Context-Aware Hybrid Attention-Enhanced YOLO Framework for Accurate Multi-Scale Scoliosis Detection Using X-Ray Images
International Journal of Medical Toxicology and Forensic Medicine,
Vol. 16 (2026),
1 January 2026
,
Page 1-10
https://doi.org/10.22037/ijmtfm.v16.52666
Abstract
Background: Scoliosis is a condition in which the spine curves abnormally, and it can be difficult to identify early because the changes are often subtle and the spine's structure is complex. Traditional manual diagnosis is time-consuming and depends on clinical expertise. Therefore, this study presents a Hybrid Attention-Enhanced YOLO Framework for Accurate Multi-Scale Scoliosis Detection (HAYOLO-MSD) approach.
Methods: The proposed model uses a hybrid CNN–Transformer backbone with C2f blocks to capture both local details and global context. A feature fusion module combines these representations for better learning. In the neck, an attention-based APAN replaces traditional FPN and PAN to focus on spinal curvature and abnormalities. The decoupled head separates classification and regression tasks, with attention-enhanced Softmax for prediction. Training is optimized using AdamW, and Grad-CAM provides interpretability by highlighting important regions.
Results: Extensive experiments on the Balanced Scoliosis X-ray Dataset demonstrate strong performance of the proposed system. The comparative result analysis demonstrates that the HAYOLO-MSD method achieves 96.56% accuracy on the Balanced Scoliosis X-ray Dataset and 95.73 accuracy on the Mendeley dataset.
Conclusion: The proposed HAYOLO-MSD approach provides an effective and accurate framework for multi-scale scoliosis detection, demonstrating strong performance on both the Balanced Scoliosis X-ray dataset and the Mendeley dataset.
- Scoliosis detection, Deep learning, X-ray images, Spinal curvature, AdamW optimizer, Explainable artificial intelligence
How to Cite
References
[1] Martrenchar P, Caille A, Dina A, Odent T. A preliminary report: screening idiopathic scoliosis (IS) with a smartphone application. J Pediatr Orthop. 2026:10-1097. [DOI: 10.1097/BPO.0000000000003228]
[2] Yeşilmen N, Baydoğan MP, Danacı Ç, Tuncer SA, Çınar A, Tuncer T. Multi-class idiopathic scoliosis detection based on Cobb angle using a hybrid ViT model. Signal Image Video Process. 2025;19(14):1225. [DOI: 10.1007/s11760-025-04810-4]
[3] Li R, Ye B, Wang W, Liu P, Xiao B, Liu Z, et al. Hierarchical contrastive learning framework for digital scoliosis screening standards, efficacy evaluation and real-world deployment. SSRN Electron J. 2026 Jan. [DOI: 10.2139/ssrn.6210365]
[4] Vishwakarma K, Nigam A. ScoliGaitX: a deep multi-modal fusion network for scoliosis assessment via gait video analysis. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision; 2026. p. 2404-13. [DOI: 10.1109/WACV61042.2026.00237]
[5] Duan X, Ma X, Zhu M, Wang L, You D, Deng L, et al. Deep learning-assisted screening and diagnosis of scoliosis: segmentation of bare-back images via an attention-enhanced convolutional neural network. J Orthop Surg Res. 2025;20(1):161. [DOI: 10.1186/s13018-025-05564-y]
[6] Mulford KL, Nugraha HK, Todderud JE, Pumford A, Milbrandt TA, Huddleston PM, et al. Predicting nonidiopathic scoliosis from plain radiographs: a deep-learning approach. JBJS Open Access. 2025;10(4):e25. [DOI: 10.2106/JBJS.OA.25.00257]
[7] Natesan G, Murugesan A, Nagarajan R. Lightweight CNN-enabled framework for automated detection of scoliosis and spondylosis using spine X-rays. J Comput Cogn Eng. 2025. [DOI: 10.47852/bonviewJCCE52026933]
[8] Takahashi S, Ichikawa S, Watanabe K, Ueda H, Arima H, Yamato Y, et al. Development and validation of transformer- and convolutional neural network-based deep learning models to predict curve progression in adolescent idiopathic scoliosis. J Clin Med. 2025;14(20):7216. [DOI: 10.3390/jcm14207216]
[9] Feng CK, Chen YJ, Dinh QT, Tran KT, Liu CY. Multi-task learning for classification and prediction of adolescent idiopathic scoliosis based on fringe-projection three-dimensional imaging. Eur Spine J. 2025;34(11):5321-33. [DOI: 10.1007/s00586-025-09340-8]
[10] Maharasi M, Senthilnayaki N, Snehaprabha K. Vertebrae landmark detection and scoliosis assessment using deep learning. In: 2024 International Conference on Communication, Computing and Internet of Things (IC3IoT); 2024. p. 1-6. [DOI: 10.1109/IC3IoT60841.2024.10550261]
[11] Salmankey. Balanced scoliosis X-ray dataset, YOLOv5 format [Internet]. Kaggle. Available from: https://www.kaggle.com/datasets/salmankey/balanced-scoliosis-x-ray-dataset-yolov5-format/
[12] Mendeley Data. Dataset xkt857dsxk/1 [Internet]. Available from: https://data.mendeley.com/datasets/xkt857dsxk/1
[13] Ma S, Lu H, Liu J, Zhu Y, Sang P. Layn: lightweight multi-scale attention YOLOv8 network for small object detection. IEEE Access. 2024;12:29294-307. [DOI: 10.1109/ACCESS.2024.3368848]
[14] Farag MM, Fouad M, Abdel-Hamid AT. Automatic severity classification of diabetic retinopathy based on DenseNet and convolutional block attention module. IEEE Access. 2022;10:38299-308. [DOI: 10.1109/ACCESS.2022.3165193]
[15] Zheng H, Wang B, Xiao M, Qin H, Wu Z, Tan L. Adaptive friction in deep learning: enhancing optimizers with sigmoid and tanh function. In: 2024 IEEE 6th International Conference on Power, Intelligent Computing and Systems (ICPICS); 2024. p. 809-13. [DOI: 10.1109/ICPICS62053.2024.10797007]
[16] Yang Y, Guo R, Wu S, Wang Y, Yang L, Fan B, et al. DecomCAM: advancing beyond saliency maps through decomposition and integration. Neurocomputing. 2024;610:127826. [DOI: 10.1016/j.neucom.2024.127826]
[17] Zhang L, Pei B, Zhang S, Lu D, Xu Y, Huang X, et al. A new method for scoliosis screening incorporating deep learning with back images. Global Spine J. 2025;15(4):2062-74. [DOI: 10.1177/21925682241282581]
[18] Lu QT, Nguyen TM. An approach to classifying X-ray images of scoliosis and spondylolisthesis based on fine-tuned Xception model. Int J Adv Comput Sci Appl. 2024;15(2). [DOI: 10.14569/IJACSA.2024.0150262]
[19] Lu QT, Nguyen TM. An approach to classifying X-ray images of scoliosis and spondylolisthesis based on fine-tuned Xception model. Int J Adv Comput Sci Appl. 2024;15(2). [DOI: 10.14569/IJACSA.2024.0150262]
[20] Yeşilmen N, Danacı Ç, Baydoğan MP, Tuncer SA, Çınar A, Tuncer T. Enhanced vision transformer with custom attention mechanism for automated idiopathic scoliosis classification. J Imaging Inform Med. 2026;39(2):1122-39. [DOI: 10.1007/s10278-025-01564-w]
- Abstract Viewed: 14 times
- PDF Downloaded: 6 times