Scalable and Robust Artificial Intelligence for Spine Alignment Assessment: A Multicentre Study Enabled by Real-Time Data Transformation (Preprint)

Guilin Chen, Nan Meng, Yipeng Zhuang, Zhe Chen, Zhen Bian, Zheng Gong, Jiawei Shi, Tao Huang, Xihe Kuang, Pengyu Lu, Cong Nie, Qifeng Yu, Zefu Chen, Hui Jiang, Zhongmin Zhang, Chaojun Zheng, Liang Yu, Haixu He, Jason Pui Yin Cheung, Jian‐Guo Zhang, Teng ZhangJournal of Medical Internet Research 2025 · DOI

TL;DR

This paper tackles the challenge of deploying AI models for scoliosis assessment across different hospitals, where image variations hinder performance. They introduce a novel "plug-and-play" data transformation method that harmonizes image characteristics in real-time, significantly improving the accuracy and robustness of their AI model, SpineHRNet+, in multicenter settings.

What Problem Does This Paper Address?

Adolescent idiopathic scoliosis (AIS) is a common spinal deformity that requires accurate assessment, typically using the Cobb angle (CA) measured from X-rays. While AI models show promise for automating this, their performance often drops when applied across different hospitals. This is due to variations in imaging equipment, protocols, and patient positioning, leading to differences in image brightness, contrast, and overall quality. This "data heterogeneity" makes it difficult to generalize AI models, hindering their widespread clinical adoption. Existing solutions often require extensive data or complex adjustments, making them impractical for real-world clinical settings.

Methods

The core of this paper is a novel Intensity Distribution-Based Data Transformation method. This approach works by analyzing the statistical profiles of pixel intensity distributions (brightness and contrast) within images from different institutions. It then uses an optimization technique to adjust each image's data to align with a reference standard. This is achieved through histogram matching, where the cumulative distribution function (CDF) of a source image is aligned to that of a target image. Crucially, this transformation is applied in real-time during the AI model's training and fine-tuning process, allowing the model to adapt dynamically to variations without requiring paired training data or extensive preprocessing. This method is designed to be "plug-and-play," meaning it can be easily integrated into existing AI pipelines with minimal calibration. The authors integrated this into their previously developed AI model, SpineHRNet+, creating an "enhanced model."

Datasets

The study utilized a multicenter retrospective dataset of 3899 full-spine radiographs from 7 hospitals: 2 in Hong Kong (QMH and DKCH) and 5 in Mainland China (PUMCH, FAH, QLH, SDPH, and HSH).

  • Training and Internal Validation: Data from the 2 Hong Kong hospitals (n=3034) were used for training and internal validation.
  • External Validation: Radiographs from the 5 mainland hospitals (n=865) served as independent external validation datasets.

The paper does not introduce new benchmark datasets but rather evaluates the model's performance on existing, real-world multicenter data.

Key Insights and Ideas

  • Data heterogeneity is a major bottleneck for AI in healthcare: The study clearly demonstrates how variations in imaging across institutions significantly degrade AI model performance, even for models that perform well in single-center settings.
  • Real-time, unsupervised transformation is key: The proposed method's ability to adapt image characteristics on-the-fly during training, without requiring paired data or manual intervention, is a significant advancement for practical AI deployment.
  • Preserving anatomical integrity is paramount: The transformation method focuses on pixel intensity alignment, ensuring that critical anatomical structures are not distorted, which is crucial for clinical diagnosis.
  • "Plug-and-play" AI is achievable: The method's ease of integration and minimal calibration requirements make it highly practical for widespread adoption in diverse clinical environments.

Key Results

The enhanced SpineHRNet+ model showed significant improvements in both internal and external validation:

  • Cobb Angle (CA) Prediction:
    • Mean absolute prediction errors were reduced across all datasets.
  • On external datasets, the mean absolute error for thoracic curves decreased by 1.35° (SD 1.72°).
  • Coefficient of determination (R²) consistently exceeded 0.90 across all cohorts, with an average improvement of 0.11 on external datasets.
  • Bland-Altman analysis showed robust agreement, with 95% limits of agreement within 7.51° across all datasets.
  • Disease Severity Grading:
    • Substantial improvements in sensitivity and negative predictive value (NPV) were observed.
  • On external datasets, average sensitivity increased by 12.57% (from 77.61% to 90.18%) and average NPV by 2.86% (from 90.30% to 93.16%).
  • Spinal Curve Type Identification:
    • Slight improvements in sensitivity were noted.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single Modality Focus: The current method addresses variability within radiographic data but does not handle cross-modality transformations (e.g., between X-ray, CT, or MRI).
  • Suboptimal Data Quality: The method does not fully resolve issues related to low resolution, class imbalance, or other forms of suboptimal image quality.
  • Missing Data: BMI data was unavailable for some external cohorts, limiting analysis of body habitus effects.

Future work could explore:

  • Developing more comprehensive solutions for content-aware enhancements.
  • Extending the transformation framework to cross-modality scenarios.
  • Enabling unified AI models across different imaging modalities for enhanced diagnostic versatility.

Summary updated

BibTeX
@article{chen2025scalable,
  author = {Guilin Chen and Nan Meng and Yipeng Zhuang and Zhe Chen and Zhen Bian and Zheng Gong and Jiawei Shi and Tao Huang and Xihe Kuang and Pengyu Lu and Cong Nie and Qifeng Yu and Zefu Chen and Hui Jiang and Zhongmin Zhang and Chaojun Zheng and Liang Yu and Haixu He and Jason Pui Yin Cheung and Jian‐Guo Zhang and Teng Zhang},
  title = {{Scalable and Robust Artificial Intelligence for Spine Alignment Assessment: A Multicentre Study Enabled by Real-Time Data Transformation (Preprint)}},
  journal = {Journal of Medical Internet Research},
  year = {2025},
  doi = {10.2196/78396},
}
Figure