Computers in Biology and Medicine 78 (2016) 120–125

Contents lists available at ScienceDirect

Computers in Biology and Medicine

http://d0010-48

n CorrDrive, C

E-m

journal homepage: www.elsevier.com/locate/cbm

Computer-Aided Detection (CADx) for Plastic Deformation Fracturesin Pediatric Forearm

Yuwei Zhou a, Uygar Teomete b, Ozgur Dandin c, Onur Osman d, Taner Dandinoglu e,Ulas Bagci f, Weizhao Zhao a,b,n

a Department of Biomedical Engineering, University of Miami, Coral Gables, FL 33146, USAb Department of Radiology, University of Miami Miller School of Medicine, Miami, FL 33136, USAc Department of General Surgery, Bursa Military Hospital, Bursa 16800, Turkeyd Department of Electrical and Electronics Engineering, Istanbul Arel University, Istanbul 34500, Turkeye Department of Physical Medicine and Rehabilitation, Bursa Military Hospital, Bursa 16800, Turkeyf Center for Research in Computer Vision, University of Central Florida, Orlando, FL 32816, USA

a r t i c l e i n f o

Article history:Received 23 March 2016Received in revised form14 September 2016Accepted 15 September 2016

Keywords:Bowing fracturePlastic deformationRadiusUlnaCADx

x.doi.org/10.1016/j.compbiomed.2016.09.01325/& 2016 Elsevier Ltd. All rights reserved.

espondence to: Department of Biomedicaloral Gables, FL 33146, USAail address: w.zhao1@miami.edu (W. Zhao).

a b s t r a c t

Bowing fractures are incomplete fractures of tubular long bones, often observed in pediatric patients, whereplain radiographic film is the non-invasive imaging modality of choice in routine radiological workflow. Dueto weak association between bent bone and distinct cortex disruption, bowing fractures may not be diag-nosed properly while reading plain radiography. Missed fractures and dislocations are common in accidentsand emergency practice, particularly in children. These missed injuries can result in more complicatedtreatment or even long-term disability. The most common reason for missed fractures is that junior radi-ologists or physicians lack expertise in pediatric skeletal injury diagnosis. Not only is additional radiationexposure inevitable in the case of misdiagnosis, but other consequences include the patient's prolongeduncomfortableness and possible unnecessary surgical procedures. Therefore, a computerized image analysissystem, which would be secondary to the radiologists’ interpretations, may reduce adverse effects and im-prove the diagnostic rates of bowing fracture (detection and quantification). This system would be highlydesirable and particularly useful in emergency rooms. To address this need, we investigated and developed anew Computer Aided Detection (CADx) system for pediatric bowing fractures. The proposed system has beentested on 226 cases of pediatric forearms with bowing fractures with respect to normal controls. Receiveroperation characteristic (ROC) curves show that the sensitivity and selectivity of the developed CADx systemare satisfactory and promising. A clinically feasible graphical user interface (GUI) was developed to serve thepractical needs in the emergency room as a diagnostic reference. The developed CADx system also has strongpotential to train radiology residents for diagnosing pediatric forearm bowing fractures.

& 2016 Elsevier Ltd. All rights reserved.

1. Introduction

Plastic bowing fractures of children are called incomplete frac-tures in the forearm (radius and/or ulna). Clinical terminology wasfirst defined by Borden in 1975 [3] based on the following uniquefeatures of fractures compared to other bone fractures as: 1) thereexists a relative elasticity in the pediatric forearm long bone, 2) If anextensive axial force is applied to the bone, 3) the bone may undergoa bowing deformation, and 4) eventually followed by fracture. Plainradiographs are usually the choice of modality for evaluating bowingfractures. Other imaging techniques, such as ultrasound, CT and MRI,can also be helpful, but are in no means the clinical standard [4,5]. Inroutine radiological evaluation of these fractures by plain films, the

Engineering, 1251 Memorial

imaging feature of the bowing fracture is excessively bended bonewith an absence of cortex disruption [6,7]. Since a bowing fracturecan be quite difficult to diagnose, radiologist may readily miss thecases with inconspicuous bending in space [1,2,8]. The consequencesmay include prolonged uncomfortableness, additional x-ray ex-posures, and possible surgery procedures.

