Fracture Detection in Pediatric Wrist Trauma X-ray

Images Using YOLOv8 Algorithm

Rui-Yang Ju1 and Weiming Cai2,*
1 NationalTaiwan University, Graduate Institute of Networking and Multimedia, Taipei City, 106335, Taiwan
2 JingjiangPeople’s Hospital, Department of Hand and Foot Surgery, Jingjiang City, 214500, China
* Corresponding author: 1318746637@qq.com

ABSTRACT
arXiv:2304.05071v5 [cs.CV] 14 Nov 2023

Hospital emergency departments frequently receive lots of bone fracture cases, with pediatric wrist trauma fracture accounting
for the majority of them. Before pediatric surgeons perform surgery, they need to ask patients how the fracture occurred and
analyze the fracture situation by interpreting X-ray images. The interpretation of X-ray images often requires a combination of
techniques from radiologists and surgeons, which requires time-consuming specialized training. With the rise of deep learning
in the field of computer vision, network models applying for fracture detection has become an important research topic. In this
paper, we use data augmentation to improve the model performance of YOLOv8 algorithm (the latest version of You Only Look
Once) on a pediatric wrist trauma X-ray dataset (GRAZPEDWRI-DX), which is a public dataset. The experimental results show
that our model has reached the state-of-the-art (SOTA) mean average precision (mAP 50). Specifically, mAP 50 of our model is
0.638, which is significantly higher than the 0.634 and 0.636 of the improved YOLOv7 and original YOLOv8 models. To enable
surgeons to use our model for fracture detection on pediatric wrist trauma X-ray images, we have designed the application
“Fracture Detection Using YOLOv8 App” to assist surgeons in diagnosing fractures, reducing the probability of error analysis,
and providing more useful information for surgery.

Introduction
In hospital emergency rooms, radiologists are often asked to examine patients with fractures in various parts of the body, such
as the wrist and arm. Fractures can generally be classified as open or closed, with open fractures occurring when the bone
pierces the skin, and closed fractures occurring when the skin remains intact despite the broken bone. Before performing
surgery, the surgeon must inquire about the medical history of the patients and conduct a thorough examination to diagnose
fracture. In recent medical imaging, three types of devices, including X-ray, Magnetic Resonance Imaging (MRI), and
Computed Tomography (CT), are commonly used to diagnose fracture1 . And X-ray is the most widely used device due to its
cost-effectiveness.
Fractures of the distal radius and ulna account for the majority of wrist trauma in pediatric patients2, 3 . In prestigious
hospitals of developed areas, there are many experienced radiologists who are capable of correctly analyzing X-ray images;
while in some small hospitals of underdeveloped regions, there are only young and inexperienced surgeons who may be unable
to correctly interpret X-ray images. Therefore, a shortage of radiologists would seriously jeopardize timely patient care4, 5 .
Specifically, some hospitals in Africa have even limited access to specialist reports6 , which badly affects the probaility of the
sucess of surgery. According to the survey7, 8 , the percentage of X-ray images misinterpreted have reached 26%.
With the advancement of deep learning, neural network models have been introduced in medical image processing9–12 .
In recent years, researchers have started to apply object detection models13–15 to fracture detection16–19 , which is a popular
research topic in computer vision (CV).
Deep learning methods in the field of object detection are divided into two-stage and one-stage algorithms. Two-stage
algorithm models such as R-CNN13 and its improved models20–26 generate location and class probabilities in two stages.
Whereas one-stage algorithm models directly produce the location and class probabilities of objects, resulting in the improvement
of the model inference speed. In addition to the classical one-stage algorithm models, such as SSD27 , RetinaNet28 , CornerNet29 ,
CenterNet30 , and CentripetalNet31 , You Only Look Once (YOLO) series algorithm models32–34 are preferred for real-time
applications35 due to the good balance between the model accuracy and inference speed.
In this paper, we first use YOLOv8 algorithm36 to train models of different sizes on the GRAZPEDWRI-DX37 dataset.
After evaluation of the model performances of YOLOv8, we train the models by using data augmentation to detect wrist
fractures in children. We compare YOLOv8 models using our training method with YOLOv7 and its improved models, and the
experimental results demonstrate that our models have the highest the mean average precision (mAP 50) value.
The contributions of this paper are summarized as follows:
 • We use data augmentation to improve the model performance of YOLOv8 model. The experimental results show that the
mean average precision of YOLOv8 model using our training method for fracture detection on the GRAZPEDWRI-DX
dataset reaches SOTA value.

• This work develops an application to detect wrist fracture in children, which aims to help pediatric surgeons interpret
X-ray images without the assistance of the radiologist, and reduce the probability of X-ray image analysis errors.

This paper is structured as follows: Section Related Work describes the deep learning methods for detecting fracture,
and describes the application of YOLOv5 model in medical image processing. Section Proposed Method introduces the
whole process of training and the architecture of our model. Section Experiments presents the improved performance of
YOLOv8 model using our training method compared with YOLOv7 and its improved models. Section Application describes
our proposed application to assist pediatric surgeons in analyzing X-ray images. Finally, Section Conclusions and Future
Work discusses the conclusions and future work of this paper.

Table 1. Experimental results of other studies on fracture detection in various parts of the body based on deep learning method.

