mal hyperplane coefficients vector w→ is generated asfollows

min12

→w�� 2

; s:t:yk Γ →w;→xk

þ b

h i≥1; 1≤k≤M ð5Þ

where Γ(·) denotes a kernel function [13]. Linear, polyno-mial, radial basis function (RBF), and sigmoid are widelyused as SVM kernels. In our tasks, we use the RBF kernelfunction which performs better than other kernels.The SVM training builds a model that is able to distin-

guish the belonging class of any future data based on the

Figure 7 An example of the LBP and HOG features. (a) An image window containing a cavity; (b) six LBP features corresponding to (a); (c) theHOG feature vector (1 × 1764) corresponding to (a).

Figure 8 The block and cell divisions in a window image.Letters b and c stand for a block and a cell, respectively.


support vectors obtained by the training dataset. Anynew feature vector x→i is classified according to the out-put of the decision function:

f →xi� � ¼

XMk¼1

αkykΓ→xk ;

→xi� � þ b ð6Þ

where αk is the Lagrange multiplier. If f →xi� �

≥0 , it

means →xi belongs to class y = 1, and if f →xi� �

< 0, it

means →xi belongs to class y = −1.An example of cavity candidate detection using GTM +

LHFC (note: LHFC includes the LBP and HOG features) isshown in Figure 9. Figure 9a shows the original CXR

image, and Figure 9b shows three detected TB cavity can-didates, C1, C2, C3. The magnified images of these candi-dates are also shown in Figure 9c. To eliminate the falsepositive candidates (C1 and C3), further contour segmenta-tion and fine feature classification are necessary.

3.3. HIEAs shown in Figure 9b, the GTM + LHFC detects a largenumber of cavity candidates some of which may be falsepositives (e.g., C1 and C3 shown in Figure 9b). In thissection, we present a technique to enhance the cavityfeature in a candidate, which will help in reducing thenumber of false positives. In order to reduce the effect

(a) (b)

(c)

(d)

(e)

(f)

Figure 9 An example of cavity candidates detection using the proposed technique. (a) Original CXR image; (b) candidate detection resultsin ULZ obtained using GTM + LHFC where the green rectangular windows (C1, C2, C3) represent the candidates, and the blue dotted contour isthe true cavity annotated by radiologists; (c) magnified candidate windows: C1–C3 (left to right); (d) HIE results of C1–C3; (e) improved fluidvector flow (IFVF) results of C1–C3 with the help of HIE; (f) final cavity detection results using fine feature classification. Red contour is thedetected cavity, while the cyan ones are the non-cavity contours.


of noise and irrelevant anatomical structures or abnor-malities, we apply the HIE to enhance the candidates.Note that the Hessian matrix has been applied in the lit-erature to enhance local patterns such as plate-like, line-like, or blob-like structures [15]. The proposed HIE hasthree steps, which are described below.

Step 1. Laplacian of Gaussian smoothed image: In thisstep, three Laplacians (in three directions) of aGaussian smoothed image, at scale σ, areobtained by convolving a cavity candidate withthe second derivative of Gaussians as follows.where I(x, y) is the candidate and G is the Gaussiankernel. Note that for a candidate of size 75 × 75,each of the three L matrices in Equation (7) willhave a size of 75 × 75. Figure 10 shows the secondderivative of a 1D Gaussian kernel. The intrinsiccharacteristic of this analysis is that the secondderivative of the Gaussian kernel at scale σ generatesa probe kernel that measures the contrast betweenthe regions inside and outside the range (−σ, σ) inthe direction of the derivative.

Lxx x; y; σð Þ ¼ σ2I x; yð Þ∗Gxx x; y; σð ÞLxy x; y; σð Þ ¼ σ2I x; yð Þ∗Gxy x; y; σð ÞLyy x; y; σð Þ ¼ σ2I x; yð Þ∗Gyy x; y; σð Þ

8<: ð7Þ

Step 2.Hessian matrix calculation: For a given σ value,the Hessian matrix corresponding to pixel(xi, yi) in the candidate is calculated asfollows

Hσ xi; yið Þ ¼ Lxx xi; yi; σð Þ Lxy xi; yi; σð ÞLyx xi; yi; σð Þ Lyy xi; yi; σð Þ