To determine the degree of excessive bending, films of the con-tralateral uninjured forearms were used as comparison standardsbecause the radius and ulna in healthy children may bend slightly[3,6,7]. This approach raises some clinical difficulties with scenariosof both forearms injured, unavailable contralateral radiographs, oravailable but not equivalent field of view (FOV). In addition, bothanteroposterior (AP) and lateral (LAT) projections are necessary, be-cause the bowing may only be visible in one projection.

In order to clarify the ambiguity between bowing and a bowingfracture, clinicians have tried different methods to describe theclassification criteria quantitatively. Angulation is one of the

Y. Zhou et al. / Computers in Biology and Medicine 78 (2016) 120–125 121

frequently used methods in describing forearm fractures [2,9–17].Angulation measures the angle at the “suspected” fracture point(called angular point) formed by the two pieces of bones on eachside of the fracture point. However, angulation may fail to providean accurate measurement when a bowing deformity does notshow an angular point. Measuring malunion of a radius in an APview radiography is another method used to describe fractures[18]. This method was originally developed for adults, but lateradopted to analyze natural bowing of the radius in healthy chil-dren in AP view films [19]. For ulnar plastic deformation, a similarmethod was proposed by Lincoln and Mubarak in 1994 [20,21].This method used the maximum ulnar bow and its location toassist in the recognition of radial-head dislocation. A 3D methodhas also been designed for complicated and severe deformities,which are usually not apparent in 2D radiographs. In these cases,CT scans were required for both the fractured arm and the con-tralateral arm [23–25]. Since then, there has been little or no im-provements in the computational aspect of bowing fracture diag-nosis, mostly due to the relatively low frequency of such cases inemergency rooms, and lack of enough interest in the domain [22].However, the nontrivial consequences of misdiagnosis make thisproblem worthy of attention.

With the rapid development of computer technology, computer-aided diagnosis/detection (CAD) has been widely used in manydiseases spanning from lung cancer to prostate cancer. One of theobjectives in CAD design is to reduce observational oversights andfalse negative rates. Radiologists/orthopedists have also adoptedCAD techniques in the diagnostic processes to detect bone tumorsor fractures. Successful CAD systems include routine clinical appli-cation for bone metastases [26], surgical planning for orbital frac-tures [27], detection of long bone fractures, and real-time surgicallocalization for long bone fractures [28–30]. For bowing fractures, aCT-based method was reported [31]. By superimposing a deformedulna onto the normal ones, reconstructed 3D CT images of forearmbones allows for measurement of the deformity in the AP, LAT, andaxial planes. Results from the method successfully elucidated me-chanisms of such an injury, indicating that traumatic bowing de-formities of the ulna involved rotation rather than bending. How-ever, there lacks systematic and quantitative assessments of thebowing fracture for both radius and ulna without comparison to theuninjured contralateral bones, particularly when only plain radio-graphic films available. Thus, it is highly desirable to create such aCAD system for pediatric forearm bowing fracture.

Herein, we propose a computerized image analysis of bowingfractures on plain-film with quantitative support for a radiologists’diagnostic decisions. The outcome of this investigation is a ComputerAided Detection (CADx) system featuring a semi-automated boneextraction preprocess followed by analyzing two quantitative assess-ment parameters. Results from statistical analysis for each projection(AP and LAT view) and each bone (radius and ulna) are also presented.The developed CADx system's sensitivity and selectivity are summar-ized and presented by receiver operating characteristics (ROC) curves.