Author Task Model Dataset mAPval 50

Guan et al.38 Thigh Fracture Detection DCFPN 3,842 thigh fracture X-ray radiographs 0.821
Wang et al.39 Thigh Fracture Detection R-CNN 3,842 thigh fracture X-ray radiographs 0.878
Guan et al.40 Arm Fracture Detection R-CNN Musculoskeletal-Radiograph (MURA)41 0.620
Wu et al.42 Bone Fracture Detection FAMO 9,040 radiographs of various body parts 0.774
Ma and Luo43 Bone Fracture Detection Faster R-CNN 1,052 bone x-ray images 0.884
Xue et al.44 Hand Fracture Detection Faster R-CNN 3,067 hand trauma x-ray images 0.700
Sha et al.45 Spine Fracture Detection Faster R-CNN 5,134 spine fractures CT images 0.733
Sha et al.46 Spine Fracture Detection YOLOv2 5,134 spine fractures CT images 0.753

Related Work
In recent years, neural networks have been widely utilized in image data for fracture detection. Guan et al.38 achieved the
average precision of 82.1% on 3,842 thigh fracture X-ray images using the Dilated Convolutional Feature Pyramid Network
(DCFPN). Wang et al.39 employed a novel R-CNN13 network ParalleNet as the backbone network for fracture detection on
3,842 thigh fracture X-ray images. In addition to thigh fracture detection, about arm fracture detection, Guan et al.40 used
R-CNN for detection on Musculoskeletal-Radiograph (MURA) dataset41 and obtained an average precision of 62.04%. Ma and
Luo43 used Faster R-CNN21 for fracture detection on a part of 1,052 bone images of the dataset and the proposed CrackNet
model for fracture classification on the whole dataset. Wu et al.42 proposed Feature Ambiguity Mitigate Operator (FAMO)
model based on ResNeXt10147 and FPN48 for bone fracture detection on 9,040 radiographs of various body parts. Qi et al.49
utilized Fast R-CNN20 with ResNet5050 as the backbone network to detect nine different types of fractures on 2,333 fracture
X-ray images. Xue et al.44 utilized the Faster R-CNN model for hand fracture detection on 3,067 hand trauma X-ray images,
achieving an average precision of 70.0%. Sha et al.45, 46 used YOLOv251 and Faster R-CNN21 models for fracture detection on
5,134 CT images of spine fractures respectively. Experiments showed that the average precision of YOLOv2 reached 75.3%,
which was higher than 73.3% of Faster R-CNN, and inference time of YOLOv2 for each CT image is 27ms, which is much
faster than 381ms of Faster R-CNN. From Table 1, it can be seen that even though most of the works using R-CNN series
models have shown excellent results, the inference speed is not satisfactory.
YOLO series models32–34 offer a balance of performance in terms of the model accuracy and inference speed, which is
suitable for mobile devices in real-time X-ray images detection. Hržić et al.52 proposed a machine learning model based
on YOLOv4 method to help radiologists diagnose fractures and demonstrated that the AUC-ROC (area under the receiver
operator characteristic curve) value of YOLO 512 Anchor model-AI was significantly higher than that of radiologists. YOLOv5
model53 , which was proposed by Ultralytics in 2021, has been deployed on mobile phones as the “iDetection” application.
On this basis, Yuan et al.54 employed external attention and 3D feature fusion techniques in YOLOv5 model to detect skull
fractures in CT images. Warin et al.55 used YOLOv5 model to detect maxillofacial fractures in 3,407 maxillofacial bone CT
images, and classified the fracture conditions into frontal, midfacial, mandibular fractures and no fracture. Rib fractures are a
precursor injury to physical abuse in children, and chest X-ray (CXR) images are preferred for effective diagnosis of rib fracture
conditions because of their convenience and low radiation dose. Tsai et al.56 used data augmentation with YOLOv5 model to
detect rib fractures in CXR images. And Burkow et al.57 applied YOLOv5 model to detect rib fractures in 704 pediatric CXR
images, the model obtained the F2 score value of 0.58. To identify and detect mandibular fractures in panoramic radiographs,

2/15
 Data
70% Augmentation Extended YOLOv8
Training Set
Training Set Algorithm

Training

20% Validation Our

Dataset Validation Set
Model
Evaluation

10% Compare
Testing Set with Model

Figure 1. Flowchart of the model training, validation and testing on the dataset. The extended training set is used to double
the number of X-ray images by data augmentation.

C C2f P5 Detect
P5 U Conv
C C2f C C2f P4 Detect
P4 U Conv
C C2f P3 Detect
P3

P2 Neck Head
P1
Bbox. Cls.

CIoU+DFL 4×reg_max nc BCE

Backbone Loss

Figure 2. The architecture of YOLOv8 algorithm, which is divided into four parts, including backbone, neck, head, and loss.

Warin et al.58 used convolutional neural networks (CNNs) and YOLOv5 model to implement it. Fatima et al.59 used YOLOv5
model to localize vertebrae, which is important for detecting spinal deformities and fractures, and obtained an average precision
of 0.94 at an IoU (Intersection over Union) threshold of 0.5. Moreover, Mushtaq et al.60 applied YOLOv5 model to localize the
lumbar spine and obtained an average precision value of 0.975. Nevertheless, relatively few researches have been reported on
pediatric wrist fracture detection using YOLOv5 model. While YOLOv8 was proposed by Ultralytics in 2023, we use this
algorithm to train the model for the first time in pediatric wrist fracture detection.

Proposed Method
In this section, we introduce the process of the model training, validation and testing on the dataset, the architecture of YOLOv8
model, and the data augmentation technique employed during training. Figure 1 illustrates the flowchart depicting the model
training process and performance evaluation. We randomly divide the 20,327 X-ray images of the GRAZPEDWRI-DX dataset
into the training, validation, and test set, where the training set is expanded to 28,408 X-ray images by data augmentation from
the original 14,204 X-ray images. We design our model according to YOLOv8 algorithm, and the architecture of YOLOv8
algorithm is shown in Figure 2.

Data Augmentation
During the model training process, data augmentation is employed in this work to extend the dataset. Specifically, we adjust
the contrast and brightness of the original X-ray image to enhance the visibility of bone-anomaly. This is achieved using the
addWeighted function available in OpenCV (Open Source Computer Vision Library). The equation is presented below:

Out put = Input1 × α + Input2 × β + γ, (1)

where Input1 and Input2 are the two input images of the same size respectively, α represents the weight assigned to the first
input image, β denotes the weight assigned to the second input image, and γ represents the scalar value added to each sum.
Since our purpose is to adjust the contrast and brightness of the original input image, we take the same image as Input1 and
Input2 respectively and set β to 0. The value of α and γ represent the proportion of the contrast and the brightness of the image

3/15
 (a)

(b)

Figure 3. Examples of pediatric wrist X-ray images using data augmentation. (a) the original images, (b) the adjusted images.

respectively. The image after adjusting the contrast and brightness is shown in Figure 3. After comparing different settings, we
finally decided to set α to 1.2 and γ to 30 to avoid the output image being too bright.

