Bone Fracture Classification Using Transfer

Learning
Shyam Gupta Dhanisha Sharma
Technicshe Universitat Dortmund B.Sc Honors Physics
Dortmund, Germany Devi Ahilya Vishwa Vidyalaya
[email protected] Indore, India
[email protected]

Abstract—The manual examination of X-ray images for frac- of automated DNN-based systems in improving diagnostic
arXiv:2406.15958v1 [eess.IV] 22 Jun 2024

tures is a time-consuming process that is prone to human error. accuracy and efficiency in bone fracture detection.
In this work, we introduce a robust yet simple training loop for In recent research, Yadav et al. (2022) proposed a novel
the classification of fractures, which significantly outperforms
existing methods. Our method achieves superior performance in hybrid model, the hybrid scale fracture network (SFNet),
less than ten epochs and utilizes the latest dataset to deliver which integrates convolutional neural networks (CNN) with
the best-performing model for this task. We emphasize the an improved canny edge algorithm for bone fracture diagnosis
importance of training deep learning models responsibly and [2]. This model efficiently identifies fractures by using multi-
efficiently, as well as the critical role of selecting high-quality scale feature fusion, achieving impressive training accuracy
datasets.
Index Terms—Efficient-net, gpu , medical images, classifica- of 99.12% which was very close to our results, respectively.
tion, pytorch, dataset The SFNet model’s innovation lies in its combination of grey
images and canny edge images for training, demonstrating
I. I NTRODUCTION superior performance compared to other deep CNN models,
thus advancing the capabilities in medical image diagnosis.
Bone fractures are a common medical condition that signifi-
The development of automated fracture detection systems
cantly impact patients’ health and quality of life. Accurate and
is crucial for enhancing computer-aided telemedicine. Ma and
timely diagnosis of fractures is crucial for effective treatment
Luo (2020) introduce Crack-Sensitive Convolutional Neural
and recovery. Traditionally, radiologists and orthopedic sur-
Network (CrackNet) within a two-stage system for fracture
geons have relied on X-ray imaging to detect fractures. How-
detection, demonstrating superior performance. Initially, Faster
ever, interpreting these images can be challenging, particularly
R-CNN identifies 20 bone regions in X-ray images, followed
when dealing with minute or complex fractures that are not
by CrackNet assessing each region for fractures. Tested on
easily visible. Misdiagnosis or delayed diagnosis can lead to
1052 images, including 526 fractured cases, their system
improper treatment, prolonged recovery times, and increased
achieved 90.11% accuracy and 90.14% F-measure, outper-
healthcare costs.
forming other two-stage systems. This approach significantly
II. P ROBLEM S TATEMENT aids hospitals lacking experienced surgeons by providing ac-
curate and efficient fracture diagnosis [10].
The problem is exacerbated in busy clinical settings, where
radiologists may face a high volume of images to review. IV. E ASE OF U SE
Additionally, subtle fractures may be overlooked due to the
A. Enhanced Diagnostic Accuracy:
limitations of human perception. Therefore, there is a critical
need for automated methods that can assist healthcare pro- The accuracy of human radiograph experts in classifying
fessionals in accurately identifying fractures in X-ray images, bone fractures varies across studies, but it typically ranges
thereby improving diagnostic accuracy and efficiency. from 77.5% to 93.5%. Our proposed algorithm significantly
improves the accuracy of fracture detection, reducing the
III. R ELATED W ORKS likelihood of missed fractures, particularly those that are
The study by [7] presents a deep neural network (DNN) minute or complex. By leveraging EfficientNet’s sophisticated
model for classifying fractured and healthy bones, addressing architecture, the model can identify subtle patterns and features
the limitations of manual X-ray diagnosis, which is time- in X-ray images that may be challenging for the human eye
consuming and error-prone. To counteract overfitting on small to detect.
datasets, data augmentation techniques were employed, sig-
nificantly enhancing model performance. The model achieved B. Time Efficiency:
a classification accuracy of 92.44% using 5-fold cross- Automated fracture classification allows for quicker analysis
validation, surpassing previous methods with accuracies of of X-ray images, enabling doctors to focus on diagnosis and
84.7% and 86%. This advancement underscores the potential treatment rather than manual image review. In emergency
settings, rapid identification of fractures can be crucial for A. Dataset - Fracatlas
timely medical intervention. The Frac-Atlas dataset is an extensive and meticulously
curated collection of X-ray images specifically designed for
C. Consistency and Reliability: the detection of bone fractures. Unlike other existing fracture
datasets, Frac-Atlas encompasses a wide variety of fracture
The algorithm provides consistent results, minimizing the types, including both common and rare fractures, as well as
variability inherent in human interpretation of X-ray images. a range of complexities from clear, easily visible fractures to
This consistency ensures that every patient receives the same subtle, minute ones. This diversity ensures that the dataset
high standard of diagnostic care. provides a comprehensive training ground for developing
robust fracture detection algorithms. Furthermore, Frac-Atlas
V. E VALUATION M ETRICS includes high-resolution images with annotated fracture lo-
cations, making it an invaluable resource for training deep
A. Precision learning models. The dataset also includes metadata such
as patient demographics, fracture types, and severity, which
can be used for more detailed analysis and model training.
TP This level of detail and variety makes Frac-Atlas a superior
Precision = (1)
TP + FP dataset for advancing the state of fracture detection technology,
providing a rich foundation for both academic research and
B. Recall practical application in clinical settings. [15]
Below are some techniques we tried to use on our dataset
TP to solve and explore the problem, we used such techniques
Recall = (2) which can help model learn differentiation between classes,
TP + FN
based on contract, edges, color diffusion, size scaling,etc. . . .
C. F1 Score
B. CLAHE Transform (Contrast Limited Adaptive Histogram
Equalization)
Precision · Recall CLAHE enhances X-ray image contrast by applying adap-
F1 = 2 · (3)
Precision + Recall tive histogram equalization on small regions, preventing noise
amplification. This technique improves the visibility of subtle
D. AUC ROC features like hairline fractures by redistributing pixel intensity
The AUC (Area Under the Curve) ROC (Receiver Operating values, making fractures more prominent and easier to detect
Characteristic) is computed from the ROC curve, which plots for both human eyes and algorithms.
the true positive rate (TPR) against the false positive rate
C. Anisotropic Diffusion Transform
(FPR) at various threshold settings.
Anisotropic Diffusion reduces image noise while preserving
important edges. By controlling diffusion to smooth homo-
E. Accuracy geneous regions more than edges, this technique enhances
fracture clarity without blurring, making fractures more dis-
TP + TN tinguishable and maintaining diagnostic quality.
Accuracy = (4)
TP + TN + FP + FN D. Canny Edge Detection
Canny Edge Detection identifies boundaries within X-ray
images by detecting sharp intensity changes. It involves noise
reduction, gradient calculation, non-maximum suppression,
VI. O BJECTIVE & M ETHODOLOGY and edge tracking to highlight bone structures and potential
fracture lines, making it easier to identify fractures in complex
This paper aims to address the challenge of bone fracture images.
classification by leveraging deep learning techniques, specifi-
cally utilizing EfficientNet for transfer learning. EfficientNet, E. Image Segmentation Transform
known for its balance between accuracy and computational Image Segmentation divides an image into multiple seg-
efficiency, presents an ideal candidate for developing a robust ments, isolating regions of interest like specific bones or
fracture detection model. The goal is to create a system that suspected fracture areas. This focused analysis reduces com-
can automatically classify X-ray images to determine the plexity and improves fracture detection accuracy, highlighting
presence or absence of fractures with high accuracy, aiding high-probability fracture areas and enhancing overall model
doctors in making quicker and more accurate diagnoses. performance.
 (a) Training Curve