2. Materials and methods

2.1. Data

With the IRB approval, plain radiographs of children aged 1–18years old were retrospectively collected. A total of 226 plain radio-graphs of forearms subjected to pediatric trauma were reviewed bytwo expert radiologists with over 10 years of experience. There were59 diagnoses of bowing fractures, and all others were reported asnormal (89 AP projections, 78 LAT projections). For the normal cases,no radius or ulna bowing deformity was found in the AP or LAT view.For abnormal cases, a bowing deformity may exist in either radius or

ulna, in either or both of the AP and LAT views. Thus, for a systematicanalysis, normal and abnormal cases are divided in to 4 groups, bybone (radius/ulna) and by field of view (AP/LAT), respectively:

Normal groups: radius in AP view, radius in LAT view, ulna inAP view, ulna in LAT viewAbnormal groups: radius in AP view, radius in LAT view, ulna inAP view, ulna in LAT view

2.2. Preprocessing for bone extraction

Pediatric forearm bone fractures include bowing fractures andbuckle fractures. This report is focused on the discussion on bowingfractures. We noticed that the radius or ulna in pediatric forearmsalways have some degrees of bending, regardless of the presence ofbowing fractures or not. This feature inspired us to investigate thecurvature of these bones. In order to measure the curvature, weextract the bowing portion and exclude the buckle portion from theradius/ulna. In other words, we need first to select the same piece ofanatomically equivalent bone from every patient. However, in orderto create a uniform processing protocol, it is necessary to definewhere the “cutting” point is, which is the first question we answer.

In this study, we designed and implemented a semi-automatedbone extraction method that is based on the radius of curvature (R).R is a measurement of a curve function's radius that best approx-imates the curve at that point, defined by an analytical equation:

( )=

[ + ]( )

dy dx

d y dxR

1 /

/,

1

2 3/2

2 2

where y is the curve functionwith respect to variable x. In other words,a point's bending degree on a curve, called curvature, is assessed by ameasurement-circle's radius. The radius is related to the curve functionby Eq. (1). This definition intuitively states that the more “bending” ofthe curve, the smaller radius of curvature; the “flatter” the curve, thelarger radius of curvature. Anatomically, the location with the max-imum radius curvature matches the location where the buckle starts.We describe the bone extraction processing as follows:

Before the measurement, the image is rotated to a reasonabledirection (for convenient and comfortable visualization). We thenuse a rectangular “crop” tool to extract either the radius bone or theulna bone. The extracted bones must have a clear outside boundaryof the radius (OBR) or outside boundary of the ulna (OBU). Theprocess is demonstrated in Fig. 1. On the extracted bone image, wefirst collect the coordinate (x, y) for each point on the OBR/OBUthrough edge detection. These coordinates will be fitted into afunction y¼ f(x), where f is a 6th-degree polynomial. We then cal-culate the radius of curvature R(x) for each point (x-coordinate) onthe boundary of the bone based on Eq. (1), which is plotted in Fig. 2.The x-coordinates for the two maximum R(x) values are the cuttingpoints for bone extraction.

Using a polynomial to perform curve fitting is a common ap-proach in science and engineering applications. The degree of apolynomial for curve fitting is regulated by the feature/nature of thecurve. If the “inherent feature/nature” of the curve does not have aknown definition, the degree of the polynomial should then bepractical in the application. Through 226 cases, we notice that R(x) monotonically increases or decreases when the function is apolynomial of 6th–degree or lower. A higher-degree polynomial willreduce the error for y ¼ f(x) curve fitting, however, will also in-crease the number of local maximums (turbulence) for R(x). In or-der to minimize the curve fitting error and maintain the robustnessof the method, we used the 6th–degree polynomial for all cases. TheR(x) curve usually appears as a double-peak distribution which iscorresponding to two buckle starting locations. The bone is auto-matically cut at these two points where RE1 or reaches a local

Fig. 1. In order to obtain useful information from the radiography image, preprocessing is needed. a) An original radiography film usually includes both the radius and ulnain the image plane. b) The whole image plane is rotated to a reasonable direction for an easy pre-processing. c) The preprocessing is to use a rectangular “crop” tool to extracteither the radius bone or the ulna bone as long as a “clear” outside boundary of radius (OBR) or outside boundary of ulna (OBU) remains.

