Deep learning model for simultaneous vertebral level localization and fracture detection on DXA vertebral fracture assessment images Authors: Sang Wouk Cho, Dongin Moon, Chang Oh Kim, Hyeon Chang Kim, Yumie Rhee, Namki Hong Published: 2026 ABSTRACT AbstractObjectivesWhile vertebral fracture assessment (VFA) on dual-energy X-ray absorptiometry (DXA) images is widely used to detect osteoporotic vertebral fractures, existing deep learning approaches typically perform either fracture detection or vertebral localization, not both. We developed VERTE-X L, a deep learning model that simultaneously localizes vertebrae (T7–L5) and detects fractures on DXA VFA images, and evaluated its performance in a community-dwelling older adult cohort.MethodsWe used DXA VFA images from 1000 randomly sampled participants (age ≥65 years) in the Korean Urban Rural Elderly (KURE) cohort, split into training (N = 800) and test (N = 200) sets. VERTE-X L employs a DINOv2-L encoder, a localization decoder for per-vertebra coordinate prediction, and a consistent rank logits (CORAL) ordinal regression module for fracture detection. Performance was evaluated at image and vertebra levels and compared with previously reported detection-only models (VERTE-X and VERTE-X model fine-tuned on VFA data) and localization-only models (VerteNet and Multi-Context Hybrid CNN Transformer).ResultsOn the held-out test set, VERTE-X L achieved an image-level area under the receiver operating characteristic curve (AUROC) of 0.952 (sensitivity 89.7%, specificity 94.2%, net present value [NPV] 98.2%) and vertebra-level AUROC of 0.957 (NPV 99.7%). Localization success (within 15 pixels) was 94.7% for fracture vertebrae. VERTE-X L outperformed VERTE-X in detection (AUROC 0.952 vs 0.913, P = 0.034) and outperformed VerteNet in localization (mean radial error 10.0 vs. 20.3 px, P < 0.001).ConclusionsThe proposed method enables simultaneous fracture detection and vertebral localization on DXA VFA images.Keywords: Vertebral fracture, Vertebral localization, Dual-energy X-ray absorptiometry vertebral fracture assessment, Artificial intelligence, Image analysis Introduction Vertebral fracture (VF) is among the most consequential manifestations of osteoporosis in an aging society, and prevalent VF predicts subsequent vertebral and non-vertebral fracture as well as mortality independent of bone mineral density (BMD) [1,2]. Rapid and accurate radiologic diagnosis is therefore essential. Semiquantitative (Genant) grading [3] and algorithm-based qualitative (ABQ) assessment [[4], [5], [6], [7]] remain the most widely used diagnostic frameworks. However, both require expert reader labeling, which is time- and labor-intensive, and are subject to substantial inter- and intra-observer subjectivity. Consequently, VF remains underdiagnosed relative to its true prevalence: in a multinational study, local radiographic reading missed 34% of vertebral fractures identified by central, expert review [8]. Dual-energy X-ray absorptiometry (DXA)-based vertebral fracture assessment (VFA), alongside conventional spine radiography, has become established as a standard tool for VF diagnosis. Recent consensus recommendations endorse VFA as a routine component of osteoporosis care, performed concurrently with bone mineral density measurement [9]. Despite widespread clinical adoption, the uneven global distribution of DXA infrastructure limits access to VFA in many community and resource-limited settings [10]. Even where DXA is available, VFA interpretation continues to depend heavily on reader experience, and the need for standardized, reproducible reporting remains an unresolved challenge [11,12]. To address these limitations, deep learning models have increasingly been applied to musculoskeletal imaging [13,14]. A convolutional neural network trained on DXA VFA images alone has shown good discriminative performance for fracture detection in a large hospital-based registry and a community-based cohort [15,16], and a hybrid convolutional neural network (CNN)-transformer architecture has achieved accurate vertebral landmark localization in lateral spine DXA images [17]. However, these studies have each focused on a single task. Recently, explainable vertebral fracture assessment (XVFA) was proposed as a method that combines vertebra detection, keypoint-based localization, and rule-based fracture grading on lateral spine DXA images, with subsequent clinical validation reported [18,19]. XVFA detects all visible vertebrae without assigning them to specific anatomical levels, and its validation was performed in a single-sex cohort of older women. Whether a single framework can achieve accurate level-specific vertebral localization and fracture detection in a community-dwelling population