International Journal for Multiscale Computational Engineering, 19(1):25–39 (2021)

ANALYSIS OF TIBIA-FIBULA BONE FRACTURE

USING DEEP LEARNING TECHNIQUE FROM X-RAY
IMAGES
Waseem Abbas,1,∗ Syed M. Adnan,1 M. Arshad Javid,2 Wakeel Ahmad,1 &
Farooq Ali1
1
Department of Computer Science, University of Engineering & Technology Taxila, Taxila,
Rawalpindi, 47080, Pakistan
2
Department of Basic Sciences, University of Engineering & Technology Taxila, Taxila,
Rawalpindi, 47080, Pakistan
*Address all correspondence to: Waseem Abbas, Department of Computer Science, University of
Engineering & Technology Taxila, Taxila, Rawalpindi, 47080, Pakistan, E-mail: [email protected]

Original Manuscript Submitted: 9/9/2020; Final Draft Received: 3/9/2021

Deep learning technologies have become a leading tool for disease diagnosis supporting and timely treatment. Many
approaches have been developed for the detection and classification of fractures in human bones. These approaches vary
from several parameters producing different detection and classification traits. Therefore, fracture detection, along with
recognizing its category, is helpful for radiologists and doctors to analyze and handle fracture cases effectively. In this
study, we employed a Faster R-CNN transfer-learning technique. The model was retrained using 50 X-ray images of
tibia-fibula bone fractures. For the evaluation of this study, we used parameters such as Kappa coefficient and the mean
average precision. The overall accuracy of this proposed method has been 97%. It is comprehensively inspected and
correlated with the earlier work on bone fractures, concerning training, detection, classification, and efficiency. The
proposed work proved to have a good impact on accurate classification and detection of fractures. Moreover, here we
analyzed six bone fracture classes: transverse, spiral, oblique, linear, comminuted, and normal. The best configuration
tends to show that this study is tremendously correct and economical with high accuracy. This work shows that the
proposed approach is an effective and useful technique for the dynamic detection, classification, and analysis of various
types of fractures. Furthermore, this approach improved the results, the run time performance, and detection quality
compared with the state-of-the-art techniques used in this area.

KEY WORDS: bone fracture, deep learning, tibia-fibula, X-ray image

1. INTRODUCTION
Machine learning is an emerging field of CAD (computer-aided diagnosis) that boasts high acceptance in the health-
care industry and provides an impressive way to help in disease diagnosis and its treatment. In most developed
countries, bone fracture is a common problem and is rapidly increasing. There are a total of 206 bones in a human
body; Tibia/fibula bone fractures more frequently happen compared to other bone fractures, because of the thin pe-
riosteum lining, the front part of the tibia-fibula covered by the skin, which is directly located under the skin fracture,
as discussed in Myint et al. (2018). The bone fracture occurs due to accidents or other bone diseases such as trauma,
stressing either a specific medical condition that weakens the bones for more details see Muchtar et al. (2018). The
overall incidence of tibia-fibula fractures reported in the Swedish Fracture Registeris (SFR) is 51.7 per 100,000, in a
year as discussed in Wennergren et al. (2018). A bone fracture is complete or partial; complete bone fracture happens
when the parts of the bone are wholly separated from each other, while an incomplete bone fracture extends partly
across the bone. It occurs to the wrist, ankle, hip, rib, tibia-fibula, chest, and knee.

1543–1649/21/$35.00 © 2021 by Begell House, Inc. www.begellhouse.com 25

26 Abbas et al.

At present, several state-of-the-art machines are available to generate digital images of living organism/organs;
these are ultrasound, X-ray computed tomography (CT), and magnetic resonance imaging (MRI). Still, X-ray tech-
nology is the mast common and oldest. It is also economical with a painless detection method. It is used to make X-ray
images of bones in the human body. Bone fractures are classified into different classes such as transverse, oblique, spi-
ral, comminuted, avulsed, segmented, impacted, torus, and greenstick, for more detail see Al-Ayyoub and Al-Zghool
(2013). Diagnosis and treatment of the bone fracture are important as the wrong diagnoses lead to dissatisfaction of
the patient and tend to have dangerous consequences, which tarnish the reputation of a diagnostic institution as well.
Moreover, manual analysis of medical images is error-prone and time-consuming. Therefore, computer vision plays
a tremendous role in automatically providing the visual semantics of medical images. Consequently, it has helped to
analyze a large amount of medical data with high accuracy of diagnosis with minimum time and effort.
Recently, machine learning techniques have improved medical diagnosis; for example, the tedious task of screen-
ing for identical findings is no longer performed by doctors and radiologists. Hence, time is saved that allows them to
interact with more patients. Past studies used hand engineering and static techniques to identify fractures as discussed
by Badgeley et al. (2019). However, this is inefficient because they require several preprocessing steps such as noise
removal, segmentation, the region growing, region merging, edge detection, feature extraction, and threshold setting;
more detail is available in Castro-Gutierrez et al. (2019). Moreover, these are static approaches that are weak in
locating the exact region of interest (bounding box) of bone fractures.
In this study, we employed a Faster R-CNN transfer-learning technique and configured it with its default param-
eters. In addition, we retrained it on 50 X-ray images of tibia-fibula bone fractures. Moreover, here we analyzed six
bone fracture classes: transverse, spiral, oblique, linear, comminuted, and normal. For the evaluation of this study,
we used mAP, kappa coefficient as evaluation parameters for defining accuracy. The best results tend to show that
this study is tremendously precise. The results of this research show that this is an efficient technique, which is bet-
ter for dynamic detection, and classification of fracture classes. Furthermore, this approach improves the run-time
performance and detection quality, even with a limited dataset as compared to earlier techniques.