� �

ð8Þ

where Lxy(xi, yi, σ) = Lyx(xi, yi, σ). A knownproblem of multi-scale analysis using Hessianmatrix is that over-blurring can occur duringthe multi-scale smoothing, which may in-crease false detections [16]. Therefore, in thisarticle, we set the σ value equal to the objectscale calculated using the method described in[17]. The object scale at every pixel is definedas the radius of the largest hyperball centeredat the pixel such that all pixels within the ballsatisfied a predefined image intensity homo-geneity criterion. Object scale represents thegeometric information (size) of the local struc-ture. Object scale at the center of a blob-likestructure is approximately equal to the radiusof the blob in pixel size.

Step 3. Image enhancement using eigenvalues of Hessianmatrix: The pixel (xi, yi) in the candidate withintensity I(xi, yi) is enhanced using the followingequation:

IE xi; yið Þ ¼ λ1j jI xi; yið Þ ð9Þwhere λ1 and λ2 are eigenvalues of Hσ(xi, yi),and |λ1| ≥ |λ2|. The intuition in Equation (9) ofusing only the largest eigenvalue for cavity en-hancement is based on the fact that the Hessianmatrix has a strong edge effect (for those strongedge points, |λ1| >> |λ2| ≈ 0) [18]. Althoughcavities are usually embedded in noisy sur-roundings due to the neighboring necrosiscaused by cavitation, the inside of a cavity (filledwith air or fluid or both) still has lower intensitythan the background. Thus, the strong edge be-tween the inside and outside of a cavity gives agood clue to indentify the contour of cavity. Dif-ferent techniques of edge enhancement wereevaluated in this study, such as contrast-limitedadaptive histogram equalization [19], fuzzy Cmeans [20], and speckle reducing anisotropicdiffusion technique [21], and the proposed HIEtechnique achieves the best performance.The enhanced window candidates C1–C3 areshown in Figure 9d. It is observed that theannular ring-like structure is greatly enhanced.

3.4. ACSActive contours or deformable models are generallydivided into two types: parametric active contours (typ-ically known as snakes) and geometric active contours(level set). The snake-based techniques are often fasterthan level sets in virtue of efficient numerical methods.In addition, the level sets produce more false detectionsdue to its multiple objects capturing ability. Therefore,in this article, we use a snake-based technique known as

Figure 10 The second derivative of a 1D Gaussian kernelprobes inside/outside contrast of the range (−σ, σ). In this

example, Gxx xð Þ ¼ x2σ4 � 1

σ2

e�

x2

2σ2 , σ = 1.


IFVF [22]. In this technique, a snake contour repre-sented by v evolves through the candidate window toreach a force balance equation Fint(v) + Fext(v) = 0,where Fint(v) is the internal force constraining contour’ssmoothness, and Fext(v) is the external force attractingthe contour toward image features.The IFVF is a fast and accurate edge-based snake tech-

nique, because of the introduction of both static and dy-namic terms in the external force.

Fext vð Þ ¼ Fstatic vð Þ þ Fdynamic vð Þ ð10ÞThe Fstatic could be a static external force which over-

comes the edge leakage problem, e.g., we use boundaryvector flow (BVF) proposed in [23] as the Fstatic. The BVFextends the capture range further to the entire image basedon simpler interpolation. Four potential functions Ψx, Ψy,Ψxy, and Ψyx are computed using line-by-line interpola-tions in the horizontal, vertical, and two diagonal direc-tions. The Fstatic is calculated as follows

Fstatic ¼ Φ1 ¼ ∇Ψ x;∇Ψ y� �

or

Fstatic ¼ Φ2 ¼ffiffiffi2

p

2∇Ψ xy þ ∇Ψ yx

� �;

ffiffiffi2

p

2∇Ψ xy � ∇Ψ yx

� ��

ð11Þ

The Fdynamic is achieved in three steps.

1. Given an HIE-enhanced candidate image, a binaryedge map B is generated using smoothing techniquespeckle reducing anisotropic diffusion [21] and theCanny edge detector [24].

