Automated Bone Fracture Detection System

using YOLOv7 with Secure Email Data

Sharing
Varshita Nukala Swarada Gade
Department of Computer Engineering, Department of Computer Engineering,
JSPM’S Rajarshi Shahu College of Engineering, JSPM’S Rajarshi Shahu College of Engineering,
Pune, India Pune, India
[email protected] [email protected]

Tanaya Sutar Shravani Walunj

Department of Computer Engineering, Department of Computer Engineering,
JSPM’S Rajarshi Shahu College of Engineering, JSPM’S Rajarshi Shahu College of Engineering,
Pune, India Pune, India
[email protected] [email protected]

Dr. Avinash Golande

Assistant Professor,
School of Engineering and Technology Pimpri Chinchwad University
Pune, India
[email protected]

Abstract—The demand for more precise and timely I. I NTRODUCTION

bone fractures diagnosis grew resulting in integration of
advanced technologies to medical imaging. This paper de- Providing accurate diagnosis of bone fractures is an
scribes the development and realization of a deep learn- integral part of optimal patient care, with a significant
ing system for automatic detection of bone fractures in role of X-ray technology in the identification and man-
X-ray images, which uses YOLOv7 model for improved agement of injuries. Historically, the diagnostic phase
diagnostic accuracy and efficiency. A secure mechanism of X-ray analysis requires engagement and expertise
to transmit medical images and reports through email
is included in the system. Furthermore, this system from a radiologist who scans each film to assess for
supports direct downloading of DICOM images as well fractures. Human observation likely varies in reliability
as creating simple diagnostic reports automatically thus depending on factors such as fatigue and cognitive
taking out serial steps that increase delay time for load, which can impact the interpretation of diagnostic
patient feedback. This innovative approach leverages images. Given these challenges, the motivation behind
cutting-edge deep learning techniques to address critical
healthcare needs, streamline diagnostic workflows, and this systems design is to establish a bone fracture
enhance patient outcomes. Real-time fracture detection detection system for X-ray images using the model
in the YOLOv7 model is possible because of its use YOLOv7, which is a real-time object detection model.
of data augmentation methods during the process of The model is established in detection of fractures
training, which assures robustness and reliability in in diagnostic images, thus improving accuracy and
different scenarios. Through the systematic methodology
of data preparation, model training, and evaluation consequently efficiency in the clinical decision-making
defined in this framework, the potential of the system as a process.
reliable asset in clinical applications is demonstrated. The The system’s motive is to reduce the manual pro-
proposed model minimized the dependency on manual cessing with neuro-scientifically expert-grounded de-
procedures, maximized the speed of clinical decisions,
sign choices to ensure effective fracture detection. This
and reinstated decision processes through standardized
results. Hence, this framework serves as a noble contri- system allows for collaborative care by ensuring that
bution to modern medical diagnostics. patient information can be shared effectively, which is
necessary for timely intervention. The primary benefit
of the system is the ability at improving the diagnostic
Index Terms—Automated Bone Fracture Detection,
Deep Learning Model, X-ray Imaging, Automated Di- accuracy. By the application of deep learning, the
agnosis, Email Data Exchange, DICOM Images, Diag- system will then be able to process the minute detail
nostic Reporting, Secure Data Transmission, Computer in X-rays to detect even very small bone fractures
Vision, Medical Imaging, Healthcare Technology, Image that would otherwise be missed by the human eye.
Analysis, Radiology Automation.
Early detection followed by prompt intervention could,
 the most promising in terms of efficiency and reliabil-
ity. For this purpose, authors have extensively analyzed
existing literature and attempted to present the most
important methodologies used, performance metrics,
and validation processes across a variety of studies.
By synthesizing findings from several studies, one will
be able to note the trends, strengths, and limitations
of current approaches in order to illustrate insights to
best practices in the development and implementation
of deep learning algorithms for fracture diagnosis.
Fig. 1. Accuracy comparison S. Shelmerdine, H. Liu, O. Arthurs, R. White, and
N. Sebire [24] AI in fracture detection has been the
dominant line of research and commercialization in
therefore, lead to better outcomes for patient care. It medicine. Children are the primary victims of the
will be used for standardization across various health inadvertent over- or under-diagnosis of injuries. This
care facilities since one human interpretation may vary is as a result of missed out fractures. The study
from another, thus reducing this variability. focuses on the AI tools’ diagnostic performance for
This technology plays an important role in advanc- pediatric fracture detection on imaging. It also aims to
ing medical research. It can support large studies over compare its performance with human readers where
X- ray imaging that may advance the understanding possible. Following a systematic review between 2011
of fracture epidemiology, risk factors, and treatment and 2021, databases including MEDLINE, Embase,
methods. Such data may also contribute to the de- and Cochrane Library were analyzed. Nine from 362
velopment of new and even newer technologies for identified eligible articles have found fracture detection
treating bone fractures that will reduce the number being most frequent with the elbow mostly studied.
of occurrences and make patients recover even better. Most of the studies were done using data from single
The integration of emerging diagnostic technologies, institutions, deep learning algorithms were usually
such as explainable AI, will further enhance trans- used, and external validation was lacking. The figures
parency and trust in the decision-making process, so for AI were from 88.8% to 97.9% in terms of accuracy.
that healthcare providers are able to understand and In those researches where AI and the readers were
validate the system’s conclusions. Also, the system compared, AI has been discovered to be more sensitive
aligns with the green Internet of Things framework, (though the difference is not significant). The analysis
that enables secure data exchange between devices pointed out AI’s unclear algos and thereby limited
while maintaining the integrity of existing healthcare genabilities due to the lack of effective validation and
infrastructure. The accuracy of medical imaging is a heterogeneous dataset. Such future research would
also supported by ongoing technological advancements benefit not only from the diverse patient populations
in the field. Studies combining traditional diagnostic of multi center studies around the world but also be the
methods with modern technological integrations bring cornerstone of real-world assessments. Hence, it can be
advancements to evolving models of investigation and argued that such future research would give a strong
data management. This system’s approach to detecting ground and potential for the tools to help doctors in
bone fractures, combining deep learning algorithms pediatric ways.
with secure data management, offers an original pro-
S. Rathor and D. P. Yadav [3] Given the frequency
gression in enhancing diagnostic methods. By stimu-
of the fractures and drawbacks of hand identification
lating diagnostic accuracy, improving workflow, and
with pictures of X-Ray, the authors of this review
promoting positive patient outcomes, this project aims
addressed the significance of automated fracture de-
to integrate cutting-edge technology into clinical prac-
tection. To differentiate between healthy and damaged
tice. As healthcare professionals adopt this modern
bones, a DNN model is suggested. Overfitting on
approach, they can support the generation of informed
a tiny dataset is eliminated using data augmentation
decisions, leading to enhanced healthcare delivery and
techniques, yielding a dataset of 4000 photos. DNN
improved patient care.
outperforms earlier techniques with accuracy of 92.4%
in differentiating between broken and healthy bone.
