B-spine: Learning B-spline Curve Representation for Robust and Interpretable Spinal Curvature Estimation

Hao Wang, Qiang Song, Ruofeng Yin, Rui MaAAAI 2024 · DOI

TL;DR

This paper introduces B-Spine, a novel deep learning pipeline for robustly estimating spinal curvature (Cobb angles) from low-quality X-ray images. By representing the spine as a B-spline curve and refining segmentation masks, B-Spine achieves superior accuracy and interpretability, particularly in challenging imaging conditions.

What Problem Does This Paper Address?

Scoliosis, a spinal deformity, requires accurate measurement of spinal curvature (Cobb angles) for diagnosis and treatment. Traditional manual measurement is time-consuming, subjective, and prone to inter-observer variability. While automated methods exist, they often struggle with low-quality X-ray images (blurry, hazy) and can lack interpretability, making it difficult to understand why a certain angle is predicted. Existing deep learning approaches often rely on precise vertebral landmarks, which are hard to annotate accurately on low-quality images, or directly regress angles from segmentation masks, obscuring the underlying geometric reasoning. This paper aims to bridge this gap by developing a robust and interpretable method that can handle low-quality images and provide insights into the estimation process.

Methods

B-Spine is a multi-stage pipeline consisting of three key modules:

  • Spine Mask Segmentation (MaskSeg): This module starts with an initial spine segmentation from a conventional network (SSformer with an SSIM loss). Crucially, it then employs a novel SegRefine network, based on unpaired image-to-image translation (CycleGAN), to refine the noisy initial mask. This unsupervised approach helps generate a higher-quality mask, making subsequent steps more robust.
  • B-Spline Centerline Prediction (CurvePred): This module learns a B-spline curve representation for the spine's centerline. It first predicts initial centerline points from the refined mask. Then, a network predicts the B-spline parameters (control points and knots). A novel resample loss is introduced to enforce explicit point supervision on the predicted B-spline curve, improving its accuracy and robustness.
  • Cobb Angle Estimation (AngleEst): This final module estimates the Cobb angles. It uses a hybrid approach:
  • Curve slope analysis: Directly estimates angles from the predicted B-spline curve by analyzing its slopes at specific points corresponding to anatomical landmarks.
  • Curve-based regression: A separate MLP regresses Cobb angles from the B-spline curve's resampled points. The final Cobb angle is a weighted combination of these two estimations, leveraging the strengths of both.

Datasets

  • AASCE2019 dataset: A public dataset containing 609 spinal AP X-ray images, with 481 for training and 128 for testing. It includes annotations for vertebral landmarks and three Cobb angles (MT, PT, TL).
  • JLU-CJUH dataset: A newly collected and annotated dataset of 584 low-quality spinal AP X-ray images. This dataset is specifically designed to be more challenging, with blurrier and hazier images, making it difficult for landmark-based methods. It includes annotations for spine segmentation masks and three Cobb angles.

Key Insights and Ideas

  • B-spline as a robust spine representation: The paper argues that B-splines, with their inherent smoothness and flexibility, are a superior representation for spinal centerlines compared to polynomial curves or simple point connections, especially in the presence of noise.
  • Unsupervised mask refinement (SegRefine): Using CycleGAN for mask refinement is a clever way to improve segmentation quality without requiring paired noisy/clean mask data, which is often difficult to obtain. This significantly boosts robustness for low-quality inputs.
  • Hybrid Cobb angle estimation: Combining direct geometric analysis (slope analysis) with a learned regression model for Cobb angles provides a more robust and accurate estimation. This approach mitigates the risk of error accumulation from earlier stages by offering a complementary estimation strategy.
  • Interpretability through B-spline: By representing the spine as a curve with control points and knots, the method offers a more interpretable output than direct regression from segmentation masks. The B-spline curve itself provides a visual representation of the spine's shape.

Key Results

  • Superior performance: B-Spine consistently outperforms representative and state-of-the-art methods on both the AASCE2019 and the challenging JLU-CJUH datasets in terms of Mean Absolute Error (MAE) for Cobb angles. For instance, on JLU-CJUH, B-Spine achieves an MAE of 4.17 for MT, 4.59 for PT, and 4.20 for TL, outperforming CasNet (4.73, 7.79, 7.28 respectively).
  • Robustness on low-quality images: Qualitative results demonstrate B-Spine's ability to produce accurate centerlines and endplate estimations even on images with noisy or partially broken masks, a common issue in low-quality X-rays.
  • Effectiveness of modules: Ablation studies confirm the significant contributions of the SegRefine network (improving mask quality) and the resample loss (enhancing B-spline learning). The hybrid Cobb angle estimation also shows improvement over using either method alone.

Limitations and Future Work

  • Error accumulation: The authors acknowledge that as a multi-stage pipeline, errors from earlier modules can accumulate, potentially impacting performance on very severe low-quality inputs.
  • Future work: The paper suggests exploring ways to improve the pipeline's robustness while maintaining interpretability, and investigating new architectures or techniques for individual modules. The authors also believe their B-spline representation and the JLU-CJUH dataset can inspire further research in this area.

Summary updated

BibTeX
@inproceedings{wang2024b,
  author = {Hao Wang and Qiang Song and Ruofeng Yin and Rui Ma},
  title = {{B-spine: Learning B-spline Curve Representation for Robust and Interpretable Spinal Curvature Estimation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i6.28346},
}
Figure