Deep Learning based Bone Fracture Prediction using

Convolutional Neural Networks: A Comparative

Study of Transfer Learning and Fine-tuning
2023 1st International Conference on Optimization Techniques for Learning (ICOTL) | 979-8-3503-2804-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICOTL59758.2023.10435047

Techniques
1st Sriram R 2nd O R Aruna
School of Computer Application Department of CSIT
Lovely Professional University Jain Deemed-to-be University
Phagwara, India Bangalore, India
[email protected] [email protected]

Abstract—Computer assisted bone fracture detection have developed an integrated approach with the hardware
system has a vital role in medical automated health- and software design to extract the requirements from an
care system. These automation system helps healthcare image within the reasonable time. The system includes an
units to detect the bone fracture accurately. In this
proposed paper, the CNN also called as convolutional image processor which is parallelly pipe lined, referred as
neural network has been proposed for the classification cyto-computer. Fracture being the most known crack or
purpose of the images that is classified into fractured commonly called as bone break, is seen across the globe
and not_fractured. These images would be used to and been faced by both youths and elders which might take
train the model to classify whether its positive or usually a month or two to heal up. The treatment for the
negative and to validate it. The proposed system has
also demonstrated to be sufficiently the robust one bone break varies according to the extremities of the bone
to retrieve the necessary information and perform the break. Using machine learning/deep learning techniques,
required analysis on important image regions. Several a researcher can detect or identify the prediction of bone
images have been considered from the dataset and fracture with an optimal range of accuracy.
validated without any hindrances. Here the evaluation
has been done by the performance measure of the
trained model with the collected dataset which contain II. Related Works
testing sets of both fractured and not_fractured images
around 600 as well as training sets of both fractured
and not_fractured around 8,863 images in total. The [3] The research describes the revolutionary generalized
accuracy rate achieved in this particular paper was over bone fracture detection system capable of detecting
86%. This system uses clinical dataset to predict the
fracture[1] and mainly focuses on hand, wrist, forearm, numerous bone structures and fracture types across the
shoulder, elbow, finger with both positive as well as body. The proposed method uses candidate patches from
negative studies. X-ray images and a discriminative learning framework
called Stacked Random Forests Feature Fusion, and
Index Terms—Machine Learning, Deep Learning, outperforms other fracture detection frameworks in terms
CNN, Neural, Network, Fracture, Detection, Model
of accuracy. Furthermore, the suggested technique catches
81.2 percent of radiologist-reported fracture results. The
I. Introduction approach described has the potential to help radiologists
in detecting bone fractures and improve patient outcomes.
Advancement in the field of computer science has bring
lot of improvements in the medical field to process the [4] The proposed paper uses the machine learning
information accurately. Image processing is one such field algorithms for diagnosing VCF including neural networks,
that has seen notable growth when identifying the images decision trees, SVM and RF. The proposed study
from various sources like computer tomography, digital ra- uses support vector machine and K nearest neighbor
diography images and scanned or clinical images, etc., For as classifiers and overall achievement or accuracy has
example [2] the system engineers from the proposed paper, been recorded upto 88.3% and 92.5% sensitivity in
diagnosing VCF. Additionally, the study achieves an
accurate diagnosis of 86.2% in detecting an abnormal
979-8-3503-2804-2/23/$31.00 ©2023 IEEE location on 25 CT scans clinical images from the

