Per the PRISMA guidelines, our review process began with the identification of 377 records through database searches (Figure 1). After removing 157 duplicate records, the remaining articles underwent title and abstract screening. During this stage, several studies were excluded because they were reviews, conference papers, or other non-journal publications. A total of 72 articles were retrieved for full-text assessment. Following a detailed evaluation, 51 studies were found to meet the inclusion criteria, specifically focusing on the application of AI techniques to CT imaging for osteoporosis or bone mineral density analysis. These eligible studies were included in the final review and subjected to comprehensive data extraction and thematic synthesis.

The scoping review ultimately included 51 studies, which are systematically summarized in Table 1, detailing their objectives, research design, imaging sources, and participant demographics. Studies were published between 2019 and 2025 and originated from diverse geographic regions, with Asia—particularly China—accounting for nearly half of all publications (49%). The majority of investigations were single-center and retrospective in nature (84.3%), whereas only a limited proportion incorporated multicenter cohorts or external validation (15.7%), highlighting a general lack of large-scale validation across institutions.

Study objectives clustered around three main topics: osteoporosis diagnosis (45.1%), opportunistic screening (39.2%) and fracture-risk prediction (15.7%). Chest CT (33.3%), lumbar spine CT (15.7%) and abdominal CT (11.8%) were the most frequently utilized imaging sources, reflecting the increasing interest in opportunistic use of routinely acquired CT data. Participant cohorts were largely adult and elderly populations (typical age ranges spanning roughly from the third to the ninth decades), with a predominance of female participants, consistent with the epidemiology of osteoporosis.

To identify the sources of heterogeneity across studies, principal component analysis (PCA) was performed on nine key variables from the 51 included studies: number of centers, mean age, proportion of female participants, CT modality, anatomical scan region, segmentation strategy, feature extraction method, reference standard, and model performance (Figure 4). The first two principal components jointly explained 35.55% of the total variance (PC1: 18.78%; PC2: 16.77%).

PC1 was primarily driven by model performance (loading = 0.49), proportion of female participants (loading = −0.47), and mean age (loading = −0.40), indicating that studies involving younger populations with a lower female proportion tended to report higher model performance.

PC2 was mainly influenced by anatomical scan region (loading = −0.53), feature extraction strategy (loading = −0.51), and number of centers (loading = 0.43). These findings suggest that single-center studies focusing on core skeletal sites such as the lumbar spine or hip and employing combined HU and radiomics features achieved higher performance, whereas multicenter studies—often designed to reflect real-world clinical scenarios such as opportunistic screening using chest CT—tended to analyze non-core regions and simpler feature sets, resulting in slightly lower performance but improved generalizability.

Together, these findings quantitatively demonstrate that population characteristics, anatomical focus, and methodological choices jointly shape reported model performance across studies.

This scoping review synthesizes 51 studies and demonstrates that AI-enhanced CT imaging has progressed from proof-of-concept toward task-specific clinical utility in osteoporosis assessment. Rather than algorithmic complexity alone, model performance across studies was primarily determined by the alignment between clinical task (diagnosis, screening, or risk prediction), anatomical target, and analytical workflow, a pattern consistently supported by both subgroup analyses and PCA findings.

Osteoporosis diagnosis represented the most technically mature application. Across studies, diagnostic performance was consistently high (AUC 0.80–0.997), particularly when lumbar spine ROIs spanning L1–L4 or T12–L3 were used. This anatomical advantage is biologically plausible, as the lumbar spine contains a high proportion of trabecular bone, which is metabolically active and sensitive to early osteoporotic changes (58). Importantly, both HU-based quantitative workflows and radiomics-based approaches achieved comparable diagnostic accuracy, but through fundamentally different mechanisms. HU-based models predominantly captured global bone mineral density loss and showed excellent agreement with DXA and QCT (R² > 0.98), making them well suited for standardized diagnosis and large-scale deployment. In contrast, radiomics-based models leveraged texture and heterogeneity features to characterize trabecular microarchitecture, enabling improved sensitivity for early osteopenia where mean HU values alone may remain within normal ranges (59). These findings suggest that HU- and radiomics-based workflows should be viewed as complementary rather than competing strategies, optimized for different stages of skeletal deterioration.

Opportunistic screening emerged as a distinct translational pathway rather than a simple extension of diagnostic modeling. Nearly 40% of included studies repurposed chest or abdominal CT scans acquired for unrelated clinical indications, confirming the feasibility of embedding osteoporosis assessment into routine imaging workflows without additional radiation exposure. However, screening scenarios imposed stricter constraints on model design, including incomplete vertebral coverage, heterogeneous scan protocols, and the need for full automation. Accordingly, HU-based and end-to-end deep learning models predominated, prioritizing robustness and scalability over maximal accuracy. The observation that non-traditional sites, such as shoulder CT, could also yield high screening performance further challenges the conventional reliance on axial skeletal regions and underscores the adaptability of CT-based AI when anatomical context is carefully considered. Nevertheless, performance variability across scan regions highlights that opportunistic screening models are inherently context-dependent and cannot be indiscriminately transferred across imaging protocols.