2. By comparing the edge map points to the currentsnake contour points (snaxels), a new control point(xc,yc) is selected by considering the point which con-tributes more to the distance between snake contourand object boundary [22]. We use the Hausdorff dis-tance to find such a point. Assuming two sets ofpoints S and O, the Hausdorff distance is thendefined as h S;Oð Þ ¼ maxo∈O mins∈S d s; oð Þf gf gwhere d(s,o) is the Euclidean distance between asnaxel s and a object boundary point o. So, the con-trol point is chosen as the point on the object bound-ary which has the Hausdorff distance value.

3. For any pixel (x,y) on the contour v, its Fdynamic(x,y)is then calculated as follows

Fdynamic x; yð Þ ¼ 1 � Bð Þδ ∇d0 x; yð Þ∇d0 x; yð Þk k ð12Þ

where δ = ±1 controls the outward or inward direc-tion. In this article, we use δ = 1, as the initial con-tour is automatically set as a small circle in thecenter of the window image with radius of 3 pixels.d’(x,y) is the Euclidean distance between points (x,y)

and (xc,yc). Note that the term (1 − B) makes the Fdy-namic zero for those points which already reach edges.Based on the edge map generated from the enhancedcandidates images using HIE, the IFVF segmentationresult of these candidates C1–C3 are shown inFigure 9e. The stopping criterion of the evolution isdetermined by computing the difference in locations(defined by the x and y coordinates) of the corre-sponding contour points between two consecutiveiterations. If it is less than a convergence threshold t,the active contour evolution will be stopped. In ourexperiments, t is empirically set to 0.05. Based on ourtests, there is no significant improvement even if t issmaller than 0.05.

3.5 CFCThe last module in our proposed technique is the CFC,which performs the fine scale feature classification.Three types of contour-based features, shape, edge andregion, are extracted for the final cavity detection. Thesefeatures include circularity measure [25], GICOV [26],and Kullback–Leibler divergence (KLD) [27] betweenthe pixel intensity distributions inside and outside thecontour. The computations of these three features areexplained below.

1. Assuming a contour has one centroid, L points areselected from the contour in L cardinal directions.The circularity of the contour is then calculated asscaled variance as follows

C ¼ var d xi; yið Þð Þmax d xi; yið Þð Þ ; i ¼ 1; 2; . . . ; L ð13Þ

where d(xi,yi) is the distance from the centroid tothe contour point (xi,yi) in the ith direction. In thisarticle, we use L = 16. The circularity feature is afeature which could effectively reduce the falsepositives.

2. Based on the observation that the inner boundary ofa cavity often has dark-to-bright transition, theGICOV value of L points on the contour is calculatedas follows

(a) For the contour point (xi,yi) in the ith direction, itsgradient in normal direction gn(xi,yi) is calculated asgn xi; yið Þ ¼ ∇I xi; yið Þ:n→ xi; yið Þ , where n

→xi; yið Þ is

the unit outward normal vector at this point.(b)The mean and standard deviation of gn, denoted by

m and s, are then calculated as m ¼ 1L

XL

i¼1

gn xi; yið Þand s2 ¼ 1

L�1

XL

i¼1

gn xi; yið Þ � mð Þ2


(c) The GICOV value of the contour is finallyachieved using following equation:

GICOV ¼ m

s=ffiffiffiL

p ð14Þ

3. Given the probability distributions, P and Q, of thepixel intensity values inside and outside the cavity,respectively, the KLD for a candidate window is cal-culated as follows

KLD ¼XB

i¼1

P ið Þ ln P ið ÞQ ið Þ ð15Þ

where B is the number of bins in the histogram spanby P and Q. The KLD compares the difference ingray level distribution between the pixels inside andoutside the contour.

Table 1 shows the above feature values corresponding tothree contours shown in Figure 9e. As in the coarse featureclassification step, we select the SVM as the fine featureclassifier in this step. Based on the feature values (Table 1),the trained SVM classifier identifies the Contour-2 as apositive and Contour-1 and Contour-3 as negatives. Thefinal detected cavity (corresponding to Contour-2) in theCXR image is shown in Figure 9f as the red contour. Theresult matches with the ground truth.

4. Performance evaluationIn this section, we evaluate our proposed coarse-to-finedual scale technique with respect to three aspects: the ef-fectiveness of candidate selection, the accuracy of contoursegmentation, and the accuracy of final cavity detection.