(b) Confusion Matrix

(a) Ineffective Transformations features without additional noise or artifacts. This streamlined
approach proved to be highly effective.
Fig. 1: We identified using transformations did not help us No center cropping! while cropping the images the fractures
identify fractures neither helped us to get good results. Thus, located at the corners went out of the convolution window. As
making our training loop simpler and easier to learn/generalize for computation, we used Tesla-P100 GPU for training, freely
on raw xray-images available on Kaggle. Our model, built using EfficientNet-B6
with pretrained weights, was trained with a batch size of 16,
with CrossEntropyLoss, optimizer as Adam and scheduler as
F. Random Central Cropping Cosine Annealing Warm Restart optimizer.
Random Central Cropping is a data augmentation technique A point to note is, while we trained on larger batch sizes,
that creates multiple variations of an X-ray image by cropping the accuracy was increasing slowly, so we figured that larger
the central region. This improves model robustness by expos- batch sizes lead to slower training improvements. This setup
ing it to different views, preventing overfitting, and enhancing facilitated efficient learning and convergence. The training pro-
generalization to new images, especially useful when labeled cess was remarkably quick, achieving an impressive training
data is limited. accuracy of 99%, with validation accuracy 97% on the training
set in just 7 epochs. The model also demonstrated robust
VII. T RAINING & R ESULTS performance on the testing set, achieving an Testing accuracy
of 96.83%.
Despite our thorough experimentation with various prepro-
cessing techniques, including CLAHE Transform, Anisotropic TABLE I: Comparison of Bone Fracture Classification Models
Diffusion Transform, Canny Edge Detection, Image Segmenta-
tion Transform, and Random Central Cropping, these methods Authors Publication Year Testing Accuracy (%)
Muhammet Emin Sahin 2023 86
did not significantly enhance the model’s ability to detect Pongsakorn Samothai et al. 2022 88
fractures. While each technique offered potential advantages D. P. Yadav et al. 2020 95
for highlighting fractures, they also introduced complexity and Basha et al. 2020 92
noise that ultimately did not improve detection accuracy. Rao, L.J. et al. 2020 90
Our method 2024 96.83
Instead, we found that normalizing the X-ray images alone
provided the best results. Normalization adjusted the pixel
intensity values to a consistent scale, simplifying the image As for some post training validation figures, we achieved
data and enhancing the model’s ability to learn relevant Test Precision: 0.9770, Test Recall: 0.9606, Test F1 Score:
 augmentation. Our approach demonstrates the high quality of
the ready-to-train images available in the FracAtlas dataset.
By emphasizing the initial step of the data science pipeline,
namely data, we achieved remarkable results. The simplicity
of our model, combined with the superior quality of the
dataset, enabled us to achieve over 99% training accuracy in
just seven epochs and a testing accuracy of 96.83
The FracAtlas dataset also provides bounding box coordi-
nates to detect the exact location of fractures. We propose
using the same network as a backbone to train a region lo-
calization algorithm for precise fracture detection. Addressing
the imbalance in the dataset remains a challenge that we must
tackle in future work.
R EFERENCES
[1] Meena, T., & Roy, S. (2022). Bone Fracture Detection Using Deep
Supervised Learning from Radiological Images: A Paradigm Shift.
Diagnostics, 12 doi-link.
(a) Classification of Fractured Bones [2] Yadav, D.P., Sharma, A., Athithan, S., Bhola, A., Sharma, B., & Dhaou,
I.B. (2022). Hybrid SFNet Model for Bone Fracture Detection and
Classification Using ML/DL. Sensors, 22. doi-link.
[3] Samothai, P., Sanguansat, P., Kheaksong, A., Srisomboon, K., & Lee, W.
(2022). The Evaluation of Bone Fracture Detection of YOLO Series. In
37th International Technical Conference on Circuits/Systems, Computers
and Communications (ITC-CSCC), pp. 1054-1057. doi-link.
[4] Santhiya, J., & Ebenezer, M.P. (2022). Bone Fracture Detection Using
Python. [Online]. Available: doi-link.
[5] Santos, K.C., Fernandes, C.A., & Costa, J.R. (2022). Feasibility of Bone
Fracture Detection Using Microwave Imaging. IEEE Open Journal of
Antennas and Propagation, 3, 836-847. doi-link.
[6] AlGhaithi, A., & al Maskari, S. (2021). Artificial intelligence application
in bone fracture detection. Journal of Musculoskeletal Surgery and
Research, 5, 4-9. doi-link.
[7] Yadav, D.P., & Rathor, S. (2020). Bone Fracture Detection and Clas-
sification using Deep Learning Approach. In International Conference
on Power Electronics & IoT Applications in Renewable Energy and its
Control (PARC), pp. 282-285. doi-link.
[8] Basha, C.Z., Reddy, M., Nikhil, K.H., Venkatesh, P.M., & Asish, A.V.
(2020). Enhanced Computer Aided Bone Fracture Detection Employing
X-ray Images by Harris Corner Technique. In Fourth International Con-
ference on Computing Methodologies and Communication (ICCMC), pp.
991-995. doi-link.
[9] Sahin, M.E. (2023). Image processing and machine learning-based bone
fracture detection and classification using X-ray images. International
Journal of Imaging Systems and Technology, 33, 853-865. doi-link.
[10] Ma, Y., & Luo, Y. (2020). Bone fracture detection through the two-stage
(b) Classification of Non-Fractured Bones system of Crack-Sensitive Convolutional Neural Network. Informatics
in Medicine Unlocked, 100452. doi-link.
Fig. 3: Results From Deployed Model [11] Zhang, X., Wang, Y., Cheng, C., Lu, L., Xiao, J., Liao, C., & Miao, S.
(2020). A New Window Loss Function for Bone Fracture Detection and
Localization in X-ray Images with Point-based Annotation. doi-link.
0.9686, Test AUC ROC: 0.9606. Which completely outper- [12] Rao, L.J., Neelakanteswar, P., Ramkumar, M., Krishna, A., & Basha,
C.Z. (2020). An Effective Bone Fracture Detection using Bag-of-Visual-
forms, existing methods of fracture detection, and provides a Words with the Features Extracted from SIFT. In International Confer-
more robust solution. ence on Electronics and Sustainable Communication Systems (ICESC),
These results underscore the effectiveness of our approach, pp. 6-10. doi-link.
[13] Yang, A.Y., & Cheng, L. (2019). Long-Bone Fracture Detection using
highlighting the power of EfficientNet-B6 and the importance Artificial Neural Networks based on Line Features of X-ray Images.
of choosing appropriate preprocessing techniques. The high In IEEE Symposium Series on Computational Intelligence (SSCI), pp.
accuracy achieved in both training and testing phases indicates 2595-2602. doi-link.
[14] Basha, C.Z., Padmaja, T.M., & Balaji, G.N. (2019). An Effective and
the model’s robustness and potential for practical application Reliable Computer Automated Technique for Bone Fracture Detection.
in clinical settings, significantly aiding in the accurate and EAI Endorsed Transactions on Pervasive Health and Technology, 5(18).
efficient diagnosis of bone fractures. doi-link.
[15] Fracatlas dataset dataset.
[16] Kaggle link
VIII. C ONCLUSION AND F UTURE SCOPES
We developed a simple yet effective method for detecting
fractures by leveraging transfer learning and the power of data