DETECTION OF NORMALITY/PATHOLOGY ON CHEST

RADIOGRAPHS USING LBP

Juan Manuel Carrillo-de-Gea and Gin´es Garc´ıa-Mateos

Dept. Inform´atica y Sistemas, Faculty of Computer Science, University of Murcia, Spain

Keywords:

Image processing, Health informatics, X-rays, Computer-aided detection (CADe), Template matching, Local

binary patterns (LBP), Histograms.

Abstract:

Since the discovery of X-rays and their applications, medical imaging has been a great help for radiologists

in the diagnosis of diseases. In recent years, there has been a great effort in the computer vision community

in the research of systems for the analysis and extraction of useful information from medical images. In this

scenario, we have designed, implemented and validated a novel method to detect normality/pathology in chest

radiographs, which constitutes the core of a computer-aided detection (CADe) system. Although the problem

addressed is very complex and little explored, our approach is completely automatic, starting from the location

of areas of interest using template matching techniques. The main novelty of our contribution is the application

of a transformation known as local binary patterns (LBP) to these areas. LBP histograms are then used as input

features for a classiﬁcation system, which is ultimately responsible for the decision of normality/pathology.

The results of our preliminary experiments are quite promising. With success rates in the best cases close to

90%, we believe that increased performance could be obtained with bigger training sets and more advanced

classiﬁcation systems, which will make these systems to be fully viable in the near future.

1 INTRODUCTION

The diagnostic imaging is based on the use of differ-

ent physical principles to visualize the internal struc-

ture, composition and functions of living organisms,

so that medical professionals have evidence available

to identify lesions in patients. Among the various ex-

isting medical imaging techniques, we can name a

few: based on X-rays (including conventional radio-

graphy), based on nuclear magnetic resonance, based

on the reﬂection of ultrasound and based on emissions

from radioisotopes (Fleckenstein and Tranum-Jensen,

2001). Figure 1 shows a sample of a chest radiograph.

Despite the subsequent emergence of new imag-

ing techniques, nowadays conventional radiology re-

mains to have a central role in health care facilities

around the world, “due to its ability of revealing some

unsuspected pathologic alterations, its non-invasivity

characteristics, radiation dose and economic consid-

erations” (Campadelli and Casiraghi, 2005, p. 1431).

Also, computer systems have been widely used

in the analysis of medical signals. Over the past 15

years many research groups have focused on the de-

velopment of computerized systems that can examine

Figure 1: Male chest radiograph in posteroanterior (from

back to front) view.

different types of medical images and extract useful

information for medical professionals (Papadopoulos

et al., 2005), speciﬁcally assisting in image analy-

sis and diagnostic decision making. Such systems

are commonly referred to as computer-aided detec-

tion/diagnosis (CAD) systems.

As we will review in section 2, there is an interest-

167

Manuel Carrillo-de-Gea J. and García-Mateos G. (2010).

DETECTION OF NORMALITY/PATHOLOGY ON CHEST RADIOGRAPHS USING LBP.

In Proceedings of the First International Conference on Bioinformatics, pages 167-172

DOI: 10.5220/0002747001670172

 SciTePress

ing amount of work done in this ﬁeld. However, our

proposal represents a new point of view, since it is a

method aimed to detect normality and pathology from

posteroanterior (PA) chest radiographs instead of be-

ing specialized in a predeﬁned number of diseases in

speciﬁc organs of the body. We should note that the

problem addressed is a little explored ﬁeld, and more

complex than most of the existing computer-aided de-

tection (CAD detection) systems.

The rest of the paper is organized as follows. In

section 2, we present a classiﬁcation of CAD systems

and review some approaches. In section 3, the pro-

posed method is described in detail, and in section 4

we expose and discuss the results of performed tests.

Finally, conclusions are presented in section 5.

2 RELATED WORK

2.1 Classiﬁcation of CAD Systems

According to (Kallergi, 2005; Papadopoulos et al.,

2005), computer-aided biomedical systems can be

classiﬁed into the ﬁrst two categories of the list be-

low. More recently, a third type of system has been

incorporated (Hao et al., 2008):

• Computer-aided detection (CAD detection or

CADe) systems, which detect signs of pathology

in medical images by extracting features that are

highly correlated with the type and nature of the

abnormality under investigation, but do not pro-

vide a medical diagnosis as such.

