Unsupervised Rib Delineation in Chest Radiographs by an

Integrative Approach

B. Buket Oğul

, Emre Sümer

and Hasan Oğul

Akgun Software Company, Ankara, Turkey

Department of Computer Engineering, Baskent University, Ankara, Turkey

Keywords: Computer Aided Diagnosis, X-ray Imaging, Medical Image Analysis.

Abstract: We address the problem of segmenting ribs in a chest radiography image as an intermediate step for

eliminating rib shadows for an effective Computer-Aided Diagnosis System (CAD). To this end, we

introduce a complete framework that takes an unprocessed x-ray image and reports the entire rib regions.

The system offers a novel strategy to fit a parabola curve to all rib seeds obtained through a log Gabor

filtering approach and extend the center curve by a problem-specific region growing technique to delineate

the entire rib, which does not necessarily follow a general parabolic model of rib cage. The visual

examinations of predicted rib delineations in a common dataset have demonstrated that the system can

achieve a reasonably good performance to be used in practice.

1 INTRODUCTION

Diagnostic image analysis is probably the biggest

contribution of computer and information sciences

to clinical practice in medicine. In the wake of

inferring abnormal structures from a medical image

without the intervention of a clinician, several

attempts have been done to develop intelligent

algorithms that can automatically analyze the raw

data from various types of modalities. X-ray

radiography is known as one of the cheapest and

least harmful of these imaging modalities. In spite of

the availability of a wide variety of more

complicated imaging techniques, it still remains as

ubiquitous in medical practice even for well-

equipped hospitals and radiology departments. On

the other hand, it is also evaluated as most difficult

modality to analyze, either manually or

computationally, due to the low resolution and high

noisy content of the input images.

Chest radiography is one of the most widely

used diagnostic tools in x-ray based imaging. A

common use is to detect potential lung nodules for

the decision of further examination using such as

computed tomography or pathology analysis. In

addition to major difficulties listed above, chest

radiography analysis suffers from normal anatomical

structures, such as bones and other overlapping

organs, which hide buried abnormalities in normal

tissue of lung. The nodules in the main tissue might

be invisible due to the shadows of quasi-parallel

bones in the ribcage. Therefore, the delineation and

suppression of ribs from the input image without any

information loss in the original lung tissue beneath is

an important computational challenge in computer-

aided diagnosis (CAD) systems for chest

radiographs. Whilst this study is motivated from our

ultimate goal of suppressing ribs for easier detection

of lung nodules, the automatic delineation of ribs has

several other benefits, such as providing a frame of

reference for locations of abnormalities or detecting

rib abnormalities, e.g. fractured or missing ribs. In

this study, we address the problem of segmenting the

anatomy of ribs in lung images obtained through the

x-ray radiography as an intermediate step of rib

shadow elimination in CAD systems. To this end,

we introduce a novel integrative approach based on

parabola curve fitting from rib seeds obtained

through a log Gabor filtering. The parabola is

extracted from a self-template. The system is

completely unsupervised, that is, it does not require

any training set of manually delineated ribs or any

knowledge-driven rules provided by an expert. The

visual examinations of predicted rib delineations in a

common dataset have demonstrated that the system

can achieve a reasonably good performance to be

used in practice.

The automated detection of ribs in chest

260

Ogul B., Sümer E. and Ogul H..

Unsupervised Rib Delineation in Chest Radiographs by an Integrative Approach.

DOI: 10.5220/0005361602600265

In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 260-265

ISBN: 978-989-758-089-5

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

radiographs is indeed an old problem in medical

image analysis. In one of the earliest studies for rib

detection Toriwaki et al. (1980) located rib edges by

a template matching approach and fitted parabolas to

these borders. This idea later has been used in many

studies. De Souza (1983) used the derivatives of

vertical profiles to detect the pixels on the rib

boundaries and fitted parabola to these. Sarkar et al.

(1997) performed a set of gray-level calculation to

predict rib border candidates. Yue et al (1995) used

a modified Hough transform to fit parabolas to rib

edges and refined the boundaries by a snake

approach. Vogelsang et al. (1998) followed a similar

strategy for fitting parabolas to the edges by a

template matching approach and refined the ribs by

new estimates of parabolas for each rib. Karargyris

et al. (2011) smoothed the detected edges using a

Savitzk-Golay filter. Lee et al. (2012) used a

knowledge-based generalized Hough transform to

locate rib borders starting from a best template

detected by a simple Sobel operator. Horvath et al.

