Ap

EGa

b

a

A

R

R

1

A

K

C

E

P

1

Ttn

T

h0

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10

jo ur nal ho me p ag e: www.int l .e lsev ierhea l t h.com/ journa ls /cmpb

utomated detection of endotracheal tubes inaediatric chest radiographs

-Fong Kaoa,∗, Twei-Shiun Jawb, Chun-Wei Lia, Ming-Chung Choua,in-Chung Liub

Department of Medical Imaging and Radiological Sciences, Kaohsiung Medical University, Kaohsiung, TaiwanDepartment of Medical Imaging, Kaohsiung Medical University Hospital, Kaohsiung, Taiwan

r t i c l e i n f o

rticle history:

eceived 26 June 2014

eceived in revised form

8 October 2014

ccepted 26 October 2014

eywords:

omputer-aided detection

ndotracheal tube

aediatric chest radiograph

a b s t r a c t

The aim of this study was to develop an automated method for the detection of endotracheal

tube and location of its tip in paediatric chest radiographs. In this method, a seed point was

first determined from the line crossing the cervical region and a line path was traced from the

seed point. Two features, Lmax and C, were determined from the path and were combined

to detect the existence of the endotracheal tube. Multiple thresholds applied to the line

path were used to determine the candidate locations for the tip, and the most suitable

location was selected from these candidates by analysing the image features. To evaluate

the performance of detection of endotracheal tube existence, support vector machine was

used to classify the images with and without endotracheal tubes on the basis of Lmax and

C. The discriminant performance of the method was evaluated using receiver operating

characteristic (ROC) analysis. To evaluate the precision of the detected tip locations, the tip

locations in paediatric chest images were annotated by a radiologist. The distance (error)

between the detected and annotated locations was used to evaluate detection precision for

the tip location. The proposed method was evaluated using 528 images with endotracheal

tubes and 816 images without endotracheal tubes. The discriminant performance in this

study, evaluated as Az (area under the ROC curve), for detecting the existence of endotracheal

tubes on the basis of the two features was 0.943 ± 0.009, and the detection error of the

tip location was 1.89 ± 2.01 mm. The proposed method obtained high performance results

and could be useful for detecting the malposition of endotracheal tubes in paediatric chest

radiographs.

© 2014 Elsevier Ireland Ltd. All rights reserved.

inside the trachea. When correctly placed, the tip of theendotracheal tube should be positioned in the mid-tracheal

. Introduction

racheal intubation, the insertion of a tube into the tracheao inflate the lungs, remains a common procedure in theeonatal intensive care unit (NICU). The main goal of tracheal

∗ Corresponding author at: Department of Medical Imaging and Radioaiwan. Tel.: +886 07 312 1101x2358; fax: +886 07 311 3449.

E-mail address: tonykao@kmu.edu.tw (E.-F. Kao).

ttp://dx.doi.org/10.1016/j.cmpb.2014.10.009169-2607/© 2014 Elsevier Ireland Ltd. All rights reserved.

intubation is to position the tube at an appropriate depth

logical Sciences, Kaohsiung Medical University, Kaohsiung 807,

region [1,2]. Malposition of the endotracheal tube may resultin complications, including hypoxaemia, pneumothorax,

m s i n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10

Fig. 1 – Overall scheme of the proposed method.

2 c o m p u t e r m e t h o d s a n d p r o g r a

lung collapse and death [3–5]. Hence confirmation of correcttracheal tube placement is an important procedure.

To ensure that the endotracheal tube is ideally placed forproper ventilation, the current gold standard for confirmingtube position is a chest radiograph [6,7]. A chest radiograph canbe used to confirm the correct tube position within the trachea,which should be just below the level of the vocal cords and wellabove the carina. However, this is often delayed after venti-lation via the tube has commenced. An important issue inthis regard is how to notify physicians at the earliest when anendotracheal tube malposition occurs. This could be achievedby automatically detecting the malposition of the endotra-cheal tube at the moment when a radiograph is acquired oruploaded to a picture archiving and communication system(PACS) and immediately notifying the physicians.

