Reproduce “Accurate Scoliosis Vertebral Landmark Localization on X-ray Images via Shape-constrained Multi-stage Cascaded CNNs” on OpenResearch

About this paper

Zhiwei Wang, Jinxin Lv, Yunqiao Yang, Yuanhuai Liang, Yi Lin, Qiang Li, Xin Li, Xin Yang

Researchers developed a system using shape-constrained multi-stage cascaded Convolutional Neural Networks for accurate localization of 68 vertebral landmarks on scoliosis X-ray images. This method achieved a Normalized Mean Squared Error of 7.2e-4, outperforming state-of-the-art approaches and setting a new benchmark for Cobb angle estimation with a Symmetric Mean Absolute Percentage Error of 21.30% in the AASCE challenge.

Key insights

• Decomposing complex global regression into sequential, local refinement steps with a cascaded CNN architecture significantly enhances accuracy and robustness in landmark localization.
• Integrating PCA-mined shape priors into the regression process effectively constrains predictions to physiologically plausible shapes, mitigating errors caused by ambiguous visual features and mutual attraction between similar anatomical structures.
• Employing a lightweight shared encoder (truncated MobileNet-V2) within a multi-stage framework allows for high accuracy and efficiency, making the solution practical for clinical deployment.

Summary by alphaXiv.