2024 International Conference on Emerging Systems and Intelligent Computing (ESIC)

DETECTION OF BONE FRACTURE USING

DEEP LEARNING
Pujitha. B *Raga Sravya. K Krishnasai. N
Student, Department of
Assistant Professor, Department of Student,Department of
ECE
ECE ECE
V R Siddhartha Engineering College
V R Siddhartha Engineering College V R Siddhartha Engineering College
Vijayawada, India
2024 International Conference on Emerging Systems and Intelligent Computing (ESIC) | 979-8-3503-4985-6/24/$31.00 ©2024 IEEE | DOI: 10.1109/ESIC60604.2024.10481535

Vijayawada, India Vijayawada, India

[email protected]
[email protected] [email protected]

Aparna. CH
Student, Department of ECE
V R Siddhartha Engineering College
Vijayawada, India
[email protected]

Abstract— This research delves into the realm of bone embarks on an exploration of the transformative role that
fracture detection in medical X-ray images by harnessing Deep Learning, particularly through Convolutional Neural
the power of Deep Learning, specifically employing the Networks (CNNs) such as DenseNet and VGG19, plays in
DenseNet and VGG19 Convolutional Neural Network (CNN) addressing the burgeoning challenge of bone fracture
architectures. Through extensive training and meticulous detection. By delving into the confluence of technological
optimization, we navigate the intricacies of these models in
the context of radiology, aiming to augment the precision
advancements, medical imaging, and its evolving
and effectiveness of fracture identification. Our landscape of healthcare, This analysis aims to illuminate
investigation unfolds against the backdrop of a how these innovative approaches are poised to enhance
comprehensive dataset comprising a diverse array of X-ray fracture diagnosis and aid to the enhancing of patient care.
images, encompassing both normal and fractured bone
instances. Rigorous experimentation and evaluation,
A. Importance of fracture identification in bones through
utilizing a spectrum of performance measures encompassing
Deep Learning:
accuracy, precision and recall reveal compelling results.
Our CNN models exhibit a superior diagnostic prowess, a) Improved Accuracy: Deep Learning (DL) models,
surpassing conventional methodologies and demonstrating specifically, Convolutional Neural Networks (CNNs) like
remarkable sensitivity and specificity in fracture detection.
Beyond these empirical achievements, we explore the
DenseNet and VGG19, offer exceptional accuracy in
clinical implications of our findings, envisioning these identifying bone fractures from medical images. This
models as indispensable tools that could empower precision is essential for reducing diagnostic errors,
radiologists to expedite accurate diagnoses, ultimately minimizing false positives, and ensuring patients receive
elevating the standard of patient care and alleviating the timely and accurate diagnoses.
burdens on healthcare practitioners. In sum, this study sets b) Speedy Diagnosis: DL models excel in the rapid
the stage for transformative advancements within the field analysis of X-ray images, significantly accelerating the
of analysis of medical images and computer-aided diagnosis,
ushering in a new era of fracture detection within the realm
diagnosis process. This speed is particularly crucial in
of radiological sciences. emergency situations, where timely medical interventions
can be life-saving.
Keywords–Deep learning,Densenet,vgg19,Medical Imaging, c) Enhanced Safety: The early identification of bone
X-ray images fractures through DL systems contributes to patient safety
by promptly detecting issues that might otherwise be
I. INTRODUCTION overlooked unnoticed. This proactive approach aids in
preventing complications and ensuring patients receive
In recent years, the global healthcare landscape has been appropriate care.
confronted with a mounting challenge: the escalating d) Cost-Effective Healthcare: DL-based fracture detection
incidence of bone fractures. This rise in fracture cases, can lead to cost savings by reducing the requirement for
attributable to diverse factors like an aging population[1], extensive and costly diagnostic procedures. Early
increased sports-related injuries, and changing lifestyles, diagnosis enables timely and cost-effective interventions,
urgent requirement for a more efficient and accurate contributing to more efficient healthcare delivery.
diagnostic tools in the realm of radiology. The traditional
e) Efficient Resource Allocation: By automating image
methods of fracture detection, reliant on the expertise of
analysis, DL systems optimize the allocation of healthcare
radiologists and manual interpretation of X-ray images,
resources. They enable healthcare professionals to focus
are facing limitations in handling the burgeoning caseload.
their expertise on complex cases, ultimately improving the
In response to this burgeoning healthcare demand, the
efficiency within the medicalcare infrastructure.
integration of cuttingedge technologies has risen as a
beacon of hope. Among these technologies, Deep f) Broadened Access to Quality Healthcare: DL-based
Learning, a subset of artificial intelligence, has attracted fracture detection has the potential to extend access to
considerable interest for its potential to transform the quality healthcare services, especially in unpreviileged
domain of bone fracture diagnosis. This review paper regions with restricted access to specialized medical