Computer-aided diagnosis/detection has been applied tochest radiographs in many aspects [8,9]. However, researchon the automated detection of endotracheal tube position hasreceived little attention. Only some related studies [10–13] onthe detection of tubes in chest radiographs had been previ-ously conducted. For the automated detection of endotrachealtube malposition, identifying the presence of endotrachealtubes and precisely locating the tip are essential steps.Besides, endotracheal tube malposition occurs more often ininfants [6,14]. Therefore, in the present study, we proposed acomputerised method for the automated detection of endotra-cheal tubes and tip positions in paediatric chest radiographsand evaluated the performance of the proposed method.

2. Materials and methods

2.1. Image database

The proposed method was evaluated using 1344 anteroposte-rior paediatric chest radiographs, including 528 images withendotracheal tubes and 816 images without endotrachealtubes. The images were obtained from 412 patients in theNICU of Kaohsiung Medical University Hospital, Taiwan, fromJanuary to July 2013. All the chest radiographs were obtainedusing a computed radiography system (FCR 5000; Fuji PhotoFilm, Tokyo, Japan) and a digital radiography system (CXDI;Canon Inc., Tokyo, Japan). The images obtained by FCR 5000were digitised with a pixel size of 0.1 mm, a matrix size of2505 × 3015, and a grey scale of 10 bits. The images obtained byCXDI were digitised with a pixel size of 0.08625 mm, a matrixsize of 2868 × 3460, and a grey scale of 12 bits. In this study, allMONOCHROME2 images were converted to MONOCHROME1images, in which the regions of greater X-ray attenuation cor-respond to lower pixel values.

2.2. Overall scheme of the proposed method

The proposed method was developed based on the assump-tion that an endotracheal tube exists in a test image. This

assumption was tested by specific criteria to determine theexistence of the endotracheal tube. Fig. 1 demonstrates theoverall scheme of the proposed method and each step isdescribed in detail as follows.

2.3. Identification of field of view

In a paediatric chest image, the field of view (FOV) is oftensmaller than the whole image, as shown in Fig. 2(a). A thresh-old, Tmask, was first determined by Otsu’s method [15] from thehistogram of the image, as shown in Fig. 2(b). The FOV maskwas further obtained by Tmask. In the FOV mask, the mass cen-tre (Cx, Cy) was first determined, and four edge points werefurther determined based on the mass centre, as shown inFig. 2(c). Only the portion of the image inside the rectangle withthe upper border at Y1, lower border at Y2, left border at X1 andright border at X2 was used for the following procedures.

2.4. Determination of the horizontal line interceptingthe cervical region

To detect the endotracheal tube in a chest radiograph, a hori-zontal line intercepting the cervical region was first obtainedas shown in Fig. 3(a). The cervical region could be located in theupper part of a paediatric chest image. In the present study,the search process was limited to the interval between 1 andHFOV/4. To determine the horizontal line, a projection profile,V(y), was first obtained by summing the pixel values of theimage along the x-coordinate. V(y) can be defined as follows(Eq. (1)):

V(y) =∑WFOV

x=1 I(x, y)

WFOV(1)

where I(x, y) is the chest image, WFOV is the width of the FOVand HFOV is the height of the FOV. Yneck, corresponding to the

narrowest portion of the cervical region, can be determined bysearching a global maximum from the projection profile V(y)between 1 and HFOV/4, as shown in Fig. 3(b). If an endotracheal

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10 3

(b) t

ti

2

Ittltmtmtl

Fdp

Fig. 2 – (a) A paediatric chest radiograph,

ube exists in a chest image, the horizontal line at Yneck willntercept the tube.

.5. Determination of seed point

f an endotracheal tube exists, the above-mentioned horizon-al line would intercept the tube. To trace the endotrachealube, a seed point was first determined from the horizontaline. To construct the template for determining the seed point,he endotracheal tube width was measured. Manual measure-

ents of 20 chest images with endotracheal tubes showed thathe average width of the tubes was 9.4 ± 2.5 pixels. This infor-

ation was used to construct the template, which consisted ofhree lines with a length of 50 pixels, as shown in Fig. 4(a). Theine V1 corresponds to the centre of the endotracheal tube and

ig. 3 – (a) A horizontal line crossing the cervical regionetermined by finding the maximum from (b) therojection profile.

he corresponding histogram and (c) FOV.

