A DUAL-PATH MULTI-TASK FRAMEWORK FOR STRICT THREE-CURVE COBB ANGLE ESTIMATION IN IDIOPATHIC SCOLIOSIS
DOI:
https://doi.org/10.37943/25XRRS8272Keywords:
Cobb angle estimation, scoliosis, deep learning, multi-task learning, dual-path fusion, medical imaging, curriculum learning, clinical decision supportAbstract
Adolescent idiopathic scoliosis management depends on reproducible Cobb angle measurement across three clinically defined spinal regions: proximal thoracic, main thoracic, and thoracolumbar/lumbar. Although manual measurement remains the reference standard, it is observer-dependent and time-consuming, with inter-observer variability exceeding five degrees even among experienced readers. Most automated deep learning approaches target a single dominant curve or use unconstrained outputs, which limits their applicability to structured clinical workflows requiring strict regional assignment. This study presents a dual-path multi-task framework for simultaneous estimation of all three regional Cobb angles from posteroanterior spinal radiographs. The architecture integrates a ConvNeXt-Tiny encoder, vertebral localization heads, direct global angle regression via soft-argmax, and a geometric tilt-aggregation pathway. A learned per-region sigmoid gate fuses the global and geometric pathways, providing a fixed but optimized balance between statistical and anatomical estimation. The model was developed on 21,294 radiographs with leakage-controlled partitioning into training (N = 17,262), validation (N = 2,016), and test (N = 2,016) subsets. Training employed a two-stage curriculum with severity-aware sampling and hard replay for difficult cases. Three independent runs (seeds 42, 52, 62) were ensembled with test-time augmentation. On the primary held-out set (N = 2,015), the ensemble achieved a mean absolute error of 2.24 degrees (proximal thoracic 2.21, main thoracic 1.97, thoracolumbar/lumbar 2.54), with near-zero Bland-Altman bias (0.03 degrees), good-to-excellent intraclass correlation coefficients (0.884–0.971), and 90.4% of predictions within 5 degrees. At the 40-degree treatment threshold, sensitivity was 0.934 and specificity was 0.994. These findings support the feasibility of strict three-curve automation for reader-in-the-loop clinical workflows.
References
Prestigiacomo, C., Hulsbosch, M. H. H. M., Bruls, V. E. J., & Nieuwenhuis, J. J. (2022). Intra- and inter-observer reliability of Cobb angle measurements in patients with adolescent idiopathic scoliosis. *Spine Deformity, 10*(1), 79-86. https://doi.org/10.1007/s43390-021-00398-0
Wang, L., Xie, C., Lin, Y., Zhou, H.-Y., Chen, K., Cheng, D., Kocev, D., Yap, C. H., Staring, M., & de Brujine, M. (2021). Evaluation and comparison of accurate automated spinal curvature estimation algorithms with spinal anterior-posterior X-ray images: The AASCE2019 challenge. *Medical Image Analysis, 72*, 102115. https://doi.org/10.1016/j.media.2021.102115
Jin, C., Wang, S., Yang, G., Li, E., & Liang, Z. (2022). A review of the methods on Cobb angle measurements for spinal curvature. *Sensors, 22*(9), 3258. https://doi.org/10.3390/s22093258
d'Astorg, H., Faron, M., Jolivet, E., Folinais, D., & Bertheau, R. C. (2023). Comparison of Cobb angle measurements for scoliosis assessment using different imaging modalities: A systematic review. *EFORT Open Reviews, 8*(6), 489-498. https://doi.org/10.1530/EOR-23-0032
Horng, M.-H., Kuok, C.-P., Fu, M.-J., Lin, C.-J., & Sun, Y.-N. (2019). Cobb angle measurement of spine from X-ray images using convolutional neural network. *Computational and Mathematical Methods in Medicine, 2019*, 6357171. https://doi.org/10.1155/2019/6357171
Suri, A., Tang, S., Kargilis, D., Taratuta, E., Kneeland, B. J., Choi, G., Fritz, J., & Carrino, J. A. (2023). Conquering the Cobb angle: A deep learning algorithm for automated, hardware-invariant measurement of Cobb angle on radiographs in patients with scoliosis. *Radiology: Artificial Intelligence, 5*(4), e220158. https://doi.org/10.1148/ryai.220158