II. R ELATED W ORKS
Additionally, certain subgroups yield accuracy rates of
M. Nandyala, P. Kanumuri, and M. Garlapati [22] 95% and 93%, demonstrating the model’s resilience.
This paper conducts a review and analysis focused S. Yang, W. Cao, B. Yin, C. Feng, S. He and G.
on the automatic diagnosis of fractures with deep Fan [25] The paper is a thorough review of diagnostic
learning application techniques. Here, an attempt has efficiency of deep learning technology for orthopedic
been made to set up which deep learning models that fractures. It introduces various researches which have
can potentially be applied for the automatic fracture used deep learning algorithms for medical imaging
detection and classification from medical images are purposes including X-rays and CT scans. The authors
examine how these algorithms performed on their
diagnostic tasks and to what accuracy these methods
were compared with the standard ones. The article
addresses the considerations that affect the diagnostic
performance including the models used, the quality of
the datasets and the imaging modalities.
J. Zou and M. R. Arshad [26] The present article
studies the customization and introduction of enhanced
YOLOv7 algorithm for predicting of whole body frac-
tures from medical images. The authors start with
the problem of good fracture detection since different
fractures can be of different sizes, at different locations
and need to have dependable and efficient tools. They
conduct a complete evaluation of the images using a
large database of medical images containing different
fractures of the bone.
Puttagunta, M., Ravi, S [27] This paper examines
the numerous DL architectures, from CNNs, to RNNs,
to hybrid models, and explains how they are effec-
tive and ineffective at dealing with different medical
imaging, whether that be X-rays, MRIs, CT scans,
or ultrasounds. The review is about how models can
be trained on huge datasets to recognize patterns that
might indicate certain diseases or abnormalities, like Fig. 2. Bone fracture detection output
tumors or fractures. In this article, the authors provide
cases where deep learning systems have been used in
a clinical setting, not to replace the radiologist or any [33] The challenges that are associated with using
other medical personnel, but to help them make the big data in healthcare are discussed in the study,
most accurate and timely diagnosis possible. including issues with data privacy, interoperability,
W. Kim, S. Kim, Y. Kim, G. Moon, Y. Jeong and the need for skilled professionals for assessing
and H. S. Choi: [30] In this study, the researchers and interpreting the data. They go over a variety of
propose a novel approach that leverages advanced big data management techniques and technologies,
object detection algorithms to automatically identify such as frameworks for processing large data sets,
and classify facial bone fractures from medical imag- data integration tools, and data storage options. These
ing data. It describes the performance metrics of the techniques and technologies can be applied to enhance
system, which compares the diagnostic accuracy with patient outcomes, manage healthcare more effectively,
conventional methods. The authors point out that their and ultimately transform the way that healthcare is
CA-FBFD system would take less time to diagnose delivered.
and would have more consistent fracture identification Aiello, M., Esposito, G., Pagliari, G [34] The au-
which would really help radiologists and other medical thors begin by providing context on the importance of
professionals in a clinical setting. DICOM as a foundational standard that governs the
L. T. Martin al. [32] It is no doubt that the increasing storage, transmission, and sharing of medical imaging
diversity of the data has the potential of pointing health data. The paper highlights DICOM’s structured data
decision-making in a new direction. However, the format, which includes not only image data but also
problems are quite profound. As such, the difficulties rich metadata. This metadata encompasses essential
include setting data sharing and interoperability stan- information such as patient demographics, imaging
dards, developing new methodological and workforce protocols, and clinical notes, facilitating better orga-
models, and, more importantly, ensuring strong data nization and retrieval of information. Such structured
stewardship and governance that will accompany the data is crucial for effective analysis in big data appli-
data to protect public health data integrity. The review cations.
including literature, environmental scan, and insights Sun, Y.; al. [35] The paper aimed at improving
from the National Commission to Transform Public detection accuracy and efficiency in various applica-
Health Data Systems inform the discussion of this tions by outlining the limitations of existing YOLO
article on the major obstacles in data sharing and reuse. models, particularly in terms of their performance in
It highlights the sector’s possibilities to close the gaps complex environments and under varying conditions.
by improved interoperability and better data ownership To counter these challenges, the authors included a
practices. number of major adjustments to the YOLOv7 frame-
S. Dash, S. K. Shakyawar , M. Sharma, and others work. Key to these includes the development of the
 Aspect Details
Parallel Backbone Architecture, known as PBA, to Task
Model Used
Bone Fracture Classification
YOLOv7 (You Only Look Once v7)
facilitate better feature extraction at several scales. Fracture Types Classified Normal vs. Fractured bones
Image Modality X-ray Images
The authors carried out extensive experiments against Detection Approach Object detection (real-time, single-stage)
Preprocessing - Resizing images to 640x640
the baseline YOLOv7 and other state-of-the-art ob- - Normalization of pixel values to [0, 1]
- Data augmentation (rotation, flipping, scaling, color variations)
ject detection models to validate the performance Training Data 500 X-ray images (fractured and normal)
Model Evaluation Metrics Precision, Recall, Accuracy and F1-Score
of the PBA-YOLOv7 model. This is accomplished Accuracy 92.44%
Precision 93%
by using benchmark datasets on the model, hence Recall 95%
F1-Score 93.44%
assessing performance on metrics such as precision, Performance High accuracy for detecting bone fractures
Real-time Application Yes
recall, and mean Average Precision (mAP). Based Training Duration 50 epochs with Adam optimizer
Loss Function Categorical Cross-Entropy
on performance, PBA-YOLOv7 performs much better
compared to its predecessors, especially in complex TABLE I
C OMPARISON WITH OTHER PAPERS
scenarios related to occlusion or change of object size.