lines V2 and V3 correspond to the neighbourhood. The lineswere separated by an interval of 5 pixels so that the lines ofthe neighbourhood would not fall within the tube accordingto the measurement of the tube width. Applying the templatewith three lines to an image, V1, V2 and V3 are the summa-tion of the pixel values on the three lines, respectively. Theseed point was detected using the following Eq. (2):

M = (V1 + V3)/2 − V2V2

(2)

Fig. 4(b) demonstrates the scanning path, which was alongthe horizontal line between (Cx − WFOV/8) and (Cx + WFOV/8).For each location in the path, scanning was performed atangles varying from −30◦ to 30◦ to overcome the tilt of the tube.Our observations in the present study revealed that the anglerange between −30◦ and 30◦ was sufficient to cover the tilt ofthe tube. Fig. 4(c) shows a typical search space (the M valuesin terms of x location and angle) obtained by scanning withthe template. When the template matched a dense line, the Mvalue was relatively high. The seed point location, Xseed, wasdetermined by searching the maximum among the M values,as shown in Fig. 4(c) and (d). In addition to the endotrachealtube, the nasogastric tube often appears in the cervical region.From our observation, it was noted that the density of theendotracheal tube in paediatric chest radiographs was higherthan that of the nasogastric tube, as shown in Fig. 5, and thatthe use of maximum value to search the seed point for theendotracheal tube worked well in this study.

2.6. Line path tracing

After determining the seed point, tracing of the line path from

the seed point was performed. A small patch of an image wasused to detect the endotracheal tube, as shown in Fig. 6(a). Theline segments in the patch were enhanced using the maskas shown in Fig. 6(b), and the line-enhanced image G(x,y) is

4 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10

Fig. 4 – (a) The template and (b) scanning path used to determine the seed point. The x location for the seed point wasdetermined by finding the maximum from (c) the search space and illustrated in (d) the image.

Fig. 5 – The densities of the endotracheal tube (ET tube) andnasogastric tube (NG tube) in a paediatric chest radiograph.

Fig. 6 – (a) The definition of regions of interest (ROI) used todetect the endotracheal tube. (b) The mask for lineenhancement was applied to the ROI to obtain (c) theline-enhanced image. (d) The line path was traced from theseed point using the line-enhanced image.

s i n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10 5

sftsiltpvYwt

2

Awtd

2Flowlatsattdl

{

wty

T

Fo

Fig. 8 – (a) The ROI above and below the candidate location

c o m p u t e r m e t h o d s a n d p r o g r a m

hown in Fig. 6(c). For a simple description of the algorithmor line path tracing, the y-coordinate of the patch image wasransformed from Yneck ∼ Cy to 1 ∼ m. A line path from theeed point (Xseed, 1) was determined using the line-enhancedmage. Along the y position from 2 to m, each x position in theine path was determined based on the previous one, itera-ively. By assuming the current y position as Ycurrent and therevious x position as Xpre, the x position with the highestalue among G(Xpre − 1, Ycurrent), G(Xpre, Ycurrent) and G(Xpre + 1,

current) was selected as the current x position. This processas repeated to find the x position for each y location from 2

o m, and the tracing path is shown in Fig. 6(d).

.7. Analysis of the line path

s the line path was obtained, the features of the line pathere analysed to determine the existence of the endotracheal

ube and location of the tip. This procedure is described inetail as follows.

.7.1. Multi-thresholding for endpoint candidatesig. 7(a) demonstrates the values along the line path (g(y)) in aine-enhanced image. At the location of the tip, the pixel valuef the line-enhanced image would drop acutely. This featureas used to determine the location of the tip. To find the tip

ocation, each pixel value could be compared with the previousverage pixel value along y. If the current value is lower thanhe previous average value by a threshold, the search wouldtop and the tip location is determined. However, choosing

proper threshold is difficult. In the present study, multiplehresholds were used to determine the potential locations ofhe tip, and a final location was selected from these candi-ates by analysing the image features around the candidate

ocations. The average pixel value can be defined as follows:

Ay = g(1) y = 2

Ay = (g(y − 1) + Ay−1)/2 y = 3, . . ., m(3)

here g(y) is the pixel values of the line-enhanced image alonghe line path and A is the average pixel value at the location

y

. The threshold can be defined by Eq. (4) as follows:

y = w × Ay (4)

ig. 7 – (a) The pixel values, g(y), of the line-enhanced image alonbtained by applying multiple thresholds to g(y) and (b) the corre

used for selecting (b) the most suitable location for the tip.

where Ty is the threshold used to compare at location y andw is the weighting factor for the threshold. Multiple thresh-olds were obtained using different w values of 0.8, 0.7, 0.6, 0.5and 0.4. For each w, the comparison was performed along yfrom 2 to m, as shown in Fig. 7(a). If g(y) was lower than Ty atlocation n, the search would stop and the location n − 1 wouldbe selected as a candidate location. Five candidate locationsfor the tip were determined by different w values, as shown inFig. 7(b). The extent of criterion (w) would result in the lengthof candidate location from the seed point.