(2013) refined the final ribs by a dynamic-

programming-based active contour algorithm. In

general, all methods have three dimensions:

detecting the position of upper and/or lower rib

borders, fitting a curve to these borders, and refining

final rib shadows. It is also worth mentioning about

two supervised approaches that used a set of training

images with manually delineated rib borders

(Ginneken et al, 2000; Loog and Ginneken (2006))

and predicts the locations of ribs from a learned

model.

On the contrary to the existing methods that

employ a curve fitting from predicted upper or lower

rib edges, our system attempts to fit a template of a

parabola curve from a seed of rib start point which

can be more effectively predicted using a robust log

Gabor filtering approach. In this case, a curve

somewhere in the middle of the rib can be identified

instead of rib borders. Entire rib shadow is

delineated by a problem-specific region growing

technique, which is also introduced in this paper.

This growing stage allows the system to detect exact

rib regions which do not necessarily follow the fitted

parabola equation extracted from the best template

in the same image, which is considered as another

major contribution of this study.

The new system is visually examined in a well-

known chest radiography set (Shiraishi et al., 2000).

It is clearly demonstrated that the rib delineation can

be successfully achieved provided that at least one

template can be detected from original image. The

paper presents the steps of a complete framework

starting from several preprocessing stages including

lung segmentation, image enhancement and edge

detection and ending with final refinement stages for

rib delineation.

2 METHODS

2.1 Image Preprocessing

In image pre-processing step, we generate Local

Contrast Enhanced (LCE) and lung segmented

images to be used for template extraction and rib

detection steps. The details of these steps are not

given in here; a comprehensive description can be

found in (Ogul et al., 2015). Template extraction

step uses the product of lung segmented LCE and

original image as input.

2.2 Rib Segmentation

In this section we describe a novel hybrid approach

for rib segmentation. The existing systems which

model the ribs only using parabola or elliptic

equations are lack of the rib shape information

because it differs from person to person. We also

know that each person have a typical rib shape

which do not show significant changes. Therefore,

rather than trying to model the ribs regardless of the

shape changes, we propose to model the ribs using

existing rib templates which are extracted from the

overall lung field.

2.3 Rib Template Extraction

The rib detection task starts with the extraction of a

rib template from the content of query image itself.

This allows the framework to exhibit entirely in an

unsupervised manner. The template extraction

method is inspired from (Lee et al., 2012), where a

locale sampling scheme technique is first employed

to enhance the rib contrast, and then edge detection

operator is used to reveal candidate edges, and

finally a simple selection scheme is applied to fix the

template which appears in the middle of the lung

fields. The overall description of their methodology

can not detail how to convert edge detected image

into binary and how to select the template that

appears in the middle of the lung, where both

strategy in fact significantly affects the continuity of

the rib structure, and in turn the quality of the final

rib template extracted. In our study, we get the best

binary image results when we use Minimum Cross

Entropy (MCE) method for thresholding. However,

even the best thresholding technique results with

UnsupervisedRibDelineationinChestRadiographsbyanIntegrativeApproach

261

many small non rib candidates and does not provide

the rib continuity. Therefore, after MCE we apply

some morphological operators such as dilation to

connect disjointed ribs, opening to remove noise and

erosion to construct final rib structure (Fig 1).

Figure 1: Rib template extraction; first candidates (left),

after elimination stage (right).

To select the best candidates, we first flag 8

connected components which have highest pixel

count. Then, a thinning morphologic operator takes

place. At the final step we remove the templates by

their orientation, euler number and eccentricity

values. Remaining templates are further reduced

after detecting rib starts until we have one template

for each of the lung field, which is detailed in

following section.

2.4 Rib Delineation

The rib delineation through extracted self-templates

involves broadly 5 consecutive steps: (i) finding the

rib starts at outmost region of ribcage by a log Gabor

filtering, (2) selecting a single best template which is

most compatible with rib cage, (3) estimating rib

thickness, (4) parabola fitting to detect all ribs, and

(5) delineating the entire ribs by a specific region

growing technique.

2.4.1 Finding Outmost Rib Starts

Because the fact that the rib edges are easily affected

from high noise levels when pure pixel-based edge

detection methods used (Karargyris et al, 2011), we

choose to use Log Gabor wavelets in order to find

rib borders. Using just the 4

scale and the degrees

45° and 135° (Fischer et al., 2009) with Otsu

thresholding method gives us not all the rib

structures from vertical middle line to the outer

borders of the lung but the initial points with real

thickness values. In here, initial points refer to the

region in which the rib structure intersects with outer

border of the lung (Fig 2). Finally, by shifting the

lung mask we can see all the ribs positioned on

initial points (Fig 2) and we can eliminate the ones

which do not appear in initial points. For further

steps we also apply thinning morphologic operation

so that the ribs are formed as a single-pixel width.

Figure 2: Log Gabor filtering to detect rib seeds.