979-8-3503-4985-6/24/$31.00 ©2024 IEEE

703
Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on August 22,2024 at 07:24:31 UTC from IEEE Xplore. Restrictions apply.
 2024 International Conference on Emerging Systems and Intelligent Computing (ESIC)

expertise. This democratization of healthcare empowers Component Analysis (PCA) to identify features that
communities and reduces healthcare disparities. contribute significantly (over 95%) in order to reduce the
dataset size. These reduced-dimensional depth features are
II. PROBLEM DESCRIPTION then fed into various classification models, each assigned
different weights based on their performance. The
The focal challenge addressed in this research experimental results demonstrate that the proposed
pertains to the accurate and timely detection of bone integrated learning algorithm approach improves fracture
fractures within the domain of radological imaging. diagnosis, achieving an mean accuracy of 81.93%. This
Traditional methods of bone fracture diagnosis finding offers valuable insights for computer-assisted
predominantly rely on X-ray imaging, a process that is orthopedists in clinical diagnosis. Furthermore, when
time-consuming and can be subject to variability in compared to standard deviation and other algorithms, the
interpretation, often contingent upon the expertise of experimental frameworks showcased both practicality and
radiologists. This problem is compounded by the resilience. Notably, the integrated learning method S-E-K
increasing prevalence of bone fractures in the modern introduced in this study surpasses alternative approaches,
population. The aim of our study is to harness the confirming its effectiveness.[3]
capabilities of deep learning, specifically utilizing
architectures such as DenseNet and VGG19, to provide a C. 3-D convolutional neural networks
robust and efficient solution. By doing so, we aspire to This paper introduces an automated approach
create a system that autonomously and consistently for identifying fractures in CT scans of the pelvic regions.
identifies bone fractures with a exceptional accuracy. This Initially, it presents a novel 3D annotation technique
innovation promises not only to expedite the diagnostic designed for efficient fracture annotation. Subsequently, a
process but also to dimnish the potential for misdiagnoses, novel method for fracture detection through automation is
thereby significantly improving patient outcomes and proposed. This method assesses specific surfaces to identify
healthcare efficiency. fractures, extracting 3D feature vectors from the surface of
the bone and employing a CNN for fracture classification. In
III. LITERATURE SURVEY experiments involving 103 participants, the accuracy during
training is 95.0%, while the accuracy during evaluation is
69.4%.[4]
The implementation of deep learning techniques
to bone fracture detection has garnered significant
D. Machine Learning Approach
attention in recent years. A plethora of studies have
This research involved the development of a deep
explored the potential of convolutional neural networks
learning framework to identify and classify of bone
(CNNs) in automating and enhancing the diagnostic
fractures. The experimentation utilized X-ray images
process. Notably, research has investigated the utilization
depicting both fractured and non-fractured human bones,
of diverse CNN architectures, such as DenseNet, VGG19,
sourced from the dataset of MURA. To mitigate the risk of
in the context of bone fracture identification. This
overfitting due to a small sample size, the dataset was
research have exhibited promising results in terms of
augmented. The developed model attained an Accuracy in
accuracy and efficiency, showcasing the transformative
classification 95% for both healthy and fractured
capability of deep learning in medical imaging.
bones.[5]
A. Image Processing Method
E. Deep Learning Approach
This study introduces a computer-based This paper presents a deep learning-based system
technique for detecting bone fractures in X-ray and CT for detecting and classifying bone fractures using X-ray
images. The process commences with noise removal images. The dataset comprises images of both fractured
through preprocessing and employs the Sobel edge bones and healthy ones from diverse sources, which were
detector for edge detection. Subsequently, the method subsequently augmented to address overfitting issues. The
conducts image segmentation to calculate the area of the developed model achieved an impressive classification
fracture. The results are evaluated depending upon GLCM accuracy of 92.44%.[6]
features, and the analysis indicates that the method
achieved a satisfactory 85% accuracy rate. However, a IV. PROPOSED METHOD
limitation of this approach is its difficulty in accurately
identifying fracture areas in some CT and X-ray images.[2] In the proposed method, our approach to detection of
fractured bones using Deep Learning is underpinned by
B. Automated Fracture Diagnosis with Integrated two highly effective Convolutional Neural Network
Learning in Computer-Aided Systems (CNN) architectures: DenseNet and VGG19. Our
This paper addresses the challenge of assisting in methodology commences with the meticulous collection
the diagnosis of fractures using single-modality of a diverse dataset of X-ray images encompassing both
normal and fractured bone samples. This dataset forms
neuroimaging. It introduces an integrated learning
the basis for our model's training. We then adapt these
classification algorithm based on deep features.
state-of-the-art CNN architectures to meet the specific
experimental procedure encompasses preprocessing the demands of fracture detection, with particular focus on
dataset of MURA, extracting depth features from diverse grayscale image processing and fine-tuning for binary
upper limb bone images, and employing Principal