A related study was conducted by Shahnaj Rah-
man, Parvin and Abdur [44], wherein the proposed Beyaz al. [16] They reported 97% convolutional
model was strengthened by integrating different imag- neural network (CNN) architecture. and the highest
ing modalities like CT scans, X-rays, and MRI scans value. A total of 1,341 femoral neck fracture images,
that make a more vast collection of images. They including 765 non-fracture images. The probability
applied advanced architectures in deep learning such as of the current version process is 0.5. Their accuracy
convolution neural networks to appropriately analyze reached 97.4%.
these different types of images. Their paper highlights
the importance of multimodal data in capturing the in-
tensity of fractures as the appearance can vary greatly In the research conducted by Mutasa al. [21] the
from one imaging method to another. main focus was on convolutional neural networks, to
Research conducted by Gupta A. [45] discusses automate the detection and classification of femoral
various optimization algorithms such as Stochastic neck fractures in medical images. This research tries
Gradient Descent and its variants. Also, the author to overcome the challenges of accurately diagnosing
studies optimizers like Adam and AdaGrad with their femoral neck fractures, which is important for deter-
individual advantages, applied in special cases. The mining the appropriate treatment and care for patients.
paper compares the performance of these optimizers Training of the model was done by using dataset of
using empirical data, providing valuable insights about annotated X-ray images which contained both normal
their effectiveness across different deep learning tasks. and fractured femoral necks. CNNs helps the system
learn spatial hierarchies of features directly from raw
A new loss function is proposed by Zheng Z al. [57]
image data. This approach helped to automate feature
named Distance-IoU Loss which aimed at improving
extraction that previously required significant exper-
the accuracy and efficiency of bounding boxes. The
tise and manual intervention. But these deep learning
authors highlight shortcomings of the existing loss
models are designed not just to detect fractures but also
functions like Intersection over Union and its vari-
to determine their type and severity, enabling doctors
ants, which often encounter convergence issues speed
to take better decisions for patient care. The results
and optimization stability. For the argument proposed,
from the study indicate that these models outperform
the authors conducted extensive experiments using
conventional diagnostic methods, sometimes even hu-
well- known benchmark datasets such as Pascal VOC
man radiologists, providing greater precision and faster
and COCO. The results showed that models using
results. Such capabilities enable consistent and reliable
Distance-IoU Loss were performing better than those
fracture detection, with reduced diagnostic errors and
using conventional IoU-based loss functions. The pa-
quicker decision-making process. The integration of
per provides a robust comparison of detection accuracy
these automated systems in clinical practice could
and training time, thus offering strong evidence to
greatly assist radiologists, particularly in low-resource
validate the effectiveness of the proposed method.
settings or high-volume environments, by facilitating
Li X al. [58] This paper introduces a novel loss
timely and effective patient care. Overall, the paper
function designed to increase the performance of
highlights the transformative potential of deep learning
dense object detection tasks, particularly in situations
in medical imaging, particularly in the diagnosis and
with class imbalance and overlapping bounding boxes.
classification of femoral neck fractures, leading to
The authors propose a ”generalized focal loss” that
improved health outcomes.
addresses these challenges by introducing two key
innovations: qualified bounding boxes and a distributed
approach to box regression. The concept of qualified The study highlights the great potential of deep
bounding boxes allows the model to prioritize certain learning methods to transform medical imaging and
boxes based on their confidence and relevance, helping enhance healthcare outcomes overall, especially when
to mitigate the impact of noise from less relevant it comes to the diagnosis and classification of femur
detections. neck fractures.
 Fig. 3. Report generation

III. P ROPOSED W ORK

We propose a thorough approach to detect object

using latest deep learning architectures, YOLOv7 in
particular. To improve detection accuracy and op-
erating efficiency, our strategy incorporates multiple
components: training techniques, model architecture,
assessment metrics, and data preparation.
The trained deep learning models assist in extract
observations/assessments from pictures. Much use of
CNN-based techniques is utilized in this research to
recognize and categorize images according to their
characteristics. This section will highlight preparation
procedures of the data for model architecture and
evaluation of the model performance. This section
incorporates all the steps of methodology including
training processes, evaluation metrics, model archi-
tecture, and data pre-processing to ensure maximized
predictability of the learnt model.
YOLOv7: This architecture is optimized for perfor-
mance and efficiency, making use of a new backbone
network and tuned anchor box settings. It achieves a
great mAP while keeping the inference time small,
so it allows real-time detection. The first task was
to generate an appropriately diverse dataset of repre-
sentative objects of interest. The collected data was Fig. 4. System Architecture
then preprocessed to have the same size; each im-
age was scaled to 640 x 640 pixels for YOLOv7
model input. Among the various data augmentation
techniques applied were random rotations, rotations,
flips, including color variations, so there was much
diversity in the very training dataset. This was done
to counter overfitting and make the model robust with
regard to changes. Our proposed system employs the
highly accurate and fast YOLOv7 architecture. The
architecture itself is grounded on innovative anchor
box strategies and innovative backbone architecture for
usage in real-time detection cases. It, while keeping
Fig. 5. Training vs Epoch and Accuracy vs Epoch
intact the positive traits of earlier YOLO versions, adds
a plethora of features to enhance its performance.