2.7.2. Selection of endpoint location from the candidatelocationsAs the candidate locations for the tip were determined, themost suitable one was selected from these candidates byanalysing the image features around the candidate locations.Fig. 8(a) shows that two regions of interest (ROI) above andbelow a candidate location were used to analyse the surround-ing features. For each candidate location, the average pixel

g the line path. The candidate locations for the tip weresponding locations in the image.

values in the two ROI (DROI1 and DROI2) were calculated fromthe original image. At the tip location, DROI1 would be smalldue to lower pixel value of the endotracheal tube and the dif-ference between the two ROI (DROI2 − DROI1) would be high at

m s i n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10

Fig. 9 – The length, Lmax, determined by the loosestthreshold (w = 0.4) used as a feature to detect the

6 c o m p u t e r m e t h o d s a n d p r o g r a

the endpoint of the tube. Eq. (5) simply combines these twofeatures as follows:

F = (Dmax − DROI1) × (DROI2 − DROI1) (5)

where Dmax is the highest pixel value in the image. Amongthe candidate locations, the one with the highest F value wasselected as the endpoint location, as shown in Fig. 8(b).

2.7.3. Endotracheal tube existence detection and tiplocalisationWithout a line segment in a chest image, the search processcannot stray considerably from the seed point even when theloosest threshold (w = 0.4) is used. This feature can be usedto indicate the existence of the line segment in the analysedregion. The length (Lmax) determined by the threshold withw = 0.4 was used as a feature to detect the existence of theendotracheal tube, as shown in Fig. 9.

In addition to the endotracheal tube, the nasogastric tubeis often noted in the analysed region, and another feature isrequired to distinguish between the endotracheal tube andother line segments. As shown in Fig. 10, the line path could beseparated into the upper and lower segments by the position pwith highest F value. For the endotracheal tube, the propertiesof the upper and lower segments are very different; however,for the nasogastric tube or other line segments crossing theanalysed region, the properties of the two segments are sim-ilar due to the continuity of the line segment. This featurewas used to distinguish between the endotracheal tube andother line segments. Eq. (6) is defined to indicate the propertydifference between the upper and lower segments as follows:

{C = 0 if p = m

C = (Mupper − Mlower)/Mupper otherwise(6)

Fig. 10 – Averages of the line-enhanced pixel values in th

endotracheal tube.

where Mupper =∑p

y=1g(y)/p, Mlower =∑m

y=p+1g(y)/(m − p),Mupper and Mlower are the average line-enhanced values ofthe upper and lower segments, respectively. The C value isclose to zero if the properties of the two segments are similar;otherwise, C has a high value due to the property difference.The C value was set to zero if p was located at the lower border(m) of the analysed region.

The two features, Lmax and C, were combined to detect the

existence of the endotracheal tube in a paediatric chest radio-graph. If the existence of the endotracheal tube was detectedby the proposed method, the location corresponding to p inthe original image was further used as the tip location.

e segments above and below the selected location, p.

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10 7

Fig. 11 – (a) Distributions of the images with and without endotracheal tubes in terms of Lmax and C. (b) The model obtainedby SVM for the combination of Lmax and C to classify the data. (c) ROC curves representing the discriminant performance ofL

2

2eLoetmuopadttmc

max, C and the combination, respectively.