Fig. 2. The boundary coordinates on the extracted bone will be fitted into a 6th orderpolynomial. a) Cropped bone of interest shows a clear outside boundary of radius (OBR).b) Boundary coordinates are plotted by the blue line. c) Calculated radius of curvature(R) for each coordinate is plotted by the red dash line, which is typically in a two-peakdistribution. d) Selected bone is returned by the auto-extraction procedure.

Table 1Total of 226 cases (the same case may have more than one view) were available forthe current study. Since the number of abnormal cases was much smaller than thenumber of normal cases for each view of radius/ulna, 35–40 normal cases wererandomly selected from each view to perform statistical analysis.

Sample size Radius (APview)

Radius (LATview)

Ulna (APview)

Ulna (LATview)

Normal 86 64 62 42Abnormal 17 15 11 13

Fig. 3. Two assessment parameters are defined in the study. After edge detection(blue line), all data points are fitted into a circle (red line). The angle (orange dotline) on the circle that covers all data points is defined as the central angle (CA). Thecentral curvature (CC) is the reciprocal of the fitted circle's radius. (For inter-pretation of the references to color in this figure legend, the reader is referred tothe web version of this article.)

Y. Zhou et al. / Computers in Biology and Medicine 78 (2016) 120–125122

maximum (Fig. 2). Outside of these two points, the buckle starts.The ulna in AP/LAT view usually shows a single-peak R(x) dis-tribution. Thus, we combine the anatomical features of ulna (beforethe ulna coronoid process) to determine the cutting point for allcases. This extraction procedure makes data collection in a con-sistent and comparable format. Table 1 shows the case number foreach group after semi-automated extraction.

2.3. Definition of measurement parameters

After cutting, the extracted bone from each subject will beanatomically equivalent for comparisons. The data points (x,y) onthe extracted bone boundary will be used for analysis. We defineand use two parameters, in terms of central angle (CA) and circlecurvature (CC), to evaluate the degree of bending of bones (Fig. 3).We briefly describe the parameters as follows:

Central Angle (CA): We first fit all data points on the extractedbone into a circle with optimal center-x, center-y and radius(linear regression). The angle subtended by the arc that coversall data points is defined as the central angle (CA).Circle Curvature (CC): The circle curvature (CC) is defined as thereciprocal of radius of curvature. The description for radius of

curvature was provided in previous section (Eq. (1)). Since thecurve function is a circle function (CA), the radius of curvature isindeed the radius of the circle. Therefore, CC¼1/R. The CC has asimilar transition trend as CA, i.e., the “flatter” the bone is(corresponding to a large R) the smaller value of CA or CC.

2.4. Comparison experiment

In order to validate the reliability of using radius of curvaturefor bone extraction and the measurement parameters (centralangle and circle curvature), we conducted a comparison study byusing a “simulated-manual” bone extraction procedure. The si-mulated extraction was based on the computed extraction pointsby adding/subtracting a random number of pixels (o10) to/fromthese points to mimic manual extraction's uncertainty. After suchan extraction procedure, all subject bones were also examined bythe CA and CC parameters.

Y. Zhou et al. / Computers in Biology and Medicine 78 (2016) 120–125 123

Finally, we use an independent two-sample t-test to evaluatethe effectiveness of each parameter for each paired group. At theend, we plot the ROC curves to find the best threshold value foreach paired group.