 IV. T RAINING DEEP LEARNING MODEL CNNs are useful for image-related tasks because
Data preparation constitutes one of the critical com- of their hierarchical extraction of features capabilities.
ponents in any deep learning project. The quality and The architecture consists of multiple layers that pro-
quantity of data directly influence model ability to cess the input image:
generalize from the data in training to unseen data. Convolutional Layers: These layers apply convo-
Dataset Collection: For this study, we collected a lutional operations to the input images using filters
dataset containing images labeled either fractured or (kernels). Each filter has potential to learn features
normal. The dataset is taken from public medical from the images, such as edges, textures, and shapes.
imaging databases groups, added also with synthet- By using multiple convolutional layers, the network
ically generated images using some of the image can learn abstract representations of the input data.
transformation techniques. The diversity of this dataset Activation Functions: After convolutional layers,
is essential for a strong model to perform well real- activation functions are applied to add non-linearity,
time in some other domains. which enables the network to learn complex patterns.
Data Augmentation: Data augmentation is a statisti- Pooling Layers: Pooling layers (e.G., max pooling)
cal methodology simply utilized to artificially enhance lessen the spatial dimensions of the feature maps
the training dataset size by implementing various generated by convolutional layers. This manner al-
transformations on the images. This consists of all lows decrease computational load and stops overfitting
rotations, scaling, flipping, and alteration of colors. while preserving crucial capabilities.
These transformations help one to augment the data, Fully Connected Layers: Very last layers of CNN
thereby solving the overfitting problem and making a are absolutely related layers, which take the flattened
robust model over the data. The augmentations were output from the preceding layers and carry out the
made on the fly throughout the training and at each kind challenge. The community outputs possibilities
epoch, the model was set to see different variations of for each class, indicating the chance of the photo-
the images. graph belonging to a specific category (healthful or
v fractured).
u n
u1 X YOLOv7 is an advanced object detection frame-
σ=t (xi − µ)2 (1)
n i=1 work that gives real-time detection skills. It is in
particular proper for programs requiring high accuracy
and speed, making it best for medical photograph
Data Splitting: To create the confusion matrix, the evaluation:
following procedure was resorted to: splitting the Single-Stage Detector: Unlike traditional -level de-
dataset into distinct levels: training, validation and test tectors, YOLOv7 strategies the whole image in a
sets, with popular data allocation proportions being unmarried forward skip, making it quicker and greater
80%: 10%: 10% for training, validating, and testing, efficient for actual-time packages.
respectively. This division should be properly able to Grid System: The YOLO model divides the input
train the model, smoothen the tuning of hyperparam- image into a grid and predicts the bounding packing
eters, and evaluate performance through the model. containers and sophistication probability for every grid
In this application we use combination of CNNs and mobile. This enables the model to stumble on multiple
YOLOv7 architecture to effectively classify and detect gadgets in the equal photo, together with identifying
bone fractures in X-ray images This technology stated different fractures.
together they allow us to use image classification and Feature Pyramid Networks (FPN): YOLOv7 incor-
object visibility And the models ability to detect subtle porates FPNs to extract features at multiple scales,
differences in bone fractures is high. which is crucial for detecting small and large fractures
within X-ray images.
x̄ − µ
t= (2) Anchor Boxes: The use of anchor boxes allows
√s
n YOLOv7 to handle different aspect ratios and sizes
ezi of fractures effectively, improving detection accuracy.
σ(zi ) = PK (3) Before feeding the X-ray images into the model,
j=1 ez j
several preprocessing steps are applied:
n K Image Resizing: The images of X-rays are resized
1 XX
L=− yi,k log(ŷi,k ) (4) to a standard input size to ensure consistency during
n i=1
k=1 training and inference.
(t+1) (t) ∂J Normalization: Pixel values are normalized to a
wij = wij − η (5)
∂wij range between 0 and 1, which helps in faster conver-
gence during the training process.
n
1X Data Augmentation: Techniques like rotation, flip-
J(θ) = L(yi , ŷi ) + λ∥θ∥2 (6)
n i=1 ping, and brightness adjustment are employed to arti-
ficially expand the training dataset, helping the model performance.
generalize better to unseen images. v
u n
There are several key steps in the education process: u1 X 2
Loss function: Categorical Go-entropy loss is used RMSE = t (Truei − Predi ) (7)
n i=1
for the type project, so that classification loss, local-
ization loss, and confidence loss are also collected for TP
α=
object recognition. TP + FP
Optimizer: An adaptive optimizer with a faster
TP
ADAM or SGD is used to prevent loss factors at an β=
education stage, so that the version matches better. TP + FN
Training stages: The model performs well in more α·β
than one stage, with frequent checks of the vali- F =2·
α+β
dation set to test overall performance and prevent
where:
exceedances.
α represents Precision
The training for interpretation is a multistep process β represents Recall
that involves collecting the sample, obtaining school F represents F1-score
data, and checking its performance in the validation
set. VI. S YSTEM C ONFIGURATION
Model compilation: The version is compiled us- The software applications used included:
ing the optimizer, loss characteristics, and evaluation • Ubuntu 22.04 LTS or Windows 10/11
metrics. The ADAM optimizer is usually chosen in • Python 3.9 or higher version
order to obtain the best knowledge of target potentials, • Django 5.0.6
while the binary cross-entropy is activated due to • TensorFlow 2.16.2
the loss characteristics, which are suitable for binary • Ultralytics YOLOv7
classification tasks Appropriateness of the model: The
school’s methodology supports sampling of training VII. T RAINING PARAMETER S ETTING
records at different stages. Each epoch goes through a In our examine, the schooling of the deep learn-
complete bypass through the education data set. The ing fashions become carried out the use of carefully
school-time version adjusts its weights entirely based decided on parameters to optimize performance and
on the estimates lost from predictions relative to true ensure green gaining knowledge of. The following
scores. The validation loss and accuracy are monitored parameters have been utilized at some point of the
to evaluate the version performance and guide the training manner:
hyperparameter tuning. Batch Size: 32 was chosen as the Batch size to
The integration of CNNs and YOLOv7 in this balance memory utilization and convergence velocity.
project provides a robust framework for the detection This size helps efficient use of GPU resources while
of fractured bones from X-ray images. By leveraging allowing the version to study from a diverse set of
the strengths of both architectures, we aim to improve examples in every training iteration.