704
Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on August 22,2024 at 07:24:31 UTC from IEEE Xplore. Restrictions apply.
 2024 International Conference on Emerging Systems and Intelligent Computing (ESIC)

classification. Leveraging the power of transfer learning, a. Gaussian Blur:

we initialize our models with pre-trained weights from Bone fracture detection datasets often contain images with
large-scale image datasets and fine-tune them on our bone varying levels of noise and texture, which can affect the
fracture dataset. Throughout this process, we apply a
accuracy of fracture detection algorithms. Gaussian blur
binary cross-entropy loss function and optimize our
can be used to smooth the images and reduce the effects of
models using the Adam algorithm
noise and texture, making it easier for the algorithms to
detect fractures accurately. This can enhance overall
V. DESIGN FLOW OF THE PROPOSED SOLUTION effectiveness of the fracture detection system. The 2D-
gaussian function for calculation of weight of pixels in a
window is given by
Data Acquisition G(x,y)=1/2π(σ^2) * (e^ -(x^2 +y ^2 ) ⁄ 2(σ^2)).
b. Normalization:
The datasets of fracture detection in bones often contain
images that have significant variations in lighting and
Pre processing contrast, which can pose a challenge to detect fractures
accurately. Normalization can find application in scaling
the pixel values in the images to a specific range, reducing
the impact of lighting variations and improving contrast.
This can enhance the visibility of fractures in the images
Data augumentation
and make it easier for the algorithms to detect them.
x normalized = (x – x minimum) ⁄ (x maximum – x
minimum)
c. Histogram Equalization:
Splitting Histogram equalization can be employed to adjust the
intensity values in detection of bone breakage datasets,
making it easier to distinguish between normal and
abnormal bone structures. This can improve the accuracy
Training model of fracture detection algorithms by enhancing the contrast
between the fracture and the surrounding bone tissue.
Additionally, histogram equalization can help to balance
the distribution of pixel values in the dataset, reducing the
Validation influence of any biases or outliers in the data.

Testing

Fig.1. Flow diagram of the proposed method[7].

VI. MODULE DESCRIPTION

A. Data Acquisition
Our dataset of X-ray images was acquired from
Github[8], a well-established platform for data sharing and
access. This dataset comprises a variety of X-ray images,
encompassing both normal bone structures and fractured
bones. Leveraging Github as our data source ensured
access to a diverse and comprehensive collection of
images, facilitating the robustness of our research. Ethical
considerations and patient data privacy were strictly
adhered to throughout the acquisition process.

B. Data Preprocessing
The raw image data needs to be processed and
cleaned to prepare it for use in the deep learning model.
This includes tasks such as adjusting the size of images,
converting them to grayscale, and normalizing the pixel
Fig 2:Image of bone before preprocessing[10]
values.[9]

705
Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on August 22,2024 at 07:24:31 UTC from IEEE Xplore. Restrictions apply.
 2024 International Conference on Emerging Systems and Intelligent Computing (ESIC)

D. Data Splitting
The dataset is divided into three fundamental
subsets: training, validation, and testing. Each subset's
size is defined by user-specified percentages. Initially, we
create separate directories for these subsets within a parent
directory. Then, for each category or class in the dataset,
we shuffle the images and distribute them in accordance
with the defined percentages into the respective training,
validation, and testing directories.This methodical
approach guarantees that we have distinct datasets for
model development, tuning, and evaluation, providing the
foundation for our bone fracture detection