2. LITERATURE REVIEW
Various artificial intelligence techniques are being used for automatic features detection and its analysis in different
fields of life from plant disease detection to COVID-19 detection from X-ray images. In one study, Tanzi et al.
(2020) artificial intelligence techniques, were used to analyze bone fractures. This technique used scale invariant
feature transform (SIFT) and Haar wavelet transforms. The SIFT method is used to detect feature points such as
compression, rotation, and scaling, while Haar wavelet transforms is used to save memory space. The study used a
dataset with a such as the total of 100 X-ray images. The model was trained with only 30 X-ray images and tested
with 70 X-ray images. The model is evaluated using state-of-the art evaluation parameters area under the curve,
sensitivity, specificity, and accuracy. The study claims an average accuracy of 94.30%. In another study, Jin et al.
(2020) developed a customized 3D UNet architecture named the FracNet model to analyze rib fractures. This model
is composed of encoder-decoder, 3D convolution, bath-normalization, nonlinearity, and max pooling. This model was
trained using the RibFrac dataset, where 420 images were used for training, and 120 images for testing. This method
achieved a detection sensitivity of 92.9% and a segmentation Dice of 71.5% on the test cohort.
Zhou et al. (2020) presented a technique for automatic fracture detection and classification of rib bones based
on Faster R-CNN. This technique was used to achieve three goals which are the robustness of the model, the frac-
ture detection and classification, and an efficient mechanism. The study claims that the Faster R-CNN performed
better than YOLO V3 in detection accuracy and detection speed. This study achieved a high precision of 91.1% and
sensitivity of 86.3%. In another study, Hržić et al. (2019) proposed a fracture detection and classification method
using X-ray images. This method used local entropy to remove noise from X-ray images. The local Shannon entropy
was computed for each pixel of the image by using a sliding 2D window. First, image segmentation was performed
on the original image, then a graph theory technique was applied to images for removing negative bone contours
and enhancing the edge detection. Finally, the difference of the extracted and estimated contour was calculated for
detection and classification of the fracture. The study reports an overall accuracy of 86.22% with a 91.16% detection
rate.

International Journal for Multiscale Computational Engineering

Analysis of Tibia-Fibula Bone Fracture Using X-Rays 27

In another study, Yang et al. (2019) proposed two line-based fracture detection schemes including standard
line-based detection and adaptive differential parameter optimized (ADPO) line-based fracture detection using X-ray
images. It differentiates the fracture line from a nonfracture line using extracted features of recognized patterns, and
the fractures are classified with the artificial neural network (ANN). The ADPO-based fracture detection technique
performed better than standard line-based fracture detection with an average accuracy of 72.89%. In another study
Castro-Gutierrez et al. (2019) proposed local binary pattern (LBP) based feature extractor and SVM for acetabulum
fracture detection and classification. This approach deals with low-resolution images in a better way by improving
image quality in the preprocessing phase. The study claims an overall accuracy of 80%.
Kim and MacKinnon (2018) aimed to identify the extent to which deep learning models pretrained on nonmedi-
cal images can be used for fracture detection and classification. In this technique the top layer of the Inception version
3 network was retrained using wrist radiographs and used for classification of fractures. This model was fully trained
using 11,112 X-ray images for eightfold and achieved an accuracy of 88%. In another study, Dimililer (2017) de-
veloped a classification system that can detect and classify the bone fractures. The system comprises two principal
stages, namely, preprocessing of images using image enhancement techniques and a classification phase using a neu-
ral network. The system was tested on bone fracture images and claimed a high classification rate. In another similar
study, Al-Ayyoub and Al-Zghool (2013) developed a system which can detect bone fracture and fracture type. The
developed system extracts features after preprocessing of the X-ray images, and then they used different classification
algorithms to detect the existence of a fracture along with its type. The shared results show that the proposed system
is accurate and efficient.
In summary, computer aided fracture detection and classification is an active field of research. However, an
efficient, accurate, and cost-effective diagnosis is still a challenge which is also subject to the qualification of a
radiologist. It is obvious from the literature that previous works lack in accuracy and reliability. Moreover, there
is no systematic approach to localize the fracture in an efficient manner, unlike the proposed technique. The main
contributions of this research are as follows:
• We employed the Faster RCNN deep transfer learning technique for fast, accurate, and automatic detection and
classification of fracture.
• Analyzed six classes of tibia-fibula bone fractures.
• The method dynamically detects the fractures.
• Experimental analysis proved that the proposed technique delivers better results as compared to the state-of
the-art existing techniques.

