FRACTURE IDENTIFICATION IN FACIAL BONE X-

RAYS: A DEEP LEARNING APPROACH WITH

TRANSFER LEARNING

MAHENDRA PERUMAN1, SENTHILKUMAR 2, TAMILSOORIYA3, and VINCENT AMALADOSS 4.

1
Sri Manakula Vinayagar Engineering College (SMVEC), Pondicherry University, Puducherry, India

ABSTRACT Facial bone fractures are significant because they can affect essential functions, even though
they are not the most common type of bone fracture. However, detecting them early and accurately is still
crucial. These fractures often occur due to trauma, accidents, or weak bones. Diagnosis typically involves
getting X-rays of the affected area, followed by consulting with a medical professional or radiologist for
confirmation. Despite X-rays being the primary tool, minor fractures may be overlooked, making timely
detection challenging. It can be hard to find skilled radiologists who can detect fractures promptly, leading
to delayed diagnosis and compromised recovery. To overcome these challenges, our approach combines
deep learning with transfer learning, using the YOLOv8 Algorithm for better accuracy and efficiency in
detecting facial bone fractures. We trained our model on a dataset of 1862 facial bone X-ray images,
categorized into three types of fractures. We then compared our model's performance with existing ones to
assess its effectiveness in detecting facial bone fractures.

INDEX TERMS Facial bone fractures, X-rays, Deep learning, Transfer learning, YOLOv8 Algorithm,
Dataset, Performance evaluation

I. INTRODUCTION
In this study, we focus solely on the identification of
fractures within the mandibular, maxillary, and nasal bones,
Facial bones are a crucial component of the human as these regions are commonly affected by trauma and
skeletal system, consisting of various types such as the fractures. To address the challenges associated with fracture
mandibular, maxilla, zygomatic, nasal, lacrimal, detection, we introduce an innovative deep-learning
palatine, inferior nasal conchae, and vomer. These approach utilizing transfer learning for facial bone fracture
bones not only provide structural support to the face but identification in X-ray images. By harnessing pre-trained
also play vital roles in functions such as chewing, models, this approach enhances the model's ability to
breathing, and facial expression. Fractures within the facial recognize intricate patterns indicative of fractures across a
skeleton, categorized into many types such as linear diverse range of X-ray images. Through this methodology,
fractures, comminuted fractures, depressed fractures, we aim to not only improve the accuracy of fracture
and Le Fort fractures, can significantly impact an identification but also expedite the diagnostic process,
individual's appearance, sensory functions, and overall leading to enhanced patient outcomes.
well-being.
Prompt and accurate detection of facial bone Through extensive training on a comprehensive
fractures is essential for initiating appropriate treatment dataset comprising various types of facial bone fractures -
and minimizing potential complications. However, specifically on mandibular, maxillary, and nasal bones,
traditional diagnostic methods relying on manual this research endeavors to provide a valuable diagnostic tool
interpretation of X-ray images can be subjective and prone for radiologists and healthcare professionals. Our goal is to
to variability in accuracy. Moreover, certain fracture types, advance the field of medical imaging analysis, facilitating
particularly hairline or complex fractures, may evade better treatment planning and improving recovery outcomes
detection through conventional means. for individuals affected by facial bone fracture.

1
 According to Statistics, the most common site of Yeonjin Jeong et al [5] proposed CA-FBFD system
facial bone fracture was the nasal bone, accounting for addresses challenges in facial bone fracture diagnosis by
half of all cases, followed by the malar bone, blowout leveraging Yolo X-S object detection with CT image
fractures, and the mandibular bone. The occurrence rate augmentation. While achieving a Accuracy percentage of
was lowest in the maxillary bone. The ratio of male to 66.67% and higher sensitivity than the baseline, limitations
female patients was approximately 7:3, with the highest exist regarding its evaluation is only on nasal fractures and
frequency of facial bone fractures observed in the 10-29 potential accessibility issues with CT imaging. Further
age group. The contributions of this paper are as follows: validation across diverse cases is essential for clinical
applicability.