Model Architecture
Our model architecture consists of backbone, neck, and head, as shown in Figure 4. In the following subsections, we introduce
the design concepts of each part of the model architecture, and the modules of different parts.
Backbone
The backbone of the model uses Cross Stage Partial (CSP)61 architecture to split the feature map into two parts. The first part
uses convolution operations, and the second part is concatenated with the output of the previous part. The CSP architecture
improves the learning ability of the CNNs and reduces the computational cost of the model.
YOLOv836 introduces C2f module by combining the C3 module and the concept of ELAN from YOLOv732 , which allows
the model to obtain richer gradient flow information. The C3 module consists of 3 ConvModule and n DarknetBottleNeck,
and the C2f module consists of 2 ConvModule and n DarknetBottleNeck connected through Split and Concat, as illustrated in
Figure 4, where the ConvModule consists of Conv-BN-SiLU, and n is the number of the bottleneck. Unlike YOLOv553 , we use
the C2f module instead of the C3 module.
Furthermore, we reduce the number of blocks in each stage compared to YOLOv5 to further reduce the computational cost.
Specifically, our model reduces the number of blocks to 3,6,6,3 in Stage 1 to Stage 4, respectively. Additionally, we adopt the
Spatial Pyramid Pooling - Fast (SPPF) module in Stage 4, which is an improvement from Spatial Pyramid Pooling (SPP)62 to
improve the inference speed of the model. These modifications lead to our model with a better learning ability and shorter
inference time.
Neck
Generally, deeper networks obtain more feature information, resulting in better dense prediction. However, excessively deep
networks reduce the location information of the object, and too many convolution operations will lead to information loss for
small objects. Therefore, it is necessary to use Feature Pyramid Network (FPN)48 and Path Aggregation Network (PAN)63
architectures for multi-scale feature fusion. As illustrated in Figure 4, the Neck part of our model architecture uses multi-scale
feature fusion to combine features from different layers of the network. The upper layers acquire more information due to the
additional network layers, whereas the lower layers preserve location information due to fewer convolution layers.
Inspired by YOLOv5, where FPN upsamples from top to bottom to increase the amount of feature information in the
bottom feature map; and PAN downsamples from bottom to top to obtain more the top feature map information. These two
feature outputs are merged to ensure precise predictions for images of various sizes. We adopt FP-PAN (Feature Pyramid-Path
Aggregation Network) in our model, and delete convolution operations in upsampling to reduce the computational cost.

4/15
 ConvModule Conv2d Bbox_Loss CIoU+DFL
20×20×512×w×r
C2f
20×20×512×w×(1+r) ConvModule Conv2d Cls_Loss BCE
20×20×512×w×r
SPPF Concat 20×20×(4+cls)

Stage 4 C2f 20×20×512×w

ConvModule P5 ConvModule

Upsample ConvModule Conv2d CIoU+DFL

Bbox_Loss
40×40×512×w
40×40×512×w C2f
C2f Concat
Stage 3 40×40×512×w ConvModule Conv2d Cls_Loss BCE
ConvModule P4 40×40×512×w
C2f Concat 40×40×(4+cls)
40×40×512×w
Upsample ConvModule
80×80×256×w
C2f Concat
Stage 2 ConvModule Conv2d Bbox_Loss
80×80×256×w CIoU+DFL
ConvModule P3
C2f
ConvModule Conv2d Cls_Loss BCE
80×80×(4+cls)
Neck Head
C2f channel=c_in
Stage 1
ConvModule P2 ConvModule ConvModule
0.5*c_out add=True/False
c_out
0.5*c_out Conv2d
Stem Layer ConvModule P1 Split Maxpool2d
0.5*c_out
ConvModule
0.5*c_out
DarknetBottleneck Maxpool2d
nx ⋯ BN
0.5*c_out ConvModule
DarknetBottleneck Maxpool2d
640×640×3 Input 0.5*c_out
SiLU
Concat Concat
0.5(n+2)*c_out
ConvModule ConvModule
Backbone C2f c_out ConvModule DarknetBottleneck SPPF

Figure 4. Detailed illustration of YOLOv8 model architecture. The Backbone, Neck, and Head are the three parts of our
model, and C2f, ConvModule, DarknetBottleneck, and SPPF are modules.

Head
Different from YOLOv5 model utilizing a coupled head, we use a decoupled head33 , where the classification and detection
heads are separated. Figure 4 illustrates that our model deletes the objectness branch and only retains the classification and
regression branches. Anchor-Base employes a large number of anchors in the image to determine the four offsets of the
regression object from the anchors. It adjusts the precise object location using the corresponding anchors and offsets. In
contrast, we adopt Anchor-Free64 , which identifies the center of the object and estimates the distance between the center and
the bounding box.

Loss
For positive and negative sample assignment, the Task Aligned Assigner of Task-aligned One-stage Object Detection (TOOD)65
is used in our model training to select positive samples based on the weighted scores of classification and regression, as shown
in Equation 2 below:

t = sα × uβ , (2)

where s is the predicted score corresponding to the labeled class, and u is the IoU of the prediction and the ground truth
bounding box.
In addition, our model has classification and regression branches, where the classification branch uses Binary Cross-Entropy
(BCE) Loss, and the equation is shown below:

Lossn = −w [yn log xn + (1 − yn ) log (1 − xn )] , (3)

where w is the weight, yn is the labeled value, and xn is the predicted value of the model.
The regression branch uses Distribute Focal Loss (DFL)66 and Complete IoU (CIoU) Loss67 , where DFL is used to expand
the probability of the value around the object y. Its equation is shown as follows:

DFL(Sn , Sn+1 ) = −((yn+1 − y) log(Sn ) + (y − yn ) log(Sn+1 )), (4)

5/15
where the equations of Sn and Sn+1 are shown below:

yn+1 − y y − yn
Sn = , Sn+1 = . (5)
yn+1 − yn yn+1 − yn

CIoU Loss adds an influence factor to Distance IoU (DIoU) Loss68 by considering the aspect ratio of the prediction and the
ground truth bounding box. The equation is shown below:

Distance22 v2
CIoULoss = 1 − IoU + + , (6)
DistanceC2 (1 − IoU) + ν

where ν is the parameter that measures the consistency of the aspect ratio, defined as follows:

4 wgt wp 2
ν= (arctan − arctan ) , (7)
π2 hgt hp
where w is the weight of the bounding box, and h is the height of the bounding box.