2.4.2 Selecting Best Template

In some cases, the method described in section 2.2.1

gives more than a one template for each of the lung

fields while sometimes we just get one template for

whole lung region. Thus we need to choose the best

template or to form a new one. When the first case

occurs, we reduce the templates by calculating their

distance to each of the thinned rib structures. The

closest template is chosen as the best template. In

cases such as a template for a lung field is found, but

calculated list for other field is empty, we form the

new template by taking the symmetry of existing

template.

Calculating the symmetry of a template is not an

easy task. It consists of mainly 3 steps: (i) First of

all, the best template must be formed as an elliptic

curve since it is comprised of scattered pixels (Fig 3,

dark blue lines on right lung). Using these non-

formed pixels, we find the coefficients for a

polynomial p(x) of degree n that is a best fit (in a

least-squares sense) for the data in y.

...)(







pxpxpxpxp

(ii) Secondly; its mirror symmetry is calculated (Fig

3, magenta template). (iii) And finally, its final

position is found with respect to the center of mass

of the region (Fig 3, yellow template). After we

compute the x and y coordinates of the vertex of the

parabola we found in step (i), we can easily calculate

the difference between this point and the center of

mass of the lung field in which this template lies on.

Since this distance should be same on the symmetric

lung field, by shifting the symmetric template as the

distance value gives us the final template on the left

lung in Fig 3 (yellow line).

2.4.3 Estimating Rib Thickness

Rib thickness (rt) value is used as a stopping

criterion while enlarging rib structure from its

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

262

center. The main point of calculating rt is to find the

longest vertical line that exists on rib structures.

Among all lines we calculate the mean of the values

except the ones which are too thin (

for JSRT < 10 px.)

or thick (>24 px.).

Figure 3: Best template selection step.

2.4.4 Parabola Fitting

Parabola fitting is very important task in which the

parameters of the parabola that fits the middle line of

each rib structure are calculated. This step is an

important basis for region growing and parabola

growing algorithms. In subsection 2.2.2.2 we

formulated both templates by calculating p

, p

values. In order to fit the curve somewhere in the

middle of the rib, we need to optimize last parameter

of the parabola. While 1

and 2

parameters are

shape related, 3

parameter has an effect on the

position of the template on the rib by moving

template up and down. For each rib, keeping first 2

parameters same, we calculated list of p

values

using each pixel pair that appears in this structure

and their mean reports us the final position of the

template for its corresponding rib.

2.4.5 Rib Region Growing

Rib delineation is finalized with a two-stage region

growing approach. The first stage is the quasi-

parallel growing of initial parabola curve fitted to a

rib. We use the pixels in the middle of the rib as

initial seed points for region growing using 0.015 as

threshold value. The grown rib structure has broken

pixels on some points. We use these broken pixels

for generating our preliminary rib borders. The main

idea is repeating to shift the middle-line curve up

and down provided that shifted curve contains at

least 50% of initially grown pixels.

The first stage of region growing is resulted with a

ticker parabola. In some cases we can reach rib

borders on region growing part, however mostly the

result is thinner structure than actual rib. In the

second stage, we consider the rib consists of 4

horizontal parts and use each part individually in

order to form them into the final rib structure. For

each horizontal part, the growing process is repeated

separately on upper or lower part (not on both) based

on the mean intensity change until the exact rib with

predicted length is formed.

3 RESULTS

We compile our method on well-known JSRT

dataset having 154 chest x-ray images with known

nodules. In 150 of 154 images, at least one template

is identified and the best one is selected for self-

template matching. In 4 of 154 images, no template

is identified. This demonstrates the success of rib

template extraction step in general.

In 65% of images, all ribs are perfectly located

with a false discovery rate of 8.5%. This detection

rate is computed as 90.1% if only one rib is allowed

to be missed, which is usually the case since the

lowest left rib is often obscured by the heart shadow.

Here, we mean by rib detection the correction

identification of the position of the rib regardless of

that its borders are perfectly located. When each rib

is considered as a separate sample, the detection

recall (or sensitivity) is 97.1% with a precision of

91.5%. The delineation (locating the exact position

of all pixels) performance is not evaluated yet since

it requires the contribution of an expert knowledge,

which is currently in progress. Nevertheless, we

discern the ability of the system to delineate the ribs

by a few examples in Figure 4.

4 CONCLUSIONS

We introduce a complete framework that takes an

unprocessed lung image obtained thorough a chest

radiography and reports the pixels of entire rib

regions. The contribution of the study is two-fold.