• Computer-aided diagnosis (CAD diagnosis or

CADx) systems, which assist clinicians in diag-

nosis. Based on the same or additional features,

such schemes characterize the identiﬁed regions

according to their pathologies.

• Computer-aided change detection (CADcd) sys-

tems, which provide detection and diagnosis as

well as evaluation of the progression of the dis-

ease by identifying the areas that have changed.

2.2 Review of Existing CAD Systems

CAD systems have attracted the interest of the com-

puter vision community in the last decade. There is an

extensive literature in the domain of breast and lung

cancer imaging. Actually, there are techniques related

to the study of CAD systems ready to detect other

diseases such as severe acute respiratory syndrome

(SARS), acute intracranial haemorrhage (AIH), colon

polyps and thyroid gland nodules. We can even ﬁnd

research works focused on the use of imaging tech-

niques different from x-rays, such as ultrasound.

On the other hand, the local binary pattern (LBP)

operator is introduced as a way to summarize the

structure of local gray levels of an image (Ojala et al.,

1996). Regarding the application of LBP or its varia-

tions to solve medical imaging problems, researchers

are getting very good results, often improving other

existing approaches (Iakovidis et al., 2008; Oliver

et al., 2007; Sørensen et al., 2008).

There are very few works dealing with the prob-

lem of normality and pathology detection without

specializing in certain diseases. Research done by

(Park et al., 2004), aimed to the detection of nor-

mal/abnormal texture on chest radiographs, repre-

sents an exception. However, they propose the uti-

lization of techniques different from ours, such as a

knowledge based lung ﬁeld extraction method, ﬁl-

tering to reduce ribs shadows, feature extraction ap-

plying a cuasi-Gabor ﬁlter in the frequency domain

and a k-nearest-neighbors classiﬁer. Moreover, in

(Sørensen et al., 2008), authors try to solve the prob-

lem of texture classiﬁcation in lung using LBP. How-

ever, in this research work, computed tomography

images are used instead of conventional radiographs.

LBP histogram similarity is used as distance measure.

Then, classiﬁcation is performed using the k-nearest-

neighbor classiﬁer with these distances as input.

3 OUR APPROACH USING LBP

In this section, we describe our CADe system spe-

cialized in PA chest radiographs. According to (Pa-

padopoulos et al., 2005), the main phases of a typical

CAD detection system are: preprocessing, segmenta-

tion, feature analysis (extraction, selection and valida-

tion), and classiﬁcation, used to reduce false positives

or to characterize abnormalities. Figure 2 represents

the structure of this kind of systems.

3.1 Digital Radiography

The input of our system is a set of DICOM (Digi-

tal Imaging and COmmunications in Medicine) dig-

ital radiological images. The images are around

3000×3000pixels with 12 bits per pixel. DICOM is a

well-known standard for exchanging medical images,

designed for handling, storage, printing and transmis-

sion of medical images.

BIOINFORMATICS 2010 - International Conference on Bioinformatics

168

Figure 2: CAD architecture, (Papadopoulos et al., 2005).

3.2 Preprocessing

We ﬁrst transform the digital radiographs into a for-

mat more suitable for easier processing, using MRI-

cro

, which is a freeware tool for efﬁcient viewing

and exporting of medical images. We chose the PNG

(Portable Network Graphics) format, which is based

on a lossless compression algorithm. This step im-

plies the reduction of DICOM image depth from 12

bits per pixel to 8 bits per pixel; however, this more

compact representation of the image data still keeps

the information necessary for our purposes.

By using LBP, the necessity of applying im-

age equalization, or other preprocessing steps, is re-

moved. As we will describe in section 3.4, in LBP im-

ages only the value of a pixel in relation to its neigh-

bors is relevant. Nevertheless, since our system input

are high resolution radiographs, the needs of time and

memory could be too demanding. Therefore, before

the segmentation stage we apply a decimation to the

input images done by super-sampling interpolation.

3.3 Segmentation

In this stage, the original image is segregated into sep-

arate parts. As in our system the main object of inter-

est are the lungs, the goal of segmentation is to deter-

mine the position and size of both lungs.

There are a variety of techniques that can be used

for segmentation. Our proposal is based on template

matching, a technique for ﬁnding those parts of an

image which match a predeﬁned template image.

http://www.sph.sc.edu/comd/rorden/mricro.html

3.3.1 Proposed Segmentation Method

We utilize patterns of the right and left lungs as input

for a template matching algorithm. It is important to

use different patterns of lungs, to include the wide va-

riety of appearances that may be caused by sex, age of

individuals and their archetype. Furthermore, a multi-

scale search allow us to locate lungs with different

sizes. The measure used for matching is a normalized

correlation coefﬁcient (NCC).