remains to be established. In this study, we propose a framework that simultaneously performs vertebral localization and fracture detection using DXA VFA images from a cohort of community-based older adults, and we benchmark its performance against previously reported single-task models using the same held-out test set. Methods Study population and design This study used data from the Korean Urban Rural Elderly (KURE) cohort, a community-based prospective cohort of adults aged 65 years or older recruited from three urban districts and one rural district in Korea between 2012 and 2015 [20]. Among 3517 participants enrolled at baseline, 3508 had an available DXA VFA image. From this analysis cohort, a stratified random subsample of 1000 participants was selected, ensuring no participant-level overlap with the data used to fine-tune the previously published VERTE-X model [16], and divided into a training set (N = 800; 5-fold cross-validation) and a held-out test set (N = 200) (Fig. 1A). Within the training set, 5-fold cross-validation was used for hyperparameter tuning, model selection, and threshold determination. Baseline characteristics, including age, sex, body mass index (BMI), prevalent vertebral fracture, surgical implants, vertebroplasty, lumbar spine and femoral neck and total hip T-scores, and osteoporosis status, did not differ significantly between the training and test sets (all P > 0.05; Table 1). This study was approved by the Institutional Review Board (IRB No. 4-2012-0172), and the requirement for informed consent was waived for retrospective review of de-identified imaging data. Overview of the proposed study. (A) Study population flowchart. Among 3517 community-dwelling older adults from the KURE (Korean Urban Rural Elderly) cohort (2012–2015), 3508 had available DXA (dual-energy X-ray absorptiometry) VFA (vertebral fracture assessment) images. A random subsample of 1000 was selected and split into training (N = 800) and held-out test (N = 200) sets. (B) Architecture of the proposed two-stage framework. Stage 1 (localization) uses a DINOv2-L–based encoder to extract patch tokens from the resized DXA VFA image, which are decoded into per-vertebra center-point coordinates (T7–L5) via a localization decoder. Stage 2 (detection) extracts region-of-interest features around each localized vertebra and predicts ordinal fracture grade probabilities using a CORAL (COnsistent RAnk Logits) ordinal regression layer. Localization accuracy is evaluated using Success@15px (proportion of vertebrae with predicted centerpoint within 15 pixels of ground truth) and Fail@40px (proportion of vertebrae with localization error ≥40 pixels) thresholds. Table 1Baseline characteristics of the study population.Table 1CharacteristicOverall (N = 1000)Train (N = 800)Test (N = 200)P-valueAge (yr)71.97 ± 4.5772.01 ± 4.6571.81 ± 4.240.554Sex, Female n %664 (66%)537 (67%)127 (64%)0.332BMI (kg/m2)24.32 ± 2.9124.35 ± 2.9424.18 ± 2.830.460Prevalent Vertebral fracture (n)133 (13%)104 (13%)29 (15%)0.576Surgical implants (n)41 (4.1%)28 (3.5%)13 (6.5%)0.056Vertebroplasty (n)21 (2.1%)15 (1.9%)6 (3.0%)0.405LS T-score−1.3 ± 1.5−1.4 ± 1.5−1.3 ± 1.40.494FN T-score−1.7 ± 1.0−1.7 ± 1.0−1.7 ± 0.90.568TH T-score−0.9 ± 0.9−0.9 ± 1.0−0.9 ± 0.90.935Osteoporosis (n)343 (34%)282 (35%)61 (31%)0.206BMI, body mass index; LS, lumbar spine; FN, femoral neck; TH, total hip. Baseline characteristics of the study population. BMI, body mass index; LS, lumbar spine; FN, femoral neck; TH, total hip. DXA equipment and image annotation DXA VFA images were acquired using a Hologic Discovery W densitometer (Hologic Inc., Marlborough, MA, USA) in single-energy mode at the time of bone mineral density measurement, using a standard lateral spine acquisition protocol covering vertebral levels T4 through L5. All DXA VFA images from the KURE cohort at the index date were annotated using the modified algorithm-based qualitative (mABQ) method, as previously described [6,7,16]. In addition to fracture grading, the center point of every visible vertebral body—from the uppermost level visualized on each scan through L5—was manually annotated to establish ground-truth vertebral locations, independent of fracture status. Model development We developed VERTE-X L, a two-stage deep learning framework that performs vertebral localization (Stage 1) followed by fracture detection (Stage 2) on a single DXA VFA image (Fig. 1B). The model extends our previously reported VERTE-X framework [21], which was developed for image-level fracture detection only, by incorporating an explicit vertebral localization module (“L”). Image encoder and localization. DXA VFA images were processed through a DINOv2-L vision transformer encoder pretrained via self-supervised learning [22] (architecture details in Supplementary Fig. 1). The resulting patch-token embeddings were passed through a convolutional decoder to generate a localization heatmap for each of 11 vertebral levels (T7–L5), the range most consistently visualized across DXA VFA scans in this cohort. The pixel coordinate of each vertebra was obtained from the heatmap peak, and the corresponding peak activation value was retained as a localization confidence score for downstream analysis. Fracture detection. For each localized vertebra, a region-of-interest patch was extracted from the encoder feature map and fracture grade was predicted using an ordinal regression layer (consistent rank logits [CORAL] [23]), chosen for its suitability to the ordinal nature of mABQ grading. Binary fracture classification was derived from the predicted cumulative grade probabilities using P (grade ≥1) as the primary detection score. Comparator models VERTE-X L was compared against two previously reported single-task models, evaluated on the same held-out test set. VERTE-X is a convolutional neural network developed for image-level prevalent vertebral fracture detection on DXA VFA and lateral spine radiographs, without vertebral-level localization [16,21]. As VERTE-X was previously developed using DXA VFA images from the same KURE cohort acquired with the same Hologic Discovery W scanner, the fine-tuned model weights from our prior work were used directly for comparison without retraining. VerteNet is a hybrid CNN-transformer model developed for vertebral landmark localization in lateral spine DXA images and does not produce a fracture probability score [17]. As pretrained weights were not publicly released, the model architecture was obtained from the authors’ public repository and retrained on the same training set used to develop VERTE-X L, providing VerteNet with domain-matched training data for a fair comparison on the held-out test set. Outcome definitions and evaluation metrics Detection performance was evaluated at both the image level (presence of any fracture in the scan) and the vertebra level (fracture status of each individual vertebra). Discrimination was summarized using the area under the receiver operating characteristic curve (AUROC). Two operating points were defined by the Youden index, at which sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were reported. Localization performance was evaluated using the Euclidean pixel distance between the predicted and ground-truth vertebral center coordinates (mean radial error, MRE). Localization success was defined as Success@15px (MRE within 15 pixels; Fig. 1B), and localization failure was defined as Fail@40px (MRE exceeding 40 pixels). Gradient-weighted class activation mapping (Grad-CAM) [24] was used to generate attention maps illustrating the image regions contributing most to each fracture detection (Fig. S2) [Supplementary Materials]. Statistical analysis Continuous variables are presented as mean ± standard deviation (SD) and were compared using Welch's two-sample t-test; categorical variables are presented as n (%) and were compared using Pearson's chi-squared test or Fisher's exact test, as appropriate. Proportions are reported with 95% confidence intervals. AUROC was compared between VERTE-X L and the comparator detection model using the paired DeLong test [25,26]. Continuous localization evaluations were compared between models using a 1000 iterations cluster bootstrap to account for clustering of vertebrae within participants. As a sensitivity analysis, detection performance was re-evaluated using an alternative fracture definition (mABQ grade ≥ 2, excluding grade 1). Analyses were performed using R (version 4.4.1; R Foundation for Statistical Computing, Vienna, Austria). Results Study population Of the 1000 participants included in the analysis (mean age 71.97 ± 4.57 years; 66% women). Prevalent vertebral fracture was present in 133 participants (13%) overall and 29 participants (15%) in the test set (200 scans with 2000 vertebrae). Baseline characteristics were well balanced between the training and test sets (Table 1), and no significant differences were observed between the 1000-participant subsample and the remaining 2508 participants of the full cohort (Table S1) [Supplementary Materials]. Image- and vertebra-level fracture detection At the image level, VERTE-X L achieved an AUROC of 0.952 (95% CI 0.904–0.999). At the optimized threshold, sensitivity was 89.7% (95% CI 72.6%–97.8%), specificity 94.2% (95% CI 89.5%–97.2%), PPV 72.2% (95% CI 54.8%–85.8%), and NPV 98.2% (95% CI 94.7%–99.6%) (Table 2). At the vertebra level (N = 2000, prevalence 1.9%), AUROC was 0.979 (95% CI 0.961-0.998). Sensitivity was 86.8%, specificity 94.6%, PPV 23.9%, and NPV 99.7%. The