• An easy way to detect facial bone fractures. Specially Kandel et la [ 1 ] studied the usage of transfer
three types of facial bone fractures like mandibular, learning, a deep learning algorithm in computer vision to
maxillary, and nasal fractures. classify musculoskeletal images to detect bone fractures.
• The dataset used for this experiment is a Customized Various CNN methods were used such as VGG, Xception,
dataset of 1862 images. ResNet, GoogleNet, InceptionResNet, and DenseNet, and
• Developed and evaluated a deep learning model the metrics accuracy and Kappa were used to evaluate the
utilizing transfer learning to enhance the accuracy of performance of all networks. The model obtained a
fracture detection in facial bone X-rays. confidence score of 95%.

The remaining sections will be organized as follows; Atsuyuki Inui et al,[4] evaluates the diagnostic
Section II covers the related task which is the literature accuracy of YOLOv8, a deep-learning model, for
review for our proposed plan. Section III describes about identifying elbow osteochondritis dissecans (OCD) lesions in
dataset, preprocessing, and Existing Methodologies. ultrasound images. Results show high accuracy in classifying
Section IV describes the Proposed Model. Section V normal and OCD images, with precision and recall rates
describes the experiment, analysis of results, and exceeding 99%. The model demonstrates effectiveness in
comparison of results. Section VI outlines the conclusion both image classification and object detection, suggesting
and prospective scope of the study. potential for mass screening in medical check-ups for
baseball elbow.
II. RELATED WORK
Rakesh et al., [3] conducted a project to detect bone
Our approach combines deep learning with transfer fractures using the Canny edge detection Bernsen algorithm.
learning using the YOLOv8 Algorithm for better The model was performed on a Kaggle-obtained dataset with
accuracy and efficiency in detecting facial bone fractures. 15000 samples of wrist fracture RGB images. The Canny de-
tection and Bernsen algorithm achieved 76.85% and
D. P. Yadav et al [2] studied that the bone is a 75.29% respectively. Rakesh et al. also concluded that
major component of the human body. Bone provides the Canny edge detection is less sensitive to partial blockage.
ability to move the body. The bone fractures are common
in the human body. The doctors use the X-ray image to Young-Dae Jeon et al., [21] presents a deep learning
diagnose the fractured bone. The manual fracture detection model based on the "You Only Look Once" (YOLO) v4
technique is time consuming and also error probability algorithm for detecting and visualizing fracture areas in
chance is high. Therefore, an automated system needs to three-dimensional skeletal images. The model, evaluated
develop to diagnose the fractured bone. using precision-recall curves and intersection over union
metrics, demonstrates high average precision values
Robert Lindsey et al [20] Describes that deep (>0.60) and promising performance in detecting fractures in
neural network was developed to aid emergency medicine the tibia and elbow. The proposed system offers an intuitive
clinicians in detecting and localizing fractures in display of fractured areas through distinctive red masks
radiographs, trained on annotations from senior orthopedic overlaid on 3D reconstructed bone images, potentially aiding
surgeons. Clinicians demonstrated improved sensitivity orthopedic surgeons in quick and accurate trauma diagnosis.
(91.5%) and specificity (93.9%) with the model's
assistance, resulting in a 47.0% reduction in
misinterpretation rates. These findings underscore the
potential of deep learning to enhance diagnostic accuracy
and patient care in emergency settings.

2
 Surendar Reddy Vinta et al [6] A hybrid model DATASET SAMPLES
utilizing deep learning algorithms YOLO NAS, Efficient
Det, and DETR3 is developed to detect hand bone and joint
fractures through X-rays. The model, trained on a dataset
of 4736 hand-bone X-ray images, achieves precise object
detection across six fracture classes, offering improved
accuracy compared to existing methods. This approach
addresses the challenges of missed fractures and delayed
diagnosis, enhancing the effectiveness of fracture
detection in hand X-ray imaging.

Yangling et al. [22] developed a system that splits

all hu- man bones into 20 different types based on human
anatomy. For the detection tasks, the Faster R-CNN FIGURE 1: MANDIBULAR FRACTURE
algorithm was used. The authors compared various
methodologies and the results show that the Crack Net is
better in terms of recall and specificity. The system
achieved 88,39% accuracy out- performing other
methods.

Enmiao et al. [23] conducted a study on the role of

Ultrasound images to assess and diagnose various bone
fractures. The risk of bias and relevancy of eligible studies
was assessed using four domains: selection of patients,
index test, reference standard, and flow and timing. The
risk of bias calculation was assessed using the QUADAS-
2 risk assessment tool. Compared to conventional
radiography, the results demonstrated a high degree of FIGURE 2: MAXILLARY FRACTURE
accuracy in fracture detection.