The point of maximum correlation, which is de-

termined by the map of results produced by the tem-

plate matching algorithm, indicates the position of the

lungs. The dimensions are those corresponding to the

chosen pattern at the scale selected.

Finally, according to the obtained locations, the

lungs are divided into regions of equal size (see ﬁgure

3). This division is required in the subsequent steps.

Figure 3: Localization of lungs in a radiograph and segmen-

tation of them into regions. In the experiments described on

section 4, we have used a grid of 3× 4 rectangular regions.

3.4 Feature Extraction

Feature extraction consists on the calculation of some

image features, looking for a description of the prop-

erties of each object. Our proposal is to use LBP

as the basic method to extract texture features in the

images (Ojala et al., 1996). This is a gray-scale in-

variant texture measure, derived from a general def-

inition of texture in a local neighborhood. Although

this technique has been used in other medical imag-

ing applications, this work represents, to the best of

our knowledge, the ﬁrst application to test of normal-

ity/pathology on radiographs.

The LBP operator was originally deﬁned for 3× 3

neighborhoods, giving 8 bit codes based on the 8 pix-

els around the central pixel. The original LBP opera-

tor labels all the pixels of the image by thresholding

the neighborhood of each pixel with the central value

DETECTION OF NORMALITY/PATHOLOGY ON CHEST RADIOGRAPHS USING LBP

169

and considering the result as a binary number. A sam-

ple of this calculation is shown in ﬁgure 4.

Figure 4: Example of a LBP calculation.

The histogram of these 2

= 256 different labels

can then be used as a texture descriptor. The lo-

cal primitives encoded by these cells include differ-

ent types of curved edges, spots, ﬂat areas, etc. The

calculation of the LBP codes can be easily done in a

single pass through the image. Formally, the LBP op-

erator takes the form shown in (1), where n iterates

on the 8 neighbors of the central pixel c, g

and g

are

the grayscale level values of c and n respectively, and

s deﬁnes the thresholding function shown in equation

(2).

LBP(x

, y

) =

∑

n=0

s(g

− g

) (1)

s(u) =



0, u ≥ 0

1, u < 0

(2)

3.4.1 Proposed Feature Extraction Method

We apply LBP technique going over each pixel in the

radiograph and accessing to its 3× 3 pixel vicinity. A

typical application to a chest radiograph is shown in

ﬁgure 5. Next, the LBP image is analyzed by regions

(see ﬁgure 6), computing the histograms of each re-

gion of the lungs (as previously found). Then we nor-

malize the histograms for comparison.

3.5 Classiﬁcation

The diagnosis is an integrated medical procedure

aimed to the identiﬁcation of a disease from its signs

and symptoms. In the same way, the classiﬁcation

stage is an essential and critical part of a CAD sys-

tem. Classiﬁcation methods range from techniques

that classify lesions according to their types, to tech-

niques that produce binary diagnostics, as is our case,

Figure 5: Application of LBP operator to a radiograph.

Figure 6: Comparison between the histogram of gray levels

and the LBP histogram calculated for the same region.

since the objective is to determine whether the radio-

graph is normal or contains some kind of pathology.

Despite the existence of a wide variety of complex

classiﬁers for analyzing medical images, since we are

limited by the number of training examples, we pro-

pose to use conventional classiﬁers.

3.5.1 Proposed Classiﬁcation Method

The main stages of our approach are:

• Calculation of distances between histograms,

which consists of obtaining distances between

histograms of each region of the lungs of the given

radiograph and those histograms corresponding in

normal and pathologic radiographs of the training

set. At the same time, minimum distances to nor-

mal and pathologic radiographs are obtained for

each region. We use the Bhattacharyya distance

for comparison between two histograms.

• Differences between minimum distances and clas-

siﬁcation of regions. For each region, the dif-

ference between the minimum distance to nor-

mal and pathologic histograms is obtained. The

values obtained indicate whether a histogram is

closer to one class or another (to normal or pathol-

ogy). Apart from calculating the differences be-

tween minimum distances, we check the sign of

these differences and classify the regions as they

are closer to the corresponding region of the nor-

BIOINFORMATICS 2010 - International Conference on Bioinformatics

170

mal or the pathologic images.