diagnostic accuracy and facilitate timely medical in- Learning Rate: The preliminary getting to know
terventions. The use of advanced image processing price was set to zero.001, which is commonly used
techniques, coupled with effective training strategies, for training deep getting to know models. To decorate
further enhances the model’s performance, paving the convergence, we employed a getting to know rate
way for its application in real-world clinical settings. scheduler that reduces the mastering charge by a issue
of zero.1 after a sure wide variety of epochs with
V. P ERFORMANCE ANALYSIS out improvement in the validation loss. This method
allows in refining the getting to know method as the
Evaluating the model’s performance is crucial to version tactics superior weights.
determine its effectiveness in classifying unseen facts. Number of Epochs: The fashions have been skilled
The evaluation is performed the usage of the take a for a complete of fifty epochs, presenting enough time
look at dataset, which remains independent of the for the version to learn from the schooling statistics
schooling and validation datasets. Evaluation Metrics: with out overfitting. Early stopping changed into car-
Key metrics together with accuracy, precision, ried out to halt training when the validation loss ceased
don’t forget, F1-rating, and Area Under the Curve to improve for three consecutive epochs.
(AUC) are computed. Accuracy shows the share Optimizer: The Adam optimizer was utilized due to
of efficiently classified instances, even as precision its adaptive learning rate capabilities, and is shown to
and consider provide perception into the version’s increase performance in various deep learning tasks.
overall performance on effective cases. The F1-rating This optimizer adjusts learning rate from the 1st and
serves as a harmonic mean of precision and do not 2nd moments of the gradients, enhancing training
forget, providing a single metric for model overall efficiency.
 Data Augmentation: To enhance model robustness with limited resources, where the prompt access to
and generalization, some data augmentation techniques diagnostic information is crucial.
like random horizontal flipping, scaling and rotation
are applied during training. This approach increases IX. C ONCLUSION
the diverseness of the training dataset, allowing the It has been established hereby that a YOLOv7-
model to explore more features. based framework works efficaciously in line with the
Loss Function: For tasks related to regression, the requirements proposed and aimed at real-time detec-
model uses Mean Squared Error (MSE) loss, while tion of bone fractures at a high accuracy of 92.44%.
for classification tasks, categorical cross-entropy loss The system, through combining the features of fracture
is employed. These loss functions were chosen based detection with CNN-based feature extraction abilities,
on the specific objectives of each task to effectively handles problems related to fracture irrespective of
measure the model’s performance. sizes and also through different complexities. The
Regularization: The L2 regularization term was strong preprocessing and optimized training conver-
added to the model weights to prevent over fitting gence enhances the value of this model to be reliably
that helped to maintain simpler model by discouraging usable and therefore deployable quite ideally in a
large weight values, thus enhancing the models ability clinical perspective.
to generalize to new, unseen data. Real-time detection automated report generation,
and secure data sharing enhance any diagnostic work-
VIII. R ESULT flow, ensure the privacy of the data of the concerned
patients, and facilitate communication of the medical
The system involves new techniques of fracture personnel around. The architecture is scalable and
detection in bones through advanced deep learning very flexible concerning the adaptation of any general
techniques combined with state-of-the-art methods in character in bigger healthcare, transforming diagnostic
medical diagnostics and which is able to achieve up to practices. In some respect, somehow, deep learning
92.44% accuracy in classification using the YOLOv7 has contributed considerably in upgrading the current
architecture making it clear that the system meets high processes surrounding medical imaging toward better
precision and reliability, making it suitable for use in patient outcomes and innovative healthcare technology
health care applications. A pipeline that integrates this solutions.
particular network with resizing, normalization, and
R EFERENCES
advanced data augmentation is in place. It can adapt to
almost any size, orientation, and imaging condition of [1] Miss. Swapna N, Mrs. Renuka Malge, “Classification and
the fractures. By using anchor box methods followed Detection of Bone Fracture Using Machine Learning,“ In-
ternational Journal for Research in Applied Science & Engi-
by Feature Pyramid Networks, the system enhances neering Technology (IJRASET), 2022; IJRASET45523. DOI:
its ability to precisely identify fractures, ensuring https://doi.org/10.22214/ijraset.2022.45523
dependable and consistent detection across complex [2] Yee Liang Thian, Yiting Li, Pooja Jagmohan, David Sia,
Vincent Ern Yao Chan, Robby T. Tan, “Convolutional Neu-
datasets. These capabilities are crucial for minimizing ral Networks for Automated Fracture Detection and Lo-
diagnostic errors and enabling timely intervention for calization on Wrist Radiographs,” Radiology: Artificial In-
patients. This YOLOv7 architecture strengthens the telligence, vol. 1, no. 1, pp. e180001, Jan. 2019. DOI:
https://doi.org/10.1148/ryai.2019180001
system’s feature extraction, allowing it to detect even [3] D. P. Yadav and S. Rathor, ”Bone Fracture Detec-
minor cracks that might have been missed through tion and Classification using Deep Learning Approach,”
manual analysis. This ability to analyze patterns in 2020 International Conference on Power Electronics &
IoT Applications in Renewable Energy and its Con-
X-ray images is a useful asset for radiologists, as it trol (PARC), Mathura, India, 2020, pp. 282-285. DOI:
helps reduce mental and decision fatigue. The system’s https://doi.org/10.1109/PARC49193.2020.236611
real-time detection capabilities, supported by extensive [4] N. Yamamoto et al., ”An automated fracture detection
from pelvic CT images with 3-D convolutional neural
training data and refined through adaptive learning rate networks,” 2020 International Symposium on Community-
scheduling and regularization techniques, make it an centric Systems (CcS), Tokyo, Japan, 2020, pp. 1-6. DOI:
effective tool for time-sensitive and crucial situations, https://doi.org/10.1109/CcS49175.2020.9231453
[5] S. Karimunnisa, P. R. Savarapu, R. K. Madupu, C. Z. Basha
such as emergency cases where rapid diagnosis can be and P. Neelakanteswara, ”Detection of Bone Fractures Au-
a matter of life or death. In addition to its diagnostic tomatically with Enhanced Performance with Better Com-
function, the system includes a secure email feature bination of Filtering and Neural Networks,” 2020 Second
International Conference on Inventive Research in Computing
for transmitting clinical images and reports, facilitating Applications (ICIRCA), Coimbatore, India, 2020, pp. 189-193.