Experiments
Dataset
Medical University of Graz provides a public dataset named GRAZPEDWRI-DX37 , which consists of 20,327 X-ray images
of wrist trauma in children. These images were collected from 6,091 patients between 2008 and 2018 by multiple pediatric
radiologists at the Department of Pediatric Surgery of the University Hospital Graz. The images are annotated in 9 different
classes by placing bounding boxes on them.
To perform the experiments shown in Table 5 and Table 6, we divide the GRAZPEDWRI-DX dataset randomly into three
sets: training set, validation set, and test set. The sizes of these sets are approximately 70%, 20%, and 10% of the original
dataset, respectively. Specifically, our training set consists of 14,204 images (69.88%), our validation set consists of 4,094
images (20.14%), and our test set consists of 2,029 images (9.98%). The code for splitting the dataset can be found on our
GitHub. We also provide csv files of training, validation and test data on our GitHub, but it should be noted that each split is
random and therefore not reproducible.

Evaluation Metric
Intersection over Union (IoU)
Intersection over Union (IoU) is a classical metric for evaluating the performance of the model for object detection. It calculates
the ratio of the overlap and union between the generated candidate bounding box and the ground truth bounding box, which
measures the intersection of these two bounding boxes. The IoU is represented by the following equation:

area(C) ∩ area(G)
IoU = , (8)
area(C) ∪ area(G)

where C represents the generated candidate bounding box, and G represents the ground truth bounding box containing the
object. The performance of the model improves as the IoU value increases, with higher IoU values indicating less difference
between the generated candidate and ground truth bounding boxes.

Precision-Recall Curve
Precision-Recall Curve (P-R Curve)69 is a curve with recall as the x-axis and precision as the y-axis. Each point represents a
different threshold value, and all points are connected as a curve. The recall (R) and precision (P) are calculated according to
the following equations:

TP TP
Recall = , Precision = , (9)
TP + FN TP + FP

where True Positive (TP ) denotes the prediction result as a positive class and is judged to be true; False Positive (FP ) denotes the
prediction result as a positive class but is judged to be false, and False Negative (FN ) denotes the prediction result as a negative
class but is judged to be false.

6/15
Table 2. Validation results of YOLOv8 for each class on the GRAZPEDWRI-DX dataset when the input image size is 1024.

mAPval mAPval
Class Boxes Instances Precision Recall
50 50-95
all 47435 9613 0.674 0.605 0.623 0.395
boneanomaly 276 53 0.505 0.094 0.110 0.035
bonelesion 45 8 0.629 0.250 0.416 0.212
fracture 18090 3740 0.885 0.903 0.947 0.572
metal 818 168 0.878 0.899 0.920 0.768
periostealreaction 3453 697 0.645 0.684 0.689 0.357
pronatorsign 567 104 0.561 0.713 0.611 0.338
softtissue 464 89 0.324 0.315 0.251 0.125
text 23722 4754 0.961 0.984 0.991 0.750

Table 3. Validation results of our model for each class on the GRAZPEDWRI-DX dataset when the input image size is 1024.

mAPval mAPval
Class Boxes Instances Precision Recall
50 50-95
all 47435 9613 0.694 0.592 0.631 0.402
boneanomaly 276 53 0.510 0.151 0.169 0.076
bonelesion 45 8 0.658 0.243 0.414 0.213
fracture 18090 3740 0.899 0.896 0.947 0.569
metal 818 168 0.898 0.890 0.924 0.780
periostealreaction 3453 697 0.721 0.654 0.700 0.359
pronatorsign 567 104 0.534 0.683 0.611 0.342
softtissue 464 89 0.367 0.236 0.241 0.120
text 23722 4754 0.961 0.981 0.991 0.754

Table 4. Model performance comparison of YOLOv8 models using SGD and Adam optimizers. For training with the SGD
optimizer, the initial learning rate is 1×10−2 ; for training with the Adam optimizer, the initial learning rate is 1×10−3 .

mAPval mAPval Speed GPU

Model Size Optimizer Best Epoch
50 50-95 RTX 3080Ti
YOLOv8s 640 SGD 56 0.611 0.389 4.4ms
YOLOv8s 640 Adam 57 0.604 0.383 4.3ms
YOLOv8s 1024 SGD 36 0.623 0.395 5.4ms
YOLOv8s 1024 Adam 47 0.625 0.399 4.9ms
YOLOv8m 640 SGD 52 0.621 0.396 4.9ms
YOLOv8m 640 Adam 62 0.621 0.403 5.5ms
YOLOv8m 1024 SGD 35 0.624 0.402 9.9ms
YOLOv8m 1024 Adam 70 0.626 0.401 10.0ms

F1-score
The F-score is a commonly used metric to evaluate the model accuracy, providing a balanced measure of performance by
incorporating both precision and recall. The F-score equation is as follows:
1 + β 2 × Precision × Recall


F-score = (10)
β 2 × Precision + Recall
When β = 1, the F1-score is determined by the harmonic mean of precision and recall, and its equation is as follows:
2 × Precision × Recall 2TP
F1 = = (11)
Precision + Recall 2TP + FP + FN

Experiment Setup
During the model training process, we utilize pre-trained YOLOv8 model from the MS COCO (Microsoft Common Objects
in Context) val2017 dataset72 . The research reports provided by Ultralytics36, 53 suggests that YOLOv5 training requires 300