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:15:27 UTC from IEEE Xplore. Restrictions apply.
 spinal column. The authors have also concluded neural networks for solving image-based problems
saying that further research is needed to develop in nusculoskeletal radiology. In all four categories,
and validate accurate methods for diagnosing VCF. applications of DL can be classified as lesion detecting,
classification, segmentation and not interpreting tasks,
[5] This paper discusses the difficulties in predicting hip expert level results have been shown for each category.
fractures in post-menopausal women with osteoporosis due
to various risk factors. The proposed paper aims to create [11] The paper provides the use of TA combined
an artificial neural network to predict hip fracture using with ML algorithms to identify patients at risk
data from longitudinal study comprising 1167 women for insufficiency fractures in standard CT scans
aged 60 and up. The information has mineral density in with high accuracy. However, they also concluded
bone, age gaps, factors related clinically, factors involving that the further research is needed to improve
in lifestyles etc., The accuracy achieved was 87 percent the identification of single vertebrae at risk.
and an AUC of over 0.97 compared to previous statistics.
[12] Here the paper supports the use of DL with an
[6] This paper presents a deep learning with the advanced data augmentation techniques for accurately
help of CNN for the prediction of average strain as an diagnosing and classifying femoral neck fractures.
alternative to physics based approaches for estimating The overall accuracy was over 86% with the use
bio mechanical quantities of interest in the context of of two class prediction namely DRR and GAN.
osteoporosis. The model is trained on a large database
of synthetically generated cancellous bone anatomies [13] This paper deals involves combining multiple
and the average execution time for the CNN model scales of features from convolutional neural
was reduced by more than 1000 times compared to network with an enhanced canny edge algorithm
the physics-based computation, making it possible for fusion. The results indicate that the SFNet
to perform real-time assessment of average strain. with canny is more than 99.12 percent accurate.

[7] The literature review discusses the implementation [14] Here the paper deals with the automation provided by
of a stress prediction system for college students using CNN and CNN trained models. Active learning minimised
machine learning techniques such as KNN and NB. The class imbalance and resulted in a useful picture classifiers.
study shows that Naive Bayes is highly efficient in the
proposed system than KNN and has a high efficacy rate. [15] The article presents fundamental technical
The proposed system also explains the overall process information about deep learning through CNN and
with proper tables and implementation. Overall, the also highlights potential difficulties associated with this
literature review highlights the importance of predicting technique and provides guidance on how to handle them.
stress levels in college students and how machine
learning techniques can be used to improve the process. [16] The paper thoroughly examined on the
performance of locating fractures in bone X-ray
[8] The proposed paper maps the complicated 3D images using YOLO-X and YOLO-R. Using these
fracture lines of intertrochanteric fractures in a 2D the higher accuracy have been seen and predicted.
parametric space. The study involved 100 patients
based on original radiographs acquired by CT scan. [17] The authors have suggested the usage of
The paper highlights the importance of understanding transfer learning for the purpose of fracture
intertrochanteric fracture distribution in orthopedics identification and classification using RPN and
due to its high morbidity and mortality. Also, the have secured over 94 percentage of accuracy.
proposed method offers a promising solution to aid
in clinical diagnosis, and prognosis applications. [18] This paper’s goal was to use an X-ray
image to classify whether or not the image has a
[9]In Ireland, the use of medical logic techniques fracture using deep learning methods with CNN
to facilitate elderly care plans is described in this infusion. The paper has shown over 77.52 percent
report with a focus on predicting how long hip accuracy with loss percent value over 0.587.
fractures will stay and where they’ll be discharged.
The models have been trained with Microsoft’s machine [19] The authors have trained a faster CNN and
learning studio, which has enabled them o publish with over 96 percent accuracy, a machine vision neural
as precipient web services on the cloud platform. network for object detection may be used to find and
diagnose distal radius fractures in anteroposterior X-ray
[10] The proposed literature supports the rapid images.
advancements of deep learning with convolutional

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:15:27 UTC from IEEE Xplore. Restrictions apply.
 III. Proposed Method for an autonomous detection. Each classified set made
the proposed system to take very particular images for
A. Architecture of the Proposed System detection to see how well the outcome occurs. There were
few more dataset which were collected as a secondary yet
a primary one with classification of fractured as well as
not_fractured with around 10,000 images in total from a
distinguished websites. Hence, collection of similar dataset
was always a tedious task and matching those to the
proposed system takes much time.

D. Transfer Learning and Fine Tuning

In deep learning, the transfer learning plays an impor-

tant role as the fine tuning technique provides the network
layer for performing the further more tasks in order to
achieve the output layer. Other such methods are feature
extraction, fine tuning/transfer learning with generative
adaptive networks and meta learning can also be studied
and implemented.

