Detection of Bone Fractures in Upper Extremities Using XceptionNet
and Comparing the Accuracy with Convolutional Neural Network
Trisha Veronica M. Y. and P. V. Pramila
Department of Computer Science and Engineering, Saveetha University, India
Keywords: Bone Fractures, Convolutional Neural Networks, Deep Learning, Health, Novel XceptionNet Model,
Upper Extremity.
Abstract: Goal of this study is to differentiate an unique XceptionNet model deep learning (DL) model to Convolutional
neural networks (CNN) in order to recognize bone fracture at the upper extremities of hands with considerably
higher accuracy. Materials and Methods: To enhance the accuracy metric of bone fracture recognition in the
upper extremity areas of hands, deep learning techniques such as the novel XceptionNet model (N=10) and
Convolutional Neural Networks (N=10) were iterated. In this work, bone fracture detection using x-rays
images dataset was used which was acquired via Kaggle. The dataset, which has a total of 9463 x-ray images,
is 181 MB in size. The Train and Val datasets were separated. There are 633 photos in the val dataset and
8987 images in the train dataset which were used to calculate accuracy for the two groups with an 80% of G
power. Results and Discussion: The classification accuracy of the novel XceptionNet model is 88.74%, which
is significantly greater than the accuracy of the CNN model, which is 72.50% for Bone fracture detection
using x-rays images dataset. It is discovered that the novel XceptionNet model and convolutional neural
networks differed statistically with a notable difference of p<0.001 (p<0.05) (2-tailed). Conclusion: The
methodology used in this paper with two deep learning models namely novel XceptionNet and convolutional
neural networks. The results reveal the usefulness of the latest methods in identifying bone fracture detection
in the upper extremities and show their value for fracture prediction at an early stage.
1 INTRODUCTION
As we are all aware, bones act as the body's
framework and as points where muscles can attach.
As society advances, people's propensity for bone
fractures increases. Broken bones can impair blood
supply to the bones and result in additional
complications, but they can also hurt the wounded
limbs as well as the surrounding soft tissues. Delays
in the diagnosis and medical management of injuries
can potentially lead to complications with fractures.
Because broken bones can seriously harm a person's
body, it is vital for people to have a clear-cut and
efficient diagnosis of injuries (Pranata et al., 2019).
The analysis and remedial measures of many diseases
have advanced thanks to new trends and technology
in radiology (Sa, Mohammed and Hefny, 2020).
However, there are 7.5 times as many patients and
radiological scans each year as there are radiologists.
The procedure for detection is required to be
completed in the very brief period of time possible
because each of these images needs to be analyzed (Li
et al., 2020). In order to alleviate the frequency of
these issues and intensify the diagnosis effectiveness,
computer-aided diagnosis (CAD) has been
recommended recently. They may serve as an
alternative perspective for doctors to support and
assist their decisions. Numerous research has been
conducted to develop CADs for use at a range of
medical contexts, including the identification of
cancer, breast lesions, cognitive tumors, and
particularly cracks and fractures of upper extremity
(Lee and Fujita, 2020). In contrast to traditional
approaches, deep learning can automatically discover
strong, complicated features eliminating the need for
feature development using unprocessed information.
Utilization of advanced machine learning in genetics,
electronic health records, and images for medicine
and pharmaceutical discovery are said to have evident
advantages in maximizing the employment in
biomedicine data and boosting health. (S. Yang et al.,
2021).
The topic of detecting bone fractures in the upper
extremity areas were covered in 4 articles published
360
Y., T. and Pramila, P.
Detection of Bone Fractures in Upper Extremities Using XceptionNet and Comparing the Accuracy with Convolutional Neural.