Finally, In the field of facial bone fracture

identification, previous models have predominantly focused
on nasal fractures with an accuracy of 66.67%, leaving
other crucial types such as mandibular and maxillary
fractures overlooked. Addressing this gap, our study
introduces a novel model capable of detecting a broader
range of fractures, including mandibular, maxillary, and
nasal fractures. Notably, our model achieved an
exceptional accuracy of 89.0%. This advancement
represents a significant leap forward in facial bone fracture
diagnosis, offering the potential to enhance patient
outcomes through early and precise detection of various
fracture types. FIGURE 3: NASAL FRACTURE

III. MATERIALS AND METHODOLOGIES Encompassing various fracture types, including nasal,
mandibular, and maxillary fractures, each image is
A. DATASET DESCRIPTION meticulously annotated for precise identification. The dataset
is meticulously divided into training, validation, and testing
The success of any model heavily depends on the quality subsets, allowing for thorough model evaluation. Its
and diversity of its dataset. Our model leverages a simplicity and clarity render it an invaluable resource for
meticulously curated custom dataset comprising 1862 researchers and practitioners in the fracture detection domain
images sourced from diverse channels and carefully
selected based on their quality and relevance.

3
 TABLE 1: Comparison table of Related Work

Authors Year Methods used Dataset Type Detection/Classification/Analysis No. of classes

Yeonjin Jeong et al 2022 YOLO X-S MODEL Public Detection 7 (only Nasal Fracture)
Kandel et al. 2020 Transfer learning Public Classification 2 (Musculo Skeletal
Fracture)
Atsuyuki Inui et al 2023 Yolo V8 Self-developed Classification 2 (Baseball Elbow
fracture)
Rakesh et al. 2021 Bersen algorithm Public Detection 2 (Wrist Fracture)

Surendar Reddy Vintra 2024 Yolo NAS Public Detection 6 (Hand Fracture)
et al

B. DATA PREPROCESSING
C. EXISTING METHODOLOGIES
The process of data preprocessing is essential for preparing
raw data effectively for real-time detections. Raw data To evaluate the model's performance, we conducted
often contains noise and inconsistencies, making experiments using various established algorithms. The
preprocessing crucial to optimize its quality and suitability dataset created for this study was used consistently across all
for machine learning and deep learning tasks. The primary implementations of existing methodologies.
steps involved in data preprocessing include cleaning,
1) CNN MODEL – Inception V3
which involves removing noise and filling in missing
values; transformation, which organizes the dataset for
InceptionV3 is widely recognized for its balanced approach
better understanding; integration, merging data from
between accuracy and processing efficiency, making it
different sources; reduction, removing inefficient and
suitable for various image recognition tasks. However, its
irrelevant data; and normalization, facilitating comparison application in detecting facial bone fractures has limitations.
across different classes or features. In our study, we While pre-training on large datasets like ImageNet provides
utilized a custom dataset comprising images from various a rich set of features, the model may struggle to capture the
sources. To enhance the dataset's quality and reliability, we intricacies of medical imaging data. Additionally, its focus
manually cleared noise and redundancies in the images and on image classification may not directly translate to precise
manually annotated them based on predefined classes. No fracture localization in facial bone X-rays. The
automated tools were employed in this annotation process, computational complexity of InceptionV3, particularly with
ensuring accuracy and consistency in dataset preparation. high-resolution medical images, can lead to longer
processing times and increased resource requirements,
ANNOTATED IMAGES hindering real-time applications. Furthermore, overfitting
may occur when trained on limited medical imaging data,
affecting generalization performance on unseen datasets.
These constraints underscore the necessity for specialized
architectures tailored explicitly for facial bone fracture
detection.

2) Yolo X-S Model

The YoloX model was chosen as the baseline for several

reasons. Firstly, it has demonstrated strong performance on
benchmark datasets like coco, indicating its efficacy across
various object detection tasks. Secondly, its anchor-free
strategy is advantageous for detecting nasal bone fractures
due to their ambiguous nature. Unlike anchor-based
methods, YoloX predicts one object per grid, making it
suitable for our task. Thirdly, YoloX utilizes IoU loss for
regression, which is well-suited for handling ambiguous
bounding boxes encountered in nasal bone fracture
detection. By focusing solely on IoU loss, the model can
learn fracture bounding boxes more flexibly.