Table 1 Fine feature values of three contours in Figure 9e

Circularity GICOV KLD

Contour-1 0.11 15.33 1.49

Contour-2 0.15 13.68 2.28

Contour-3 0.69 15.26 0.28

Figure 11 Sample histograms of cavity properties. (a) Histogram of diameter; (b) histogram of circularity; (c) histogram of wall thickness offour categories: “Thick” (≥16 mm), “Intermediate” (4–15 mm), “Thin” (<4 mm), and “Uncertain” (wall not discernible).


4.1. Experimental dataset and parameters configurationA cavity dataset of 35 CXR images containing 50 cavitiesis obtained from the University of Alberta Hospital. Allthe images were independently read by three experi-enced chest radiologists who are specialized in TB diag-nosis. The presence of TB cavities was confirmed by theagreement of at least two radiologists. The sample histo-grams of cavity properties such as diameter, circularity,and wall thickness are shown in Figure 11. From the his-tograms, it can be seen that the cavities vary in dia-meters while their circularities range mainly from 0.15to 0.2 and most of them have intermediate thickness.For computational efficiency, the original CXR imagesare resized as 512 × 512 (or close to this size) with afixed pixel spacing [0.8 mm, 0.8 mm]. Since all the cav-ities are located in the ULZ, a similar preprocessing pro-cedure as described in [7] was applied to segment thetarget lung region, which reduces the processing area toa smaller rectangular bounding box. Figure 12 shows anexample of the target area.The proposed cavity detection technique is implemen-

ted in MATLAB 2007b on an Intel Pentium 4 CPU 2.8GHz with 2 GB RAM computer. All the parameters inthe proposed technique are listed in Table 2. The SVMclassifiers in both coarse and fine feature classificationare built using LIBSVM software [28]. To train the SVMclassifiers, we applied the ‘leave-one-out’ method [29]since the size of samples with cavities is small. For ex-ample in LHFC, to detect the candidate regions in oneof the 30 CXR images, we use the remaining 29 CXRimages for the training. The training set contains the

LBP and HOG feature vectors extracted from windowswith and without cavities (positive and negative samples)in these 29 CXR images. Note that the negative samplesfor training were selected from the contralateral positionof the positive samples based on the approximate sym-metry of the lung field. The SVM classifier in CFC istrained in a similar way.

4.2. Effectiveness of candidate selectionThe proposed coarse feature classification technique forcandidate detection is evaluated by the MR, which is cal-culated as follows

MR ¼ of Cavities Excluded from CandidatesTotalof True Cavities

� 100%

ð16Þ

A preliminary experiment using only GTM for candi-date detection has already been reported in [30]. We an-ticipate that by integrating with other novel techniques,a better result can be obtained. Thus, we used differentcombinations of LBP and/or HOG features together withGTM, and checked whether the MR could be reduced.Table 3 shows our test results.From the results, we observe that the HKG frame-

work for TB cavity detection [7] missed more cavitiesthan our proposed approach. HKG is based on anadaptive thresholding on the mean-shifted clusteredimage for candidate detection. Its high MR is due totwo reasons. First, the mean-shift clustering approxi-mates nearest neighbors intensities and space infor-mation but neglects the texture. Second, the adaptivethreshold, which is a quadratic polynomial of theGICOV feature, is not suitable for modeling all

Figure 12 An example of the target area. The enhancedsubimage inside the green rectangle is the result of thepreprocessing procedure.

Table 2 Parameters configuration in the proposedtechnique

Modules Parameters names Parameters values

GTM Template size 75 × 75 pixels

Wall thickness σ [6,20]

Aspect ratio a/b [1, 1.6]

Rotation angle θ {0°,45°,90°,135°}

LHFC Window size 75 × 75 pixels

Cell size 8 × 8 pixels

Block size 2 × 2 cells

Block overlap 0.5

SVM parameters Default values in LIBSVMsoftware [28]