low vertebra-level PPV reflects the low prevalence of fracture at the individual-vertebra level (Table 2). Table 2Fracture detection and localization performance of the VERTE-X L model on the held-out test set.Table 2Image-level detectionn (scans)200AUROC (95% CI)0.952 (0.904-0.999)Sensitivity, % (95% CI)89.7 (72.6–97.8)Specificity, % (95% CI)94.2 (89.5–97.2)PPV, % (95% CI)72.2 (54.8–85.8)NPV, % (95% CI)98.2 (94.7–99.6)Vertebra-level detectionan (vertebrae)2000AUROC (95% CI)0.979 (0.961-0.998)Sensitivity, % (95% CI)86.8 (79.2-95.2)Specificity, % (95% CI)94.6 (93.6-95.6)PPV, % (95% CI)23.9 (19.0-28.6)NPV, % (95% CI)99.7 (99.6-99.9)Fracture vertebra localizationn (fractured vertebrae)38Mean radial error (mean ± SD)8.4 ± 14.5Success@15px, % (95% CI)94.7 (82.3-99.4)Fail@40px, % (95% CI)5.3 (0.6-17.7)AUROC, area under the receiver operating characteristic curve; CI, confidence interval; Fail@40px, proportion of vertebrae with mean radial error ≥ 40 pixels; NPV, negative predictive value; PPV, positive predictive value; Success@15px, proportion of vertebrae with predicted centerpoint; SD, standard deviation.aCluster bootstrap 95% CI (1000 iterations, participant-level resampling). Fracture detection and localization performance of the VERTE-X L model on the held-out test set. AUROC, area under the receiver operating characteristic curve; CI, confidence interval; Fail@40px, proportion of vertebrae with mean radial error ≥ 40 pixels; NPV, negative predictive value; PPV, positive predictive value; Success@15px, proportion of vertebrae with predicted centerpoint; SD, standard deviation. Cluster bootstrap 95% CI (1000 iterations, participant-level resampling). Of the 38 fracture vertebrae, 32 (84.2%) were both correctly detected and localized within 15 pixels of the ground-truth center point. Across all 200 test scans, the mean number of false-positive vertebral alerts was 0.53 per scan. Sixty-eight scans (34.0%) contained at least one false-positive vertebral alert, with a mean of 1.5 false-positive alerts among these scans. Vertebral localization Among the 38 fracture vertebrae in the test set, MRE was 8.4 ± 14.5 pixels. Localization success rate was 94.7% (95% CI 82.3%–99.4%) at the 15-pixel threshold, with a failure rate (>40 pixels) of 5.3% (95% CI 0.6–17.7; Table 2). Across all 2000 vertebrae, MRE was 10.05 ± 20.37 pixels and Success@15px was 92.2% (95% CI 90.9%–93.3%). At the scan level, which requires correct localization of every vertebra within a scan, Success@15px was 83.0% (95% CI 77.1%–87.9%) overall, 75.9% (95% CI 56.5%–89.7%) for fracture-containing scans, and 84.2% (95% CI 77.9%–89.3%) for non-fracture scans (Table S2) [Supplementary Materials]. In neither case did vertebra-level localization failure propagate to an image-level false negative, indicating that the two-stage architecture does not silently suppress fracture-positive signals at the scan level when localization fails (Table S3) [Supplementary Materials]. At the individual-vertebra level, localization error was numerically higher at L1 and L3, the levels with the greatest morphological variability and the highest concentration of fractures (Table S4) [Supplementary Materials]. Localization confidence scores were stratified into tertiles to examine their relationship with localization accuracy. Vertebrae in the lowest confidence tertile had nearly 2.3-fold higher MRE (15.94 ± 28.27 px) and a 12.6–percentage point lower Success@15px (83.8% [95% CI 80.8%–86.5%]) compared with the mid and high confidence tertiles (MRE 7.1 px, Success@15px 96.4%; (Table S5) [Supplementary Materials]). Comparison with single-task models VERTE-X L significantly outperformed the detection-only comparator model (VERTE-X) at the image level (AUROC 0.952 vs. 0.913, DeLong P = 0.034; sensitivity 89.7% vs. 79.3%; NPV 98.2% vs. 96.3%). Restricting the fracture definition to mABQ grade ≥ 2 improved AUROC for both VERTE-X L (0.952 to 0.992) and the comparator detection model (0.913 to 0.968; Supplementary Table 6). For localization, VERTE-X L significantly outperformed the VerteNet across all 2000 vertebrae (MRE 10.0 ± 20.4 vs. 20.3 ± 39.6; Success@15px 92.2% vs. 81.5%; Fail@40px 7.2% vs. 18.1%; Table 3), with the difference in MRE remaining significant after scan-level cluster bootstrap adjustment (mean difference 11.19 px, 95% CI 5.95–16.71, P < 0.001). Table 3Comparison of the proposed model against published state-of-the-art models on the same test set.Table 3MetricVERTE-X L (Detection + Localization)VERTE-X (Detection)VerteNet (Localization)p value (vs. Ref)Image-level detectionn (with fracture/total scans)29/20029/20029/200AUROC (95% CI)0.952 (0.904-0.999)0.913 (0.853-0.974)0.034Sensitivity, %89.7079.30Specificity, %94.2091.80PPV, %72.2062.20NPV, %98.2096.30Vertebral