Zou, L., Guo, L., Zhang, R., Ni, L., Chen, Z., He, X., & Wang, J. (2023). VLTENet: A deep-learning-based vertebra localization and tilt estimation network for automatic Cobb angle estimation. *IEEE Journal of Biomedical and Health Informatics, 27*(6), 3002-3013. https://doi.org/10.1109/JBHI.2023.3258361
Wang, M. X., Kim, J. K., Choi, J.-W., Park, D., & Chang, M. C. (2024). Deep learning algorithm for automatically measuring Cobb angle in patients with idiopathic scoliosis. *European Spine Journal, 33*(11), 4155-4163. https://doi.org/10.1007/s00586-023-08024-5
Shi, C., Meng, N., Zhuang, Y., Cheung, J. P. Y., Zhao, M., Huang, H., Lin, Y., & Zhang, T. (2025). Accurate Cobb angle estimation via SVD-based curve detection and vertebral wedging quantification. *IEEE Journal of Biomedical and Health Informatics, 29*(12), 8607-8614. https://doi.org/10.1109/JBHI.2025.3600647
Chen, R., Xi, Y., Wang, T., Wang, A., Ma, Z., Liang, M., Yuan, S., Zang, L., & Fan, N. (2026). Advances in fusion level selection and surgical approaches for adolescent idiopathic scoliosis based on the Lenke classification system: A narrative review. *BMC Surgery, 26*(1), 102. https://doi.org/10.1186/s12893-025-03481-9
Zhu, Y., Yin, X., Chen, Z., Zhang, H., Xu, K., Zhang, J., & Wu, N. (2025). Deep learning in Cobb angle automated measurement on X-rays: A systematic review and meta-analysis. *Spine Deformity, 13*(1), 19-27. https://doi.org/10.1007/s43390-024-00954-4
He, K., Gkioxari, G., Dollar, P., & Girshick, R. (2017). Mask R-CNN. *Proceedings of the IEEE International Conference on Computer Vision*, 2961-2969. https://doi.org/10.1109/ICCV.2017.322
Zhang, Y., & Yang, Q. (2022). A survey on multi-task learning. *IEEE Transactions on Knowledge and Data Engineering, 34*(12), 5586-5609. https://doi.org/10.1109/TKDE.2021.3070203
Crawshaw, M. (2020). Multi-task learning with deep neural networks: A survey. *arXiv preprint arXiv:2009.09796*. https://doi.org/10.48550/arXiv.2009.09796
Soviany, P., Ionescu, R. T., Rota, P., & Sebe, N. (2022). Curriculum learning: A survey. *International Journal of Computer Vision, 130*(6), 1526-1565. https://doi.org/10.1007/s11263-022-01611-x
Jimenez-Sanchez, A., Mateus, D., Kirchhoff, S., Navab, N., & Ballester, M. A. G. (2019). Medical-based deep curriculum learning for improved fracture classification. *Proceedings of Medical Image Computing and Computer-Assisted Intervention*, 694-702. https://doi.org/10.1007/978-3-030-32226-7_77
Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., & Xie, S. (2022). A ConvNet for the 2020s. *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 11976-11986.
Loshchilov, I., & Hutter, F. (2019). Decoupled weight decay regularization. *Proceedings of the International Conference on Learning Representations*.
Tarvainen, A., & Valpola, H. (2017). Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. *Advances in Neural Information Processing Systems, 30*.
Kendall, A., & Gal, Y. (2017). What uncertainties do we need in Bayesian deep learning for computer vision? *Advances in Neural Information Processing Systems, 30*.
Gal, Y., & Ghahramani, Z. (2016). Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. *Proceedings of the 33rd International Conference on Machine Learning*, 1050-1059.
Tejani, A. S., Klontzas, M. E., Gatti, A. A., Mongan, J. T., & Kahn, C. E. (2024). Checklist for Artificial Intelligence in Medical Imaging (CLAIM): 2024 update. *Radiology: Artificial Intelligence, 6*(4), e240300. https://doi.org/10.1148/ryai.240300
Mongan, J., Moy, L., & Kahn, C. E. Jr. (2020). Checklist for Artificial Intelligence in Medical Imaging (CLAIM): A guide for authors and reviewers. *Radiology: Artificial Intelligence, 2*(2), e200029. https://doi.org/10.1148/ryai.2020200029
Lakshminarayanan, B., Pritzel, A., & Blundell, C. (2017). Simple and scalable predictive uncertainty estimation using deep ensembles. *Advances in Neural Information Processing Systems, 30*.
Downloads
Published
How to Cite
Section
License
Copyright (c) 2026 Articles are open access under the Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.