Abstract
Bone fracture is a very common medical issue faced
by medical practitioners, and fracture detection is quite crucial
for treatment. Till now, the most popular diagnosing method
for bone fracture is X-ray imaging. However, in many cases,
an X-ray image of a fractured bone is prone to human error.
Advancements in artificial intelligence (AI) have demonstrated
the potential to achieve high accuracy in tasks like X-ray
image classification of fractured bone. In this work, we have
incorporated a deep learning-based approach to facilitate the
diagnosis of bone fracture. We have utilized different pre-trained
state-of-the-art (SOTA) convolutional neural networks (CNN),
namely, DenseNet, VGG, ResNet, Xception, and EfficientNet
as feature extractors, to extract higher-level feature space representations of X-ray images. We have incorporated multiple
classifiers, namely Logistic Regression, Random Forest, Linear
Regression, XGBoost, and a custom feed-forward network (FFN)
to discriminate among the higher-level representations extracted
by the aforementioned pre-trained feature extractors. Among
the several feature extractor and classifier combinations we
have experimented with, DenseNet169 combined with the custom
FFN produced the best results, reporting an overall accuracy
of 99.48%, Receiver Operating Characteristic Area Under the
Curve (ROC-AUC) score of 99.99%, precision of 99.48%, recall
of 99.48%, and F1 score of 99.48%. The results demonstrate that
combining an appropriate pretrained SOTA CNN model and a
classifier can achieve high classification accuracy in the bonefracture X-ray image classification task. Such a method previews
a promising avenue in helping medical practitioners to achieve
better diagnostic performance with patient benefit.
Index Terms—Bone Fracture, Deep Learning, Transfer Learning, X-ray Image Classification