localizationn (vertebrae)20002000Mean radial error (px, mean ± SD)10.1 ± 20.420.3 ± 39.60.026Success@15px, %92.2081.50Fail@40px, %7.2018.10AUROC, area under the receiver operating characteristic curve; CI, confidence interval; Fail@40px, proportion of vertebrae with mean radial error ≥40 pixels; NPV, negative predictive value; PPV, positive predictive value; Success@15px, proportion of vertebrae with predicted centerpoint; SD, standard deviation. Comparison of the proposed model against published state-of-the-art models on the same test set. AUROC, area under the receiver operating characteristic curve; CI, confidence interval; Fail@40px, proportion of vertebrae with mean radial error ≥40 pixels; NPV, negative predictive value; PPV, positive predictive value; Success@15px, proportion of vertebrae with predicted centerpoint; SD, standard deviation. Failure case characterization and visualization All false negatives at the image level occurred in vertebrae with only grade 1 fractures, with model scores near zero and indistinguishable from true negatives. Of 10 false-positive scans, 8 had only marginally elevated scores near the threshold boundary. Representative Grad-CAM visualizations of true-positive, true-negative, false-negative, and false-positive cases are shown in Fig. 2 and (Fig. S4) [Supplementary Materials]. True-positive cases showed attention concentrated at the endplate region of the fractured vertebra, including in the presence of vertebroplasty. True-negative cases showed diffuse, low-intensity attention without a focal hotspot. The false-negative and false-positive cases showed attention distributed across multiple vertebral levels without a clear focal site. Representative Grad-CAM visualizations illustrating model attention across three clinically relevant scenarios. Left panels show vertebra-level localization labels overlaid on the DXA VFA image; right panels show the corresponding Grad-CAM attention map. (A) True positive: an L1 fracture correctly identified, with attention sharply localized to the fractured vertebral body. (B) True negative: a scan with no vertebral fracture, showing diffuse, low-intensity attention without a focal hotspot, consistent with the absence of a fracture signal. (C) True positive: a T12 vertebra with vertebroplasty cement correctly identified as fractured, with attention concentrated precisely on the augmented vertebral body, indicating that the model attends to the true site of structural abnormality rather than being confounded by the cement artifact. Discussion In this study, VERTE-X L achieved high discriminative performance for both image- and vertebra-level fracture detection while simultaneously providing accurate vertebra-level localization, outperforming previously reported single-task comparator models for both detection and localization on the community-based older adults cohort. These findings demonstrate the potential of a single framework to deliver clinically useful vertebral fracture localization and detection on DXA VFA images. Beyond its discriminative performance, the image-level PPV of 72.2% should be interpreted in the context of a community-based screening population, in which vertebral fractures are relatively uncommon, and false-positive findings may increase the number of scans requiring radiologist review. In this setting, VERTE-X L may be better positioned as a prioritization and triage tool rather than as a standalone diagnostic system. By identifying scans that are more likely to contain vertebral fractures, the model could help direct radiologists’ attention to cases warranting further assessment, while positive predictions would still require expert confirmation. This potential role is consistent with AI-assisted screening approaches that have been incorporated into fracture liaison service workflows in other imaging modalities [27,28]. In addition, the vertebral localization output provides information on the suspected fracture level, which may further support targeted review of model-positive scans. The localization confidence score derived from the localization decoder correlated with localization error, stratifying localization quality into clinically meaningful tiers. This suggests a potential role as an automated quality-control flag, allowing low-confidence vertebrae to be triaged for manual review without requiring additional model components or inference cost, since the score is obtained directly from a single forward pass through the existing heatmap decoder. Such confidence-aware triage may be particularly relevant in this pipeline, where unreliable localization can propagate downstream and affect the region-of-interest extraction that underlies fracture grading. Several prior approaches to vertebral