easier communication between healthcare profession- DOI: https://doi.org/10.1109/ICIRCA48905.2020.9183085
als. This feature ensures strong data protection and [6] A. Savaris, A. A. Gimenes Marquez Filho, R. Rodrigues
Pires de Mello, G. B. Colonetti, A. Von Wangenheim, and
enhances collaborative efforts within medical teams. D. Krechel, ”Integrating a PACS Network to a Statewide
The system speeds up patient management workflows Telemedicine System: A Case Study of the Santa Catarina
by minimizing delays in sharing important informa- State Integrated Telemedicine and Telehealth System,” 2017
IEEE 30th International Symposium on Computer-Based Med-
tion, resulting in quicker and more efficient treatment. ical Systems (CBMS), Thessaloniki, Greece, 2017, pp. 356-
These efficiencies are especially beneficial in settings 357. DOI: https://doi.org/10.1109/CBMS.2017.128
 [7] C. Z. Basha, M. R. K. Reddy, K. H. S. Nikhil, P. Validation: A Systematic Review and Meta-Analysis,” Int. J.
S. M. Venkatesh, and A. V. Asish, ”Enhanced Computer Intell. Syst. Appl. Eng., vol. 11, no. 2, pp. 192–200, 2023.
Aided Bone Fracture Detection Employing X-Ray Images DOI: https://doi.org/10.46328/ijisae.4781.
by Harris Corner Technique,” 2020 Fourth International [23] A. V. Tigade and S. M. Shaikh, ”A Hybrid Approach
Conference on Computing Methodologies and Communi- for Classification of Bone Fracture using DWT,
cation (ICCMC), Erode, India, 2020, pp. 991-995. DOI: GLCM, and KNN,” J. Comput. Theor. Nanosci., vol.
https://doi.org/10.1109/ICCMC48092.2020.ICCMC-000184 18, no. 4, pp. 1419–1425, 2021. [Online]. Available:
[8] Rao, Kschandra & Srinivas, Y & Chakravarty, S., ”A https://www.naturalspublishing.com/files/published/1r886bfj97787b.pdf.
systematic approach to diagnosis and categorization of [24] S. Shelmerdine, R. White, H. Liu, O. Arthurs, and N. Se-
bone fractures in X-Ray imagery,” International Journal bire, ”Artificial intelligence for radiological paediatric fracture
of Healthcare Management, vol. 1, pp. 1-12, 2022. DOI: assessment: a systematic review,” Insights Imaging, vol. 13,
https://doi.org/10.1080/20479700.2022.2097765 2022, Art. no. 10. DOI: https://doi.org/10.1186/s13244-022-
[9] Rini Widyaningrum, Enny Itje Sela, Reza Pulungan, Anin- 01234-3.
dita Septiarini, ”Automatic Segmentation of Periapical Ra- [25] S. Yang, B. Yin, W. Cao, C. Feng, G. Fan, and
diograph Using Color Histogram and Machine Learning S. He, ”Diagnostic accuracy of deep learning in
for Osteoporosis Detection,” International Journal of Den- orthopaedic fractures: a systematic review and meta-
tistry, vol. 2023, Article ID 6662911, pp. 9, 2023. DOI: analysis,” Clin. Radiol., vol. 75, 2020, Art. no. 10. DOI:
https://doi.org/10.1155/2023/6662911 https://doi.org/10.1016/j.crad.2020.05.021.
[10] Antani S, Long LR, Thoma GR, et al. ”Anatomical shape [26] J. Zou and M. R. Arshad, ”Detection of whole body
representation in spine x-ray images,” VIIP 2003. Proceedings bone fractures based on improved YOLOv7,” Biomed. Sig-
of the 3rd IASTED International Conference on Visualization, nal Process. Control, vol. 91, p. 105995, 2024. DOI:
Imaging and Image Processing, Benalmadena, Spain, Sep. https://doi.org/10.1016/j.bspc.2024.105995.
2003. [27] M. Puttagunta and S. Ravi, ”Medical image analysis based
[11] W. Abbas et al., ”Lower Leg Bone Fracture Detection on deep learning approach,” Multimed. Tools Appl., vol. 80,
and Classification Using Faster RCNN for X-Rays Im- pp. 24365–24398, 2021. DOI: https://doi.org/10.1007/s11042-
ages,” 2020 IEEE 23rd International Multitopic Confer- 021-10707-4.
ence (INMIC), Bahawalpur, Pakistan, 2020, pp. 1-6. DOI: [28] H. Zhang and Y. Qie, ”Applying Deep Learning to Medical
https://doi.org/10.1109/INMIC50486.2020.9318052. Imaging: A Review,” Appl. Sci., vol. 13, p. 10521, 2023. DOI:
[12] L. Yang, S. Gao, P. Li, J. Shi, and F. Zhou, ”Recogni- https://doi.org/10.3390/app131810521.
tion and Segmentation of Individual Bone Fragments with [29] L. Tanzi, E. Vezzetti, A. Aprato, A. Audisio, and A. Massè,
a Deep Learning Approach in CT Scans of Complex In- ”Computer-Aided Diagnosis System for Bone Fracture Detec-
tertrochanteric Fractures: A Retrospective Study,” J. Digit. tion and Classification: A Review on Deep Learning Tech-
Imaging., vol. 35, no. 6, pp. 1681–1689, Dec. 2022. DOI: niques,” 2020.
https://doi.org/10.1007/s10278-022-00669-w. [30] G. Moon, S. Kim, W. Kim, Y. Kim, Y. Jeong, and
[13] Y. Ma and Y. Luo, ”Bone fracture detection H.-S. Choi, ”Computer Aided Facial Bone Fracture Di-
through the two-stage system of Crack-Sensitive agnosis (CA-FBFD) System Based on Object Detection
Convolutional Neural Network,” Informatics in Medicine Model,” IEEE Access, vol. 10, pp. 79061–79070, 2022. DOI:
Unlocked, vol. 22, 2021, Article ID 100452. DOI: https://doi.org/10.1109/ACCESS.2022.3192389.