DOI: 10.5220/0012772300003739
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 1st International Conference on Artificial Intelligence for Internet of Things: Accelerating Innovation in Industry and Consumer Electronics (AI4IoT 2023), pages 360-366
ISBN: 978-989-758-661-3
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

by IEEE and over 16,800 papers according to Google
Scholar over the past five years. The study by
(Agarwal et al., 2022) concentrates on the
classification of photographs from X-rays and
assessment of lower limb broken bones using a
variety of mechanisms for undertaking, such as
approach to envision refining, unsharp masking and
harris corner recognition for features extraction, and
classification methods, such as decision tree(DT) for
determining whether a bone is fractured or
unfractured and KNN algorithm for determining a
particular type of cracked bone, with an accuracy of
92% and 83%, independently. Region Proposal
Network (RPN) accelerated Region-based
Convolutional Neural Networks (R-CNN) Deep
learning (DL) model for cracked bone diagnosis and
categorization was proposed by (Abbas et al. 2020)
Additionally, the model's topmost layer was retaught
through 50 x-rays and the Inception v2 (version 2)
network blueprint. They assessed the effectiveness of
the suggested framework when it comes to both
category and determination. Overall accuracy of their
system in terms of classification and detection is 94%.
Because bone fractures result in various mechanical
vibration responses, the research by (Yoon et al.,
2021) investigates the transverse vibration responses
of bones. The technique of modal guarantee to
evaluate the longitudinal resonance remarks of both
solid and damaged joints criteria was developed and
utilized to detect bone fractures. In the work by
(Raisuddin et al., 2021), For the purpose of
diagnosing wrist fractures, they established and
evaluated a cutting-edge DL-based source dubbed
DeepWrist, and they evaluated it against two test sets:
one for the general population and one difficult test
set that only included instances that needed CT
confirmation. Their findings show that while a widely
used and effective method, such as DeepWrist,
performs almost perfectly on the hard test set, it does
far worse on the general independent sets of test
typical precision is 0.99 (0.99-0.99) Vs 0.64. (0.46–
0.83). The study "Bone fracture detection in X-ray
images using convolutional neural networks"
(Bagaria, Wadhwani, 2022) is in my opinion the
finest one.
Convolutional neural networks (CNN) do
remarkably well when classifying images that are
highly similar to the dataset. Convolutional Neural
Networks, on the other hand, typically struggle to
categorize images if there is even a slight tilt or
rotation. This problem was examined in this work,
and it was resolved by training the system using a
method known as data augmentation, or image
augmentation. The research's primary objective is to
determine the efficiency of two novel DL (deep
learning) models XceptionNet and CNN in predicting
the upper extremity bone fractures.
2 MATERIALS AND METHOD
Proposed experiment took place in the programming
lab of the Saveetha School of Engineering, SIMATS.
Convolutional neural networks and XceptionNet are
the two deep learning methods selected for this
research. The experiment was repeated using the
aforementioned algorithms 10 times. The iteration's
sample size is 10. The G-power for the available data
samples was calculated to be 80%, and Alpha was
adjusted to 0.05.
The testing environment for the planned study
was a Lenovo ideapad FLEX 15 IWL with 8.00 GB
RAM, an Intel i7 8th gen processor, 475 GB of
storage capacity, and Windows 11OS.
Bone fracture detection using x-rays images is the
open source dataset used in this research project,
which was acquired via Kaggle (Sairam 2022).
Multiple joints in the upper extremity regions of
hands are included in this dataset. The dataset, which
has a total of 9463 x-ray images, is 181 MB in size.
The Train and Val datasets were separated. There are
633 photos in the val dataset and 8987 images in the
train dataset.The Train set of data was further divided
into subcategories for fractured and not-fractured
data. The not fractured set of data included x-ray
images of healthy bone while the fractured
subcategory included images of bone fractures in the
upper extremities.
2.1 XceptionNet
Depth Wise Separable Convolutions are used in the
XceptionNet deep convolutional neural network
architecture. Researchers at Google created it.
Convolutional neural networks' Inception modules,
according to Google, act as a transitional stage
between the regular convolution process and the
depthwise separable convolution process (a
depthwise convolution followed by a pointwise
convolution).
2.2 Convolutional Neural Networks
(CNN)
A CNN (Convolutional Neural Network) is a kind of
ANN employed in deep learning that is often applied
to image, text, object, and recognition categorization.
Convolutional neural networks, usually referred to as
Detection of Bone Fractures in Upper Extremities Using XceptionNet and Comparing the Accuracy with Convolutional Neural
361

convnets or CNN, are a well-known technique in
computer vision applications. The class of deep
neural networks that are employed in the evaluation
of visual imagery. To identify objects from picture
and video data, this kind of architecture is studied.