IV. Results and Discussion

This proposed system uses csv file with dataset that

has been described above and also one can give a specified
Figure 1. Proposed Architecture pathway to the code to make the compilation more easier.
As the code runs, one could get the prediction done.
B. Methods and Methodologies
A. Use of Tensorflow
At first, the dataset collection can be done and the data
gathering must be of a diverse one with both fractured and This proposed system uses tensorflow as a library in
not_fractured images. After the preprocessing step which python and also supports machine learning based projects
converts and shapes the raw images, we split the data or systems especially in the field of image or numerical
into trained set and valid set as from the collection. It is predictions and analysis.
normal to have the trained set as high upto 60 or 70 pecent
with remaining as validation sets. In a validation stage,
the images of valid one are compared with the trained B. Epoch in Deep Learning
set. Further said libraries like tensorflow from python used
for model evaluation and analysis. The detailed process of Generally, an epoch is something to deal with a count
collection of the dataset has been given below. or an entire pass through the training dataset which we
consider for the proposed system. Considering one such
image which is need to be trained and the output is to be
C. Dataset Collection
predicted. Generally by considering deep learning stages,
an image must have input layer, hidden layer and finally
Collecting N number of dataset for the said prediction an output layer. So one epoch is nothing but traversing
of bone fracture can be very much time taking task as throughout the layers and bringing up with many desired
the dataset collection varies from one end to another end values.
in many different forms. As in the collection process, the
author need to stick on to the background of collecting
the images if the proposed is for mammals, or others C. Results
without confusing with the concepts. At first, considering
the paper called "MURA: Large Dataset for Abnormality The epoch took 50 on 50 and validated the loss and
Detection in Musculoskeletal Radiographs [1] " the dataset accuracy where the loss occurred was around 0.2930
has been collected. The dataset also includes large number which rounds up to 29 percentage. The accuracy was
of collection of images of fracture in bone over 40,000 from higher which meets up the requirement as the number
more than 14,000 studies. This has made the path easier goes around 0.8769 or of 89 percentage. Hence, by using

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:15:27 UTC from IEEE Xplore. Restrictions apply.
 deep learning method, we’ve achieved over 89 percent of
accuracy traversing throughout the processes of image
processing. The sample images given below were used to
implement and the images that were allowed to run the
proposed system focuses more on the positive side as the
accuracy rate was high in total. The model can also include
with much number of dataset from various points in which
the implementation can happen in more easier way with-
out any confusions. As the results were also divided in the
proper set of table below as metrics and values, this allows
the further researchers to even pay much attentions on the
future research. The basic system requirement would be
using windows as an operating system and python and
Figure 4. This figure[1] portrays the negative set of an hand X-ray
libraries for running and developing models, as well as
Pycharm IDE has been used for implementation purposes.

D. Dataset Samples

The dataset that the proposed system considered has

both training sets and validation sets. Taking the images
which are classified into positive and negative, the predic-
tion analysis happens. Here are some images given below
which classifies the dataset[1] into negative set and positive
set of the collected fracture images from various parts of
the hands.
Figure 5. This figure[1] portrays the positive set of an shoulder X-ray

Figure 2. This figure[1] portrays the negative set of an elbow X-ray

Figure 6. This figure[1] portrays the negative set of an wrist X-ray

Figure 3. This figure[1] portrays the negative set of an finger X-ray

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:15:27 UTC from IEEE Xplore. Restrictions apply.
 Statistic 1st Feature 2nd Feature 3rd Feature 4th Feature
Mean 2.57 13.24 0.53 7.09
Median 2.50 13.05 0.55 7.00
Std Dev 0.83 2.13 0.20 2.35
Min 1.00 10.00 0.20 3.00
Max 4.00 20.00 0.80 12.00

Table II
Descriptive Statistics

G. Confusion Matrix

The below mentioned table presents the confusion ma-

trix of the trained machine learning model. This includes
the actual labels and predicted labels for each class. The
Figure 7. This figure[1] portrays the positive set of an wrist X-ray
table also includes the total number of observations in
each class, including number that has true negatives and
positives with false negatives and positives.
E. Model Evaluation Metrics
Predicted: Negative Predicted: Positive
This table presents the evaluation metrics of the trained Actual: Negative 200 40
machine learning model. The metrics has such as F1-score, Actual: Positive 35 225
recall, accuracy, precision. Also, these metrics are com-
monly used to evaluate the performance of a classification Table III
model. This table also includes the values of these metrics Descriptive Statistics
class 0 and 1 for every available class.