4
Fourthly, YoloX employs a decoupled head structure to
mitigate interference between bounding box regression and 4) Yolo V8
class classification, resulting in more accurate predictions.
Unlike coupled head structures, which output results Ultralytics' YOLOv8 model represents a
together, decoupled heads independently produce significant advancement in object identification and image
classification and regression results, reducing the trade-off segmentation, offering improvements in speed, accuracy,
between the two tasks. Finally, YoloX uses a multi-positive and adaptability compared to its predecessors. One notable
strategy to address data imbalance issues common in enhancement is the adoption of anchor-free detection, which
medical datasets like fracture data. This sampling strategy eliminates the reliance on pre-defined anchor boxes for
considers predicted bounding boxes in a 3x3 region predicting object bounding boxes. This approach enhances
centered on the fracture location as positive samples, performance efficiency by allowing for more flexible
alleviating the problem of insufficient positive samples. detection of objects with varying sizes and shapes.
Additionally, YOLOv8 achieves high accuracy while
However, YoloX also has limitations. Its computational maintaining fast inference speeds, making it suitable for a
complexity may pose challenges for real-time applications wide range of object detection tasks.
or resource-constrained environments. Additionally, while
the anchor-free strategy is beneficial for certain tasks As a single forward pass neural network, YOLOv8
like nasal bone fracture detection, it may not perform efficiently predicts bounding boxes and class probabilities
optimally in scenarios requiring precise localization of in an image by dividing it into a grid of cells. Following
objects. Moreover, the multi-positive strategy, while prediction, the model employs a non-maximum suppression
effective in mitigating data imbalance, may introduce noise (NMS) algorithm to filter out overlapping bounding boxes,
or false positives in the predictions, especially in cases of ensuring that only the most confident predictions are
overlapping fractures or complex anatomical retained. In the context of facial bone identification, this
structures. These limitations should be considered streamlined process enables accurate detection based on
when deploying YoloX for fracture detection tasks. annotated images from the dataset, leveraging convolutional
and pooling layers within the model's architecture.
3) Yolo V4
Activation functions play a crucial role in
In utilizing Yolo V4 for facial fracture detection, optimizing model performance, and YOLOv8 utilizes Leaky
several limitations need consideration. Firstly, while Yolo ReLU, Swish, and Tanh functions to mitigate the vanishing
V4 is known for its high accuracy and efficiency in object gradient problem commonly encountered in classification
detection tasks, its performance may vary depending on tasks. These activation functions contribute to the model's
factors such as dataset quality, model configuration, and ability to effectively identify facial bone fractures by
training methodology. Secondly, the effectiveness of Yolo enhancing gradient flow and promoting efficient learning.
V4 in detecting facial fractures may be impacted by the
complexity and variability of fracture patterns, particularly Overall, YOLOv8's advancements in speed,
in cases where fractures are small, subtle, or obscured by accuracy, and adaptability, coupled with its streamlined
other anatomical structures. architecture and effective use of activation functions, make
it a powerful tool for facial bone identification tasks,
Additionally, Yolo V4's reliance on anchor boxes offering significant advantages over previous models.
for object localization may pose challenges in accurately Activation functions have a noteworthy influence on the
delineating fractured regions, especially in scenarios where performance of the model. In the YOLOv8 algorithm, Leaky
fractures exhibit irregular shapes or orientations. ReLU, Swish, and Tanh are the most commonly used as they
Furthermore, the computational resources required for give the best results for classification tasks by avoiding the
training and inference with Yolo V4 may be significant, vanishing gradient problem.
limiting its accessibility for resource-constrained
environments or real-time applications. Moreover, Yolo • Leaky ReLU function:
V4 may struggle with detecting facial fractures in images x = max (a ∗ x, x)
with poor quality or low resolution, potentially leading
to false negatives or inaccurate predictions. Finally, the ▪ Swish activation function:
generalization ability of Yolo V4 to detect facial fractures x = x ∗ sigmoid (beta ∗ x)
across diverse patient populations, imaging modalities, and
clinical settings remains to be fully evaluated, highlighting ▪ Tanh activation function:
the need for comprehensive validation studies in real-world
healthcare environments. These limitations underscore the
importance of careful consideration and evaluation when
employing Yolo V4 for facial fracture detection tasks.

