Femur Fracture Detection Based on Deep Learning Model YOLOv8

Dongru Xie, Hongjian Yu

, Zhijiang Du, Hao Wang, Xiangyu Shen and Zhenyi Wang

State Key Laboratory of Robotics and System, Harbin Institute of Technology, Harbin, China

Keywords: Femur Fracture, X-Ray Images, Object Detection, Image Pre-Process, Deep Learning.

Abstract: Femur fracture occurs in various circumstances like car accidents, high-altitude fall incidents, tumour illness,

and elderly falls. For better recognition and treatment, physicians need to search the X-ray images for

fracture detail. However, some X-ray images were unclear to diagnose, and some were taken from the side

position, which is difficult to detect the fracture. This study uses the YOLOv8 model to help physicians with

femur fracture detection by utilizing deep-learning models. The performance of YOLOv8 is 42.35% in

AP50:95, 84.24% in mAP50, and 25.45% in mAP75 on the private dataset is from Shenzhen University

General Hospital. The result shows that the YOLOv8 detection model is competitive and faster on the

personal femur fracture dataset than YOLOv3 and YOLOv5 models.

INTRODUCTION

Bone fractures are regular in hospitals due to car

accidents, high-altitude fall incidents, tumor illness,

and elderly falls. Physicians use medical images, for

example, X-ray images, to search for the detail of

the fracture. However, due to the angle of images

taken can vastly alter the fracture info, the deep

learning model YOLOv8 is used to detect femur

fracture.

YOLOv8 is a multi-scale detection model. It uses

three detection heads to classify. According to the

basic of the deep learning model, YOLOv8 can learn

simple features like straight lines and oblique lines

at a low level. At a high level, it can learn more

complex features like the femur, femur shaft, and

femoral head. Figure 1 shows the anatomy of the

femur. The fracture can occur at the femoral shaft,

femoral distal, and femoral proximal.

Femoral proximal mainly include the femoral

neck, intertrochanteric, and subtrochanteric femoral

fractures. Femoral neck fractures occur

predominantly in the elderly, typically resulting

from low-energy falls, and may be associated with

osteoporosis.

Femoral shaft fractures are among the most

common fractures seen in orthopedic practice. The

femur is the most prominent bone in the body and

one of the primary weight-bearing bones of the

lower limbs, and unless treated appropriately,

fractures can lead to long-term morbidity and

disability.

Distal Femur often is unstable and comminuted

and tends to have a bimodal distribution, occurring

in elderly or younger multiple-injured patients.

Figure 1: The anatomy of the femur.

This research used the YOLOv8 model to detect

the fracture for better detection. This deep learning

model has three detection heads, which are 8x, 16x,

and 32x. Different scales can detect different scales

of objects. For instance, if 8x can detect the femur,

then 16x can detect the approximate location where

190

Xie, D., Yu, H., Du, Z., Wang, H., Shen, X. and Wang, Z.

Femur Fracture Detection Based on Deep Learning Model YOLOv8.

DOI: 10.5220/0012277300003807

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Seminar on Artiﬁcial Intelligence, Networking and Information Technology (ANIT 2023), pages 190-193

ISBN: 978-989-758-677-4

the femur is, and 32x can use the bounding box to

locate the fracture.

The accuracy score obtained from the fracture

detection performed by Rashid et al., using a 28-

layer dilated CNN and long short-term memory

(DCNN-LSTM) on 965 wrist X-ray images, is

88.24% (Rashid, 2023). The result of fracture

detection performed by Jia et al. on 1227 sternum

fracture X-ray images from the collection of sternal

radiographs and hospital diagnostic reports, 0.71

mAP, was obtained using the cascade R- CNN

method (Jia, 2022). The AP score of Guan et al. was

62.04% with a two-stage R-CNN method developed

for fracture detection based on nearly 4000 arm

fracture X-ray images using Resnet backbone.[Guan

B, 2020] Wang et al. carried out fracture detection

procedures(WrisNet), achieving a 56.6% score of

AP, using the model inspired by Faster-RCNN,

mainly composed of ResNeXt-TA and FPN for a

total of 4346 hairline fractures in hand X-rays

images.(WANG W.) Lu et al. developed automated

universal fractures detection in X-ray images using a

modified Ada-ResNeSt backbone network and the

AC-BiFPN detection method based on the part of

the MURA dataset. They achieved an AP score of

68.4% on 30000 X-ray images. (LU S, 2022) Guan

et al. achieved an AP score of 88.9% using a

balanced FPN-ResNeXt model developed for

fracture detection in a 3842 thighbone X-ray

radiographs dataset. (Guan B, 2022) Yadav et al.

used a deep learning model to detect and classify X-

ray images of human fracture bone and healthy bone.