ACS Snake evolution direction δ 1

Convergence threshold t 0.05

CFC SVM parameters Default values in LIBSVMsoftware [28]


shapes, especially when the boundary of a cavity isweak. Figure 13 compares the detection results ofHKG and our technique. The green boxes representcavity regions reported by the classifier. In Figure 13a,HKG cannot identify both cavities due to the failureof mean-shift clustering in the noisy ULZ. Our tech-nique is able to identify the two cavities (Figure 13b).Figure 13c is yet another example showing the adap-tive threshold value used in HKG unable to identifythe cavity. However, our technique is able to detectall cavities correctly (Figure 13d).Using the same parameter values for LBP and HOG

as in the literature, we found that a combination ofLBP and HOG together with GTM achieved better

performance. Our finding is consistent with the resultsin human detection using LBP and HOG features [10].HOG performs poorly when the background is clut-tered with noises. LBP is able to alleviate this defi-ciency. It can filter out noises following the uniformpattern estimation. However, if LBP is used alone with-out HOG, the entire ULZ will be extracted if someother abnormalities are also present in the area. In thatcase, the HOG helps to reduce the false positives basedon the available edge information. Figure 14 illustratesthe complementary effect of LBP and HOG. The win-dow reported by the classifier should contain acomplete cavity in order to be qualified as a positivecandidate. Note that in the first row second column

Table 3 Candidates detection results

HKG [7] GTM [30] GTM + LBP GTM + HOG GTM + LBP + HOG

Number of cavities 50 50 50 50 50

Number of candidates 170 164 315 229 160

Number of missing cavities 22 18 10 17 7

MR (%) 44 36 20 34 14

(a) (b)

(c) (d)

Figure 13 Comparison of candidates detection between HKG [7] and the proposed technique; (a, c) the results of HKG, (b, d) generatedfrom the proposed technique. Green regions in the images are cavity candidates regions reported by the classifier, and blue dotted contours arethe true cavities annotated by radiologists.


when using only HOG, no reported window contains acomplete cavity. The HOG performs poorly when thebackground is cluttered with noises, and the edge infor-mation is no longer reliable. Similarly, in the secondrow first column, when using only the LBP, the smallcavity is missing because no reported window containsthe complete small cavity, and only the larger cavity isfully contained in a reported window.The above test results show that combining the

LBP and HOG features for capturing the texture andgradient information around the cavity region, andusing the GTM for shape recognition, contributes tothe low MR of the proposed coarse feature classifica-tion technique.

4.3. Accurate contour segmentationWe evaluate segmentation accuracy using the followingTanimoto measure (TMM) [7]:

TMM ¼ Rc \ Rg

�� Rc [ Rg

�� ð17Þ

where Rc denotes the region enclosed by the contourgenerated by the segmentation techniques, such asDBC-GVF [7] and our IFVF [22]; Rg denotes the region

of a TB cavity that is enclosed by the ground-truth con-tour manually drawn by radiologists; and ‖.‖ denotes thecardinality (number of pixels). TMM = 0 indicates thatthe segmented contour has no intersection with theground truth, while TMM = 1 indicates that the segmen-ted contour is identical to the exact cavity. To improvethe segmentation accuracy, we apply the HIE on thecandidates before segmentation.The performance of the DBC-GVF and the IFVF tech-

niques with and without the HIE is shown in Table 4.Note that around 10% accuracy improvement is achievedfor both DBC-GVF and IFVF when HIE is incorporated.The results are also more robust as demonstrated by thelower standard deviations of the TMM. Figure 15 pre-sents subjective comparison of different segmentationtechniques. With the HIE, the segmented contours arecloser to the ground truth compared to the same techni-ques without the HIE.Note that image patterns, even without cavities, may

generate close to ring-like shape after the HIE step.Figure 16 shows some of these cases. For example, theimage in the bottom row contains a pattern similar toa cavity. To eliminate this type of candidates, the finescale feature classification step in our approach is ne-cessary. The accuracy of our final cavity detection isevaluated in the next section.

GTM+LBP GTM+HOG GTM+LBP+HOG

(a) (b) (c)

(d) (e) (f)

Figure 14 Comparison of candidate detection in the coarse feature classification step using (a, d) GTM + LBP, (b, e) GTM + HOG, (c, f)GTM + LBP + HOG. Note that in the first row HOG misses the cavity but LBP is able to detect it. In the second row, LBP misses the small cavitybut HOG can detect it. In both rows our technique is able to detect all the cavities.