5
 ▪ Feature Extraction: After preprocessing, feature
IV. PROPOSED MODEL extraction techniques are applied to capture relevant
information from the X-ray images. This involves
1. Description:
identifying key features such as bone structures, density
The proposed model for fracture identification in variations, and fracture patterns using methods like edge
facial bone X-rays integrates deep learning detection, texture analysis, and morphological
methodologies, specifically employing the YOLOv8 operations with labeling
algorithm. YOLOv8 stands out as a cutting-edge object
detection and image segmentation model renowned for its
rapid processing, precision, and versatility. In this
instance, the YOLOv8 algorithm will be trained using a
dataset comprising 1862 facial bone X-ray images,
categorized into three distinct types of fractures like
Mandibular, Maxillary and Nasal Fractures Through
the utilization of transfer learning, the pre-trained
YOLOv8 model will be fine-tuned, aiming to enhance
accuracy and efficiency in identifying facial bone
fractures compared to conventional approaches.
Furthermore, the proposed model will undergo
comprehensive evaluation to gauge its proficiency in
precisely detecting fractures within facial bone X-rays.

2. Proposed Architecture: ▪ Model Creation: With the extracted features, a deep

learning model is constructed to learn patterns and
relationships within the data for fracture detection.
The model architecture, such as YOLOv8, is selected
based on factors like performance and efficiency. It is
then trained on a labeled dataset of X-ray images,
where the model learns to associate extracted features
with fracture labels and optimize its parameters for
accurate predictions.

Input X-ray Image: The process begins with obtaining X-

ray images of facial bones from medical imaging devices
or databases. These images serve as the foundation for
fracture detection.
Preprocessing: The acquired X-ray images undergo
preprocessing to enhance their quality and prepare them
for analysis. This includes resizing the images to a
standardized format, normalizing pixel intensity values,
reducing noise artifacts, and adjusting contrast levels for
optimal visualization.

6
 TABLE 2: Comparison between Proposed and
existing models
▪ Prediction: Once the model is trained, it is ready to S. No
make predictions on new, unseen X-ray images. The Author Method Accuracy%
trained model takes preprocessed X-ray images as input
and generates predictions regarding the presence and 1 Yeonjin YOLO X-S 66.67% (only
Jeong et al MODEL nasal fracture)
location of fractures. These predictions are represented
as bounding boxes or segmentation masks overlaid on
2 Proposed Customized 89.0% (Mandibular,
the original images, providing valuable insights for Dataset with Maxillary, nasal
healthcare professionals in diagnosing and treating YOLO V8 fractures)
facial bone fractures.
OUTPUT

MANDIBULAR FRACTURE WITH PRECISION

V) EXPERIMENT AND RESULTS PERCENTAGE OF 92%

The dataset utilized in this project is a Custom

compilation, comprising a diverse selection of high-quality
X-ray images. With a substantial number of samples
allocated for both training and testing, the dataset is well-
suited for comprehensive evaluation and validation.
Evaluation metrics including accuracy, precision,
sensitivity (recall), and F1 score were employed to
gauge performance across various aspects. The model's
overall prediction accuracy and the classification accuracy
specific to the validation (training) data were meticulously
measured. Additionally, the counts of true positives, true
negatives, false positives, and false negatives were
analyzed to provide a comprehensive understanding of
each model's performance in fracture identification and
classification.

MAXILLARY FRACTURE WITH PRECISION

PERCENTAGE OF 96%

7
 F1 – Confidence Curve

It showcases the relationship between

the confidence level of the model's predictions and its
performance metrics, such as accuracy or precision.
As the confidence level increases, typically, the
model's performance improves. The curve helps
visualize how changes in confidence thresholds affect
the model's ability to make accurate predictions.
NASAL FRACTURE WITH PRECISION
PERCENTAGE OF 92%

The precision percentages obtained for

Mandibular, Maxillary, and Nasal fractures are
impressive, with values of 92%, 96%, and 92%,
respectively. These high precision rates reflect the
model's accuracy in correctly identifying these
specific types of facial bone fractures.