2.5. GUI and processing procedure

We implemented the above mentioned method into a graphicaluser interface (GUI). Fig. 4 illustrates the GUI, which facilitates thefollowing functions: (1) input/output of radiographic images, (2) imagepre-processing, (3) image extraction, and (4) calculation of statisticalparameters and data analysis. The GUI was originally created in Ma-tLab, but compiled and built into a standalone application, which canbe executed on any platform running any operation system.

3. Results

3.1. Statistical analysis

We first measured and calculated two parameters for all cases.Data show that obvious differences exist between normal groupsand bowing fracture groups for both parameters. 95% confidenceintervals (CI) of CA and CC for each group are listed in Table 2.

Independent two-sample t-test was used to examine the statistics forthe collected data. Outcomes show that both CA and CC parameters caneffectively distinguish bowing fracture from normal bone in all pairedcomparisons (po0.05 for all cases). However, the CA and CC parameterscannot consistently show the same results on the bones by the simu-lated-manual extraction. Statistical results are shown in Table 3.

3.2. ROC curve

After we confirmed that CA and CC were sensitive in detectingbowing deformity, we can determine the best threshold value for eachpaired group by constructing the ROC curves. We used half of thesamples in each group to plot ROC curves and the other half of thesamples to test the thresholds. For the normal group, data were ran-domly selected. However, because the abnormal group's sample size istoo small to support random selection, we first ranked them in a nu-merical order (based on the values of CA or CC), and then selected everyother one as the initial data set. In Fig. 5, ROC curves for CA and CCwereplotted at various threshold settings, ranging from 0 to 45° for CA (stepsize is 1°), and 0-250 k/pixel for CC (step size is 5 k/pixel). The ROC

Fig. 4. Graphical User Interface (GUI) and procedure flowchart present the operating fundisplay the original film image. The “Operating area” allows and displays all image proceprocess. The “Function buttons” are associated with graphical editing tools, such as rese

curve of CA is obviously closer to the upper left corner than the ROCcurve of CC, indicating that CA is a better predictor for bowing fracture.

Based on the ROC curve of CA, a threshold value can be de-termined for this CADx system. In this study, we used the medianvalue between normal and abnormal groups as a threshold (e.g.,13.59° for radius in AP view). We adopted this mechanism, definedas the optimal area under the curve (AUC), to determine allthreshold values in this study. Table 4 shows the best thresholdvalue determined by AUC and accuracy rate tested by the otherhalf of data for each paired group. Both the AUC and accuracy rateof the test show that thresholds generated from the ROC curves ofCA are greatly credible, better than CC, which even reached point(0, 1) in ROC space for radius and ulna in LAT view. CA is alsooverwhelmingly better than CC in the accuracy rate.

4. Discussion

Pediatric forearm bowing fractures can be easily misdiagnosed ifthe radiographic images were not read by a pediatric-radiologist. Inthe CADx approach, the first difficulty encountered is the un-biaseddata acquisition. The developed bone extraction method based onradius of curvature calculation has shown to be a reliable approach.In this study, we investigate the effectiveness of two parameters toquantitatively assess bowing deformity of children forearm withoutcomparing with the contralateral arm. We demonstrate that both CAand CC can effectively differentiate bones with bowing deformityfrom normal bones. A comparison study between the proposed boneextraction method and the simulation method through CA and CCparameters further validated the method's robustness (Table 3). Interms of sensitivity and selectivity, CA is a more reliable parameterthan CC. In addition, CA is also a “scale-invariant” parameter, which isindependent of image resolution. In plain radiography, FOV is asso-ciated with the size of the object. Therefore, CC is impacted by theimage scale. Our next development for the CADx system will gen-erate an age-based database by using both CA and CC parameters.

In addition to the rotation tool associated with the developed GUI,a manual segmentation tool and a graphical editing tool are alsoprovided. These manual tools are used to handle special cases, such asoverlapped bones. The CADx system has two potential immediateapplications. First, it can be applied to an emergency room as a di-agnostic reference. Pediatric forearm trauma may require orthopedicand radiology specialists for final diagnosis. However, an emergencyroom usually lacks such specialists. This CADx system can provide a

ctions and procedures. The “Original image window” is used to open image file andssing procedures and outcomes. The “Output windows” show the results after eacht, undo, erase, crop, and assessment parameter calculations.