NLP (neural language processing), video or image
recognition, and other applications utilize it.
2.3 Statistical Analysis
Statistical Package for the Social Sciences (SPSS)
version 26 has been used to do the crucial statistical
calculations. SPSS was developed to undertake
statistical analysis for the data collected. It currently
offers an extensive library of AI statistics, open
source scalability, and the capability to compare the
average accuracy of various algorithms. The
independent variable is the accuracy of the novel
XceptionNet model, while the dependent variable is
adequacy. To ascertain the fact that empirical proof
exists that demonstrates the ability of the
corresponding populations to be substantially
different, the Independent Samples T-test was utilized
to gauge the pair of independent groups, XceptionNet
and Convolutional Neural Networks.
Table 1: This table displays the Pseudocode and steps
performed.
S.No
Model
Pseudocode
1.
XceptionNet
Input: Training and Testing data
Output: PerformanceAccuracy
1. Import dataset and required
packages
2. Pre-process the data
3. Divide the dataset into train and test
data
4. Define the model
5. Fit the model on the training and
testing dataset
6. Train and test the model
7. Evaluate the model
8. Plot the graph using matplotlib
9. Derive the confusion matrix for the
model’s metrics
10. Report performance
2.
CNN
Input: Data from Training and Testing
Output: Value of accuracy
1. Importing dataset and required
packages
2. Pre-process the data
3. Divide the dataset into Training and
Testing sets
4. Fit the CNN model on the training
and testing dataset
5. Evaluate the CNN model
6. Plot the graph using matplotlib
7. Derive Confusion matrix
8. Calculate the accuracy score
Table 2: Sample values for the groups i.e, CNN and novel
XceptionNet model.
S.NO
ACCURACY
CNN
XceptionNet
1
59.38
78.45
2
63.00
84.40
3
64.44
87.52
4
64.13
88.11
5
67.56
88.31
6
65.33
88.45
7
67.62
88.49
8
68.94
88.60
9
69.25
88.38
10
72.50
88.74
Average
66.21
86.94
Table 3: Performance metrics of XceptionNet and
Convolutional neural network.
S.NO
METRICS
XCEPTION
CNN
1
Accuracy
88.74%
72.50%
2
Precision
88.45%
88.82%
4
F1-score
89.68%
89.11%
5
Sensitivity
90.94%
89.4%
6
Specificity
87.86%
89.73%
Table 4: Statistical analysis of diverse samples examined
between novel XceptionNet and CNN deep learning
models. The mean accuracy of XceptionNet is 88.74% and
CNN is 72.50% Standard Deviation(SD) of XceptionNet is
3.24 and CNN is 3.73. The T-test for comparison for
XceptionNet standard error mean (std. mean) is 1.02 and
CNN is 1.17.
Group
N
Mean
Std.
Deviation
Std. Mean Error
Accuracy
XceptionNet
10
86.94
3.24
1.02
CNN
10
66.21
3.73
1.17
Loss
XceptionNet
10
30.15
6.57
2.07
CNN
10
60.26
3.40
1.07
AI4IoT 2023 - First International Conference on Artificial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry
and Consumer Electronics
362

Table 5: The data set is subjected to an Isolated Sample T-Test, with the confidence interval set at 95% and the corresponding
significance level set to p<0.001 (p<0.05) (2-tailed), groups are statistically significant.
Leven’s Test of
Equality of
Variances
T-test for Equality of Means
95% Confidence
Interval of the
Difference
F
Sig
t
df
Sig (2-tailed)
Mean Difference
Std Error difference
Lower
Upper
Accuracy
Equal Variance assumed
0.606
0.004
13.25
18
<.001
20.73
1.56
17.44
24.01
Equal Variance not assumed
13.25
17.66
<.001
20.73
1.56
17.43
24.02
Loss
Equal Variance assumed
1.150
0.298
-12.86
18
<.001
-30.10
2.34
-35.02
-25.19
Equal Variance not assumed
-12.86
13.51
<.001
-30.10
2.34
-35.14
-25.07
Figure 1: Conceptual representation of XceptionNet Deep learning model’s architecture.