.8. Evaluation of the proposed method

.8.1. Discriminant performance of the endotracheal tubexistence detection

max and C were combined for use in detecting the existencef the endotracheal tube. The discriminant performance wasvaluated for each feature and combination of the two fea-ures. For the combination of the two features, support vector

achine (SVM) [16,17] with radial basis function kernel wassed in the present study to classify the cases with and with-ut endotracheal tubes based on the Lmax and C values. Torepare the training set, 100 cases with endotracheal tubesnd 100 cases without endotracheal tubes were selected ran-omly from the whole data. The randomly selected cases werehen used to train the classifier and create a model. Finally,

he prediction values for the remaining 1144 cases were deter-

ined using the obtained model. The prediction values for theombination of the two features as well as the values of the

two features themselves were analysed by receiver operatingcharacteristic (ROC) analysis [18,19]. The discriminant perfor-mance was evaluated as Az, the area under the ROC curve. Thedata obtained were analysed and plotted using SPSS statisticalsoftware (version 14.0).

2.8.2. Error of tip detectionA radiologist annotated the tip location for each paediatricchest image with endotracheal tube used in the presentstudy by employing the software developed in-house. Theseannotated locations were then used to evaluate the proposedmethod. Using the manually annotated locations as the goldstandards, the errors of the detections are defined by the dis-tance between the detected locations (xd, yd) and the gold

standards (xs, ys) as follows:

Errtip =√

(xd − xs)2 − (yd − ys)2 (7)

8 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10

Fig. 12 – Distribution of the detection errors for the tip location.

3. Results

The proposed method was evaluated using 528 images withendotracheal tubes and 816 images without endotrachealtubes. The success rate of FOV determination by threshold,Tmask, was 99.1% (1333/1344). Among the images with endotra-cheal tube, the detection accuracy of the seed point was 94.3%(498/528). Fig. 11(a) reveals the distributions in terms of Lmax

and C for the images with and without endotracheal tubes.The results indicated that the paediatric chest images withand without endotracheal tubes can be separated effectivelyby the proposed method. For the combination of two features,SVM was used to obtain a model, as shown in Fig. 11(b), andthe prediction value for each test case was computed usingthe model. The prediction values for the combination of thetwo features as well as the values of the two features wereanalysed by ROC analysis. The discriminant performance wasevaluated as the area under the ROC curve, Az. Fig. 11(c) showsthe ROC curves and Az values for Lmax, C and the combinationof the two features, which were 0.806 ± 0.014, 0.881 ± 0.012 and0.943 ± 0.009, respectively.

In addition to the detection of endotracheal tube existence,detection of the tip location is another procedure that is nec-essary for analysing the position of the endotracheal tube. Toevaluate the precision of the detected locations, the locationsof the tip in 528 paediatric chest images with endotrachealtubes were annotated by a radiologist. The distance (error)between the detected and annotated locations was used toevaluate the detection precision of the proposed method.Fig. 12 shows the distribution of the errors for the 528 images.Most of the cases had errors close to zero, with 85.6% (452/528)of the cases having small errors within 5 mm. The cases withErr greater than 15 mm were due to the failure in detec-

tip

ting the existence of the endotracheal tube. The average errorfor the cases successful in the endotracheal tube existencedetection was 1.89 ± 2.01 mm.

4. Discussion

In the present study, an automated method for detecting theendotracheal tube as well as the tip location in paediatricchest radiographs was proposed. In this method, a seed pointwas determined from a line crossing the cervical region anda line path was traced from the seed point. The two fea-tures Lmax and C determined from the path were combinedto detect the existence of the endotracheal tube. Multiplethresholds applied to the line path were used to determinethe candidate locations, and the tip location was selectedfrom these candidates by analysing the image features aroundthem. The proposed method was evaluated using 528 imageswith endotracheal tubes and 816 images without endotrachealtubes. The discriminant performance of detecting the exist-ence of the endotracheal tube based on the two features was0.943 ± 0.009, and the detection error of the tip location was1.89 ± 2.01 mm. Although the proposed method focused on thepaediatric chest radiographs, a similar method and conceptcan be applied to adult chest radiographs.

