Differentiation of benign and malignant vertebral fractures using a convolutional neural network to extract CT-based texture features

Sophia S. Goller, Sarah C. Foreman, Jon Rischewski, Jürgen Weißinger, Anna-Sophia Dietrich, David Schinz, Robert Stahl, Johanna Luitjens, Sebastian Siller, Vanessa F. Schmidt, Bernd Erber, Jens Ricke, Thomas Liebig, Jan S. Kirschke, Michael Dieckmeyer, Alexandra S. GersingEuropean 2023 · DOI

Abstract

Three-dimensional CT-based global texture feature (TF) skewness assessed using a convolutional neural network (CNN)-based framework showed significant difference between benign and malignant thoracolumbar vertebral fractures (VFs). This automated approach may contribute significantly to the clinical diagnostic work-up of patients with VFs.

Introduction

In an aging population, vertebral fractures (VFs) are a frequent clinical finding. However, determining the underlying cause—whether "benign" (osteoporotic) or "malignant" (metastatic)—remains a significant diagnostic challenge. On standard CT scans, these two types of fractures often appear remarkably similar, as both involve the breakdown of bone architecture and loss of density.

CT scans of thoracolumbar spines showing malignant (A) and benign (B) fractures, illustrating the visual similarity that complicates diagnosis.

For patients with a history of cancer, misdiagnosis can lead to unnecessary invasive biopsies or delays in critical treatment. Radiologists currently rely on subjective visual inspection. This study investigates whether "radiomics"—the extraction of quantitative texture features—can provide a more objective and reliable diagnostic tool.

Methods

The researchers developed a multi-step automated pipeline to extract quantitative data from routine clinical images:

The four-step automated workflow: 1) Spine CT acquisition, 2) Deep-learning-based segmentation, 3) Radiomics feature extraction, and 4) Statistical comparison.
  1. Automated Segmentation: A CNN-based framework was used to automatically identify, label, and segment vertebrae from T1 to L6, eliminating the need for tedious manual tracing.
  2. Texture Analysis (TA): Eight 3D texture features (TFs) were extracted, including first-order (global variance and skewness), second-order (energy and entropy), and higher-order features (Short-Run Emphasis and Run Percentage) to characterize bone microstructure.
  3. Statistical Modeling: Multivariate regression models, adjusted for age and sex, were employed to identify features that significantly correlate with fracture type.

Datasets

The study analyzed a retrospective cohort of 409 patients from two institutions (2010–2020). The reference standard ("ground truth") was established via biopsy or at least three months of clinical imaging follow-up. In total, 228 benign and 308 malignant fractured vertebrae were evaluated.

Key Results and Insights

The study identified global skewness—a measure of the asymmetry in bone density distribution—as the primary differentiator between fracture types.

  • Significant Differentiator: Global TF skewness was significantly higher in the benign group (0.70) compared to the malignant group (0.59) (p=0.017). This suggests that osteoporotic bone maintains a specific structural "asymmetry" that is disrupted when a tumor infiltrates the bone.
  • Non-Significant Features: Other common radiomic features, such as entropy (randomness) and energy (uniformity), did not show statistically significant differences in this specific cohort.
  • Clinical Feasibility: The results demonstrate that 3D texture analysis can be performed on routine, non-specialized CT scans, making it a viable candidate for integration into standard clinical workflows without requiring additional manual labor.

Limitations and Future Work

While promising, several steps remain before clinical implementation:

  • Retrospective Design: Future prospective studies are needed to validate these findings and minimize selection bias.
  • Contrast Agents: The impact of intravenous contrast agents on texture values requires further investigation.
  • Multimodal Models: Combining radiomic features with clinical data (e.g., BMI, smoking status, cancer history) into a unified "radiomics-clinical model" would likely enhance diagnostic accuracy further.

Summary updated

Abstract

Three-dimensional CT-based global TF skewness assessed using a CNN-based framework showed significant difference between benign and malignant thoracolumbar VFs and may therefore contribute to the clinical diagnostic work-up of patients with VFs.
BibTeX
@inproceedings{goller2023differentiation,
  author = {Sophia S. Goller and Sarah C. Foreman and Jon Rischewski and Jürgen Weißinger and Anna-Sophia Dietrich and David Schinz and Robert Stahl and Johanna Luitjens and Sebastian Siller and Vanessa F. Schmidt and Bernd Erber and Jens Ricke and Thomas Liebig and Jan S. Kirschke and Michael Dieckmeyer and Alexandra S. Gersing},
  title = {{Differentiation of benign and malignant vertebral fractures using a convolutional neural network to extract CT-based texture features}},
  booktitle = {European},
  year = {2023},
  doi = {10.1007/s00586-023-07838-7},
}
Figure