8
GRAPHS:

In the training phase, the model undergoes continuous evaluation through various loss and
metric calculations. The training loss includes box loss, representing regression errors in bounding boxes, class
loss, indicating errors in class predictions, and dynamic feature learning (DFL) loss, accounting for losses
during dynamic feature learning. Alongside, metrics such as precision, recall, and mean average precision
(mAP) at different Intersection over Union (IoU) thresholds are computed to gauge the model's performance.
Validation loss metrics, including box loss, class loss, and DFL loss, are also monitored to ensure the model's
generalization on unseen data. Additionally, adjustments in learning rates for different parameter groups are
made throughout the training process to optimize convergence and prevent overfitting.

We trained the model based on the iterations, here we print the loss parameters & Accuracy from each Iteration.
For a iteration of 10 times.

9
VI. CONCLUSION
[7] R. Lindsey, A. Daluiski, S. Chopra, A. Lachapelle, M.
The project represents a significant advancement in Mozer, S. Sicular, D. Hanel, M. Gardner, A. Gupta, R.
the field of facial bone fracture detection through the Hotchkiss et al., “Deep neural network improves
innovative integration of deep learning with transfer learning fracture detection by clinicians,” Proceedings of the
techniques. By leveraging the YOLOv8 algorithm, the National Academy of Sciences, vol. 115, no. 45, pp.
model demonstrates remarkable accuracy and efficiency in 11 591–11 596, 2018.
identifying mandibular, maxillary, and nasal fractures,
addressing a critical gap in existing methodologies. [8] J. Olczak, F. Emilson, A. Razavian, T. Antonsson, A.
The precision percentages obtained for mandibular, Stark, and M. Gordon, “Ankle fracture classification
maxillary, and nasal fractures further validate the model's using deep learning: automating detailed ao
effectiveness in fracture detection. Additionally, the foundation/orthopedic trauma association (ao/ota) 2018
confidence curve and detailed analysis of loss metrics and malleolar fracture identification reaches a high degree
performance indicators offer valuable insights into the of correct classification,” Acta Orthopaedical, pp. 1–7,
model's performance and learning dynamics. Overall, the 2020.
project's findings underscore its potential as a reliable and
practical solution for automated fracture detection in facial [9] N. Tajbakhsh, Y. Hu, J. Cao, X. Yan, Y. Xiao, Y. Lu,
bone X-rays, with implications for improving diagnostic J. Liang, D. Terzopoulos, and X. Ding, “Surrogate
accuracy and patient care in clinical settings. supervision for medical image analysis: Effective deep
learning from limited quantities of labeled data,” in
REFERENCES IEEE Intl. Symposium on Biomedical Imaging. IEEE,
2019, pp
[1] 1. Kandel I, Castelli M, Popovicˇ A. Musculoskeletal
[10] O. Ranneberger, P. Fischer, and T. Brox, “U-net:
Images Classification for Detection of Fractures Using Convolutional networks for biomedical image
Transfer Learning. J Imaging. 2020 Nov. DOI: segmentation,” in Intl. Conf. on Medical Image
10.3390/jimaging6110127. PMID: 34460571; PMCID: Computing and Computer-assisted Intervention.
PMC8321195. Springer, 2015, pp. 234–241.