7/15
Table 5. Quantitative comparison of fracture detection when the input image size is 640. Speed means the total time of
validate per image, and the total time includes the preprocessing, inference, and post-processing time.

mAPval mAPval Speed CPU Speed GPU

Model PARAMS FLOPs
50 50-95 Intel Core i5 RTX 3080Ti
YOLOv5n 0.589 0.339 \ 2.8ms 1.77M 4.2B
YOLOv8n 0.601 0.374 67.4ms 2.9ms 3.01M 8.1B
Ours 0.605 0.379 111.3ms 3.4ms 3.01M 8.2B
YOLOv5s 0.601 0.357 \ 3.3ms 7.03M 15.8B
YOLOv8s 0.604 0.383 191.5ms 4.3ms 11.13M 28.5B
Ours 0.612 0.392 285.1ms 4.9ms 11.13M 28.7B
YOLOv5m 0.613 0.371 \ 4.0ms 20.89M 48.0B
YOLOv8m 0.621 0.403 536.4ms 5.5ms 25.84M 78.7B
Ours 0.629 0.404 685.9ms 5.1ms 25.84M 78.7B
YOLOv5l 0.620 0.379 \ 5.6ms 46.15M 107.8B
YOLOv8l 0.624 0.403 1006.3ms 7.4ms 43.61M 164.9B
Ours 0.637 0.406 1370.8ms 7.2ms 43.61M 164.9B

Table 6. Quantitative comparison of fracture detection when the input image size is 1024. Speed means the total time of
validate per image, and the total time includes the preprocessing, inference, and post-processing time.

mAPval mAPval Speed CPU Speed GPU

Model PARAMS FLOPs
50 50-95 Intel Core i5 RTX 3080Ti
YOLOv5n 0.600 0.347 \ 3.2ms 1.77M 4.2B
YOLOv8n 0.605 0.387 212.1ms 3.3ms 3.01M 8.1B
Ours 0.608 0.391 260.4ms 4.4ms 3.01M 8.1B
YOLOv5s 0.622 0.371 \ 4.4ms 7.03M 15.8B
YOLOv8s 0.625 0.399 519.5ms 4.9ms 11.13M 28.5B
Ours 0.631 0.402 717.1ms 6.2ms 11.13M 28.5B
YOLOv5m 0.624 0.380 \ 7.1ms 20.89M 48.0B
YOLOv8m 0.626 0.401 1521.5ms 10.0ms 25.84M 78.7B
Ours 0.635 0.411 1724.4ms 9.4ms 25.85M 78.7B
YOLOv5l 0.626 0.378 \ 11.3ms 46.15M 107.8B
YOLOv8l 0.636 0.404 2671.1ms 15.1ms 43.61M 164.9B
Ours 0.638 0.415 3864.5ms 13.6ms 43.61M 164.9B

epochs, while training YOLOv8 requires 500 epochs. Since we use pre-trained model, we initially set the total number of
epochs to 200 with a patience of 50, which indicate that the training would end early if no observable improvement is noticed
after waiting for 50 epochs. In the experiment comparing the effect of the optimizer on the model performance, we notice that
the best epcoh of all the models is within 100, as shown in Table 4, mostly concentrated between 50 and 70 epochs. Therefore,
to save computing resources, we adjust the number of epochs for our model training to 100.
As the suggestion36 of Glenn, for model training hyperparameters, the Adam73 optimizer is more suitable for small custom
datasets, while the SGD74 optimizer perform better on larger datasets. To prove the above conclusion, we train YOLOv8
algorithm models using the Adam and SGD optimizers, respectively, and compare the effects on the model performance. The
comparison results are shown in Table 4.
For the experiments, we choose the SGD optimizer with an initial learning rate of 1×10−2 , a weight decay of 5×10−4 , and
a momentum of 0.937 during our model training. We set the input image size to 640 and 1024 for training on a single GPU
GeForce RTX 3080Ti 12GB with a batch size of 16. We train the model using Python 3.8 and PyTorch 1.8.2, and recommend
readers to use Python 3.7 or higher and PyTorch 1.7 or higher for training. It is noteworthy that due to GPU memory limitations,
we choose 3 worker threads to load data on GPU GeForce RTX 3080Ti 12GB when training our model. Therefore, using GPUs
with larger memory and more computing power can effectively increase the speed of model training.

Ablation Study
In order to demonstrate the positive effect of our training method on the performance of YOLOv8 model, we conduct an
ablation study on YOLOv8s model by calculating each evaluation metric for each class, as shown in Table 2. Among all

8/15
 (a) (b)

Figure 5. Detailed illustration of the validation at the input image size of 1024, (a) is our model, and (b) is YOLOv8 model.

classes, YOLOv8s model has good accuracy in detecting fracture, metal and text, with mAP 50 of each above 0.9. On the
opposite, the detection ability of bone-anomaly is poor, with mAP 50 of 0.11. Therefore, we increase the contrast and brightness
of X-ray images to make bone-anomaly easier to detect. Table 3 presents the predictions of YOLOv8s model using our
training method for each class. Compared with YOLOv8s model, the mAP value predicted by the model using our training
method for bone-anomaly increased from 0.11 to 0.169, an increase of 53.6%. Figure 5 also shows that our model has a
better performance in detecting bone-anomaly, which enables the improvement of the overall model performance. From the
ablation study presented above, we demonstrate that the model performance can be improved by using our training method
(data augmentation). In addition to the data enhancement, researchers can also improve model performance by adding modules
such as the Convolutional Block Attention Module (CBAM)70 .

Experimental Results
Before training our model, in order to choose an optimizer that has a more positive effect on the model performance, we
compare the performance of models trained with the SGD74 optimizer and the Adam73 optimizer. As shown in Table 4, using
the SGD optimizer to train the model requires less epochs of weight updates. Specifically, for YOLOv8m model with an input
image size of 1024, the model trained with the SGD optimizer achieves the best performance at the 35th epoch, while the best

9/15
 (a) (b)
Figure 6. Examples of pediatric wrist fracture detection on X-ray images. (a) manually labeled images, (b) predicted images.

