Accuracy of Low-Dose Chest CT-Based Artificial Intelligence Models in Osteoporosis Detection: A Systematic Review and
Huang Ya'nan1, Zhou Jianfeng2, Tang Wei1
1Department of Radiology, Shaoxing People's Hospital (Shaoxing Hospital of Zhejiang University), No. 568 Zhongxing North Road, Shaoxing, 312000, Zhejiang, China.
Abstract:
The purpose of this study is to systematically review and evaluate the accuracy of low-dose chest CT-based artificial intelligence in osteoporosis screening. A systematic literature search for relevant studies up to 13th December 2024 was performed in the PubMed, Scopus, Web of Science, and Cochrane Library databases. This meta-analysis was conducted in accordance with the PRISMA-DTA statement. Modified QUADAS-2 was used to assess the methodological quality of the studies. Quantification bias metrics were extracted to evaluate the performance of the AI models for vertebrae segmentation and labeling based on low-dose chest CT images. Pooled sensitivity, specificity, and area under the curve (AUC) were calculated. To assess publication bias, Egger test and funnel plot were conducted. Meta-regression and subgroup analysis were performed to explore potential heterogeneity. Eight studies suitable for the analysis were included. The pooled Dice similarity coefficient (DSC) for automatic vertebrae segmentation was 0.92 (95% CI 0.88-0.97). For the diagnosis of abnormal (osteoporosis + osteopenia) or osteoporosis participants, respectively, pooled sensitivities were 0.90 (95% CI 0.88-0.91) and 0.86(95% CI 0.82-0.89); pooled specificities were 0.90 (95% CI 0.88-0.91) and 0.93 (95% CI 0.92-0.94); and summary receiver operating characteristic (SROC) curves were 0.9653 and 0.9676. Meta-regression and subgroup analyses identified potential sources of heterogeneity, including result source (external dataset vs. internal dataset), ROI annotations (one radiologist vs. two radiologists), model developed with or without radiomics, and VBs segmentation output (included lumbar spine vs. only thoracic spine) (P < 0.05). The low-dose chest CT-based AI model shown promise information for identifying patients with osteoporosis or osteopenia who need further evaluation. Further prospective multi-center, multi-dataset studies are still required to assess the complementary role of the AI model in osteoporosis and osteopenia diagnosis through low-dose chest CT images.