Figure 2: Novel XceptionNet and CNN - training accuracy as well as validation accuracy.
Detection of Bone Fractures in Upper Extremities Using XceptionNet and Comparing the Accuracy with Convolutional Neural
363

Figure 3: Novel XceptionNet and CNN - training loss and validation loss.
Figure 4: Confusion matrix (CM) for XceptionNet deep
learning model.
Figure 5: Confusion matrix (CM) of CNN deep learning
model.
Figure 6: Bar graph presents the contrast between Mean
accuracy of novel XceptionNet and Convolutional neural
network (CNN) deep learning model in bone fracture
detection using x-rays dataset. XceptionNet does produce
better results with standard deviation (SD). X Axis:
XceptionNet vs Convolutional neural network and Y axis:
Mean Accuracy of detection SD = ±2 and confidence
interval of 95%.
3 RESULTS
The experimental results are based on the recital of
Novel XceptionNet and CNN DL models, which are
both measured in terms of accuracy. The XceptionNet
has a 89.44 percent accuracy, whereas the CNN has a
54.44 percent accuracy. Table 1represents the steps of
the pseudocode performed for XceptionNet and CNN
models. The innovative XceptionNet and
Convolutional neural networks, two deep learning
models, are contrasted for accuracy in Table 2. In the
third or Table 3 the metrics of XceptionNet and
Convolutional neural network models are shown.
Precision, F1 score, specificity and sensitivity values
AI4IoT 2023 - First International Conference on Artificial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry
and Consumer Electronics
364

obtained for XceptionNet are 88.45%, 99.68%,
87.86% and 90.94%. Table 4 examines the summary
statistics from separate samples utilizing the cutting-
edge CNN and XceptionNet deep learning models.
The novel XceptionNet model has a mean accuracy
of 88.74% compared to the CNN model's 72.50%.
Relating the XceptionNet to convolutional neural
networks, the XceptionNet has a standard deviation
of 3.24. The CNN standard error mean is 1.17,
whereas the XceptionNet standard error mean is 1.02,
according to the T-test. Table 5 demonstrates a
statistical independent sample t-test comparing the
novel XceptionNet and CNN (Convolutional Neural
Network), of a 95% confidence interval(CI). p<0.001
(p<0.05) (2-tailed) has been calculated as the
significant value for accuracy and groups are
statistically significant.
Fig. 1 displays the conceptual representation of
XceptionNet deep learning model’s architecture. The
training and validation accuracy of the innovative
XceptionNet and CNN deep learning models are
shown in Fig. 2, where x is the axis that represents the
epochs and y-axis displays accuracy performance.
Fig. 3 depicts training and validation loss of the
XceptionNet and CNN (Convolutional Neural
Networks) deep learning models. Fig. 4 and Fig. 5
shows the pictorial representation of confusion
matrices for XceptionNet and Convolutional neural
networks deep learning models. A vertical bar chart
depicting the average accuracy of the novel
XceptionNet and CNN models for identifying bone
fractures in upper extremity areas of hands using x-
ray images dataset can be observed in Fig. 6. Novel
XceptionNet appears to get better results with
standard deviation. Mean of Detection SD = ±2 and
X Axis: XceptionNet Vs. Convolutional Neural
Network (CNN).
4 DISCUSSIONS
The novel XceptionNet and Convolutional Neural
Networks (CNN) approaches are used and contrasted
to be able to improve the accuracy rate of bone
fracture detection in the upper extremity regions of
hands. In the proposed research, observations
revealed that the unique XceptionNet deep learning
model worked substantially better than the
Convolutional Neural Networks Technique at
detecting bone fractures in x-ray datasets. The novel
XceptionNet model has an accuracy rate of 88.74%
for recognizing fractures in the upper extremity
regions compared to Convolutional Neural Networks
Technique's accuracy rate of 72.50%. The outcomes
demonstrate that Novel XceptionNet performs more
accurately than Convolutional Neural Networks in
terms of accuracy.
According to a research, categorizing broken and
healthy bone with a block attention module,
convolutional neural network (CNN), faster R-CNN,
and FPN provided recall, precision, accuracy,
sensitivity, and specificity values of 0.789, 0.894,
0.853, 0.789, and an AUC of 0.920 (T.-H. Yang et al.