First, a parabola equation extracted from a rib border

template is fitted into rib seeds which are defined as

the start points of the ribs closer to the outmost

(right or left) part of the ribcage. The previous

methods largely depend on the quantitative and

qualitative value of detected rib borders. Most often,

the method employed for edge detection may fail to

UnsupervisedRibDelineationinChestRadiographsbyanIntegrativeApproach

263

cover all ribs in the ribcage or report incorrectly

some other curves that resembles a rib boundary.

Since the present method does not require the

detection of any border but only its outmost starting

point, the detection rate can be substantially

increased by log Gabor filtering approach. Second

contribution is a new region growing technique that

is introduced to extend rib center to build the entire

rib until rib boundaries are reached. The technique

allows each individual rib to grow independently

from its initial parabola equation. In that sense, the

difference between the curvative structure of upper

and lower ribs can be elaborated. Based on the

visual examinations of a common dataset of lung

images, we can argue that the results are very

promising to continue with a suppression method

that follows the rib delineation.

The future work will involve: (1) a quantitative

evaluation the results after a manual delineation of

rib borders by an observer, (2) a qualitative

evaluation of the results by a committee of radiology

experts, (3) improving the algorithm that smoothes

the rib boundaries by a dynamic programming

approach, and finally (4) applying a data-driven

bone suppression technique to make the main tissue

more visible so that the abnormalities beneath the

ribs can be easily identified.

ACKNOWLEDGEMENTS

This study was supported by Turkey Ministry of

Science, Technology and Industry by the grant

number 379.STZ.2013-2, and Akgun Software

Company.

REFERENCES

De Souza P., Automatic rib detection in chest radiographs,

Comput. Vis., Graphics, Image Process., vol. 23,

pp.129 -161, 1983.

Fischer, S., Redondo, R. and Cristobal, G. “How to

construct Log-Gabor filters”, Open Access Digital

CSIC Document, 2009.

Horváth A., Orbán G.G., Horváth A., Horváth G., An X-

ray CAD system with ribcage suppression for

improved detection of lung lesions, Periodica

Polytechnica Electrical Engineering and Computer

Science, vol. 57, pp. 19-33, 2013.

Karargyris, A., Antani, S., Thoma, G., Segmenting

anatomy in chest x-rays for tuberculosis

screening, Engineering in Medicine and Biology

Society,EMBC, 2011 Annual International Conference

of the IEEE, On page(s): 7779 - 7782.

Lee J.S., Wangb J.W., Wuc H.H., Yuand M.Z., A

nonparametric-based rib suppression method for chest

radiographs Computers & Mathematics with

Applications, Volume 64, Issue 5, September 2012,

Pages 1390–1399.

Loog, M.; Ginneken, B. "Segmentation of the posterior

ribs in chest radiographs using iterated contextual

pixel classification", Medical Imaging, IEEE

Transactions on, On page(s): 602 - 611 Volume: 25,

Issue: 5, May 2006.

Ogul B.B., Kosucu P., Özcam A., Kanik S.D, Lung

Nodule Detection in X-Ray Images: A New Feature

Set, 6th European Conference of the International

Federation for Medical and Biological Engineering,

IFMBE Proceedings Volume 45, 2015, pp 150-155.

Sarkar S. and Chaudhuri S.,Detection of rib shadows in

digital chest radiographs, Proc. ICIAP (2), pp.356 -

363 1997.

Shiraishi J, Katsuragawa S, Ikezoe J, Matsumoto T,

Kobayashi T, Komatsu K, Matsui M, Fujita H, Kodera

Y, and Doi K.: Development of a digital image

database for chest radiographs with and without a lung

nodule: Receiver operating characteristic analysis of

radiologists detection of pulmonary nodules. AJR 174;

71-74, 2000.

Toriwaki J., Hasegawa J, Fukumura T, Tagaki Y, (1980),

Computer analysis of chest photofluorograms and its

application to automated screening, Automedica 3, 63-

81.

van Ginneken B. and ter Haar Romeny V, Loew M. H.

and Sonka M., Automatic delineation of ribs in frontal

chest radiographs, Image Process., vol. 3979, pp.825

-836 2000.

Vogelsang F., Weiler F., Dahmen J., Kilbinger M.W.,

Wein B.B., Guenther R.W., Detection and

compensation of rib structures in chest radiographs for

diagnostic assistance, SPIE Proceedings Vol.3338:

Medical Imaging 1998: Image Processing.

Yue Z., Goshtasby A. and Ackerman L. V., Automatic

detection of rib borders in chest radiographs, IEEE

Trans. Med. Imag., vol. 14, no. 3, pp.525 -536 1995.

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

264



Figure 4: Example chest radiography images (left) and delineated ribs by our system (right).

UnsupervisedRibDelineationinChestRadiographsbyanIntegrativeApproach

265