fracture detection in other imaging modalities, such as computed tomography (CT), have adopted a two-stage architecture in which vertebral localization and fracture classification are performed by separate, independently trained networks with no shared parameters [29,30]. A similar pipeline structure is employed by XVFA [19], which combines a bounding box detector, a keypoint regressor, and a separate fracture classifier on lateral spine DXA images. While this approach offers strong explainability through rule-based Genant grading, it does not assign detected vertebrae to specific anatomical levels and relies on independently optimized components without a shared feature space. In such pipelines, information learned by the localization network is not directly accessible to the downstream classification network beyond the cropped region of interest. In contrast, VERTE-X L performs both tasks within a single framework built on a shared DINOv2-L encoder, allowing localization- and detection-relevant representations to be learned jointly from a common feature space rather than passed through a hard, one-way interface between independently optimized models. DINOv2 [22] is a self-supervised vision foundation model whose representations have demonstrated strong cross-task and cross-modality transferability in radiologic imaging, including classification and segmentation tasks across X-ray, CT, and magnetic resonance imaging (MRI) [31]. Accordingly, the pretrained encoder weights used in VERTE-X L may also be applicable to downstream tasks beyond fracture detection and localization on DXA VFA images, such as osteoporosis classification or vertebral morphometry, without requiring a fully separate model to be trained from scratch. Qualitative review of Grad-CAM attention maps revealed a notable difference between VERTE-X L and the original VERTE-X model: whereas VERTE-X attention was concentrated directly on the vertebral body [16,21], VERTE-X L attention was more frequently localized to the intervertebral space, particularly at the endplate region. In non-fractured scans, attention was distributed broadly across the visualized spine, whereas in fractured scans, attention was sharply restricted to the regions immediately above and below the fracture site. A similar focal pattern was observed even in the presence of vertebroplasty or surgical instrumentation. This pattern closely parallels the visual process underlying the mABQ method, in which endplate depression is the primary feature used to identify fracture, rather than overall vertebral body height [[4], [5], [6], [7]]. We speculate that joint training with an explicit localization objective may have encouraged some degree of task specialization within the shared encoder, wherein the localization module learns features related to overall vertebral body position and structure, while the detection module attends more specifically to local morphological cues at the endplate. This pattern is consistent with prior observations in multi-task learning architectures, where task-specific branches sharing a common encoder have been shown to develop distinct attention patterns aligned with each task's underlying objective, such as geometry-oriented versus semantics-oriented attention in joint depth-estimation and segmentation networks [32,33]. Failure case analysis showed that all false negatives at the Youden threshold occurred at the grade 1 fracture boundary and that excluding grade 1 fractures improved AUROC, sensitivity, and NPV. This pattern is consistent with prior deep learning models for vertebral fracture detection, which have similarly reported substantially lower discriminative performance for grade 1 fractures compared with grade 2 or grade 3 fractures using Genant grading [30,34]. It also mirrors a well-recognized challenge in human reader performance using mABQ-based classification: inter-reader agreement for the full ordinal grading scale, which requires distinguishing mild from no fracture, has been reported as only fair to moderate (Fleiss kappa 0.29–0.54), whereas agreement improves substantially when mild deformities are grouped with non-fracture and only clinically significant fractures are considered (kappa 0.68–0.86) [35]. These findings suggest that VERTE-X L's errors concentrate at a genuinely ambiguous clinical boundary shared by both AI models and expert human readers, rather than reflecting a systematic model deficiency. This study has several limitations. First, the model was drawn from a single community cohort of Korean older adults using a single DXA manufacturer and was not evaluated in a sex-stratified manner due to the small number of fracture-positive men in the test set. Differences in vertebral morphology and BMD between