In contrast to diagnosis and screening, fracture risk prediction remained the least mature and most heterogeneous application. Although multimodal models integrating imaging features with clinical variables consistently outperformed imaging-only approaches, overall performance was modest (AUC 0.70–0.92). This gap reflects a fundamental conceptual distinction: fracture risk is a prospective, multifactorial outcome driven not only by bone density and microstructure, but also by fall mechanics, muscle strength, neuromuscular coordination, and comorbidities (60, 61). CT-derived skeletal features capture only one component of this risk cascade, explaining why even anatomically matched, high-quality imaging fails to achieve diagnostic-level accuracy. These findings suggest that fracture prediction should be framed not as an extension of osteoporosis diagnosis, but as a distinct modeling problem requiring broader phenotypic integration.

Beyond task-specific differences, substantial heterogeneity was observed across studies, and the PCA analysis provides quantitative insight into its underlying drivers. The first principal component linked model performance with population characteristics, particularly age distribution and sex composition. Studies involving younger cohorts and lower proportions of female participants consistently reported higher accuracy, a pattern likely attributable to reduced imaging confounders. In elderly populations—especially postmenopausal women—degenerative changes, osteophytes, vertebral deformities, and prior fractures introduce complex imaging patterns that challenge both density-based and texture-based models (62). This observation highlights that reported performance metrics cannot be interpreted independently of population context.

The second principal component reflected methodological trade-offs between anatomical focus, feature strategy, and study scale. Single-center studies targeting core skeletal regions (lumbar spine or proximal femur) and employing richer feature representations tended to achieve higher peak performance. In contrast, multicenter studies—often designed to approximate real-world opportunistic screening—frequently analyzed non-core regions and relied on simpler features, resulting in slightly lower accuracy but improved generalizability. These findings reveal an inherent tension between methodological optimization and clinical realism, suggesting that maximal performance achieved under controlled conditions may not translate directly to routine practice.

Geographical and study-design bias further contributed to heterogeneity. Nearly half of the included studies were conducted in China, and most adopted single-center retrospective designs, reflecting the availability of large-scale CT archives and centralized healthcare systems. While this setting facilitates rapid model development and high internal performance, it may underrepresent interethnic variability and real-world practice patterns, potentially inflating reported accuracy. This imbalance underscores the need for prospective, multicenter, and cross-ethnic validation to ensure robust generalizability.

Additional heterogeneity arose from inconsistent ROI definitions, segmentation strategies, and reference standards. Even when identical vertebral levels were analyzed, validation against DXA versus QCT produced divergent performance estimates, reflecting fundamental differences in what these reference standards measure. Without harmonized methodological frameworks, cross-study comparisons risk conflating true model capability with study-specific design choices.

Despite strong technical performance, translation into routine clinical practice remains limited. Most studies prioritized algorithmic accuracy over prospective validation, workflow integration, or regulatory considerations. End-to-end deep learning models dominate the field due to their efficiency and automation, yet their limited interpretability may hinder clinician trust and adoption, particularly in screening contexts where false positives can have downstream consequences.

To bridge this gap, future research should build on our findings to address three critical gaps:

First, standardization: Adopt consensus protocols for ROI selection (e.g., L1-L4 for diagnosis, T7-L1 for chest CT screening) and dual reference standards (DXA + QCT) to enable cross-study benchmarking. Publicly available annotated datasets with standardized imaging parameters would accelerate this alignment.

Second, context-aware model design: Match workflows to clinical purpose—HU-based/end-to-end models for screening, radiomics for early diagnosis, and multimodal models (integrating imaging, clinical data, and fall risk factors) for fracture prediction—rather than pursuing a “one-size-fits-all” approach. Explainable AI (XAI) should be integrated into end-to-end models to highlight key decision-making features (e.g., HU thresholds, trabecular texture), enhancing clinical acceptance.

Third, prospective multicenter validation: Conduct large-scale studies (n > 10,000) to evaluate real-world performance, particularly for opportunistic screening in lung cancer or abdominal CT programs. Integration with PACS/EMR systems and attention to regulatory compliance are essential for routine deployment.

Ultimately, the greatest value of AI-enhanced CT may lie not in replacing existing standards, but in complementing them—enabling large-scale opportunistic screening, refining risk stratification, and identifying individuals who would benefit most from targeted DXA, QCT, or preventive intervention.

Our scoping review has several limitations. First, despite the use of a comprehensive search strategy, relevant studies published in non-English languages or indexed in non-major databases may have been missed. Second, considerable heterogeneity across included studies—particularly in CT acquisition parameters, reference standards, and AI modeling approaches—limited direct comparison and precluded quantitative synthesis. Third, this review focused solely on CT-based applications and did not include AI studies using other imaging modalities, such as X-ray or MRI, which may offer complementary insights. Finally, most included studies were retrospective and single-center, potentially limiting the generalizability of our findings to broader populations and real-world clinical settings.