# Artificial intelligence in osteoporosis assessment using CT imaging: a scoping review

## TL;DR

This review maps how AI is used in CT scans to assess osteoporosis, highlighting performance and methodological challenges.

## Contribution

A systematic scoping review of AI applications in CT-based osteoporosis assessment, emphasizing methodological diversity and performance trends.

## Key findings

- AI models for osteoporosis diagnosis and screening achieved high accuracy (AUC 0.781–0.997).
- Fracture-risk prediction models showed lower accuracy (AUC 0.702–0.92) compared to diagnostic models.
- Methodological diversity and inconsistent validation limit generalizability and cross-study comparisons.

## Abstract

This scoping review aimed to systematically summarize and map current research on the application of artificial intelligence (AI) in CT-based osteoporosis assessment, with a focus on methodological approaches, anatomical target regions, and reported algorithmic performance across existing studies.

PubMed, EMBASE, and Web of Science databases were searched for studies published between January 1995 and December 2025. Eligible studies applied AI, machine learning, or deep learning techniques to CT images for osteoporosis classification, bone mineral density (BMD) estimation, or fracture-risk prediction. Data extraction covered study characteristics, imaging sources, analytical workflows, and validation methods.

A total of 51 studies were included. Most were retrospective (84.3%) and single-center (84.3%), with nearly half conducted in China. Study objectives clustered around osteoporosis diagnosis (45.1%), opportunistic screening (39.2%), and fracture-risk prediction (15.7%). Diagnostic and screening models generally achieved high performance (AUC 0.80–0.997 and 0.781–0.99, respectively), whereas fracture-risk prediction showed more modest accuracy (AUC 0.702–0.92). Across studies, technical workflows varied widely, encompassing Hounsfield Units (HU)-based quantitative analyses, radiomics-based models, end-to-end deep learning, and multimodal approaches. Such methodological diversity, combined with inconsistent validation strategies, limits direct comparison and reduces overall generalizability.

Current evidence shows that AI-enhanced CT can achieve diagnostic and screening performance comparable to DXA and QCT, although fracture-risk prediction still requires improvement through multimodal data integration. However, methodological heterogeneity and the lack of standardized workflows limit cross-study comparability and clinical translation. Integrating AI into routine CT pipelines may reduce screening costs, enable earlier detection and intervention, and help mitigate the global burden of osteoporosis.