S. No Metric Value
I Recall 0.87 V. Conclusion and Future Scope

II Accuracy 0.86 Fracture detection might be a challenging task as not

everyone goes through this. A normal bone break or
III F1-Score 0.86 dislocation might take up to 4 weeks of healing if a person
IV Precision 0.85 concentrates much on his health as the doctor would
advice. But if the condition is worse, one would need to
V AUC ROC Score 0.93 make sure that a POP is attached to take it over 6 weeks
of time and on some rare case scenario, a fracture or major
Table I bone break could take over 6 months span in an individual
Results of model evaluation on test set
for an heal. There are various dataset which can be easily
availed for fracture and real time survey is also appreciated
with a proper authorization and precautions. In future, a
research can be done on this same field, observing closely
about much severe bone breaks and could come up with
F. Descriptive Statistics of the dataset
a real time application including various algorithms in
order to predict or detect a fracture. Also, not only human
The below demonstrated table provides the descriptive
fracture but even with animals can be encouraged. That
statistics of the dataset used to train the machine learning
could help a veterinary doctors to come up with new ideas
model. This includes various statistics measures such as
of curing in a very short time.
mean, standard deviation (SD), min, max and quartiles
for each feature. These statistics provide insights into the
distribution and variability of the data, and can help in A. Contribution towards future research
understanding the characteristics of the dataset.
• This paper can be a key for future researches with
the same research area but with different dataset and
models perhaps