Table 295% of CI (confidence interval) of CA and CC for each group are presented. Note that 1) CA was measured and calculated unit of degrees, 2) CC was measured in unit of pixels.However, since curvature is reciprocal of radius of curvature, CC's unit is 1/pixel after calculation. A constant of 1000 (k) was used to scale three parameters in a similar“comparable” level for data reading only.

Type Radius (AP view) Radius (LAT view) Ulna (AP view) Ulna (LAT view)

CA (unit: °) normal 10.2770.91 (n¼40) 8.4271.00 (n¼40) 7.0970.98 (n¼37) 6.5571.05 (n¼39)abnormal 17.6471.64 (n¼17) 17.8373.41 (n¼15) 22.9375.56 (n¼11) 20.7172.34 (n¼13)

CC (unit: 1000/pixel) normal 36.8174.91 (n¼40) 30.4275.22 (n¼40) 34.4676.05 (n¼37) 29.9678.40 (n¼39)abnormal 77.40726.94 (n¼17) 66.66720.61 (n¼15) 99.55742.32 (n¼11) 71.34714.34 (n¼13)

Table 3Comparisons for assessment parameters CA and CC yield significant difference (po0.05)between normal and abnormal groups (all views) consistently by the radius curvaturebased bone extractionmethod. However, the CA and CC parameters lost sensitivity whenthe bones were extracted by simulated manual extraction (o710 pixel variations onthe automated extraction points) between normal and abnormal groups.

Analysis Parameter p-value

Radius (AP) Radius(LAT)

Ulna (AP) Ulna(LAT)

Central Angle (CA) o 0.001 o 0.001 o 0.001 o 0.001Based on curvatureextraction

Central Angle (CA) o 0.001 0.074 o 0.001 o 0.001Based on simulationextraction

Circle Curvature (CC) 0.006 0.002 0.006 o 0.001Based on curvatureextraction

Central Curvature (CC) 0.237 0.594 0.068 0.086Based on simulationextraction

Table 4Threshold values obtained from the ROC curves are presented, where AN (accuratenormal) is the accuracy rate for the normal group (negative accuracy), AA (accurateabnormal) is the accuracy rate for the abnormal group (positive accuracy), and AUCis the area under curve, for each paired group.

Threshold AN AA AUC

CA (°) Radius(AP) 13 90% 100% 0.941Ulna(AP) 13 94% 100% 0.992Radius(LAT) 14 95% 86% 0.898Ulna(LAT) 14 100% 100% 1

CC (kP) Radius(AP) 53 90% 66% 0.845Ulna(AP) 53 67% 60% 0.763Radius(LAT) 55 90% 57% 0.806Ulna(LAT) 45 80% 100% 0.886

Y. Zhou et al. / Computers in Biology and Medicine 78 (2016) 120–125124

“second opinion” to radiologists. Second, the system can serve as atraining platform. Our current system has included more than twohundreds cases of plain radiography images from the database in theDept. of Radiology, University of Miami Miller School of Medicine. We

Fig. 5. Receiver operating characteristic curves show the sensitivity (TPR: true positive raCC. Optimal threshold value for each paired group can be determined by using ROC cumeans CA is a better predictor for bowing fracture compared to CC.

will include more cases to build a training system for radiology re-sidents. Residents will randomly select cases from the database, givetheir trial diagnoses, and run the system to verify their trials.

