HealthiVert-GAN is a novel framework for synthesizing pseudo-healthy vertebral CT images from fractured vertebrae. By simulating pre-fracture states, it enables interpretable quantification of vertebral compression fractures (VCFs) through Relative Height Loss of Vertebrae (RHLV). The model integrates a two-stage GAN architecture with anatomical consistency modules, achieving state-of-the-art performance on both public and private datasets.
🚀 Key Features
Two-Stage Synthesis: Coarse-to-fine generation with 2.5D sagittal/coronal fusion.
HealthiVert-Guided Attention Module (HGAM): Focuses on non-fractured regions via Grad-CAM++.
Iterative Synthesis: Generates adjacent vertebrae first to minimize fracture interference.
RHLV Quantification: Measures height loss in anterior/middle/posterior regions for SVM-based Genant grading.
🛠️ Architecture
Workflow
Workflow
Preprocessing:
Spine Straightening: Align vertebrae vertically using SCNet segmentation.
De-pedicle: Remove vertebral arches for body-focused analysis.
Masking: Replace target vertebra with a fixed-height mask (40mm).
Two-Stage Generation:
Coarse Generator: Outputs initial CT and segments adjacent vertebrae.
Refinement Generator: Enhances details with contextual attention and edge loss.
Iterative Synthesis:
Step 1: Synthesize adjacent vertebrae.
Step 2: Generate target vertebra using Step 1 results.
RHLV Calculation:
RHLV = \frac{H_{syn} - H_{ori}}{H_{syn}}
Segments vertebra into anterior/middle/posterior regions for detailed analysis.
SVM Classification: Uses RHLV values to classify fractures into mild/moderate/severe.
🔑 Key Contributions
comparisons
Interpretable Quantification Beyond Black-Box Models
Traditional end-to-end fracture classification models suffer from class imbalance and lack interpretability. HealthiVert-GAN addresses these by synthesizing pseudo-healthy vertebrae and quantifying height loss (RHLV) between generated and original vertebrae. This approach achieves superior performance (e.g., 72.3% Macro-F1 on Verse2019) while providing transparent metrics for clinical decisions.
Height Loss Distribution Mapping for Surgical Planning
HealthiVert-GAN generates cross-sectional height loss heatmaps that visualize compression patterns (wedge/biconcave/crush fractures). Clinicians can use these maps to assess fracture stability and plan interventions (e.g., vertebroplasty) with precision unmatched by single-slice methods.
Anatomic Prior Integration
Unlike conventional inpainting models, HealthiVert-GAN introduces adjacent vertebrae height variations as prior knowledge. The Self-adaptive Height Restoration Module (SHRM) dynamically adjusts generated vertebral heights based on neighboring healthy vertebrae, improving both interpretability and anatomic consistency.
1@misc{zhang2025healthivertgannovelframeworkpseudohealthy,
2 title={HealthiVert-GAN: A Novel Framework of Pseudo-Healthy Vertebral Image Synthesis for Interpretable Compression Fracture Grading},
3 author={Qi Zhang and Shunan Zhang and Ziqi Zhao and Kun Wang and Jun Xu and Jianqi Sun},
4 year={2025},
5 eprint={2503.05990},
6 archivePrefix={arXiv},
7 primaryClass={eess.IV},
8 url={https://arxiv.org/abs/2503.05990},
9}