5-fold cross-validation was implemented on 4000

augmented datasets and got 92.44 % accuracy for

the healthy and the fractured bone. (Xue L, 2021)

ParallelNet is proposed by Wang et al. for detection

tasks on thigh bone fracture based on multiple

backbone networks. The dataset contains 3842 X-

way radiographs; the result is 87.8% AP50 and

49.3% AP75. (WANG M, 2021) Chin et al.

proposed an Auxiliary Classifier Generative

Adversarial Network (AC-GAN) model to label the

position of the fracture. The result shows an

accuracy of 91.2% (Chiun-Li Chin, 2019).

ANOTHER SECTION OF YOUR

PAPER

YOLOv8 is the latest deep-learning model in the

YOLO series. The structure of this model is shown

in Figure 2. The detail of the model will be

explained in this section.

Downsample 4X

Downsample 16X

Downsample 8X

Downsample 32X

Downsample 16X

Down sample 8X

Downsample 16X

Downsample 32X

C2f & ConvModule

C2f & SPPF

Upsample & Concat

C2f & Upsample & Concat

C2f

2x ConvModule

Femur

Frac ture

C2f & ConvModule

Head1

ConvModule

Head2

Head 3

BACKBO NE BODY

Figure 2: The structure of YOLOv8 model.

2.1 Yolov8 Model

Like YOLOv5, YOLOv8 provides different size of

models based on the scaling factor to meet the needs

of different scenarios. The Backbone and Neck part

of the model refers to the YOLOv7 ELAN design

idea. The C3 structure from YOLOv5 has been

replaced with a richer C2f structure of gradient flow,

and the number of channels has been adjusted for

models of different scales. It is a fine-tuning of the

model structure. It is no longer a brainless set of

parameters to apply to all models, dramatically

improving the model's performance. Compared with

YOLOv5, the Head part has changed a lot. It has

been replaced with the current mainstream

decoupling head structure, which separates the

classification and detection heads. The

TaskAlignedAssigner positive sample allocation

strategy is adopted in the calculation, and the data

enhancement part of the Distribution Focal Loss

training is introduced. The Mosiac enhanced

operation can effectively improve the accuracy

introduced from YOLOX.

2.2 Study Dataset

The X-ray images of femur fracture include 312

fracture images collected from Shenzhen University

General Hospital between 2019 to 2023, ranging

from femoral neck fracture to distal femur fracture.

The physicians from Shenzhen University General

Hospital checked the data labeling of femur X-ray

images.

EXPERIMENTS

This study used local PC to train the deep-learning

model for femur fracture detection. The graphics

card of the local PC is 12 GB Nvidia GTX3060.

The following conﬁguration is used in all

machine learning models for femur fracture

detection: the epoch of data training is 300 times, the

Femur Fracture Detection Based on Deep Learning Model YOLOv8

191

initial learning rate is set as 0.01, and the final

learning rate is 0.01 of the initial learning rate.

Warmup epochs are 3. The parameters used to

optimizer weight decay are 5e-4, initial warmup

momentum is 0.8.

The bounding box outputs performed with the

YOLOv8 model in femur fracture X-ray images are

provided as a dataset sample in Figure 3 below.

Figure 3: Sample of Detection result of YOLOv8 model.

In order to explain the detection result of

YOLOv8, this research adds other deep learning

models to compare the result of YOLOv8.

YOLOv3 uses more profound and more accurate

Darknet-53 as the backbone and shifts from multi-

category to multi-label classification, removing

softmax and using binary cross entropy instead.

