Automated fracture identification in X-Rays using deep neural networks

Al-Hussaini, Mohammed Abdulla Salim and Al-Musharfi, Hilal and Al Husaini, Yousuf Nasser and Abdulghafor, Rawad and Habaebi, Mohamed Hadi and Abrar, Mohammad (2026) Automated fracture identification in X-Rays using deep neural networks. In: 10th International Conference on Mechatronics Engineering (ICOM’26), 8-9 September 2026, Kuala Lumpur, Malaysia.

PDF - Published Version Restricted to Repository staff only Download (1MB) | Request a copy

Abstract

Accurate and timely fracture identification from X-ray images is essential for supporting clinical diagnosis and reducing missed fractures in emergency and orthopedic practice. Recent advances in deep learning have shown promising results for automated medical image analysis; however, identifying the most suitable architecture for fracture detection remains a significant challenge. This study presents a comprehensive comparative framework for automated fracture identification using the publicly available FracAtlas dataset, which contains 4,085 X-ray images, including 719 fractured and 3,366 non-fractured cases. Eight state-of-the-art deep learning models were evaluated, including ResNet50, DenseNet121, EfficientNet-B0, EfficientNet-B3, ConvNeXt-Tiny, ViT-B16, Swin-Tiny, and a proposed attention-based CNN. In addition, a Soft Voting Ensemble was developed to improve classification robustness by integrating the predictions of multiple models. The models were trained and evaluated using 3-fold crossvalidation, and their performance was assessed using Accuracy, Precision, Recall, Specificity, F1-score, ROC-AUC, PR-AUC, Matthews Correlation Coefficient (MCC), and Cohen's Kappa. The experimental results demonstrated that the Soft Voting Ensemble achieved the best overall performance with an accuracy of 89.03%, ROC-AUC of 0.909, PR-AUC of 0.780, MCC of 0.635, and Cohen's Kappa of 0.634, outperforming all individual models. Among the standalone architectures, EfficientNet-B3 achieved the highest accuracy (88.32%) and specificity (93.40%), while DenseNet121 obtained the highest recall (75.52%). Overall, CNN-based models consistently outperformed the evaluated transformer-based architectures in both convergence stability and classification performance. Future work will focus on validating the proposed framework on larger multi-center clinical datasets, integrating explainable artificial intelligence techniques, and developing lightweight models suitable for real-time clinical deployment.

Item Type:	Proceeding Paper (Plenary Papers)
Uncontrolled Keywords:	Bone Fracture Detection, X-ray Image Classification, Deep Neural Networks, Soft Voting Ensemble.
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7800 Electronics. Computer engineering. Computer hardware. Photoelectronic devices > TK7885 Computer engineering
Kulliyyahs/Centres/Divisions/Institutes (Can select more than one option. Press CONTROL button):	Kulliyyah of Engineering > Department of Electrical and Computer Engineering
Depositing User:	Dr. Mohamed Hadi Habaebi
Date Deposited:	10 Nov 2026 11:08
Last Update:	10 Nov 2026 11:30
Queue Number:	2026-10-Q6054
URI:	http://irep.iium.edu.my/id/eprint/132481
Indexed In:	Google Scholar

Actions (login required)

View Item