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:15:27 UTC from IEEE Xplore. Restrictions apply.
 • This paper enhances the usage of clinical dataset for future directions of deep learning in musculoskeletal
the whole fracture detection which also can ensure the radiology,” Skeletal radiology, vol. 49, no. 2, pp. 183–
accuracy of bone fracture prediction using CNN 197, 2020.
• In future, a research can be done in the field of frac- [11] U. J. Muehlematter, M. Mannil, A. S. Becker, K. N.
ture prediction using hybrid models which enhances Vokinger, T. Finkenstaedt, G. Osterhoff, M. A. Fis-
the accuracy level as well as other metrics cher, and R. Guggenberger, “Vertebral body insuf-
ficiency fractures: detection of vertebrae at risk on
standard ct images using texture analysis and ma-
References chine learning,” European radiology, vol. 29, pp. 2207–
2217, 2019.
[1] P. Rajpurkar, J. Irvin, A. Bagul, D. Ding, T. Duan, [12] S. Mutasa, S. Varada, A. Goel, T. T. Wong, and M. J.
H. Mehta, B. Yang, K. Zhu, D. Laird, R. L. Ball Rasiej, “Advanced deep learning techniques applied
et al., “Mura: Large dataset for abnormality detec- to automated femoral neck fracture detection and
tion in musculoskeletal radiographs,” arXiv preprint classification,” Journal of Digital Imaging, vol. 33, pp.
arXiv:1712.06957, 2017. 1209–1217, 2020.
[2] Sternberg, “Biomedical image processing,” Computer, [13] D. P. Yadav, A. Sharma, S. Athithan, A. Bhola,
vol. 16, no. 1, pp. 22–34, 1983. B. Sharma, and I. B. Dhaou, “Hybrid sfnet model
[3] Y. Cao, H. Wang, M. Moradi, P. Prasanna, and T. F. for bone fracture detection and classification using
Syeda-Mahmood, “Fracture detection in x-ray images ml/dl,” Sensors, vol. 22, no. 15, p. 5823, 2022.
through stacked random forests feature fusion,” in [14] B. A. Monchka, J. T. Schousboe, M. J. Davidson,
2015 IEEE 12th international symposium on biomed- D. Kimelman, D. Hans, P. Raina, and W. D. Leslie,
ical imaging (ISBI). IEEE, 2015, pp. 801–805. “Development of a manufacturer-independent convo-
[4] H. Yousefi, E. Salehi, O. S. Sheyjani, and lutional neural network for the automated identifica-
H. Ghanaatti, “Lumbar spine vertebral compression tion of vertebral compression fractures in vertebral
fracture case diagnosis using machine learning fracture assessment images using active learning,”
methods on ct images,” in 2019 4th International Bone, vol. 161, p. 116427, 2022.
Conference on Pattern Recognition and Image [15] K. Yasaka, H. Akai, A. Kunimatsu, S. Kiryu, and
Analysis (IPRIA). IEEE, 2019, pp. 179–184. O. Abe, “Deep learning with convolutional neural
[5] T. P. Ho-Le, J. R. Center, J. A. Eisman, T. V. network in radiology,” Japanese journal of radiology,
Nguyen, and H. T. Nguyen, “Prediction of hip frac- vol. 36, pp. 257–272, 2018.
ture in post-menopausal women using artificial neu- [16] P. Samothai, P. Sanguansat, A. Kheaksong,
ral network approach,” in 2017 39th Annual In- K. Srisomboon, and W. Lee, “The evaluation
ternational Conference of the IEEE Engineering in of bone fracture detection of yolo series,” in
Medicine and Biology Society (EMBC). IEEE, 2017, 2022 37th International Technical Conference on
pp. 4207–4210. Circuits/Systems, Computers and Communications
[6] C. F. Ciuşdel, A. Vizitiu, F. Moldoveanu, C. Suciu, (ITC-CSCC). IEEE, 2022, pp. 1054–1057.
and L. M. Itu, “Towards deep learning based esti- [17] W. Abbas, S. M. Adnan, M. A. Javid, F. Majeed,
mation of fracture risk in osteoporosis patients,” in T. Ahsan, S. S. Hassan et al., “Lower leg bone
2017 40th International Conference on Telecommuni- fracture detection and classification using faster rcnn
cations and Signal Processing (TSP). IEEE, 2017, for x-rays images,” in 2020 IEEE 23rd International
pp. 659–665. Multitopic Conference (INMIC). IEEE, 2020, pp.
[7] S. Sinha and R. Sriram, “An educational based intel- 1–6.
ligent student stress prediction using ml,” in 2022 3rd [18] N. Singh, G. K. Rajput, P. Mittal, and P. Madan,
International Conference for Emerging Technology “Analytical comparison of deep learning algorithms
(INCET). IEEE, 2022, pp. 1–7. for fracture detection,” in 2022 11th International
[8] Y. Fu, R. Liu, Y. Liu, and J. Lu, “Bone fracture Conference on System Modeling & Advancement in
visualization and analysis using map projection and Research Trends (SMART). IEEE, 2022, pp. 1231–
machine learning techniques,” in Smart Health: Inter- 1238.
national Conference, ICSH 2018, Wuhan, China, July [19] E. Yahalomi, M. Chernofsky, and M. Werman, “De-
1–3, 2018, Proceedings 6. Springer, 2018, pp. 53–58. tection of distal radius fractures trained by a small
[9] M. Elbattah and O. Molloy, “Using machine learning set of x-ray images and faster r-cnn,” in Intelligent
to predict length of stay and discharge destination for Computing: Proceedings of the 2019 Computing Con-
hip-fracture patients,” in Proceedings of SAI Intelli- ference, Volume 1. Springer, 2019, pp. 971–981.
gent Systems Conference (IntelliSys) 2016: Volume 1.
Springer, 2018, pp. 207–217.
[10] P. Chea and J. C. Mandell, “Current applications and

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on February 15,2025 at 10:15:27 UTC from IEEE Xplore. Restrictions apply.