Using artificial intelligence for distinguishing benign and malignant vertebral compression fractures by computed tomography: a scoping review

Abstract

Objective: The purpose of this scoping review is to compile the evidence regarding the use of computed tomography (CT) to distinguish between benign and malignant vertebral compression fractures (VCFs) in artificial intelligence (AI) applications. Methods: We systematically searched PubMed/MEDLINE, DOAJ, and ScienceDirect through October 2025 for studies reporting AI-based diagnostic approaches (machine learning with radiomics features, deep learning with convolutional neural networks, or hybrid models) applied to CT images in adult patients with VCFs. Eligible studies required diagnostic performance metrics and definitive reference standards. Results: Six retrospective studies from Asia and Europe have been found, involving 1,767 participants with 3,408 vertebrae, including 1,257 benign and 1,442 malignant vertebral fractures. The diagnostic performance of AI algorithms is high, achieving an area under the curve (AUC) score from 0.76 to 0.99. Deep learning approaches achieved performance comparable to experienced radiologists, while radiomics-based methods provided interpretable quantitative features. Hybrid models combining both approaches showed synergistic benefits. Three studies with external validation confirmed reasonable generalizability, though with some performance reduction in independent datasets. Conclusions: AI-assisted evaluation of CT scans also has good potential in differentiating benign from malignant VCFs, which could reduce unnecessary MRI referrals and improve the initial evaluation time. However, the data regarding this are limited because of retrospective studies with variable methods and limited multicenter data, which are still required to implement this in routine clinical practice.