Enhancing vertebral fracture prediction using multitask deep learning computed tomography imaging of bone and muscle

Sung Hye Kong, Saemee Choi, Wonwoo Cho, Sung Bae Park, Seung Shin Park, Jaegul Choo, Jung Hee Kim, Sang Wan Kim, Chan Soo ShinEuropean 2025 · DOI

Abstract

Question: Vertebral fracture risk remains underestimated in many individuals undergoing CT scans for other reasons, highlighting the need for improved opportunistic prediction tools.
Findings: A multitask deep learning model integrating both bone and muscle features from CT scans demonstrated superior performance compared to bone-only and traditional clinical models, including FRAX.
Clinical relevance: The proposed model enables accurate vertebral fracture risk prediction using routinely acquired CT scans, facilitating early identification and intervention without the need for additional tests.

Introduction

Osteoporosis-related vertebral fractures (VFs) are a significant public health burden, often serving as a precursor to further debilitating skeletal events. Despite their clinical importance, VFs are frequently underdiagnosed because many high-risk patients do not undergo dedicated screening via Dual-energy X-ray absorptiometry (DXA), the current diagnostic gold standard.

However, millions of patients undergo abdominal computed tomography (CT) scans annually for unrelated indications, such as oncological staging or emergency evaluations. These routine scans contain substantial "opportunistic" data regarding bone mineral density and muscle composition. This study explores the development of a multitask deep learning framework designed to extract this hidden information to both detect existing fractures and predict future risk.

Methodology

Multitask Learning Framework

The researchers implemented a Multitask Learning (MTL) approach using a 3D deep learning architecture. The core of the system is a ConvNeXT Tiny backbone, which serves as a robust feature extractor. This backbone is connected to two specialized "heads":

  1. Classification Head: Designed for the cross-sectional detection of existing vertebral fractures.
  2. Hazard Head: Designed for longitudinal risk prediction, utilizing survival analysis to estimate the probability of a new fracture over a five-year window.

Input and Feature Integration

Unlike previous models that focused exclusively on bone, this framework processes 27 CT patches per patient that capture both the vertebrae and the surrounding musculature. This allows the model to account for osteosarcopenia—the synergistic decline of bone and muscle health. By training on both detection and prediction tasks simultaneously, the model leverages shared parameters, using the features of current fractures to better inform the prediction of future ones.

Datasets and Validation

The study utilized a large-scale retrospective cohort across three major institutions:

  • Development: 2,553 patients (ages 50–80) from Seoul National University Bundang Hospital (SNUBH).
  • External Validation: 1,506 patients from two independent hospitals (Seoul National University Boramae Hospital and Seoul National University Hospital).
  • Longitudinal Analysis: A subset of 3,283 total patients was followed for the 5-year fracture prediction task.

Results

Superior Predictive Performance

The MTL model demonstrated high accuracy in both tasks. For the detection of existing fractures, it achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.82 in the development set and 0.80 in external testing.

In the 5-year prediction task, the "Bone + Muscle" MTL model achieved a c-index of 0.68, which significantly outperformed the clinical standard FRAX tool (c-index: 0.66, p < 0.01). Longitudinal testing showed the model maintained superior AUROCs at various intervals (2-year: 0.79; 3-year: 0.71; 5-year: 0.71) compared to models relying solely on bone data.

The Role of Muscle and Clinical Data

A key insight from the results is that the inclusion of muscle data consistently improved model performance. Interestingly, the addition of traditional clinical variables (e.g., smoking status, steroid use) to the image-based model did not yield significant improvements. This suggests that the physiological impact of these risk factors is already manifest in the CT-derived bone and muscle features.

Discussion and Clinical Relevance

The ability to perform "free" opportunistic screening is the primary clinical advantage of this model. By running in the background of existing clinical workflows, the tool can flag high-risk individuals who might otherwise be overlooked.

Limitations

The study is limited by its retrospective nature and its focus on a primarily Korean population. Furthermore, while the 5-year prediction window is clinically useful, extending the horizon to 10 years would better align with standard FRAX outputs.

Conclusion

By integrating bone and muscle features through a multitask deep learning architecture, this study provides a powerful tool for opportunistic vertebral fracture screening. The model's superior performance over traditional clinical tools highlights the potential for AI to transform routine imaging into a proactive diagnostic resource, enabling earlier intervention and improved patient outcomes in osteoporosis management.

Summary updated

Abstract

Question Vertebral fracture risk remains underestimated in many individuals undergoing CT scans for other reasons, highlighting the need for improved opportunistic prediction tools. Findings A multitask deep learning model integrating both bone and muscle features from CT scans demonstrated superior performance compared to bone-only and traditional clinical models, including FRAX. Clinical relevance The proposed model enables accurate vertebral fracture risk prediction using routinely acquired CT scans, facilitating early identification and intervention without the need for additional tests.
BibTeX
@inproceedings{kong2025enhancing,
  author = {Sung Hye Kong and Saemee Choi and Wonwoo Cho and Sung Bae Park and Seung Shin Park and Jaegul Choo and Jung Hee Kim and Sang Wan Kim and Chan Soo Shin},
  title = {{Enhancing vertebral fracture prediction using multitask deep learning computed tomography imaging of bone and muscle}},
  booktitle = {European},
  year = {2025},
  doi = {10.1007/s00330-025-12049-3},
}
Figure