[2] Yadav DP, Sharma A, Athithan S, Bhola A, Sharma [11] J. Prijs, Z. Liao, M.-S. To, J. W. Verjans, P. Jutte, V.
B, Dhaou IB. Hybrid SFNet Model for Bone Fracture Stirler, J. Olczak, M. Gordon, D. Guss, C. DiGiovanni,
Detection and Classification Using ML/DL. Sensors R. Jaarsma, F. IJpma, and J. Doornberg, “Development
(Basel). 2022 Aug DOI: 10.3390/s22155823. PMID: and external validation of automated detection,
35957380; PMCID: PMC9371081. classification, and localization of ankle fractures: Inside
the black box of a convolutional neural network
[3] Rakesh, Y., & Akilandeswari, A. (2023). Evaluation of (CNN),” Submitted work, 2021.
Bernsen algorithm for bone fracture detection based on
edge detection. Journal of Survey in Fisheries Sciences, [12] S. Pereira, R. Meier, V. Alves, M. Reyes, and C. A.
10(1S), 2060-2068. Silva, “Automatic brain tumor grading from mri data
using convolutional neural networks and quality
[4] Inui A, Mifune Y, Nishimoto H, Mukohara S, Fukuda S, assessment,” in Understanding and Interpreting
Kato T, Furukawa T, Tanaka S, Kusunose M, Takigami Machine Learning in Medical Image Computing
S, et al. Detection of Elbow OCD in the Ultrasound Applications. Springer, 2018, pp. 106–114.
Image by Artificial Intelligence Using YOLOv8. Applied
Sciences.2023: https://doi.org/10.3390/app13137623 [13] S. Soffer, A. Ben-Cohen, O. Shimon, M. M. Amitai, H.
Greenspan, and E. Klang, “Convolutional neural
[5] Moon, G., Kim, S., Kim, W., Kim, Y., Jeong, Y., & Choi, networks for radiologic images: a radiologist’s guide,”
H.-S. (2022). Computer aided facial bone fracture Radiology, vol. 290, no. 3, pp. 590–606, 2019.
diagnosis (CA-FBFD) system based on object detection
model. Medical Imaging Review, 10(3), 123-135. DOI: [14] E. Ozkaya, F. E. Topal, T. Bulut, M. Gursoy, M.
10.1109/ACCESS.2022.3192389 Ozuysal, and Z. Karakaya, “Evaluation of an artificial
intelligence system for diagnosing scaphoid fracture on
direct radiography,” European Journal of Trauma and
[6] Medaramatla, S. C., Samhitha, C. V., Pandey, S. D., & Emergency Surgery, pp. 1–8, 2020.
Vinta, S. R. (2024). Detection of Hand Bone Fractures in
X-ray Images using Hybrid YOLO NAS. Medical
Imaging Review, 15(2), 150-165. DOI
10.1109/ACCESS.2024.3379760

10
[15] D. W. Langerhuizen, A. E. J. Bulstra, S. J. Janssen, D.
Ring, G. M. Kerkhoffs, R. L. Jaarsma, and J. N.
Doornberg, “Is deep learning on par with human
observers for detection of radiographically visible and
occult fractures of the scaphoid?” Clinical Orthopaedics
and Related Research, 2020.

[16] Qi, Y., Zhao, J., Shi, Y., Zuo, G., Zhang, H., Long, Y.,
Wang, F., & Wang, W. (Year). Ground Truth Annotated
Femoral X-Ray Image Dataset and Object Detection
Based Method for Fracture Types Classification. IEEE,
Volume 8, 2020. DOI: 10.1109/ACCESS.2020.3029039

[17] Srinivas, P., Ahmed, J., & Chander, M. B. Maxillofacial

Fracture Detection System in Accident Victims using
Convolution Neural Network (CNN) Transfer Learning.

[18] Kosrat Dlshad Ahmed, Roojwan Hawezi, Detection of

bone fracture based on machine learning techniques,
Measurement: Sensors, Volume 27, 2023, 100723, ISSN
2665-9174,
https://doi.org/10.1016/j.measen.2023.100723.

[19] H. Guly, “Diagnostic errors in an accident and emergency

department,” Emergency Medicine Journal, vol. 18, no.
4, pp. 263–269, 2001.

[20] Robert Lindsey, Aaron Daluiski, Sumit Chopra ,

Alexander Lachapelle , Michael Mozer, Serge Sicular ,
Douglas Hanel , Michael Gardner , Anurag Gupta ,
Robert Hotchkiss , Hollis Potter . Deep neural network
improves fracture detection by clinicians. PMID:
30348771 PMCID: PMC6233134 DOI:
10.1073/pnas.1806905115

[21] Young-Dae Jeon,Min-Jun Kang,Sung-Uk Kuh,Ha-

Yeong Cha,Moo-Sub Kim,Ju-Yeon You,Hyeon-Joo
Kim,Seung-Han Shin,Yang-Guk Chung andDo-Kun
Yoon Deep Learning Model Based on You Only Look
Once Algorithm for Detection and Visualization of
Fracture Areas in Three-Dimensional Skeletal Images
Link: https://www.mdpi.com/2075-4418/14/1/11

[22] Bone Fracture Detection Through the Two-stage System

of Crack-Sensitive Convolutional Neural Network
Yangling Ma, Yixin Luo. November 2020

[23] Enmiao Li, Qingsong Tan, Role of ultrasound

imaging to assess and diagnose various body
fractures: Systemic review and meta-analysis,
Journal of Radiation Research and Applied
Sciences, Volume 15, Issue 3, 2022,
https://doi.org/10.1016/j.jrras.2022.08.008.

11
12