Table 7. Evaluation of wrist fracture detection with other state-of-the-art (SOTA) models on the GRAZPEDWRI-DX dataset.

mAPval
Model Precision Recall F1
50
YOLOv553 0.682 0.581 0.607 0.626
YOLOv732 0.556 0.582 0.569 0.628
YOLOv732 + CBAM70 0.709 0.593 0.646 0.633
YOLOv732 + GAM71 0.745 0.574 0.646 0.634
YOLOv836 0.694 0.679 0.623 0.636
Ours 0.734 0.592 0.635 0.638

performance of the model trained with the Adam optimizer is at the 70th epoch. In terms of mAP and inference time, there
is not much difference in the performance of the models trained with the two optimizers. Specifically, when the input image
size is 640, the mAP value of YOLOv8s model trained with the SGD optimizer is 0.007 higher than that of the model trained
with the Adam optimizer, while the inference time is 0.1ms slower. Therefore, according to the above experimental results
and the suggestion by Glenn36, 53 , for YOLOv8 model training on a training set of 14,204 X-ray images, we choose the Adam
optimizer. However, after using data augmentation, the number of X-ray images in the training set extend to 28,408, so we
switch to the SGD optimizer to train our model.
After using data augmentation, our models have a better mAP value than that of YOLOv8 model, as shown in Table 5 and
Table 6. Specifically, when the input image size is 640, compared with YOLOv8m model and YOLOv8l model, the mAP 50
of our model improves from 0.621 to 0.629, and from 0.623 to 0.637, respectively. Although the inference time on the CPU
is increased from 536.4ms and 1006.3ms to 685.9ms and 1370.8ms, respectively, the number of parameters and FLOPs are
the same, which means that our model can be deployed on the same computing power platform. In addition, we compare the
performance of our model with that of YOLOv7 and its improved models. As shown in Table 7, the mAP value of our model is
higher than those of YOLOv732 , YOLOv7 with Convolutional Block Attention Module (CBAM)70 and YOLOv7 model with
Global Attention Mechanism (GAM)71 , which demonstrates that our model has obtained SOTA performance.
This paper aims to design a pediatric wrist fracture detection application, so we use our model for fracture detection. Figure
6 shows the results of manual annotation by the radiologist and the results predicted using our model. These results demonstrate
that our model has a good ability to detect fractures in single fracture cases, but metal puncture and dense multiple fracture
situations badly affects the accuracy of prediction.

10/15
 Figure 7. Example of using the application “Fracture Detection with YOLOv8 Application” on macOS operating system .

Application
After completing model training, we utilize a Python library that includes the Qt toolkit, PySide6, to develop a Graphical User
Interface (GUI) application. Specifically, PySide6 is the Qt6-based version of the PySide GUI library from the Qt Company.
According to the model performance evaluation results in Table 5 and Table 6, we choose our model with YOLOv8s
algorithm and the input image size of 1024, to perform fracture detection. Our model is exported to onnx format, and is applied
to the GUI application. Figure 7 depicts the flowchart of the GUI application operation on macOS. As can be seen from the
illustration, our application is named “Fracture Detection Using YOLOv8 App”. Users can open and predict the images, and
save the predictions in this application. In summary, our application is designed to assist pediatric surgeons in analyzing
fractures on pediatric wrist trauma X-ray images.

Conclusions and Future Work

Ultralytics proposed the latest version of YOLO series (YOLOv8) in 2023. Although there are relatively few research works
on YOLOv8 model for medical image processing, we apply it to fracture detection and use data augmentation to improve the
model performance. We randomly divide the dataset, consisting of 20,327 pediatric wrist trauma X-ray images from 6,091
patients, into training, test, and validation sets to train the model and evaluate the performance.
Furthermore, we develop an application named "Fracture Detection Using YOLOv8 App" to analyze pediatric wrist trauma
X-ray images for fracture detection. Our application aims to assist pediatric surgeons in interpreting X-ray images, reduce the
probability of misclassification, and provide a better information base for surgery. The application is currently available for
macOS, and in the future, we plan to deploy different sizes of our model in the application, and extend the application to iOS
and Android. This will enable inexperienced pediatric surgeons in hospitals located in underdeveloped areas to use their mobile
devices to analyze pediatric wrist X-ray images.
In addition, we provide the specific steps for training the model and the trained model in our GitHub. If readers wish to use
YOLOv8 model to detect fracture in other parts of the body except the pediatric wrist, they can use our trained model as the
pre-training model, which can greatly improve the performance of the model.