3. MATERIAL AND METHOD

This study, employed the Faster R-CNN transfer-learning-based technique for accurate detection and classification of
tibia-fibula bone fractures. The proposed method used VGG-16 architecture as a base network to generate a convolu-
tional feature map to produce the proposals for fracture area detection and was followed by classification.

3.1 The Architecture of Faster R-CNN

The overall architecture of the proposed model with a configuration for tibia-fibula bone fracture detection and clas-
sification is shown in Fig. 1. The model comprised an input layer, convolutional layers, feature map layer, a region
of interest (ROI) layer, pooling layer, classifier layer, region proposal network (RPN) layer, and an output layer. The
proposed model comprises 13 convolutional layers, 13 ReLu layers, and 4 pooling layers. The input X-ray images
are established as P × Q (height × width) in different sizes 1365 × 2048 to 540 × 400, and M × N (height × width)
in fixed size, which is 224 × 224, which is forwarded to a pre-prepared CNN up to a transitional layer, done with
a convolutional highlight map. We used RPN to identify the area of specific fracture in the X-ray image. After this,
the area of interest is fed to ROI pooling to find out the required area and this classified the input images. The region
of proposal is predicted at each sliding window. After this, the offset value of the bounding box is predicted. This

Volume 19, Issue 1, 2021

28 Abbas et al.

FIG. 1: Architecture of Faster R-CNN for fracture analysis

network is less time-consuming as compared to other networks and has a high detection efficiency. The first layer
of the proposed model is the input layer, which accepted the X-ray image of the dimension P × Q and automati-
cally converted it into M × N. The X-ray image is cropped to a specific size and fed to the convolution layer. In the
convolutional layer, a kernel with a specific size convolves on a received image. The proposed model has a default
configuration. The convolutional layer minimizes the size of the input X-ray image, which reduces the effort and
computation cost of the model. The produced size of the convolution layer is shown in Eq. (1).

(i − k) + 1
outputsize = , (1)
s
where i is the input dimension, k is the kernel dimension, and s is the stride dimension.
The pooling layer reduces the X-ray image size for further input to the next layer; it works as down-sampling.
Max pooling was used in the whole model for the best results. The pooling layer output is shown in Eq. (2).

(i − p) + 1
outputsize = , (2)
s
where i is the input dimension, p is the pooling dimension, and s is the stride size. ReLu is used as a activation
function; its gradient is zero or one, and it ignores the negative values, providing much faster computation for proposed
work as compared to other activation functions. The ReLu activation function is shown in Eq. (3).

y = max(0, x). (3)

At the first convolutional layer, the configuration is 224 × 224 × 64, similarly, at the second convolutional layer
112 × 112 × 128, at the third convolutional layer 56 × 56 × 256, and at the fourth convolutional layer 28 × 28 ×
512. The feature map contains complete information for the X-ray image while keeping the location of the fractures.
The feature map includes all features of the input image, such as if a fracture is on the right side of the X-ray image
also, convolutional layers are activated so that these fractures are on the right part of the X-ray image in a feature map
with the output 28 × 28 × 512 which is fed to the next layer.
The RPN layer is an intermediate layer having a filter (kernel) size 512, 3 × 3 filter, stride 1, and padding also 1.
The classification layer has filter size 18 (9 × 2), 1 × 1, stride 1 and padding 0. The regression layer has filter size
36 (9 × 4), 1 × 1, stride 1, and padding 0. In this layer, there is a small window slide on the feature map layer
made by building a small network for classification of the fractures, which means whether any fracture is present