E. Model Training
A deep learning model, like a convolutional
neural network (CNN), undergoes training on the training
dataset through backpropagation and gradient descent.
The objective is to minimize the loss function and
enhance the correctness of the model. We have developed
the model using CNN and trained the model by using
VGG19 and densenet architectures.
DenseNet : The architecture of DenseNet
comprises multiple dense blocks, with each dense block
being a fusion of a convolutional layer, batch
normalization, and ReLU activation.[12]. Every dense
block receives the combined feature representations from
Fig 3:Image of bone after preprocessing all preceding blocks as input and produces a new set of
feature maps that are passed onto the next dense block.
C. Data Augumentation Subsequent to the final dense block, the architecture
It is a pivotal technique in computer vision and includes a global average pooling layer, followed by a
deep learning, especially in the context of image datasets. fully connected layer, and finally, a softmax activation
It involves the application of various transformations to function for the classification process.We used the
original images to generate additional training samples. In DenseNet121 architecture to develop a bone fracture
our project, we employ the `ImageDataGenerator` from detection model. We froze the layers of the pre-trained
Keras to perform data augmentation, enhancing the model and added a custom classification layer with a
robustness and effectiveness of our bone fracture sigmoid activation function. We compiled the model using
detection model. binary cross-entropy loss and Adam optimizer with a
The key augmentation techniques applied include: learning rate of 1e-4.
1.Rotation: Randomly rotating images by up to 10 degrees
to simulate variations in X-ray image orientation.
2.Width and Height Shift: Adjusting image width and
height by up to 10% of their original dimensions,
mimicking shifts in bone positioning.
3.Shear Transformation: Applying shear transformations
to introduce image deformations, aiding the model in
handling common image distortions.
4.Brightness Adjustment: Randomly modifying image
brightness within a defined range, accounting for
variations in image exposure.
5.Horizontal and Vertical Flipping: Mirroring images Fig 4:Architecture of Densenet[13].
horizontally and vertically, diversifying the dataset with
VGG19: VGG19 represents a CNN structure
mirror-image versions.
comprising 19 layers, encompassing 16 convolutional
6.Fill Mode: Utilizing the 'nearest' fill mode to address
layers and the layers which are fully connected are 3. The
gaps or regions that may emerge after transformations.
initial 13 layers focus on convolutional operations for
Augmentation of data not only escalates
feature extraction, succeeded by 5 max pooling layers,
dataset size but also exposes the model to diverse
which serve to decrease the spatial extents of the features.
conditions, enhancing its ability to generalize and make
Conclusively, the architecture concludes with 3 layers
precise predictions on unseen X-ray images. In
which are fully connected dedicated to classification.[14].
radiological imaging tasks like bone fracture detection,
VGG19 also uses ReLU activation function after each
where imaging conditions can vary, data augmentation
convolutional and fully connected layer. The training
serves a pivotal function in model training and
process implements deep learning utilizing the VGG19
performance improvement.[11]

706
Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on August 22,2024 at 07:24:31 UTC from IEEE Xplore. Restrictions apply.
 2024 International Conference on Emerging Systems and Intelligent Computing (ESIC)

architecture for the classification of images on a custom identification while upholding a strong balance between
dataset. The process first sets up data generators for the correctly identified true positives and incorrectly
training and validation datasets and then loads the pre- identified false positives. The findings demonstrate the
trained VGG19 model from the Keras applications potential clinical relevance of VGG19 as a powerful tool
module. A custom classification layer is added on top of for radiologists and healthcare practitioners, offering
the VGG19 model, and the pre-trained layers in VGG19 precise bone fracture detection capabilities.This
are frozen. The model is then compiled using the achievement highlights the importance of architecture
optimizer known as adam employing a specified learning selection in deep learning applications, as it can
rate of 0.0001. significantly impact model performance. While DenseNet
showed commendable results, the slightly superior
performance of VGG19 underscores the necessity for
considering specific architectural nuances within the
domain of medical image analysis.

Table-1: Performance of Densenet architecture

Metric Training Validation Testing
Accuracy 89.2% 91.58% 90.25%
Precision 89% 92.7% 91.14%
Recall 88.97% 91.18% 87.94%

