A Computed Tomography–Based Fracture Prediction Model With Images of Vertebral Bones and Muscles by Employing Deep Learning: Development and Validation Study

Sung Hye Kong, Wonwoo Cho, Sung Bae Park, Jaegul Choo, Jung Hee Kim, Sang Wan Kim, Chan Soo ShinJournal of Medical Internet Research 2024 · DOI

Introduction

As global populations age, fragility fractures have emerged as a significant public health burden. While Dual-energy X-ray Absorptiometry (DXA) remains the gold standard for measuring bone mineral density (BMD), it is significantly underutilized. Statistics indicate that nearly 60% of patients who suffer major osteoporotic fractures have never undergone proper screening. This gap in care necessitates more accessible and efficient screening methods.

The Concept of "Opportunistic CT"

This study explores the potential of "opportunistic CT"—leveraging the millions of abdominal CT scans performed annually for unrelated reasons, such as oncology follow-ups or appendicitis. These scans contain rich data regarding a patient’s skeletal and muscular health that typically goes unanalyzed by radiologists focused on the primary diagnostic task.

The researchers focused on the "bone-muscle unit" hypothesis. This theory suggests that osteoporosis (bone loss) and sarcopenia (muscle loss) are deeply interconnected. By analyzing both tissues simultaneously, a more comprehensive picture of fracture risk can be obtained compared to evaluating bone alone.

Methodology: A Hybrid Deep Learning Architecture

To process these complex 3D volumes, the team developed a hybrid CNN-RNN architecture featuring an attention mechanism:

  • CNN Encoder: A ResNeXt-50 backbone was used to extract spatial features from individual 2D axial CT slices.
  • RNN Decoder: To capture the 3D structural context, the researchers treated a sequence of 14 equidistant slices as a "video" stream. A Gated Recurrent Unit (GRU) integrated these slices, providing 3D insights without the computational intensity of a full 3D CNN.
  • Attention Module: This component directed the model’s focus toward the most relevant anatomical regions: the vertebral body and the paravertebral muscles.
  • Interpretability: GradCAM (Gradient-weighted Class Activation Mapping) was utilized to generate heatmaps, ensuring the model prioritized bone and muscle tissues rather than irrelevant background data.

Validation and Performance

The model was developed using a dataset of 1,214 patients from Seoul National University Bundang Hospital (SNUBH) and validated against an external "real-world" cohort of 495 patients from Seoul National University Boramae Hospital.

The results demonstrated that the "Bone + Muscle" model was the most effective predictor:

  • Predictive Accuracy: In external validation, the combined model achieved an AUROC of 0.827, surpassing the bone-only model (0.815).
  • Comparison to Clinical Standards: The image-based model significantly outperformed standard clinical models (AUROC 0.749) and the widely used FRAX (Fracture Risk Assessment Tool) scores (AUROC 0.810).
  • Sensitivity: The model correctly identified 70% of future fracture cases up to five years before they occurred.

Discussion and Future Directions

A key insight from the study is that CT images appear to "encode" clinical variables. Adding clinical data (like age or BMI) to the image-based model did not significantly improve performance, suggesting that a patient's metabolic health is already physically reflected in their bone calcification and muscle texture.

While promising, the study's retrospective nature means prospective trials are required to integrate this into real-time clinical workflows. Future work may also involve differentiating between specific muscle groups (e.g., psoas vs. multifidus) and extending the prediction window to 10 years to align with international standards like FRAX.

Conclusion

By utilizing deep learning to analyze the bone-muscle unit in opportunistic CT scans, clinicians may soon be able to identify high-risk patients who would otherwise go unscreened, potentially preventing thousands of debilitating fractures through early intervention.

Summary updated

Abstract

The model using the images of the vertebral bone and paravertebral muscle showed better performance than that using the images of the bone-only or clinical variables. Opportunistic CT screening may contribute to identifying patients with a high fracture risk in the future.
BibTeX
@article{kong2024computed,
  author = {Sung Hye Kong and Wonwoo Cho and Sung Bae Park and Jaegul Choo and Jung Hee Kim and Sang Wan Kim and Chan Soo Shin},
  title = {{A Computed Tomography–Based Fracture Prediction Model With Images of Vertebral Bones and Muscles by Employing Deep Learning: Development and Validation Study}},
  journal = {Journal of Medical Internet Research},
  year = {2024},
  doi = {10.2196/48535},
}
Figure