International Journal for Multiscale Computational Engineering

Analysis of Tibia-Fibula Bone Fracture Using X-Rays 29

or not. Also, the regression-bounding box locates where the fracture is present. Localization information is provided
by the position of a sliding window. More excellent localization information for the sliding window is supplied by
box regression. The RPN layer is used to locate the region of interest by creating a bounding box. RPN is a total
convolution network that is used to predict the target area from each input X-ray image and shows the score on it.
The real target value was represented by probability, and to create a high-value region of interest box, the RPN was
the end-to-end training for classification and detection.
The convolutional features region of interest was used as the input to the ROI pooling layer. It made a bounding
box around the fracture with its class name (type). It only reduced the error of the feature map such as average pooling
and max pooling. It was also used to split the feature into the equatorial region; after, that, we used max-pool for an
entire area; hence, the input size did not max-pool the output of max pooling. The classifier layer is the final layer.
It is used to classify the input X-ray image with its class name by generating a bounding box around it. There exist
two output layers; i.e., the Soft-max activation layer is used for an object classification of fracture types and the linear
activation function for bounding boxes coordinates regression.
The deep, fully convolutional network (DFCN) is used for detection and segmentation. The DFCN is an ordinary
CNN, but the main difference is that the last fully connected layer is replaced with another convolution layer. A box
is generated around the fracture that plays a vital role in the identification and classification of input X-ray images.
Multiple bounding boxes are used at each position. The position of the bounding box is 320 × 320 of an image size of
600 × 800. There are three scales of the box: 128 × 128, 256 × 256, and 512 × 512 represented by three colors with
a ratio of 1:1, 1:2, and 2:1. Each bounding box has its features of prediction with a single-scale feature and multiscale
feature.
The labels of the bounding box are positive and negative. It depends upon the value of IoU (intersection over
union). In the case that IoU is more than 0.7 then its detection is positive for ground truth; on the other hand, if IoU is
less than 0.3, the label is assigned as unfavorable to all ground-truth boxes. A clear set of boxes increased speed and
accuracy.

3.2 Evaluation Parameters

3.2.1 Overall Accuracy
The overall accuracy is calculated by Eq. (4) as discussed in Ranganathan (2017).
x
overallacuracy = , (4)
y
where x is the right classified value and y is the overall value.

3.2.2 Kappa Coefficient

The overall performance of the model is evaluated by the Kappa coefficient, for more details see Myint et al. (2018).
Cohen’s Kappa coefficient is less than or equal to 1, 0, or less value, showing the classifier is inefficient. There is
no standardized way to interpret its values. However, according to some scheme, a value less than 0 indicates null
agreement, 0–0.20 as slight agreement, 0.21–0.40 as fair agreement, 0.41–0.60 as modest, 0.61–0.80 as considerable
and 0.81–0.99 as nearly perfect agreement. Here 1 is for perfect agreement. We used a confusion matrix that contains
all classes of fracture and is shown in Eq. (6).
obseredaccuracy − expedtedagreement
K= , (5)
1 − expectedagreement
x0 − Xe 1 − x0
K= =1− , (6)
1 − xe 1 − xe
where x0 is the observed accuracy and xe is the expected agreement. It essentially tells us a way that is much better
for how a category is acting over the performance of a classifier that arbitrary guesses indiscriminately, consistent
with the frequency of every class.

Volume 19, Issue 1, 2021

30 Abbas et al.

3.2.3 Mean Average Precision (mAP)

Mean average precision is calculated using Eq. (7).

1 X T P (x)
mAP = x ∈ classes . (7)
|classes| T P (x) + F P (x)
T P (x) is true positive: probability is a predicted bounding box for class x, tab is a ground truth bounding box
of class x, and IoU is an intersection over union which is (pred-bb, gt-bb) ≥ 0.5. F P (x) is false positive: a bounding
box for class x prediction is pred-bb, there is no ground truth for class x where gt-bb and IoU is an intersection over
union which is (pred-bb, gt-bb) ≤ 0.5. For a class x to compute IoU between any predicted box and ground box take
the best overlap. The highest overlap is between the box of truth predicted and any ground.

3.3 Experimental Setup

3.3.1 Dataset
The proposed model was trained and tested on X-ray images of tibia-fibula bone fracture. The dataset comprises 50
X-ray images. The dataset was collected from the Islamabad Diagnosis Center in Islamabad, Pakistan. The datasets
which are used to support the finding are not publicly available due to certain protected information about health. The
supportive dataset is accessible from the associated author by reasonable request.

3.3.2 Implementation Details

In this study, all the experiments are performed using cross-validation to produce the best evaluation results. Training
data labeled using labelImg and each fracture class contain nine images, but the normal class contains only five images
for training. This research work used two-way training; the first RPN classifier and regressor were trained and the
second was trained jointly at the same time. The second training approach is 1.5 times faster with similar accuracy.
The backpropagation loss is saved after every eight steps as a checkpoint. We trained our model on a Dell laptop,
which had the following specifications, Intel CORE i3 CPU, with 4 GB of memory. The training is stopped when loss
reached at 0.005 and here the total steps are 4k. The total time taken to train the model is 72 hours. The proposed
method is implemented in Python 3.7 with the Tensorflow framework.