5. Conclusion

This study introduced a semi-automated bone extractionmethod which reduced the error in manual measurement andcreated normalized data sets for statistical comparisons. We de-fined and investigated two parameters to assess bowing fracturesin pediatric forearm. The results show that CA and CC can effec-tively detect bowing in radius/ulna by optimal threshold values to

te) and selectivity (FPR: false positive rate) for the assessment parameters of CA andrve. ROC curves of CA are obviously closer to the upper left corner than CC, which

Y. Zhou et al. / Computers in Biology and Medicine 78 (2016) 120–125 125

distinguish bowing deformity from normal bone. The optimalthreshold values, derived from ROC, improve diagnostic reliabilityand accuracy. The CADx system can be potentially applied inemergency rooms and be used as residency training software.

Conflict of interest statement

All authors (Yuwei Zhou, Uygar Teomete, Ozgur Dandin, OnurOsman, Taner Dandinoglu, Ulas Bagci and Weizhao Zhao) declarethat they have no conflict of interest in submitting this manuscript.

Ethical standards statement

The nature of the study is a retrospective investigation of pe-diatric forearm bone fractures. Statement of informed consent wasnot applicable since radiographic images were already generatedand were provided for the study by removing identities and re-maining age and diagnosis only. All authors have been CITI (col-laborative institutional training initiative) certified regardless theyworked on programming only, tested the program, or accessednon-identity radiographic images.

References

[1] L. Romano, A. Pinto, Errors in Radiology, Springer, 2012.[2] S.P. Jadhav, L.E. Swischuk, Commonly missed subtle skeletal injuries in chil-

dren: a pictorial review, Emerg. Radiol. 15 (6) (2008) 391–398.[3] S. Borden IV, Roentgen recognition of acute plastic bowing of the forearm in

children, Am. J. Roentgenol. Radium Ther. Nucl. Med 125 (3) (1975) 524–530.[4] F. Checo, M.C. Lee, Pediatric forearm fractures, Tech. Orthop. 24 (3) (2009)

163–171.[5] O. Ackermann, et al., Ultrasound diagnosis of juvenile forearm fractures, J.

Med. Ultrason. 37 (3) (2010) 123–127.[6] J.E. Crowe, L.E. Swischuk, Acute bowing fractures of the forearm in children: a

frequently missed injury, Am. J. Roentgenol. 128 (6) (1977) 981–984.[7] A.A. Sawkar, L.E. Swischuk, S.P. Jadhav, Plastic bending fractures in children,

Contemp. Diagn. Radiol. 34 (20) (2011) 1–6.[8] J.E. Aponte Jr., A. Ghiatas, Acute plastic bowing deformity: a review of the

literature, J. Emerg. Med 7 (2) (1989) 181–184.[9] C. Creasman, D.J. Zaleske, M.G. Ehrlich, Analyzing forearm fractures in chil-

dren. The more subtle signs of impending problems, Clin. Orthop. Relat. Res188 (1984) 40–53.

[10] J.A. Roberts, Angulation of the radius in children’s fractures, J. Bone Jt. Surg. Br.68 (5) (1986) 751–754.

[11] D. Vittas, E. Larsen, S. Torp-Pedersen, Angular remodeling of midshaft forearmfractures in children, Clin. Orthop. Relat. Res 265 (1991) 261–264.

[12] A. Gulec, O. Buyukbebeci, Traumatic bowing of the forearm in a child, Eur. J.Radio. 23 (3) (1996) 247–249.

[13] P. Vorlat, H. De Boeck, Bowing fractures of the forearm in children: a long-termfollowup, Clin. Orthop. Relat. Res 413 (2003) 233–237.

[14] L. Gordon, et al., Acute plastic deformation of the ulna in a skeletally matureindividual, J. Hand Surg. Am. 16 (3) (1991) 451–453.

[15] W.E. Sanders, J.D. Heckman, Traumatic plastic deformation of the radius andulna. A closed method of correction of deformity, Clin. Orthop. Relat. Res. 188(1984) 58–67.

[16] D. Rowland, E. Baird, Common upper limb injuries in childhood, Surg. (Oxf.) 32(1) (2014) 9–16.