• Classiﬁcation of radiographs with or without in-

formation on discrimination of regions, which

is aimed at the ﬁnal classiﬁcation of the radio-

graph. We have analyzed three different clas-

siﬁcation methods, which will be compared in

the experiments: greater difference in absolute

value (GDAV), discrete voting (DV) and contin-

uous voting (CV). GDAV method consists on go-

ing through all regions of the lungs looking for

the difference between the minimum distances (as

calculated above) that is larger in absolute value.

Once obtained, the class of the radiograph is the

same as that of the region concerned. DV method

is a voting in which each region, depending on

its class, contributes with one positive or nega-

tive vote. When all the votes are collected, if

the number exceeds a certain threshold the image

is considered as normal and otherwise as patho-

logic. CV method is similar to DV method, but

now there is a global sum reﬂecting the accumula-

tion of the differences between the minimum dis-

tances of the input image to the normal and to the

pathologic images, for each region of the lungs.

At the end, if the sum exceeds a certain threshold

the image is considered as pathologic and other-

wise as normal.

Note that the threshold established for voting

methods is determinant to control the compromise be-

tween false positives (FP) and false negatives (FN).

In turn, the three methods have two variants: using

discrimination matrix or not. This matrix weights

the relative importance of each region of the grids in

which the lungs are divided. When no discrimination

matrix is used, we have no information about the level

of discrimination of each region for the determination

of normality and pathology, so it is assumed that all

the regions of the grid have exactly the same impor-

tance. The discrimination matrix is obtained automat-

ically through training with the same set of normal

and pathologic radiographs available.

4 EXPERIMENTAL RESULTS

The clinical performance of medical tests, includ-

ing medical imaging, is normally determined by es-

timating indexes for true positives (TP), true nega-

tives (TN), false positives (FP), false negatives (FP),

sensitivity, speciﬁcity and accuracy, among others

(Kallergi, 2005). In addition, confusion matrices as

the one shown in table 1 are often used.

We have 25 normal (12 males and 13 females) and

23 pathologic (13 males and 10 females) high resolu-

Table 1: Confusion matrix.

Prediction Outcome

Normal Pathologic

Actual Normal TN FP

Value Pathologic FN TP

tion DICOM images of chest radiographs to perform

tests. With this input, we have conducted an extensive

battery of tests using the leave-one-out technique.

In each test, we measured the following values:

sensitivity SENS =

TP+FN

(probability of giving a

correct response to the pathologic cases), speciﬁcity

SPEC =

TN+FP

(probability of obtaining a correct re-

sponse to the cases with no pathology) and accuracy

ACC =

TP+TN

TP+TN+FP+FN

(equivalent to the proportion

of successes over the total).

The experimental tests performed have produced

the results summarized in table 2.

Table 2: Results of the tests using the GDAV, DV and CV

classiﬁers (in this order). WOD and WD means without

discrimination and with discrimination matrix, respectively.

M, F and A means using male, female and all radiographs,

respectively.

WOD WD

M F A M F A

SENS 0.31 0.4 0.39 0.54 0.4 0.39

SPEC 0.67 0.85 0.96 0.83 1 0.88

ACC 0.48 0.65 0.69 0.68 0.74 0.65

SENS 0.46 0.6 0.96 0.92 0.7 0.78

SPEC 0.67 0.77 0.2 0.5 0.77 0.8

ACC 0.56 0.69 0.56 0.72 0.74 0.79

SENS 0.46 0.7 0.7 0.54 0.7 0.52

SPEC 0.67 0.54 0.44 0.75 1 0.88

ACC 0.56 0.61 0.56 0.64 0.87 0.71

4.1 Discussion of the Results

Given the considerable amount of data available, we

will focus our attention on the accuracy obtained in

each of the experiments. We will analyze and dis-

cuss methods that do not use discrimination matrix

and those that use it.

For the methods without discrimination informa-

tion, there is a general low levelof precision, as shown

in ﬁgure 7. This is more pronounced in the voting

methods. GDAV method has an interesting accuracy

rate, especially for female radiographs and when the

radiographs are considered all together. Male radio-

graphs are the most difﬁcult to classify in this set of

data, but we believe that this fact is not generalizable.

Finally, although there are some cases with less than

DETECTION OF NORMALITY/PATHOLOGY ON CHEST RADIOGRAPHS USING LBP

171

50% accuracy (on male radiographs), the average ac-

curacy is around 55% with peaks close to 70%.