4. RESULTS
4.1 Training
In this stage, the top layer of Faster R-CNN is retrained using inception v2 (Version 2) networks. The training is
continued until loss reaches 0.0005%. We used stochastic gradient descent (SGD) to train the proposed method to
minimize convolution layer filters, proposal region weights, and fully connected layer weights. Stochastic gradient
descent (SGD) performs an update of the parameters for each example of training x(i) and label y(i).

β = β − η.∇βk (β : xi , y i ), (8)
where β is the learned rate and η is the new learning rate.

4.1.1 Learning Rate

The learning rate of a model, concerns new learned information override’s the old one. The character η, is used to
represent the learning rate. It was not too large or too small. If the learning rate is too low, the model cannot be
made learn or stop training. The learning rate should be diversified for the best training of the model.The step-based
learning rate, according to the predefined step, is computed using Eq. (9) and is shown in Fig. 2. Here, the learning

International Journal for Multiscale Computational Engineering

Analysis of Tibia-Fibula Bone Fracture Using X-Rays 31

rate is a tuning parameter in an optimization algorithm that determines the step size at each iteration while moving
toward a minimum of loss function.

ηn = η0 df loor([1+n]/r) . (9)
In the above equation, ηn is the learning, at iteration η, η0 ; d is the learning rate, and r is the drop rate. Here the
floor function is used to convert all values into zeros, which are less than 1.

4.1.2 Classification Loss

Classification loss is the measure of predictive imprecision of the classification function. When we compare, it may
be happen that a model of the same loss type is a better predictive model indicating a lower loss. Classification loss is
like the cross-entropy loss when we transform the x values via the sigmoid function before applying the cross-entropy
loss. We desire to apply the sigmoid function. This function is typically the expected value of this model, while labels
are ground truth values. The classification loss is computed using Eq. (10) and results are shown in Fig. 3.
2
x
Ijobj (zj (b) − ẑj (b))2 .
X
classification loss = (10)
j=0

In the above equation, Ijobj = 1 when an object is present in cell j; otherwise 0. ẑj (b) = 1 denotes the probability
of class b in the cell j.

4.1.3 Localization Loss

Localization loss is a loss between ground truth and model detection. The coordinate’s correction transformation is
identical to what R-CNN does in bounding box regression. Where the model suggests whether the bounding box with
coordinates is matched to the ground-truth value with coordinates of any object are predicted correction terms. The
localization loss is given by Eq. (11) and results are shown in Fig. 4.

2 2
S X
B S X
B
obj obj √ √
X X p p
2 2
λcoord Ijk [(yj − ŷj ) + (zj − ẑj ) ] + λcoord Ijk [( gj − ĝj )2 + ( mj − m̂j )2 ]. (11)
j=0 k=0 j=0 k=0

In the above equation, roman on-line a boundary box kth is located in a cell j for detecting the object, else 0. λcoord
represents loss increasing in the boundary box coordinates.

FIG. 2: Learning rate of R-CNN model

Volume 19, Issue 1, 2021

32 Abbas et al.

FIG. 3: Classification loss

FIG. 4: Localization loss

4.1.4 RPN Loss

The RPN loss of the model describes that an object is detected or not to each bounding box. When we train the RPN,
we give a class label to each class. A box that has an IoU measure higher than 0.7, considers a positive detection of
fracture, and if the IoU value is less than 0.3 it considers a negative detection of the fracture. If the box is neither
negative nor positive, do not consider it the training objective. The first part of Eq. (12) shows the loss of the clas-
sification over two classes, i.e., if is an object or not an object. The second part of the equation shows the bounding
boxes, regression loss if an object is detected. RPN loss as shown in Fig. 5.

1 X 1 X ∗
L(qi , ri ) = (Lcls qi ,qi∗ ) + λ q Lreg (ri , ri∗ ). (12)
Zcls i Zreg i i

In the above equation, i is used to represent the box as an index. Moreover, qi shows the prediction probability of
the box i being an object. If there is an object, the value of ground-truth is given by qi∗ being equal to 1 else 0 (zero).
ri is used to show the detected box coordinates and ri∗ is used to show the ground-truth box value, which is associated
with the positive box. Lcls is used to show classification loss, which is log loss two classes (object vs. not object).
Lcls (ri , ri∗ ) = R(ri , ri∗ ) is used to show the regression loss, where loss function (smooth L1) is, and represents the
regression loss. The regression loss is given by qi∗ Lreg when a positive box qi∗ = 1 otherwise qi∗ = 0. The cls and
reg layers contain upon the output qi∗ , ri∗ . The Zcls and Zreg are used to represent the normalized and equal weights.