4.4. Accuracy of final cavity detectionBefore performing the final cavity detection, 160 can-didate contours are divided into cavity and non-cavitycontours. Candidate region reported by the classifieras highlighted by the green windows in Figure 12 maynot contain true cavities. Also, even if a reported win-dow contains the entire cavity, its segmented contourmay not be the same as the ground truth. To evaluate

the accuracy of the final contour classification, weneed to impose a value TMM > 0.7 (based on the seg-mentation accuracy of 67.1% reported in Table 4), inorder to qualify a candidate to be a true cavity; other-wise it is considered as non-cavity. Three contour-based features (circularity, GICOV, and KLD) areextracted from the candidate contours for the finalcavity classification. To evaluate the performance of

Ground Truth

HIE Result

DBF-GVF without HIE

DBF-GVF with HIE

IFVF without HIE

IFVF with HIE

Figure 15 Cavity segmentation result comparison using different edge-based snakes with and without HIE. From top to bottom, thecavity is more and more difficult to identify. Blue contours are the true cavities annotated by radiologists. Green contours are the computersegmentation results.

Table 4 Segmentation accuracy evaluation

DBC-GVF without HIE (%) DBC-GVF with HIE (%) IFVF without HIE (%) IFVF with HIE (%)

Average of TMM 55.1 64.6 56.8 67.1

Standard division of TMM 15.8 12.6 12.2 9.3

Mean of TMM 58.2 64.9 59.3 66.1


classification, sensitivity, specificity, and accuracy arecalculated as follows.

Sensitivity ¼ Number of Correctly�Detected Cavity ContoursTotal Number of Cavity Contours

� 100% ð18Þ

Specificity ¼ Number of Correctly�Detected Non�Cavity ContoursTotal Number of Non�Cavity Contours

� 100% ð19Þ

Accuracy ¼ Number of Correctly�Detected ContoursTotal Number of Candidates Contours

� 100% ð20Þ

Candidateswithout Cavity

HIE Result

IFVFwith HIE

Figure 16 Segmentation results of candidates without cavity.


For our sample size, we use cross-validation method[29] for the SVM classification. The classification resultfor the 160 candidate contours is shown in Table 5. Itcan be observed that the detection accuracy is increasedby more than 8% in our approach after adding KLDfeature. Figure 17 shows cavity detection results ofHKG [7] and the proposed technique, which demon-strate that our technique can detect more true cav-ities and detect fewer false cavities. As illustrated in

Figure 17, the proposed cavity detection system iden-tifies all cavities annotated by the radiologists andthere is only one false alarm. The presence of cavitiesin the upper half of the lungs, especially when thereare multiple or bilateral cavities, should raise suspi-cion of TB in the appropriate epidemiologic and/orclinical context. Unfortunately, in practice, a lot ofthese findings are not mentioned in the radiologist’s report,because the epidemiologic or clinical information, neces-sary to raise suspicion, is not provided by the orderingphysician on the requisition. This is often the case in geo-graphic regions where TB rate is low. Based on the clini-cian’s perspective, a relatively higher false positive rate isbetter than false negatives because the latter can cause aninfectious TB to spread. Even with false positives, cliniciansfind automatic cavity detection system helpful in reducing

Original CXR with Ground Truth HKG [6] Cavity Detection Result Proposed Cavity Detection Result

Figure 17 Cavity detection comparison between HKG [7] and the proposed technique. Blue dotted contours are the true cavitiesannotated by radiologists. Red contours are the detected cavities, while the cyan ones are the non-cavity contours.

Table 5 Cavity detection evaluation

Sensitivity (%) Specificity (%) Accuracy (%)

Circularity + GICOV [7] 62 46 54

Circularity + GICOV + KLD 70 60 65


a large number of true negatives and radiograph examina-tions. This is beneficial given the limited radiologists avail-able particularly in remote communities and developingcountries.The radiologists also classified the true cavity contours

into two categories: E-Group and D-Group, containingcavities which are ‘easy’ or ‘difficult’ to identify, respect-ively. The D-Group contains cavities even radiologistsfound them difficult to identify without other demo-graphic or additional information. False cavity contourswere then combined with each of these two groups. Thecross-validation SVM classification results of thesegroups are shown in Tables 6 and 7. Observe that onaverage the classification accuracy in each group ishigher than the result reported in Table 5. The perform-ance of the E-Group is significantly improved by addingthe KLD feature. In the D-Group, although the intensityvariation inside and outside a cavity changes only slightlymaking it very difficult to identify the contour even forradiologists, there is still improvement in the detectionresult. This shows that the classifier can perform betterif trained using more specific knowledge.