Table-2: Performance of VGG19 architecture

Metric Training Validation Testing
Fig 5:Architecture of VGG19[15].
Accuracy 93.25% 94.1% 95.93%
Precision 92.64% 93.6% 94.44%
F. Model evaluation:
Recall 90.68% 91.11% 89.47%
In assessing the acheivement of our bone fracture
detection model, we employed a comprehensive suite of
metrics, including accuracy, precision, recall. VIII. CONCLUSIONS
1. Accuracy measures the overall correctness of the Our exploration of deep learning and the
model's predictions: selection of VGG19 as the optimal architecture for bone
Accuracy = (Number of Correct Predictions) ⁄ (Total fracture detection, achieving an impressive 95.93%
Number of Predictions) accuracy, have opened exciting possibilities. While we've
It signifies the proportion of correctly identified cases out made significant strides, the journey to more effective
of complete instances count in the dataset.[16] diagnostic tools continues.Future extensions in this field
2. Precision quantifies the model's ability to make may encompass the advancement of enhanced algorithms
correct positive predictions: capable of detecting fractures in varying imaging
Precision = (True Positives) ⁄ (True Positives + False conditions, such as low-quality scans. Additionally, fine-
Positives) tuning models for specific types of fractures, such as
Precision is the measure of the proportion of affirmative stress fractures or pediatric bone fractures, could yield
predictions made by the model that were actually specialized diagnostic solutions. The integration of multi-
correct.[17] modal data sources, including X-rays, CT scans, and
3. Recall measures the model's capability to recognize patient medical histories, could further enhance the
all relevant instances: precision of diagnosis. Furthermore, global collaboration
Recall = (True Positives) ⁄ (True Positives + False among researchers, healthcare institutions, and technology
Negatives) innovators may drive the evolution of standardized
fracture detection models for widespread adoption.Our
Recall indicates the percentage of actual positive cases
study serves as a stepping stone in the ongoing quest to
correctly recognized by the model.[17]
boost the accuracy and accessibility of bone fracture
detection. We anticipate that these potential future
extensions will contribute to more efficient, precise, and
VII. RESULTS AND DISCUSSION
accessible medical diagnostics, benefiting patients and
medical practitioners alike.
In our bone fracture detection project, we
employed two well-established convolutional neural REFERENCES
network architectures, DenseNet and VGG19, to appraise
[1] S. Myint, A. S. Khaing& H. M. Tun, Detecting Leg Bone Fracture
their performance. Notably, VGG19 emerged as the in X-ray Images, International Journal of Scientific & Research,
superior candidate, exhibiting exceptional results on the vol. 5,2016.
testing dataset with an accuracy of 95.93%. Precision and [2] Peruri, Srinivasulu & Vamsi, Jollu & Kattubadi, Drutesh &
recall values of 94.44% and 89.47%, respectively, Prudhvi, Gandham. Bone fracture detection using Image
Processing. 5. 329-334,2020.
underscored VGG19's ability to provide accurate fracture [3] F. Yang and B. Ding, "Computer Aided Fracture Diagnosis Based

707
Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on August 22,2024 at 07:24:31 UTC from IEEE Xplore. Restrictions apply.
 2024 International Conference on Emerging Systems and Intelligent Computing (ESIC)

on Integrated Learning," IEEE 3rd International Conference on

Information Systems and Computer Aided Education (ICISCAE),
Dalian, China, 2020
[4] N. Yamamoto et al., "An automated fracture detection from pelvic
CT images with 3-D convolutional neural networks," 2020
International Symposium on Community-centric Systems (CcS),
Tokyo, Japan, 2020.
[5] W. Abbas et al., "Lower Leg Bone Fracture Detection and
Classification Using Faster RCNN for X-Rays Images," 2020 IEEE
23rd International Multitopic Conference (INMIC), Bahawalpur,
Pakistan, 2020.
[6] D. P. Yadav and S. Rathor, "Bone Fracture Detection and
Classification using Deep Learning Approach," International
Conference on Power Electronics & IoT Applications in
Renewable Energy and its Control (PARC), Mathura, India, 2020.
[7] https://www.codecademy.com/article/deep-learning-workflow
[8] https://github.com/eliashossain001/Bone-Fracture-
Detection/tree/main/Notebooks
[9] Gonzalez, R. C., & Woods, R. E. Digital Image Processing (3rd
ed.). Pearson Education,2018.
[10] https://github.com/eliashossain001/Bone-Fracture-
Detection/blob/2f00682dfc0ea81c614863e602642420f9db881e/Dat
aset/valid/XR_WRIST/patient11185/study1_positive/image1.png.
[11] Krizhevsky, A.; Sutskever, I., Hinton, G.E.,”ImageNet
classification with deep convolutional neural networks”.In
Proceedings of the 25th International Conference on Neural
Information Processing Systems, Stateline,NV, USA, 3–6
December 2012.
[12] Gao Huang, Zhuang Liu, Laurens van der Maaten, Kilian Q.
Weinberger,"Densely Connected Convolutional Networks
(DenseNet)",2016.
[13] https://images.app.goo.gl/PkHJgs4xsRrordeC6
[14] Karen Simonyan, Andrew Zisserman,"Very Deep Convolutional
Networks for Large-Scale Image Recognition",2014.
[15] https://images.app.goo.gl/vpa8XwiG4k4qehuE8
[16] https://deepchecks.com/how-to-check-the-accuracy-of-your-
machine-learning-model/
[17] https://www.analyticsvidhya.com/blog/2020/09/precision-recall-
machine-learning/

708
Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on August 22,2024 at 07:24:31 UTC from IEEE Xplore. Restrictions apply.