To determine the tip location, each pixel value along theline path was compared with the previous average value,iteratively, and the search process was stopped at the locationat which the pixel value acutely dropped. The reductionextent of the pixel value at the tip location varied from imageto image due to image contrast or noise, and the use of a fixedthreshold may pose problems. To overcome these problems,multiple thresholds were used to determine the candidatelocations for the tip, and finally the fittest one was selectedfrom these candidates. For a very strict threshold, from theseed point to the endpoint, any point lesser than the thresholdwould interrupt the search process and result in failure inreaching the endpoint. For a very loose threshold, the search

process may continue and not stop at the endpoint. Basedon the success of seed point detection in the present study,all the tip locations could be detected with w in the range of

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10 9

F dete

0m

mctTeatialatscoowecpt

lwiirdIudtsa

classified as an endotracheal tube with the proposed method.Although the proposed method obtained high perfor-

mance results, some aspects can be considered for further

ig. 13 – Improper radiological positioning caused failure in

.8–0.4. The proposed method using multiple thresholds wasore adaptive to different image conditions.Two main processes, detecting the existence of the line seg-

ent and ruling out the line segment without endpoint, wereombined to detect the existence of the endotracheal tube inhis study, which was achieved by the two features, Lmax and C.he Lmax feature is the length between the seed point and thendpoint determined by using the loosest threshold (w = 0.4)nd is sensitive to the existence of the line segment. Withouthe line segment in the analysed region, the endpoint search-ng process cannot stray considerably from the seed point,nd the length (Lmax) would be very small. In contrast, theength is relatively large when a line segment exists in thenalysed region. The C feature, which is defined based on con-inuity, is sensitive to the distinguishability between the lineegments with and without endpoints. In addition to endotra-heal tube, other line segments such as nasogastric tubes areften present in paediatric chest images and need to be ruledut. The endotracheal tube, a line segment with an endpoint,ould obtain a high C value, whereas the other lines without

ndpoints would obtain low C values. In the present study, theombination of these two features obtained high discriminanterformance for detecting the existence of the endotrachealube.

Failure to determine the seed point and further obtainingow Lmax values was the main cause for false-negative cases

here the endotracheal tubes were not detected. As shownn Fig. 13(a) and (b), the radiographic positioning for thesemages was not suitable, and the line crossing the cervicalegion was not determined properly. This caused failures inetermining the seed point and in tracing the line segment.

n addition to the endotracheal tube existence detection, fail-re in seed point detection also caused failure in tip locationetection. The false-positive cases, in which the nasogastric

ubes were incorrectly detected as endotracheal tubes, arehown in Fig. 14. The contrast between the nasogastric tubend background was very different in the upper and lower

rmining the horizontal line crossing the cervical region.

regions, which resulted in a high C value. Thus, the tube was

Fig. 14 – Nasogastric tube classified as an endotrachealtube due to the contrast of the tube differing in the upperand lower regions.

m s i

r

10 c o m p u t e r m e t h o d s a n d p r o g r a

improvement. First, determining the seed point improperlyresulted in missing the endotracheal tube as well as incorrecttip location detection. A few paediatric chest examinationsexhibited improper radiological positioning. To overcomedetection failure in an improper radiological positioningimage, the existence of the cervical portion could first be deter-mined. Different algorithms for determining the seed pointcould be used according to the image with or without cervicalregion. For the images without cervical region, the algorithmfor detecting the seed point needs to be developed further. Sec-ond, some nasogastric tubes were detected as endotrachealtubes by the proposed method due to high C values. Continuityfeatures need to be further investigated to improve the perfor-mance for distinguishing the line segments with and withoutendpoints. However, the final goal of this study is to identifythe malposition of the endotracheal tube. To analyse the posi-tion of the endotracheal tube automatically, the locations ofsome anatomic structures, such as the carina, clavicle and firstthoracic vertebral body [20], need to be determined as refer-ence positions. The algorithms for detecting these anatomicstructures need to be further developed and combined withthe proposed method to detect the malposition of endotra-cheal tubes in paediatric chest radiographs.

5. Conclusions

In the present study, we proposed an automated method fordetecting the endotracheal tube and location of its tip in pae-diatric chest radiographs. In this method, a seed point wasdetermined first from a line crossing the cervical region andwas used as the initial point to trace the line path. Two fea-tures, Lmax and C, determined from the path were combinedto detect the existence of the endotracheal tube, and multiplethresholds were used to determine the tip location. The pro-posed method obtained high performance results and couldbe useful for detecting the malposition of the endotrachealtubes in paediatric chest radiographs.

Acknowledgement

This work was partially supported by Kaohsiung Medical Uni-versity under Grant No. NSYSUKMU101-029.

e f e r e n c e s

[1] G.M. Schmölzer, M. O’Reilly, P.G. Davis, P.Y. Cheung, C.C.Roehr, Confirmation of correct tracheal tube placement innewborn infants, Resuscitation 84 (2013) 731–737.