5. ConclusionsIn this article, we proposed an efficient coarse-to-finedual scale feature classification technique for TB cavitydetection in chest radiographs. Experimental resultsdemonstrate that the proposed technique outperformsexisting methods in three aspects. First, a lower MR isachieved because in the proposed method local cavityregion-related coarse features, such as geometric, tex-tural, and gradient features, are taken into consideration.Second, edge-based segmentation becomes more accur-ate by incorporating HIE to enhance the contours.Third, the final cavity detection accuracy is greatlyincreased by introducing the fine scale feature classifica-tion using three types of contour-related features, whichincludes shape, edge, and region. This study contributesin the development of CAD systems for infectious TBdiagnosis, because of the higher detection rate and lowerMR compared to other techniques. Future work will

focus on exploring novel algorithms to model othercharacteristics of infectious TB.

Competing interestsThe authors declare that they have no competing interests.

Author details1Department of Electrical and Computer Engineering, University of Alberta,Edmonton, AB T6G 2V4, Canada. 2Department of Computing Science,University of Alberta, Edmonton, AB T6G 2E8, Canada. 3Division of PulmonaryMedicine, Department of Medicine, University of Alberta, Edmonton, AB T6G2V2, Canada.

Received: 14 December 2011 Accepted: 28 November 2012Published: 8 January 2013

References1. BV Ginneken, BMTH Romeny, MA Viergever, Computer-aided diagnosis in

chest radiography: a survey. IEEE Trans. Med. Imag. 20(12), 1228–1241 (2001)2. BV Ginneken, L Hogeweg, M Prokop, Computer-aided diagnosis in chest

radiography: beyond nodules. Eur. J. Radiol. 72(2), 26–30 (2009)3. RC Hardie, SK Rogers, T Wilson, A Rogers, Performance analysis of a new

computer aided detection system for identifying lung nodules on chestradiographs. Med. Image Anal. 12(3), 240–258 (2008)

4. World Health Organization, Epidemiology, in Global Tuberculosis Control:Epidemiology, Strategy, Financing, WHO Press, Geneva, 2009, pp. 6–33

5. R Long, E Ellis et al., Canadian Tuberculosis Standards, 6th edn. (Public HealthAgency of Canada, 2007)

6. A Lau, R Long et al., The public health consequences of smear positivepulmonary tuberculosis in patients with typical and atypical chest radiographs,in 15th Annual International Union Against Tuberculosis and Lung Disease, NorthAmerican Region Conference, Vancouver, BC, 2011

7. R Shen, I Cheng, A Basu, A hybrid knowledge-guided detection techniquefor screening of infectious pulmonary tuberculosis from chest radiographs.IEEE Trans. Biomed. Eng. 57(11), 2646–2656 (2010)

8. BS Reddy, BN Chatterji, An FFT-based technique for translation, rotation,and scale-invariant image registration. IEEE Trans. Image Process. 5(8),1266–1271 (1996)

9. YH Lin, CH Chen, Template matching using the parametric template vectorwith translation, rotation and scale invariance. Pattern Recognit. 41(7),2413–2421 (2008)

10. X Wang, TX Han, S Yan, An HOG-LBP Human Detector with Partial OcclusionHandling (in Proceedings of the ICCV, Kyoto, 2009), pp. 32–39

11. T Ojala, M Pietikainen, D Harwood, A comparative study of texture measureswith classification based on feature distributions. Pattern Recognit. 29(1),51–59 (1996)

12. N Dalal, B Triggs, Histograms of oriented gradients for human detection, in inProceedings of the IEEE CVPR, ed. by, Vol. 1st, San Diego, 2005, pp. 886–893