The network architecture of YOLOv5 consists of

three parts: CSPDarknet as the backbone, PANet as

the neck, and Yolo Layer Head. The data is first

input to CSPDarknet for feature extraction and then

input to PANet for feature fusion. Finally, the Yolo

layer outputs the detection results.

Based on the same parameters, the detection result

of three deep learning models is shown in the table

below.

Table 1: Detection result of YOLOv8, YOLOv5, and

YOLOv3 models.

Model mAP50-95 mAP50 mAP75

YOLOv8 0.4235 0.8424 0.2545

YOLOv5 0.3627 0.7578 0.3146

YOLOv3 0.4079 0.8116 0.3666

Table 1 shows that YOLOv8 has better overall

detection results than YOLOv5 and YOLOv3

Models on fracture detection. When the results

mentioned above are examined, it depicts that

YOLOv8 has an improvement in mAP50-95 and

mAP50 values. YOLOv8 can be used in femur

fracture detection in the private dataset.

CONCLUSION

This study aims to support physicians and intern

doctors in medical image detection and solve the

problem when X-ray images are deficient in clinical

needs and require physicians to retake X-rays of

femur fractures. In addition, the result shows that the

YOLOv8 model can detect femur fracture better

than other deep-learning models.

ACKNOWLEDGMENTS

This work was financially supported by Key-Area

Research and Development Program of Guangdong

Province (No.2020B0909020002) and Self-Planned

Task (No.SKLRS202211B) of State Key Laboratory

of Robotics and System (HIT).

REFERENCES

Rashid T, Zia M S, Najam UR R, et al. A Minority Class

Balanced Approach Using the DCNN-LSTM Method

to Detect Human Wrist Fracture [J]. Life-Basel, 2023,

13(1). https://doi.org/10.3390/life130 10133

Jia Y, Wang H, Chen W, et al. An attention-based cascade

R-CNN model for sternum fracture detection in X-ray

images [J]. CAAI Transactions on Intelligence

Technology, 2022, 7(4): 658-70. https://doi.org/

10.1049/cit2.12072

Guan B, Zhang G, Yao J, et al. Arm fracture detection in

X-rays based on improved deep convolutional neural

network [J]. Computers & Electrical Engineering,

2020, 81. https://doi.org/ 10.1016/j.compeleceng.2019

.106530

Wang W, Huang W, Lu Q, et al. Attention mechanism-

based deep learning method for hairline fracture

detection in hand X-rays [J]. Neural Computing &

ANIT 2023 - The International Seminar on Artiﬁcial Intelligence, Networking and Information Technology

192

Applications, 2022, 34(21): 18773-85. https://

doi.org/10.1007/s00521-022-07412-0

Lu S, Wang S, Wang G. Automated universal fractures

detection in X-ray images based on deep learning

approach [J]. Multimedia Tools and Applications,

2022, 81(30): 44487-503. https://doi.org/10.1007/

s11042-022-13287-z

Guan B, Yao J, Wang S, et al. Automatic detection and

localization of thighbone fractures in X-ray based on

improved deep learning method [J]. Computer Vision

and Image Understanding, 2022, 216.

https://doi.org/10.1016/j.cviu.2021.103345

Xue L, Yan W, Luo P, et al. Detection and localization of

hand fractures based on GA_Faster R-CNN [J].

Alexandria Engineering Journal, 2021, 60(5): 4555-

62. https://doi.org/10.1016/j.aej.2021.03.005

Wang M, Yao J, Zhang G, et al. ParallelNet: multiple

backbone network for detection tasks on thigh bone

fracture [J]. Multimedia Systems, 2021, 27(6): 1091-

100. https://doi.org/10.1007/s00530-021-00783-9

Chiun-Li Chin, Yong-Long Lin, Yu-Chieh Liu. Various

Types Fracture Labeling In Bone Radiographs Using

Modified AC-GAN[C], Proceedings of the

International Conference on Technologies and

Applications of Artficial Intelligence (TAAI),

Kaohsiung, TAIWAN, 2019 Nov. 21-23.

https://doi.org/10.1109/TAAI48200.2019.8959863

Femur Fracture Detection Based on Deep Learning Model YOLOv8

193