https://doi.org/10.1016/j.imu.2020.100452. [31] H. Tan, H. Xu, N. Yu, Y. Yu, H. Duan, Q. Fan, and Z. T.,
[14] M. Wang, J. Yao, G. Zhang, et al., ”ParallelNet: Multiple ”The value of deep learning-based computer aided diagnostic
backbone network for detection tasks on thigh bone fracture,” system in improving diagnostic performance of rib fractures
Multimedia Systems, vol. 27, pp. 1091–1100, 2021. DOI: in acute blunt trauma,” *BMC Med. Imaging*, vol. 23, no. 1,
https://doi.org/10.1007/s00530-021-00783-9. Art. no. 55, Apr. 2023. DOI: https://doi.org/10.1186/s12880-
[15] W. Abbas, S. Adnan, J. Dr., and W. Ahmad, ”Analysis 023-01012-7.
of Tibia-Fibula Bone Fracture Using Deep Learning Tech- [32] L. T. Martin, C. Nelson, D. Yeung, J. D. Acosta, N.
nique of X-Ray Images,” International Journal for Mul- Qureshi, T. Blagg, and A. Chandra, ”The Issues of In-
tiscale Computational Engineering, vol. 19, 2021. DOI: teroperability and Data Connectedness for Public Health,”
https://doi.org/10.1615/IntJMultCompEng.2021036137. *Big Data*, vol. 10, no. S1, pp. S19-S24, Sep. 2022. DOI:
[16] S. Beyaz, K. Açıcı, and E. Sümer, ”Femoral neck fracture https://doi.org/10.1089/big.2022.0207.
detection in X-ray images using deep learning and genetic [33] S. Dash, S. K. Shakyawar, M. Sharma, and others,
algorithm approaches,” Jt. Dis. Relat. Surg., vol. 31, no. 2, pp. ”Big data in healthcare: management, analysis and future
175-183, 2020. DOI: https://doi.org/10.5606/ehc.2020.72163. prospects,” *J. Big Data*, vol. 6, Art. no. 54, 2019. DOI:
[17] R. M. Jones, A. Sharma, R. Hotchkiss, et al., ”Assessment of a https://doi.org/10.1186/s40537-019-0217-0.
deep-learning system for fracture detection in musculoskeletal [34] M. Aiello, G. Esposito, G. Pagliari, and others, ”How does
radiographs,” npj Digit. Med., vol. 3, no. 144, 2020. DOI: DICOM support big data management? Investigating its use
https://doi.org/10.1038/s41746-020-00352-w. in medical imaging community,” *Insights Imaging*, vol. 12,
[18] M. Dupuis, L. Delbos, R. Veil, and C. Adamsbaum, ”Ex- Art. no. 164, 2021. DOI: https://doi.org/10.1186/s13244-021-
ternal validation of a commercially available deep learning 01081-8.
algorithm for fracture detection in children,” Diagn. Interv. [35] Y. Sun, Y. Li, S. Li, Z. Duan, H. Ning, and Y. Zhang, ”PBA-
Imaging., vol. 103, no. 3, pp. 151–159, Mar. 2022. DOI: YOLOv7: An Object Detection Method Based on an Improved
https://doi.org/10.1016/j.diii.2021.10.007. YOLOv7 Network,” *Appl. Sci.*, vol. 13, Art. no. 10436,
[19] F. Hardalaç, F. Uysal, O. Peker, et al., ”Fracture Detection 2023. DOI: https://doi.org/10.3390/app131810436.
in Wrist X-ray Images Using Deep Learning-Based Object [36] K. Y. Zha, L. Yan, and G. Zeng, ”A Comprehensive Review of
Detection Models,” Sensors, vol. 22, no. 3, p. 1285, 2022. AI Techniques for Bone Fracture Detection,” arXiv preprint,
DOI: https://doi.org/10.3390/s22031285. 2022. [Online]. Available: https://arxiv.org/abs/2207.02696.
[20] Y. Pranata, K.-C. Wang, J.-C. Wang, et al., ”Deep [37] C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao,
learning and SURF for automated classification and de- ”YOLOv7: Trainable Bag-of-Freebies Sets New State-
tection of calcaneus fractures in CT images,” Com- of-the-Art for Real-Time Object Detectors,” in *Proc.
put. Methods Programs Biomed., vol. 171, 2019. DOI: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR)*,
https://doi.org/10.1016/j.cmpb.2019.02.006. Vancouver, BC, Canada, 2023, pp. 7464-7475. DOI:
[21] S. Mutasa, S. Varada, A. Goel, T. T. Wong, and M. J. Rasiej, https://doi.org/10.1109/CVPR52729.2023.00721.
”Advanced Deep Learning Techniques Applied to Automated [38] S. Medaramatla, C. Samhitha, S. Pande, and S. Vinta, ”De-
Femoral Neck Fracture Detection and Classification,” J. Digit. tection of Hand Bone Fractures in X-ray Images using
Imaging., vol. 33, no. 5, pp. 1209–1217, Oct. 2020. DOI: Hybrid YOLO NAS,” *IEEE Access*, vol. 1, 2024. DOI:
https://doi.org/10.1007/s10278-020-00364-8. https://doi.org/10.1109/ACCESS.2024.3379760.
[22] M. Nandyala, P. Kanumuri, and M. Garlapati, ”Automatic [39] A. Kheaksong, P. Sanguansat, P. Samothai, T. Din-
Diagnosis of Fracture using Deep Learning and External dam, K. Srisomboon, and W. Lee, ”Analysis of Mod-
 ern Image Classification Platforms for Bone Fracture De- [57] Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, and D. Ren,
tection,” in *Proc. 6th Int. Conf. Inf. Technol. (In- ”Distance-IoU loss: Faster and better learning for bound-
CIT)*, Nonthaburi, Thailand, 2022, pp. 471-474. DOI: ing box regression,” CoRR, 2019. [Online]. Available:
https://doi.org/10.1109/InCIT56086.2022.10067836. http://arxiv.org/abs/1911.08287.