International Journal for Multiscale Computational Engineering

Analysis of Tibia-Fibula Bone Fracture Using X-Rays 33

4.1.5 Objectness Loss

The model is trained in a such way that if an object is detected with a correct bounding box, then it has a perfect IoU
score, otherwise a smaller score.

2
S X
B
obj
X
λobj Ijk [(bj − b̂j )2 . (13)
j=0 k=0

obj
In the above equation, b̂j is used for the object in the cell. The core of a box k in a cell j, Ijk = 1. If the jth
boundary box is in cell j, then an object is detected in the bounding box, otherwise not. Objectness loss is given by
Eq. (13) and is shown in Fig. 6.

2
S X
B
noobj
X
λnoobj Ijk [(bj − b̂j )2 . (14)
j=0 k=0

noobj obj
Ijk is the complement of Ijk , the confidence score of the box k in the cell j. λnoobj when only background
is detected, but the object is not detected. The weight loss of the factor λnoobj is default 0.5.

FIG. 5: RPN loss

FIG. 6: Objectness loss

Volume 19, Issue 1, 2021

34 Abbas et al.

4.2 Model Generated Output

The proposed model detects fractures along its area and classifies it into a predicted class. The detection and calcifi-
cation of tibia-fibula as shown by Figs. 7–12, where (a) part show the input given to the model and (b) part show the
output.

5. DISCUSSION
This study considers the tibia-fibula bone fracture problem for detection, classification, and analysis. We employed a
Faster R-CNN deep learning technique with transferlearning for tibia-fibula bone fracture analysis. We evaluated the
proposed method with the Kappa coefficient as shown in Table 1. Overall accuracy has been found to be 98% and the
Kappa coefficient is computed as 97% as shown in Table 1. The mean average precision (mAP) also obtained good
accuracy by using the values in Tables 2 and 3.
The high accuracy of the proposed technique of fracture detection and its analysis provide’s a higher efficiency
than earlier methods for detection, classification, and analysis of different fracture types as discussed in Muchtar et al.
(2018), Castro-Gutierrez et al. (2019), and Yang et al. (2019). Before this, the proposed technique was pre-trained
upon a nonmedical dataset. Now, it is used for medical purposes and we employed this proposed model of bone
fracture X-ray images. Earlier proposed techniques for bone fracture analysis have been static, but for the first time,
we introduced a dynamic method. Moreover, the proposed technique also detects fractures from live streaming-like
video. The proposed method provides better results than earlier ones, as shown in Figs. 7–12.
The proposed technique shows excellent performance in detection, classification, and analysis of the fracture
with its different classes. The findings show that the tibia-fibula fracture study has enhanced interpretation, which is
helpful for radiologists to identify and examine tibia-fibula patients. An effective approach comes from our proposed
study. The small size datasets, which are not publicly available, can be used for training. The overall performance
of the proposed method is excellent from every aspect. During the training, the learning rate remains constant, as
shown in Fig. 2, but losses in the start of training are much higher, but with passes the time loss decreases as shown
in Figs. 3–6. The model is fully trained in 40k steps as shown in Fig. 7 and loss remains only 0.0005%. Similarly, in
the Hržić et al. (2019) study using the segmentation technique, and local entropy and for fracture classification (not
based on deep learning) reported an accuracy of 91.16% for fracture classification in the X-ray image of a child’s

(a) (b)
FIG. 7: Oblique displaced fracture

International Journal for Multiscale Computational Engineering

Analysis of Tibia-Fibula Bone Fracture Using X-Rays 35

(a) (b)
FIG. 8: Comminuted fracture

(a) (b)
FIG. 9: Oblique nondisplaced fracture

ulna and radius bones. This study detects the fracture manually, but the proposed technique automatically classify, as
shown in Fig. 11, this demonstrates transfer learning, which was earlier trained upon the ImageNet dataset, which is
nonmedical and we trained it upon the medical images using the Inception v2 network.
Furthermore, this study minimizes the radiologist’s involvement in detecting, classifying, and analyzing fractures.
This level of accuracy is beneficial for the radiologist in workflow optimization and minimization of error, which
happened in manual detection and classification analysis. It improves the diagnosis and productivity of the radiologist
by saving time. This study reduces the patient’s loss, which he faced for reporting errors or delays. This research is