11/15
References
1. Fractures, health, hopkins medicine. https://www.hopkinsmedicine.org/health/conditions-and-diseases/fractures (2021).
2. Hedström, E. M., Svensson, O., Bergström, U. & Michno, P. Epidemiology of fractures in children and adolescents:
Increased incidence over the past decade: a population-based study from northern sweden. Acta orthopaedica 81, 148–153
(2010).
3. Randsborg, P.-H. et al. Fractures in children: epidemiology and activity-specific fracture rates. JBJS 95, e42 (2013).
4. Burki, T. K. Shortfall of consultant clinical radiologists in the uk. The Lancet Oncol. 19, e518 (2018).
5. Rimmer, A. Radiologist shortage leaves patient care at risk, warns royal college. BMJ: Br. Med. J. (Online) 359 (2017).
6. Rosman, D. et al. Imaging in the land of 1000 hills: Rwanda radiology country report. J. Glob. Radiol. 1 (2015).
7. Mounts, J., Clingenpeel, J., McGuire, E., Byers, E. & Kireeva, Y. Most frequently missed fractures in the emergency
department. Clin. pediatrics 50, 183–186 (2011).
8. Erhan, E., Kara, P., Oyar, O. & Unluer, E. Overlooked extremity fractures in the emergency department. Ulus Travma Acil
Cerrahi Derg 19, 25–8 (2013).
9. Adams, S. J., Henderson, R. D., Yi, X. & Babyn, P. Artificial intelligence solutions for analysis of x-ray images. Can.
Assoc. Radiol. J. 72, 60–72 (2021).
10. Tanzi, L. et al. Hierarchical fracture classification of proximal femur x-ray images using a multistage deep learning
approach. Eur. journal radiology 133, 109373 (2020).
11. Chung, S. W. et al. Automated detection and classification of the proximal humerus fracture by using deep learning
algorithm. Acta orthopaedica 89, 468–473 (2018).
12. Choi, J. W. et al. Using a dual-input convolutional neural network for automated detection of pediatric supracondylar
fracture on conventional radiography. Investig. radiology 55, 101–110 (2020).
13. Girshick, R., Donahue, J., Darrell, T. & Malik, J. Rich feature hierarchies for accurate object detection and semantic
segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 580–587 (2014).
14. Ju, R.-Y., Lin, T.-Y., Jian, J.-H., Chiang, J.-S. & Yang, W.-B. Threshnet: An efficient densenet using threshold mechanism
to reduce connections. IEEE Access 10, 82834–82843 (2022).
15. Ju, R.-Y., Lin, T.-Y., Jian, J.-H. & Chiang, J.-S. Efficient convolutional neural networks on raspberry pi for image
classification. J. Real-Time Image Process. 20, 21 (2023).
16. Gan, K. et al. Artificial intelligence detection of distal radius fractures: a comparison between the convolutional neural
network and professional assessments. Acta orthopaedica 90, 394–400 (2019).
17. Kim, D. & MacKinnon, T. Artificial intelligence in fracture detection: transfer learning from deep convolutional neural
networks. Clin. radiology 73, 439–445 (2018).
18. Lindsey, R. et al. Deep neural network improves fracture detection by clinicians. Proc. Natl. Acad. Sci. 115, 11591–11596
(2018).
19. Blüthgen, C. et al. Detection and localization of distal radius fractures: Deep learning system versus radiologists. Eur.
journal radiology 126, 108925 (2020).
20. Girshick, R. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, 1440–1448 (2015).
21. Ren, S., He, K., Girshick, R. & Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks.
Adv. neural information processing systems 28 (2015).
22. Cai, Z. & Vasconcelos, N. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference
on computer vision and pattern recognition, 6154–6162 (2018).
23. Lu, X., Li, B., Yue, Y., Li, Q. & Yan, J. Grid r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, 7363–7372 (2019).
24. Pang, J. et al. Libra r-cnn: Towards balanced learning for object detection. In Proceedings of the IEEE/CVF conference on
computer vision and pattern recognition, 821–830 (2019).
25. Zhang, H., Chang, H., Ma, B., Wang, N. & Chen, X. Dynamic r-cnn: Towards high quality object detection via dynamic
training. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings,
Part XV 16, 260–275 (Springer, 2020).

12/15
26. He, K., Gkioxari, G., Dollár, P. & Girshick, R. Mask r-cnn. In Proceedings of the IEEE international conference on
computer vision, 2961–2969 (2017).
27. Liu, W. et al. Ssd: Single shot multibox detector. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam,
The Netherlands, October 11–14, 2016, Proceedings, Part I 14, 21–37 (Springer, 2016).
28. Lin, T.-Y., Goyal, P., Girshick, R., He, K. & Dollár, P. Focal loss for dense object detection. In Proceedings of the IEEE
international conference on computer vision, 2980–2988 (2017).
29. Law, H. & Deng, J. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on
computer vision (ECCV), 734–750 (2018).
30. Duan, K. et al. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference
on computer vision, 6569–6578 (2019).
31. Dong, Z. et al. Centripetalnet: Pursuing high-quality keypoint pairs for object detection. In Proceedings of the IEEE/CVF
conference on computer vision and pattern recognition, 10519–10528 (2020).
32. Wang, C.-Y., Bochkovskiy, A. & Liao, H.-Y. M. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time
object detectors. arXiv preprint arXiv:2207.02696 (2022).
33. Ge, Z., Liu, S., Wang, F., Li, Z. & Sun, J. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021).
34. Wang, C.-Y., Yeh, I.-H. & Liao, H.-Y. M. You only learn one representation: Unified network for multiple tasks. arXiv
preprint arXiv:2105.04206 (2021).
35. Ju, R.-Y., Chen, C.-C., Chiang, J.-S., Lin, Y.-S. & Chen, W.-H. Resolution enhancement processing on low quality images
using swin transformer based on interval dense connection strategy. Multimed. Tools Appl. 1–17 (2023).
36. Glenn, J. Ultralytics yolov8. https://github.com/ultralytics/ultralytics (2023).
37. Nagy, E., Janisch, M., Hržić, F., Sorantin, E. & Tschauner, S. A pediatric wrist trauma x-ray dataset (grazpedwri-dx) for
machine learning. Sci. Data 9, 222 (2022).
38. Guan, B., Yao, J., Zhang, G. & Wang, X. Thigh fracture detection using deep learning method based on new dilated
convolutional feature pyramid network. Pattern Recognit. Lett. 125, 521–526 (2019).
39. Wang, M. et al. Parallelnet: Multiple backbone network for detection tasks on thigh bone fracture. Multimed. Syst. 1–10
(2021).
40. Guan, B., Zhang, G., Yao, J., Wang, X. & Wang, M. Arm fracture detection in x-rays based on improved deep convolutional
neural network. Comput. & Electr. Eng. 81, 106530 (2020).
41. Rajpurkar, P. et al. Mura dataset: Towards radiologist-level abnormality detection in musculoskeletal radiographs. In
Medical Imaging with Deep Learning (2018).
42. Wu, H.-Z. et al. The feature ambiguity mitigate operator model helps improve bone fracture detection on x-ray radiograph.
Sci. Reports 11, 1–10 (2021).
43. Ma, Y. & Luo, Y. Bone fracture detection through the two-stage system of crack-sensitive convolutional neural network.
Informatics Medicine Unlocked 22, 100452 (2021).
44. Xue, L. et al. Detection and localization of hand fractures based on ga_faster r-cnn. Alex. Eng. J. 60, 4555–4562 (2021).
45. Sha, G., Wu, J. & Yu, B. Detection of spinal fracture lesions based on improved yolov2. In 2020 IEEE International
Conference on Artificial Intelligence and Computer Applications (ICAICA), 235–238 (IEEE, 2020).
46. Sha, G., Wu, J. & Yu, B. Detection of spinal fracture lesions based on improved faster-rcnn. In 2020 IEEE International
Conference on Artificial Intelligence and Information Systems (ICAIIS), 29–32 (IEEE, 2020).
47. Xie, S., Girshick, R., Dollár, P., Tu, Z. & He, K. Aggregated residual transformations for deep neural networks. In
Proceedings of the IEEE conference on computer vision and pattern recognition, 1492–1500 (2017).
48. Lin, T.-Y. et al. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision
and pattern recognition, 2117–2125 (2017).
49. Qi, Y. et al. Ground truth annotated femoral x-ray image dataset and object detection based method for fracture types
classification. IEEE Access 8, 189436–189444 (2020).
50. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference
on computer vision and pattern recognition, 770–778 (2016).