[2] J. Peterson, N. Johnson, K. Deakins, D. Wilson-Costello, J.E.Jelovsek, R. Chatburn, Accuracy of the 7-8-9 Rule forendotracheal tube placement in the neonate, J. Perinatol. 26(2006) 333–336.

n b i o m e d i c i n e 1 1 8 ( 2 0 1 5 ) 1–10

[3] O. da Silva, D. Stevens, Complications of airwaymanagement in very-low birth-weight infants, Biol. Neonate75 (1999) 40–45.

[4] T.A. Leone, W. Rich, N.N. Finer, Neonatal intubation: successof pediatric trainees, J. Pediatrics 146 (2005)638–641.

[5] C.P. O’Donnell, C.O. Kamlin, P.G. Davis, C.J. Morley,Endotracheal intubation attempts during neonatalresuscitation: success rates, duration, and adverse effects,Pediatrics 117 (2006) e16–e21.

[6] L.R. Kuhns, A.K. Poznanski, Endotracheal tube position inthe infant, J. Pediatrics 78 (1971) 991–996.

[7] G.J. Hauser, M.M. Pollack, C.J. Sivit, G.A. Taylor, D.I. Bulas, C.J.Guion, Routine chest radiographs in pediatric intensive care:a prospective study, Pediatrics 83 (1989) 465–470.

[8] B. van Ginneken, B.M. ter Haar Romeny, M.A. Viergever,Computer-aided diagnosis in chest radiography: a survey,IEEE Trans. Med. Imaging 20 (2001)1228–1241.

[9] M.A. Hasan, S.L. Lee, D.H. Kim, M.K. Lim, Automaticevaluation of cardiac hypertrophy using cardiothoracic arearatio in chest radiograph images, Comput. MethodsPrograms Biomed. 105 (2012) 95–108.

[10] Z. Huo, S. Li, M. Chen, J. Wandtke, Computer-aidedinterpretation of ICU portable chest images: automateddetection of endotracheal tubes, Proc. SPIE 6915 (2008),69152J-1–69152J-8.

[11] C. Sheng, L. Li, W. Pei, Automatic detection of supportingdevice positioning in intensive care unit radiography, Int. J.Med. Robot. Comput. Assist. Surg. 5 (2009) 332–340.

[12] B. Ramakrishna, M. Brown, J. Goldin, C. Cagnon, D.Enzmann, Catheter detection and classification on chestradiographs: an automated prototype computer-aideddetection (CAD) system for radiologists, Proc. SPIE 7963(2011), 796333-1–796333-6.

[13] B. Ramakrishna, M. Brown, J. Goldin, C. Cagnon, D.Enzmann, An improved automatic computer aided tubedetection and labeling system on chest radiographs, Proc.SPIE 8315 (2012), 83150R-1–83150R-7.

[14] E.A. Harris, K.L. Arheart, D.H. Penning, Endotracheal tubemalposition within the pediatric population: a commonevent despite clinical evidence of correct placement, Can. J.Anaesth. 55 (2008) 685–690.

[15] N. Otsu, A threshold selection method from gray-levelhistograms, IEEE Trans. Syst. Man Cybern. 9 (1979) 62–66.

[16] C.C. Chang, C.J. Lin, LIBSVM: a library for support vectormachines, ACM Trans. Intell. Syst. Technol. 2 (2011),27:1-27:27. Software available athttp://www.csie.ntu.edu.tw/∼cjlin/libsvm

[17] G. García, A. Tapia, M. De Blas, Computer-supporteddiagnosis for endotension cases in endovascular aorticaneurysm repair evolution, Comput. Methods ProgramsBiomed. 115 (2014) 11–19.

[18] C.E. Metz, Basic principles of ROC analysis, Semin. Nucl.Med. 8 (1978) 283–298.

[19] J.A. Hanley, B.J. McNeil, The meaning and use of the areaunder a receiver operating characteristic ROC curve,

Radiology 143 (1982) 29–36.

[20] M.P. Blayney, D.R. Logan, First thoracic vertebral body asreference for endotracheal tube placement, Arch. Dis. Child.Fetal Neonatal Ed. 71 (1994) F32–F35.