2022). In another study, “You Look Only Once”
(YOLO) models served as algorithms for skeletal
segmentation and lesion diagnosis and an accuracy of
90% was reported for bone fracture diagnosis
(Faghani et al. 2023). A convolutional neural network
(CNN) was used to locate the fracture ROI, and a
second CNN was used to identify and segment
different types of fragments within the ROI. This
deep learning model was implemented into this
cascaded architecture. For segmenting CT images,
this model produced average dice coefficients of
90.5% and average mean accuracy value of 89.40%
for identifying individual fragments (L. Yang et al.
2022). On the MobileNet network, a two-step
categorization technique was found to have a 73.42%
accuracy rate for bone fracture prediction. (El-
Saadawy et al. 2020). The accuracy of the AC-BiFPN
detection methods and modified Ada-ResNeSt
backbone network was reported to be 68.4% in
another study (Lu, Wang, and Wang 2022).
Sometimes, the mistake that occurred in the train
set of data and the error that occurred in the testing
data set are very different. It happens in complicated
models when there are too many parameters in
comparison to the quantity of observations. Future
research can assess a model's effectiveness based on
how well it performs on a test data set rather than how
well it performs on the training data that was provided
to it.
5 CONCLUSION
In this paper, computer diagnostic methods for X-ray
image-based bone detection in upper extremities were
demonstrated. The dataset was initially trained using
the XceptionNet deep learning model. The fracture's
position is determined after the initial stage. Patterns
tested on a set of images and the results were
evaluated according to the characteristics. In
comparison to convolutional neural networks (CNN),
which had an accuracy rate of 72.50%, the analysis
revealed that the results were satisfactory and that the
XceptionNet technique had a high accuracy of
88.74%.
Detection of Bone Fractures in Upper Extremities Using XceptionNet and Comparing the Accuracy with Convolutional Neural
365

REFERENCES
Abbas, Waseem, Syed M. Adnan, M. Arshad Javid, Fahad
Majeed, Talha Ahsan, Haroon Zeb, and Syed Saqlian
Hassan. (2020). “Lower Leg Bone Fracture Detection
and Classification Using Faster RCNN for X-Rays
Images.” 2020 IEEE 23rd International Multitopic
Conference (INMIC).
https://doi.org/10.1109/inmic50486.2020.9318052.
Agarwal, Shikha, Manish Gupta, Jitendra Agrawal, and
Dac-Nhuong Le. (2022). Swarm Intelligence and
Machine Learning: Applications in Healthcare. CRC
Press.
Bagaria, Rinisha, Sulochana Wadhwani, and A. K.
Wadhwani. (2022). “Bone Fracture Detection in X-Ray
Images Using Convolutional Neural Network.” SCRS
Conference Proceedings on Intelligent Systems, April,
459–66.
El-Saadawy, Hadeer, Manal Tantawi, Howida A. Shedeed,
and Mohamed F. Tolba. (2020). “A Two-Stage Method
for Bone X-Rays Abnormality Detection Using
MobileNet Network.” Advances in Intelligent Systems
and Computing. https://doi.org/10.1007/978-3-030-
44289-7_35.
Faghani, Shahriar, Francis I. Baffour, Michael D. Ringler,
Matthew Hamilton-Cave, Pouria Rouzrokh, Mana
Moassefi, Bardia Khosravi, and Bradley J. Erickson.
(2023). “A Deep Learning Algorithm for Detecting
Lytic Bone Lesions of Multiple Myeloma on CT.”
Skeletal Radiology. https://doi.org/10.1007/s00256-
022-04160-z.
G. Ramkumar, R. Thandaiah Prabu, Ngangbam Phalguni
Singh, U. Maheswaran, Experimental analysis of brain
tumor detection system using Machine learning
approach, Materials Today: Proceedings, (2021), ISSN
2214-7853, https://doi.org/10.1016/j.matpr.2021.01.246.
Kumar M, M., Sivakumar, V. L., Devi V, S.,
Nagabhooshanam, N., & Thanappan, S. (2022).