13/15
51. Redmon, J. & Farhadi, A. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision
and pattern recognition, 7263–7271 (2017).
52. Hržić, F., Tschauner, S., Sorantin, E. & Štajduhar, I. Fracture recognition in paediatric wrist radiographs: An object
detection approach. Mathematics 10, 2939 (2022).
53. Glenn, J. Ultralytics yolov5. https://github.com/ultralytics/yolov5 (2022).
54. Yuan, G., Liu, G., Wu, X. & Jiang, R. An improved yolov5 for skull fracture detection. In Exploration of Novel Intelligent
Optimization Algorithms: 12th International Symposium, ISICA 2021, Guangzhou, China, November 20–21, 2021, Revised
Selected Papers, 175–188 (Springer, 2022).
55. Warin, K. et al. Maxillofacial fracture detection and classification in computed tomography images using convolutional
neural network-based models. Sci. Reports 13, 3434 (2023).
56. Tsai, H.-C. et al. Automatic rib fracture detection and localization from frontal and oblique chest x-rays. In 2022 10th
International Conference on Orange Technology (ICOT), 1–4 (IEEE, 2022).
57. Burkow, J. et al. Avalanche decision schemes to improve pediatric rib fracture detection. In Medical Imaging 2022:
Computer-Aided Diagnosis, vol. 12033, 597–604 (SPIE, 2022).
58. Warin, K. et al. Assessment of deep convolutional neural network models for mandibular fracture detection in panoramic
radiographs. Int. J. Oral Maxillofac. Surg. 51, 1488–1494 (2022).
59. Fatima, J., Mohsan, M., Jameel, A., Akram, M. U. & Muzaffar Syed, A. Vertebrae localization and spine segmentation
on radiographic images for feature-based curvature classification for scoliosis. Concurr. Comput. Pract. Exp. 34, e7300
(2022).
60. Mushtaq, M., Akram, M. U., Alghamdi, N. S., Fatima, J. & Masood, R. F. Localization and edge-based segmentation of
lumbar spine vertebrae to identify the deformities using deep learning models. Sensors 22, 1547 (2022).
61. Wang, C.-Y. et al. Cspnet: A new backbone that can enhance learning capability of cnn. In Proceedings of the IEEE/CVF
conference on computer vision and pattern recognition workshops, 390–391 (2020).
62. He, K., Zhang, X., Ren, S. & Sun, J. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE
transactions on pattern analysis machine intelligence 37, 1904–1916 (2015).
63. Liu, S., Qi, L., Qin, H., Shi, J. & Jia, J. Path aggregation network for instance segmentation. In Proceedings of the IEEE
conference on computer vision and pattern recognition, 8759–8768 (2018).
64. Tian, Z., Shen, C., Chen, H. & He, T. Fcos: A simple and strong anchor-free object detector. IEEE Transactions on Pattern
Analysis Mach. Intell. 44, 1922–1933 (2020).
65. Feng, C., Zhong, Y., Gao, Y., Scott, M. R. & Huang, W. Tood: Task-aligned one-stage object detection. In 2021 IEEE/CVF
International Conference on Computer Vision (ICCV), 3490–3499 (IEEE Computer Society, 2021).
66. Li, X. et al. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Adv.
Neural Inf. Process. Syst. 33, 21002–21012 (2020).
67. Zheng, Z. et al. Enhancing geometric factors in model learning and inference for object detection and instance segmentation.
IEEE Transactions on Cybern. 52, 8574–8586 (2021).
68. Zheng, Z. et al. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI
conference on artificial intelligence, vol. 34, 12993–13000 (2020).
69. Boyd, K., Eng, K. H. & Page, C. D. Area under the precision-recall curve: point estimates and confidence intervals. In
Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2013, Prague, Czech
Republic, September 23-27, 2013, Proceedings, Part III 13, 451–466 (Springer, 2013).
70. Woo, S., Park, J., Lee, J.-Y. & Kweon, I. S. Cbam: Convolutional block attention module. In Proceedings of the European
conference on computer vision (ECCV), 3–19 (2018).
71. Liu, Y., Shao, Z. & Hoffmann, N. Global attention mechanism: Retain information to enhance channel-spatial interactions.
arXiv preprint arXiv:2112.05561 (2021).
72. Lin, T.-Y. et al. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference,
Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, 740–755 (Springer, 2014).
73. Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014).
74. Ruder, S. An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747 (2016).

14/15
Author contributions statement
R.J. conceived and conducted the experiments, and W.C. provided knowledge about the fracture and analyzed the results. All
authors have reviewed the manuscript.

Funding
The authors did not receive support from any organization for the submitted work.

Competing interests
The authors have no financial or proprietary interests in any material discussed in this article.

Ethics approval
This research does not involve human participants and/or animals.

Data availability
The datasets analysed during the current study are available at Figshare under https://doi.org/10.6084/m9.figshare.
14825193.v2. The implementation code and the trained model for this study can be found on GitHub at https:
//github.com/RuiyangJu/Bone_Fracture_Detection_YOLOv8.

15/15