13. C Cortes, V Vapnik, Support-vector networks. Mach. Learn. 20(3), 273–297 (1995)14. B Li, M Meng, Computer-aided detection of bleeding regions for capsule

endoscopy images. IEEE Trans. Biomed. Eng. 56(4), 1032–1039 (2009)15. Y Sato, S Nakajima et al., 3D multiscale line filter for segmentation and

visualization of curvilinear structure in medical images. Med. Image Anal 2,143–168 (1998)

16. J Liu, JM White, RM Summers, Automated Detection of Blob Structures byHessian Analysis and Object Scale (Proceedings of the ICIP, Hong Kong,2010), pp. 841–844

17. PK Saha, JK Udupa, Scale-based image filtering preserving boundarysharpness and fine structure. IEEE Trans. Med. Imag. 20, 1140–1155 (2001)

18. H Zhang, M Wan, Z Bian, Complementary tensor-driven image coherencediffusion for oriented structure enhancement. EURASIP J. Adv. SignalProcess. 70(2011) (2011)

19. K Zuiderveld, Contrast limited adaptive histogram equalization, in ChapterVIII.5, Graphics Gems IV,Academic Press, Cambridge, MA, 1994, pp. 474–485

20. M Gil, EG Sarabia et al., Fuzzy c-means clustering for noise reduction,enhancement and reconstruction of 3D ultrasonic images, in Proceedings ofthe ETFA, Barcelona, 1999, pp. 465–472

21. Y Yu, ST Acton, Speckle reducing anisotropic diffusion. IEEE Trans. ImageProcess. 11(11), 1260–1270 (2002)

Table 6 Cavity detection evaluation of E-Group


Circularity + GICOV [7] 65 78.2 71.6

Circularity + GICOV + KLD 78.8 86.8 82.8

Table 7 Cavity detection evaluation of D-Group


Circularity + GICOV [7] 57.6 88 72.8

Circularity + GICOV + KLD 69.4 81.6 755


22. T Xu, I Cheng, M Mandal, An Improved fluid vector flow for cavitysegmentation in chest radiographs, in Proceedings of the ICPR, Istanbul,2010, pp. 3376–3379

23. KW Sum, PYS Cheung, Boundary vector field for parametric active contours.Pattern Recognit. 40(6), 1635–1645 (2007)

24. J Canny, A computational approach to edge detection. IEEE Trans. PatternAnal. Mach. Intell. 8(6), 679–714 (1986)

25. C Di Ruberto, A Dempster, Circularity measures based on mathematicalmorphology. Electron. Lett. 36(20), 1691–1693 (2000)

26. G Dong, N Ray, S Acton, Intravital leukocyte detection using the gradientinverse coefficient of variation. IEEE Trans. Med. Imag. 24(7), 910–924 (2005)

27. C Bishop, Pattern Recognition and Machine Learning, Chap. 1 (Springer, NewYork, 2006)

28. CC Chang, CJ Lin, LIBSVM: A Library for Support Vector Machines. (2011).available at http://www.csie.ntu.edu.tw/~cjlin/libsvm/

29. T Hastie, R Tibshirani, J Friedman, The Elements of Statistical Learning: DataMining, Inference, and Prediction, Chap. 7, 2nd edn. (Springer, New York, 2009)

30. T Xu, I Cheng, M Mandal, Automated cavity detection of infectiouspulmonary tuberculosis in chest radiographs, in Proceedings of the IEEEEMBC, Boston, 2011, pp. 5178–5181

doi:10.1186/1687-5281-2013-3Cite this article as: Xu et al.: Novel coarse-to-fine dual scale techniquefor tuberculosis cavity detection in chest radiographs. EURASIP Journal onImage and Video Processing 2013 2013:3.

Submit your manuscript to a journal and benefi t from:

7 Convenient online submission

7 Rigorous peer review

7 Immediate publication on acceptance

7 Open access: articles freely available online

7 High visibility within the fi eld

7 Retaining the copyright to your article

Submit your next manuscript at 7 springeropen.com


Novel coarse-to-fine dual scale technique for tuberculosis cavity detection in chest radiographs

Documents

Transcript of Novel coarse-to-fine dual scale technique for tuberculosis cavity detection in chest radiographs