[40] K. Y. Yigzaw, S. Olabarriaga, A. Michalas, L. Marco-Ruiz, [58] X. Li, W. Wang, L. Wu, S. Chen, X. Hu, J. Li,
C. Hillen, Y. Verginadis, M. T. Oliveira, D. Krefting, T. J. Tang, and J. Yang, ”Generalized focal loss:
Penzel, J. Bowden, J. Bellika, and T. Chomutare, ”Health data Learning qualified and distributed bounding boxes
security and privacy: Challenges and solutions for the future,” for dense object detection,” 2020. [Online]. Available:
*Computer Systems Science and Engineering*, 2022. DOI: https://doi.org/10.48550/arXiv.2006.04388.
https://doi.org/10.1016/B978-0-12-823413-6.00014-8. [59] G. Yang, J. Wang, Z. Nie, H. Yang, and S. Yu, ”A lightweight
[41] K. P. Kumar, B. R. Prathap, M. M. Thiruthuvanathan, H. YOLOv8 tomato detection algorithm combining feature en-
Murthy, and V. J. Pillai, ”Secure approach to sharing digi- hancement and attention,” *Agronomy*, vol. 13, no. 7, Art. no.
tized medical data in a cloud environment,” *Data Science 1824, 2023. DOI: https://doi.org/10.3390/agronomy13071824.
and Management*, vol. 7, no. 2, pp. 108-118, 2024. DOI: [60] Y. Li, Q. Fan, H. Huang, Z. Han, and Q. Gu, ”A modified
https://doi.org/10.1016/j.dsm.2023.12.001. YOLOv8 detection network for UAV aerial image recogni-
[42] Z. Zukarnain, A. Muneer, N. Atirah, and A. Almo- tion,” *Drones*, vol. 7, no. 5, Art. no. 304, 2023. DOI:
hammedi, ”Medi-Block Record Secure Data Sharing in https://doi.org/10.3390/drones7050304.
Healthcare System: Issues, Solutions and Challenges,” *Com- [61] S. Saxena, ”Binary cross entropy/log loss for binary
puter Systems Science and Engineering*, 2022. DOI: classification,” Analytics Vidhya, 2021. [Online]. Available:
https://doi.org/10.32604/csse.2023.034448. https://www.analyticsvidhya.com/blog/2021/03/binary-cross-
[43] H. Guo, Y. Wang, H. Ding, and J. Chen, ”Bone Frac- entropy-log-loss-for-binary-classification/.
ture Detection Based on an Improved YOLOv7 Model,”
*IEEE Access*, vol. 11, pp. 7635-7645, 2023. DOI:
https://doi.org/10.1109/ACCESS.2023.10476607.
[44] S. Parvin and A. Rahman, ”A real-time human bone fracture
detection and classification from multi-modal images using
deep learning technique,” *Appl. Intell.*, vol. 54, pp. 1-17,
2024. DOI: https://doi.org/10.1007/s10489-024-05588-7.
[45] A. Gupta, ”A comprehensive guide on optimizers in
deep learning,” *Analytics Vidhya*, 2024. [Online].
Available: https://www.analyticsvidhya.com/blog/2021/10/a-
comprehensive-guide-on-deep-learning-optimizers/.
[46] D. Kingma and J. Ba, ”Adam: A method for
stochastic optimization,” in *Proc. Int. Conf. Learn.
Representations (ICLR)*, 2015. [Online]. Available:
https://doi.org/10.48550/arXiv.1412.6980.
[47] S. Ruder, ”An overview of gradient descent
optimization algorithms,” 2016. [Online]. Available:
https://doi.org/10.48550/arXiv.1609.04747.
[48] T. Wu and Y. Dong, ”YOLO-SE: Improved YOLOv8 for
remote sensing object detection and recognition,” *Appl.
Sci.*, vol. 13, no. 24, Art. no. 12977, 2023. DOI:
https://doi.org/10.3390/app132412977.
[49] P. Das, A. Chakraborty, R. Sankar, O. K. Singh, H. Ray,
and A. Ghosh, ”Deep learning-based object detection al-
gorithms on image and video,” in *Proc. 2023 3rd Int.
Conf. Intell. Technol. (CONIT)*, pp. 1–6, 2023. DOI:
https://doi.org/10.1109/CONIT59222.2023.10205601.
[50] H. Lou, X. Duan, J. Guo, H. Liu, J. Gu, L. Bi, and
H. Chen, ”DC-YOLOv8: Small size object detection algo-
rithm based on camera sensor,” 2023. [Online]. Available:
https://doi.org/10.20944/preprints202304.0124.v1.
[51] M. Hussain, ”YOLO-v1 to YOLO-v8, the rise of YOLO
and its complementary nature toward digital manufac-
turing and industrial defect detection,” *Machines and
Tooling*, vol. 11, no. 7, Art. no. 677, 2023. DOI:
https://doi.org/10.3390/machines11070677.
[52] G. Deo, J. Totlani, and C. Mahamuni, ”A survey on
bone fracture detection methods using image processing
and artificial intelligence (AI) approaches,” 2023. DOI:
https://doi.org/10.1063/5.0188460.
[53] T. Meena and S. Roy, ”Bone fracture detection using deep
supervised learning from radiological images: A paradigm
shift,” *Diagnostics*, vol. 12, Art. no. 2420, 2022. DOI:
https://doi.org/10.3390/diagnostics12102420.
[54] J. Redmon and A. Farhadi, ”YOLOv3: An incremental
improvement,” CoRR, 2018. [Online]. Available:
https://doi.org/10.48550/arXiv.1804.02767.
[55] A. Mehra, ”Understanding YOLOv8 architecture, applications,
and features,” Labellerr, 2023. [Online]. Available:
https://www.labellerr.com/blog/understanding-yolov8-
architecture-applications-features/.
[56] K. He, X. Zhang, S. Ren, and J. Sun, ”Spatial pyramid
pooling in deep convolutional networks for visual recognition,”
in *Proc. Comput. Vis. ECCV*, 2014, pp. 346–361. DOI:
https://doi.org/10.1007/978-3-319-10578-92 3.