Figure 7: Accuracy obtained with different classiﬁers with

and without using discrimination matrix.

Regarding discrimination methods, the degree of

accuracy achieved is clearly better, highlighting the

classiﬁcation of female radiographs again. In most

cases, voting methods perform better than GDAV

method. It is remarkable that with the appropriate

threshold, the CV method classiﬁes women radio-

graphs with success rates very close to 90% of accu-

racy. With all the radiographs the best is DV method,

reaching an accuracy near to 80%. In general, the av-

erage accuracy is around 70% considering all cases.

5 CONCLUSIONS

As a main general conclusion, the utilization of the

LBP texture analysis operator has emerged as an ef-

fective means for the study of radiographic images,

and more particularly chest radiographs, which are

the focus of this paper. LBP is a method that collects

interesting advantages as a small number of features

and a high power of discrimination.

The use of template matching technique at the

stage of segmentation to locate the lungs has been

satisfactory, despite its known limitations. The great

variability among individuals in terms of shape and

size of the lungs makes it difﬁcult to achieve an opti-

mal location of these structures.

Our approach to the problem has been imple-

mented and tested as a ﬁrst step towards a system

for the detection of normality and pathology in radio-

graphs which can be implanted in a hospital. The ra-

tios of success achieved, near 90% in the best cases,

are quite promising, although the method designed

should not be considered closed. Clearly, higher accu-

racy rates should to be obtained to achieve a practical

CAD system. To work in its development the use of

much larger sets of test will be essential.

ACKNOWLEDGEMENTS

This work has been supported by the Spanish MEC

and European FEDER grants TIN2006-15516-C04-

03 and Consolider CSD2006-00046.

REFERENCES

Campadelli, P. and Casiraghi, E. (2005). Lung Field Seg-

mentation in Digital Postero-Anterior Chest Radio-

graphs. In PRIA, volume 3687 of LNCS, pages 736–

745. Springer Heidelberg.

Fleckenstein, P. and Tranum-Jensen, J. (2001). Anatomy in

Diagnostic Imaging. Blackwell Science, 2 edition.

Hao, B., Leung, C. K.-S., Camorlinga, S., Reed, M. H.,

Bunge, M. K., Wrogemann, J., and Higgins, R. J.

(2008). A Computer-Aided Change Detection Sys-

tem for Paediatric Acute Intracranial Haemorrhage. In

C3S2E ’08, pages 109–111, New York. ACM.

Iakovidis, D. K., Keramidas, E. G., and Maroulis, D.

(2008). Fuzzy Local Binary Patterns for Ultrasound

Texture Characterization. In IAR, volume 5112 of

LNCS, pages 750–759. Springer Heidelberg.

Kallergi, M. (2005). Evaluation Strategies for Medical-

Image Analysis and Processing Metodologies. In

Medical Image Analysis Methods, chapter 12, pages

433–471. CRC Press.

Ojala, T., Pietik¨ainen, M., and Harwood, D. (1996). A Com-

parative Study of Texture Measures with Classiﬁca-

tion Based on Feature Distributions. Pattern Recogni-

tion, 29(1):51–59.

Oliver, A., Llad, X., Freixenet, J., and Mart, J. (2007). False

Positive Reduction in Mammographic Mass Detection

Using Local Binary Patterns. In MICCAI ’07, volume

4791 of LNCS, pages 286–293. Springer Heidelberg.

Papadopoulos, A. N., Plissiti, M. E., and Fotiadis, D. I.

(2005). Medical-Image Processing and Analysis for

CAD Systems. In Medical Image Analysis Methods,

The Electrical Engineering and Applied Signal Pro-

cessing Series, chapter 2, pages 51–86. CRC Press,

Boca Raton, FL, USA.

Park, M., Jin, J. S., and Wilson, L. S. (2004). Detection

of Abnormal Texture in Chest X-rays with Reduction

of Ribs. In VIP ’05, pages 71–74, Darlinghurst, Aus-

tralia. Australian Computer Society, Inc.

Sørensen, L., Shaker, S. B., and de Bruijne, M. (2008).

Texture Classiﬁcation in Lung CT Using Local Bi-

nary Patterns. In MICCAI ’08, volume 5241 of LNCS,

pages 934–941. Springer Heidelberg.

BIOINFORMATICS 2010 - International Conference on Bioinformatics

172