[17] S. Carson, et al., Pediatric upper extremity injuries, Pediatr. Clin. North Am. 53(1) (2006) 41–67.

[18] E.H. Schemitsch, R.R. Richards, The effect of malunion on functional outcomeafter plate fixation of fractures of both bones of the forearm in adults, J. BoneJt. Surg. 74 (7) (1992) 1068–1078.

[19] M. Firl, L. Wunsch, Measurement of bowing of the radius, J. Bone Jt. Surg. Br. 86(7) (2004), 1047-9.47-.

[20] T.L. Lincoln, S.J. Mubarak, “ Isolated” traumatic radial-head dislocation, J. Pe-diatr. Orthop. 14 (4) (1994) 454–457.

[21] S. Sano, et al., Radial head dislocation with ulnar plastic deformation in chil-dren: an osteotomy within the middle third of the ulna, J. Shoulder Elb. Surg.17 (5) (2008) 768–771.

[22] W.-K. Jeong, et al., Factors affecting assessment of ulnar bowing in radio-graphy, J. Pediatr. Orthop. 32 (1) (2012) 48–53.

[23] L.M. Fayad, F. Corl, E.K. Fishman, Pediatric skeletal trauma: use of multiplanarreformatted and three-dimensional 64-row multidetector CT in the emer-gency department, RadioGraphics 29 (1) (2009) 135–150.

[24] D.A.R. Storelli, et al., The use of computer-aided design and 3-dimensionalmodels in the treatment of forearm malunions in children, Tech. Hand Up.Extrem. Surg. 19 (1) (2015) 23–26.

[25] T. Murase, et al., Three-dimensional corrective osteotomy of malunited frac-tures of the upper extremity with use of a computer simulation system, J.Bone Jt. Surg. 90 (11) (2008) 2375–2389.

[26] O. Tokuda, et al., Investigation of computer-aided diagnosis system for bonescans: a retrospective analysis in 406 patients, Ann. Nucl. Med 28 (4) (2014)329–339.

[27] X. Fan, et al., Late reconstruction of the complex orbital fractures with com-puter-aided design and computer-aided manufacturing technique, J. CraniofacSurg. 18 (3) (2007) 665–673.

[28] P.A. Grutzner, N. Suhm, Computer aided long bone fracture treatment, Injury35 (Suppl 1) (2004), S-A57-A64.

[29] H.Y. Chai, et al., Gray-level co-occurrence matrix bone fracture detection,WTOS 10 (1) (2011) 7–16.

[30] L. Joskowicz, et al., FRACAS: a system for computer-aided image-guided longbone fracture surgery, Comput. Aided Surg. 3 (6) (1998) 271–288.

[31] E. Kim, et al., Three-dimensional analysis of acute plastic bowing deformity ofulna in radial head dislocation or radial shaft fracture using a computerizedsimulation system, J. Shoulder Elb. Surg. 21 (12) (2012) 1644–1650.

Yuwei Zhou, M.S., is graduate student with the Department of Biomedical En-gineering, College of Engineering, University of Miami, USA

Uygar Teomete, M.D., is assistant professor with the Department of Radiology,University of Miami Miller School of Medicine, USA

Ozgur Dandin, M.D., is general surgeon with the Department of General Surgery,Bursa Military Hospital, Turkey

Onur Osman is associate professor with the Department of Electrical and Elec-tronics Engineering, Istanbul Arel University, Turkey

Taner Dandinoglu, M.D., is physical therapy and rehabilitation specialist with theDepartment of Physical Medicine and Rehabilitation, Bursa Military Hospital,Turkey

Ulas Bagci, Ph.D., is assistant professor with the Center for Research in ComputerVision, School of Computer Science, University of Central Florida, USA

Weizhao Zhao, Ph.D., is professor with the Departments of Biomedical Engineeringand Radiology, University of Miami, USA