Interpretable Vertebral Fracture Quantification via Anchor-Free Landmarks Localization

Alexey Zakharov, Maxim Pisov, Alim Bukharaev, Alexey Petraikin, Sergey Morozov, Victor Gombolevskiy, Mikhail Belyaev

Abstract

Vertebral body compression fractures are early signs of osteoporosis. Though these fractures are visible on Computed Tomography (CT) images, they are frequently missed by radiologists in clinical settings. Prior research on automatic methods of vertebral fracture classification proves its reliable quality; however, existing methods provide hard-to-interpret outputs and sometimes fail to process cases with severe abnormalities such as highly pathological vertebrae or scoliosis. We propose a new two-step algorithm to localize the vertebral column in 3D CT images and then detect individual vertebrae and quantify fractures in 2D simultaneously. We train neural networks for both steps using a simple 6-keypoints based annotation scheme, which corresponds precisely to the current clinical recommendation. Our algorithm has no exclusion criteria, processes 3D CT in 2 seconds on a single GPU, and provides an interpretable and verifiable output. The method approaches expert-level performance and demonstrates state-of-the-art results in vertebrae 3D localization (the average error is 1 mm), vertebrae 2D detection (precision and recall are 0.99), and fracture identification (ROC AUC at the patient level is up to 0.96). Our anchor-free vertebra detection network shows excellent generalizability on a new domain by achieving ROC AUC 0.95, sensitivity 0.85, specificity 0.9 on a challenging VerSe dataset with many unseen vertebra types.

Introduction

Vertebral compression fractures (VCFs) are critical early indicators of osteoporosis, yet they are frequently missed by radiologists—sometimes at rates exceeding 50%. This diagnostic gap often occurs because CT scans are typically reviewed in axial (top-down) slices, whereas fractures are most apparent in the sagittal (side) plane.

Existing AI solutions for automated fracture detection face significant hurdles. Many models operate as "black boxes," providing probability scores without verifiable measurements. Furthermore, geometric complexities like scoliosis or severe deformities often break standard detection algorithms. Finally, many methods rely on precise bone segmentation, which frequently fails on the very bones that are most severely fractured.

Methodology

The researchers developed a two-stage "localize then quantify" pipeline designed for both robustness and clinical interpretability.

Step 1: Spine Localization and Straightening

The first stage employs a 3D UNet-like architecture to find the spinal centerline. By utilizing a soft-argmax layer, the model generates probability maps to pinpoint the spine's location accurately, even in volumes containing non-spinal anatomy like the head or legs.

3D Spine Localization using Soft-Argmax

Once the centerline is identified, the system "straightens" the spine by interpolating the 3D volume along this curve. This step transforms complex, curved spinal anatomy into a standardized 2D sagittal image where vertebrae are aligned vertically, simplifying the subsequent detection task.

Step 2: Anchor-Free Landmark Localization and Quantification

Instead of traditional rectangular bounding boxes, which poorly fit the irregular shapes of vertebrae, the model predicts six keypoints for each vertebral body. These points correspond to the anterior, middle, and posterior heights ($h_a, h_m, h_p$).

To handle overlapping predictions, the authors introduced Butterfly-IoU, a metric that calculates the overlap of hexagonal shapes formed by these keypoints. This approach is significantly more accurate for deformed or fractured bones than standard rectangular IoU. Using these landmarks, the system automatically calculates the Genant scale, the clinical standard for grading fracture severity based on height ratios.

End-to-End Pipeline: From Spine Straightening to Fracture Quantification

Key Insights

  • Ditching the Box: Moving from anchor-based bounding boxes to anchor-free landmarks allows the model to focus on the specific points required for medical diagnosis, reducing noise introduced by irregular vertebral shapes.
  • Straightening as Pre-processing: Mathematically "un-curving" the spine makes the model robust to patient positioning and severe anatomical deformities like scoliosis.
  • Verifiable AI: The output is a set of points and height ratios, allowing radiologists to visually confirm the model's measurements, fostering trust in the system's diagnostic suggestions.

Results and Performance

The pipeline demonstrated exceptional performance across multiple datasets, including LungCancer500, VerSe-2020, and various private clinical sets:

  • Accuracy: Patient-level ROC AUC reached up to 0.96 for fracture identification.
  • Speed: The entire 3D CT volume is processed in approximately 2 seconds on a single GPU.
  • Generalizability: On the challenging VerSe dataset, the model maintained a high ROC AUC of 0.95 without fine-tuning, proving its ability to handle diverse scanner types and unseen vertebral regions.

Conclusion

While the system shows state-of-the-art results, limitations remain in handling atypical anatomy (like the C2 vertebra) and extreme image artifacts from metal implants. Future work will focus on expanding training data to include more diverse cervical and lower lumbar cases to further enhance robustness across the entire spinal column. Nonetheless, this interpretable landmark-based approach represents a significant step toward reliable, real-time automated vertebral fracture screening.

Summary updated

Abstract

Vertebral body compression fractures are early signs of osteoporosis. Though these fractures are visible on Computed Tomography (CT) images, they are frequently missed by radiologists in clinical settings. Prior research on automatic methods of vertebral fracture classification proves its reliable quality; however, existing methods provide hard-to-interpret outputs and sometimes fail to process cases with severe abnormalities such as highly pathological vertebrae or scoliosis. We propose a new two-step algorithm to localize the vertebral column in 3D CT images and then detect individual vertebrae and quantify fractures in 2D simultaneously. We train neural networks for both steps using a simple 6-keypoints based annotation scheme, which corresponds precisely to the current clinical recommendation. Our algorithm has no exclusion criteria, processes 3D CT in 2 seconds on a single GPU, and provides an interpretable and verifiable output. The method approaches expert-level performance and demonstrates state-of-the-art results in vertebrae 3D localization (the average error is 1 mm), vertebrae 2D detection (precision and recall are 0.99), and fracture identification (ROC AUC at the patient level is up to 0.96). Our anchor-free vertebra detection network shows excellent generalizability on a new domain by achieving ROC AUC 0.95, sensitivity 0.85, specificity 0.9 on a challenging VerSe dataset with many unseen vertebra types.
BibTeX
@article{zakharov2023interpretable,
  author = {Alexey Zakharov and Maxim Pisov and Alim Bukharaev and Alexey Petraikin and Sergey Morozov and Victor Gombolevskiy and Mikhail Belyaev},
  title = {{Interpretable Vertebral Fracture Quantification via Anchor-Free Landmarks Localization}},
  journal = {Medical Image Analysis},
  year = {2023},
  url = {https://openreview.net/forum?id=w4nkM1GRo},
}
Figure