Volume 19, Issue 1, 2021

36 Abbas et al.

(a) (b)
FIG. 10: Linear fracture

(a) (b)
FIG. 11: Multiple fracture

primarily for the detection of fracture, where the detecting feature is missing rather than misdiagnosed. The proposed
technique provides detection, classification, and analysis of fracture types accurately. In these comparative studies,
the proposed technique is robust in the performance of fracture analysis under various conditions of the X-ray images
which were comparable to those of the traditional methods.
This technique successfully detects the fracture area of the bone and makes it under the rectangle within the input
image. Although there are many techniques to identify the fracture area of bone using image processing and machine
learning, these are not enough. The proposed technique detect six classes of fractures. Also, we compare our model
accuracy with the other state-of-the-art techniques as shown in Table 4. To summarize, the proposed technique pre-
sented in this paper is improved from the earlier work because we employed Faster R-CNN and retrained the top layer
of it, and operated end-to-end on the deep neural network. Therefore, in distinction to the earlier techniques, the pro-
posed technique presents potential benefits, such as efficiency, high accuracy, consistent interpretation, instantaneous
reporting results, reproducibility, and an accurate analysis tool in this domain.

International Journal for Multiscale Computational Engineering

Analysis of Tibia-Fibula Bone Fracture Using X-Rays 37

(a) (b)
FIG. 12: Normal

TABLE 1: Confusion matrix tibia-fibula bone fracture

Normal Transverse Spiral Oblique Linear Comminuted Classification Producer
overall accuracy
Normal 5 5 100%
Transverse 9 9 100%
Spiral 9 9 100%
Oblique 9 9 100%
Linear 9 9 100%
Comminuted 1 9 10 85.71%
Truth overall 5 9 9 9 10 9 51
User accuracy
100% 100% 100% 100% 85.71% 100%
(recall)
Overall accuracy 98.04%
Kappa coeffcient 97.5%

TABLE 2: Ground truth values of tibia-fibula bone fractures

Image Name Width Height Class XMin YMin XMax YMax
input1.jpg 590 979 Transverse 201 473 327 604
input2.jpg 540 400 Spiral 249 193 263 274
input3.jpg 642 809 Oblique 363 220 448 374
input4.jpg 1365 2048 Linear 828 1280 997 1523
input5.jpg 540 400 Comminuted 463 319 485 338
input6.jpg 550 410 Normal 343 323 470 330

6. CONCLUSIONS
This study presents the Faster R-CNN-based method for automatic identification of tibia-fibula bone fractures along
with classification of its different fracture types (classes). This study analyzed six different classes of fractures, i.e.,

Volume 19, Issue 1, 2021

38 Abbas et al.

TABLE 3: Predicted values of tibia-fibula bone fractures

Image Name Width Height Class XMin YMin XMax YMax
output1.png 557 639 Transverse 181 444 308 571
output2.png 513 384 Spiral 222 182 240 266
output3.png 631 681 Oblique 360 219 442 378
output4.png 334 527 Linear 167 229 203 348
output5.png 541 425 Comminuted 72 217 96 232
output6.png 549 412 Normal 343 323 470 330

TABLE 4: Results comparison with state-of-the-art techniques

Reference Method Dataset Accuracy
Tanzi et al. (2020) Preprocessing and deep CNN Wrist radiographs 93.3%
Jin et al. (2020) Customized 3D UNet architecture Rib fracture 92.2%
X-ray images of child
Hržić et al. (2019) Local entropy and segmentation 91.16%
ulna and radius bones
Convolutional neural
Yang et al. (2019) Hip radiographs 78%
networks (CNNs)
Feature extraction with LBP
X-ray images of
Badgeley et al. (2019) (local binary pattern) and 80%
acetabulum fracture
the SVM (support vector machine)
Convolutional neural X-ray images of
Castro-Gutierrez et al. (2019) 79.3%
networks (CNNs) and ANNs osteoporosis
deep convolutional
Kim and MacKinnon (2018) Wrist radiographs 95%
neural networks
X-ray images of tibia/fibula
Muchtar et al. (2018) Scan-line algorithm 87.5%
fracture
Recurrent neural networks Vertebral fracture on
Tomita et al. (2018) 89.2%
(RNN) and CNN’s a CT scan
Harris corner detection X-ray images of a lower
Myint et al. (2018) 82%
Algorithm and KNN leg bone
X-ray images of tibia-fibula
— Proposed method 97%
fractures

transverse, oblique, spiral, segmented, comminuted, and normal. We employed a deep transfer learning technique,
which can detect fractures with the bounding box, and its type. This method eliminates preprocessing and reduces
the training complexity of detection and classification. The proposed method was found accurate with an average
accuracy of 97%. Future work of this study is to localize the dimensions of the fracture, and to find the applicability
of this study on other long bones such as legs and arms.

