Deep learning model for simultaneous vertebral level localization and fracture detection on DXA vertebral fracture assessment images

Abstract

Objectives

While vertebral fracture assessment (VFA) on dual-energy X-ray absorptiometry (DXA) images is widely used to detect osteoporotic vertebral fractures, existing deep learning approaches typically perform either fracture detection or vertebral localization, not both. We developed VERTE-X L, a deep learning model that simultaneously localizes vertebrae (T7–L5) and detects fractures on DXA VFA images, and evaluated its performance in a community-dwelling older adult cohort.

Methods

We used DXA VFA images from 1000 randomly sampled participants (age ≥65 years) in the Korean Urban Rural Elderly (KURE) cohort, split into training (N = 800) and test (N = 200) sets. VERTE-X L employs a DINOv2-L encoder, a localization decoder for per-vertebra coordinate prediction, and a consistent rank logits (CORAL) ordinal regression module for fracture detection. Performance was evaluated at image and vertebra levels and compared with previously reported detection-only models (VERTE-X and VERTE-X model fine-tuned on VFA data) and localization-only models (VerteNet and Multi-Context Hybrid CNN Transformer).

Results

On the held-out test set, VERTE-X L achieved an image-level area under the receiver operating characteristic curve (AUROC) of 0.952 (sensitivity 89.7%, specificity 94.2%, net present value [NPV] 98.2%) and vertebra-level AUROC of 0.957 (NPV 99.7%). Localization success (within 15 pixels) was 94.7% for fracture vertebrae. VERTE-X L outperformed VERTE-X in detection (AUROC 0.952 vs 0.913, P = 0.034) and outperformed VerteNet in localization (mean radial error 10.0 vs. 20.3 px, P < 0.001).

Conclusions

The proposed method enables simultaneous fracture detection and vertebral localization on DXA VFA images.