Investigation on Durability Behavior of Fiber
Reinforced Concrete with Steel Slag/Bacteria beneath
Diverse Exposure Conditions. Advances in Materials
Science and Engineering, 2022.
Lee, Gobert, and Hiroshi Fujita. (2020). Deep Learning in
Medical Image Analysis: Challenges and Applications.
Springer Nature.
Li, Xuechen, Linlin Shen, Xinpeng Xie, Shiyun Huang,
Zhien Xie, Xian Hong, and Juan Yu. (2020). “Multi-
Resolution Convolutional Networks for Chest X-Ray
Radiograph Based Lung Nodule Detection.” Artificial
Intelligence in Medicine 103 (March): 101744.
Lu, Shuzhen, Shengsheng Wang, and Guangyao Wang.
(2022). “Automated Universal Fractures Detection in
X-Ray Images Based on Deep Learning Approach.”
Multimedia Tools and Applications 81 (30): 44487–
503.
Pranata, Yoga Dwi, Kuan-Chung Wang, Jia-Ching Wang,
Irwansyah Idram, Jiing-Yih Lai, Jia-Wei Liu, and I-Hui
Hsieh. (2019). “Deep Learning and SURF for
Automated Classification and Detection of Calcaneus
Fractures in CT Images.” Computer Methods and
Programs in Biomedicine 171 (April): 27–37.
Raisuddin, Abu Mohammed, Elias Vaattovaara, Mika
Nevalainen, Marko Nikki, Elina Järvenpää, Kaisa
Makkonen, Pekka Pinola, et al. (2021). “Critical
Evaluation of Deep Neural Networks for Wrist Fracture
Detection.” Scientific Reports 11 (1): 6006.
S. G and R. G, (2022) "Automated Breast Cancer
Classification based on Modified Deep learning
Convolutional Neural Network following Dual
Segmentation," 3rd International Conference on
Electronics and Sustainable Communication Systems
(ICESC), Coimbatore, India, 2022, pp. 1562-1569, doi:
10.1109/ICESC54411.2022.9885299.
Sa, Abdelaziz Ismael, A. Mohammed, and H. Hefny.
(2020). “An Enhanced Deep Learning Approach for
Brain Cancer MRI Images Classification Using
Residual Networks.” Artificial Intelligence in Medicine
102 (January).
https://doi.org/10.1016/j.artmed.2019.101779.
Sairam, Vuppala Adithya. (2022). “Bone Fracture
Detection Using X-Rays.”
https://www.kaggle.com/vuppalaadithyasairam/bone-
fracture-detection-using-xrays.
“SPSS Software.” n.d. Accessed March 31, (2021).
https://www.ibm.com/analytics/spss-statistics-
software.
Yang, Lv, Shan Gao, Pengfei Li, Jiancheng Shi, and Fang
Zhou. (2022). “Recognition and Segmentation of
Individual Bone Fragments with a Deep Learning
Approach in CT Scans of Complex Intertrochanteric
Fractures: A Retrospective Study.” Journal of Digital
Imaging 35 (6): 1681–89.
Yang, Sijie, Fei Zhu, Xinghong Ling, Quan Liu, and Peiyao
Zhao. (2021). “Intelligent Health Care: Applications of
Deep Learning in Computational Medicine.” Frontiers
in Genetics.
https://doi.org/10.3389/fgene.2021.607471.
Yang, Tai-Hua, Ming-Huwi Horng, Rong-Shiang Li, and
Yung-Nien Sun. (2022). “Scaphoid Fracture Detection
by Using Convolutional Neural Network.” Diagnostics
(Basel, Switzerland) 12 (4).
https://doi.org/10.3390/diagnostics12040895.
Yoon, Gil Ho, Yeon-Jun Woo, Seong-Gyu Sim, Dong-
Yoon Kim, and Se Jin Hwang. (2021). “Investigation of
Bone Fracture Diagnosis System Using Transverse
Vibration Response.” Proceedings of the Institution of
Mechanical Engineers. Part H, Journal of Engineering
in Medicine 235 (5): 597–611.
AI4IoT 2023 - First International Conference on Artificial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry
and Consumer Electronics
366