Korean and Western populations, as well as differences in image characteristics across DXA manufacturers, may limit the direct applicability of VERTE-X L to other ethnic groups or scanner platforms without additional fine-tuning. External validation in diverse populations and across multiple DXA manufacturers is needed for future work. Second, the number of fracture vertebrae was small in several regional and subgroup analyses, limiting the precision of stratum-specific estimates; these findings should be considered hypothesis-generating. Third, the model was developed and evaluated for vertebral levels T7–L5, the range most consistently visualized in this cohort. Although fractures at T4–T6 were rare in the annotated subsample, protocols with broader upper thoracic coverage would require additional validation. In addition, a formal inter-reader reliability assessment was not performed. Fourth, the retrained comparator model may underperform relative to its original reported performance due to differences in training data and hyperparameter optimization. Fifth, this study evaluated prevalent fracture detection using a cross-sectional design, but the prognostic value of VERTE-X L–detected fractures for predicting incident fractures or mortality, as has been shown for prior VERTE-X models [16,21], remains to be evaluated. Conclusions To summarize, VERTE-X L achieved high discriminative performance for fracture detection at both the image and vertebra level, while simultaneously providing accurate vertebra-level localization within the same framework. VERTE-X L outperformed previously reported single-task comparator models in both detection and localization using the same holdout test set. By utilizing a single, shared framework for both tasks, VERTE-X L demonstrates the potential to deliver clinically useful vertebral fracture localization and detection on DXA VFA images. CRediT author statement Sang Wouk Cho: Conceptualization, Software, Methodology, Formal analysis, Investigation, Writing-original draft, Visualization, Writing-review and editing. Dongin Moon: Methodology, Formal Analysis, Investigation. Chang Oh Kim: Writing-review and editing, Resources, Data Curation. Hyeon Chang Kim: Writing-review and editing, Resources, Data Curation. Yumie Rhee: Writing-review and editing, Supervision, Resources, Data Curation. Namki Hong: Conceptualization, Project Administration, Funding Acquisition, Investigation, Resources, Methodology, Writing-review and editing. Data sharing statements Data sharing requests will be considered for the research purpose upon a written request to the corresponding author. If agreed, deidentified participant data and/or deep learning model weights will be made available, subject to a data sharing agreement. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work, the authors used Claude (Anthropic PBC, San Francisco, CA, USA) for language editing assistance, including grammar checks and sentence readability improvement. The authors reviewed and edited the content as needed and take full responsibility for the content of the publication. Conflicts of interest No potential conflict of interest relevant to this article was reported. Acknowledgments This work was supported by the Korea Disease Control and Prevention Agency (2013-E63007-01, 2013-E63007-02, 2016-ER6302-00; 2016-ER6302-01; 2016-ER6302-02; 2019-ER6302-00; 2019-ER6302-01), the Korea National Institute of Health (KNIH) research project (2023ER080700; RS-2026-25592627; RS-2026-25548321), a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (HI22C0452) to NH, a Young Medical Scientist Research Grant through the Seokchunnanum Foundation (SCY2507P) to NH and a grant of International Cooperation (R&D) (Joint Research) with Switzerland through the National Research Foundation of Korea(NRF), funded by the Ministry of Science and ICT, Republic of Korea (RS-2023-00231864) to NH. The funding sources had no involvement in study design, data collection, analysis, interpretation of data, writing of the report, and decision to submit the paper. We appreciate the effort of the participants and investigators of the KURE cohort study. ORCID Sang Wouk Cho: 0000-0001-6780-1070. Dongin Moon: 0009-0004-0457-2847. Chang Oh Kim: 0000-0002-0773-5443. Hyeon Chang Kim: 0000-0001-7867-1240. Yumie Rhee: 0000-0003-4227-5638. Namki Hong: 0000-0002-8246-1956. Footnotes This article is part of a special issue entitled, “AI in AFOS,” published in Osteoporosis and Sarcopenia. Supplementary data to this article can be found online at https://doi.org/10.1016/j.afos.2026.09.002. Supplementary data The following is the Supplementary data to this article: Multimedia component 1 References Associated Data Supplementary Materials Multimedia component 1