REFERENCES
Al-Ayyoub, M. and Al-Zghool, D., Determining the Type of Long Bone Fractures in X-Ray Images, WSEAS Trans. Inf. Sci. Appl.,
vol. 10, no. 8, pp. 261–270, 2013.
Badgeley, M.A., Zech, J.R., Oakden-Rayner, L., Glicksberg, B.S., Liu, M., Gale, W., McConnell, M.V., Percha, B., Snyder, T.M.,
and Dudley, J.T., Deep Learning Predicts Hip Fracture Using Confounding Patient and Healthcare Variables, NPJ Dig. Med.,
vol. 2, no. 1, pp. 1–10, 2019.
Castro-Gutierrez, E., Estacio-Cerquin, L., Gallegos-Guillen, J., and Obando, J.D., Detection of Acetabulum Fractures Using X-
ray Imaging and Processing Methods Focused on Noisy Images, in Proc. of 2019 Amity International Conference on Artificial
Intelligence (AICAI), IEEE, pp. 296–302, 2019.

International Journal for Multiscale Computational Engineering

Analysis of Tibia-Fibula Bone Fracture Using X-Rays 39

Dimililer, K., IBFDS: Intelligent Bone Fracture Detection System, Procedia Comput. Sci., vol. 20, pp. 260–267, 2017.
Hržić, F., Štajduhar, I., Tschauner, S., Sorantin, E., and Lerga, J., Local-Entropy Based Approach for X-Ray Image Segmentation
and Fracture Detection, Entropy, vol. 21, no. 4, p. 338, 2019.
Jin, L., Yang, J., Kuang, K., Ni, B., Gao, Y., Sun, Y., Gao, P., Ma, W., Tan, M., and Kang, H., Deep-Learning-Assisted Detection
and Segmentation of Rib Fractures from CT Scans: Development and Validation of Fracnet, EBioMedicine, vol. 62, p. 103106,
2020.
Kim, D. and MacKinnon, T., Artificial Intelligence in Fracture Detection: Transfer Learning from Deep Convolutional Neural
Networks, Clin. Radiol., vol. 73, no. 5, pp. 439–445, 2018.
Muchtar, M., Simanjuntak, S., Rahmat, R., Mawengkang, H., Zarlis, M., Sitompul, O., Winanto, I., Andayani, U., Syahputra,
M., and Siregar, I., Identification of Tibia and Fibula Bone Fracture Location Using Scanline Algorithm, J. Phys.: Conf. Ser.,
vol. 978, no. 1, p. 012043, 2018.
Myint, W.W., Tun, K.S., and Tun, H.M., Analysis on Leg Bone Fracture Detection and Classification Using X-Ray Images, Mach.
Learn. Res., vol. 3, no. 3, pp. 49–59, 2018.
Ranganathan, P., The Effects of Land Use Change on Carnivore Use of Wildlife Dispersal Routes in Ranthambhore Tiger Reserve,
India, PhD, Duke University, Durham, NC, 2017.
Tanzi, L., Vezzetti, E., Moreno, R., and Moos, S., X-Ray Bone Fracture Classification Using Deep Learning: A Baseline for
Designing a Reliable Approach, Appl. Sci., vol. 10, no. 4, p. 1507, 2020.
Tomita, N., Cheung, Y.Y., and Hassanpour, S., Deep Neural Networks for Automatic Detection of Osteoporotic Vertebral Fractures
on Ct Scans, Comput. Biol. Med., vol. 98, pp. 8–15, 2018.
Wennergren, D., Bergdahl, C., Ekelund, J., Juto, H., Sundfeldt, M., and Möller, M., Epidemiology and Incidence of Tibia Fractures
in the Swedish Fracture Register, Injury, vol. 49, no. 11, pp. 2068–2074, 2018.
Yang, A.Y., Cheng, L., Shimaponda-Nawa, M., and Zhu, H.Y., Long-Bone Fracture Detection Using Artificial Neural Networks
Based on Line Features of X-Ray Images, in Proc. of 2019 IEEE Symposium Series on Computational Intelligence (SSCI),
IEEE, pp. 2595–2602, 2019.
Zhou, Q.Q., Wang, J., Tang, W., Hu, Z.C., Xia, Z.Y., Li, X.S., Zhang, R., Yin, X., Zhang, B., and Zhang, H., Automatic Detection
and Classification of Rib Fractures on Thoracic CT Using Convolutional Neural Network: Accuracy and Feasibility, Kor. J.
Radiol., vol. 21, no. 7, pp. 869–879, 2020.

Volume 19, Issue 1, 2021