End-to-end deep learning pipeline for opportunistic vertebral fracture and Schmorl’s node detection from routine CT scans

1 Introduction

Osteoporotic vertebral compression fractures (VCFs) are highly prevalent, yet frequently underdiagnosed in aging populations [1]. Since patients undergoing chest or abdominal CT scans are typically not evaluated for bone health, vertebral abnormalities—such as fractures and Schmorl’s nodes [2]—often go unnoticed. Opportunistic detection offers a promising strategy to address this care gap by repurposing existing clinical imaging for secondary screening of skeletal conditions [3, 4].

In this context, we introduce a fully automated machine learning pipeline for vertebral assessment using chest and abdominal CT scans originally acquired for non-spinal clinical indications. The system automatically localizes vertebrae, detects fractures, assesses their severity, and identifies the presence of Schmorl’s nodes. By enabling opportunistic fracture detection, the pipeline contributes to reducing the diagnostic gap in osteoporosis care and supports earlier clinical intervention.

Our pipeline consists of two main stages, segmentation and classification. The segmentation model is based on YOLOv11 and trained using a combination of three public datasets [CTSpine1K [5], VerSe2019 [6], VerSe2020 [7]], providing robust anatomical generalization. It isolates individual vertebrae from full CT volumes with high accuracy. The classification stage is powered by a second YOLOv11 model trained on a curated dataset of 5,000 CT scans from AUBMC, annotated by radiology experts using the Genant fracture grading scale [8].

The classification pipeline includes a triage classifier (normal, fractured, unknown), a fracture severity grader, and a Schmorl’s node detector. Multi-label outputs are generated for each vertebra, allowing clinicians to interpret complex scenarios such as co-occurrence of Schmorl’s nodes and vertebral fractures. To promote real-world usability, the entire workflow is embedded into a web-based interface that enables visualization and confidence score review.

This work contributes to the field by combining opportunistic imaging analysis with robust deep learning methodology, advancing the development of practical AI tools that can enhance musculoskeletal diagnosis in routine clinical settings. To our knowledge, this is among the first deployed systems combining Genant-based severity grading and Schmorl’s node detection in an end-to-end opportunistic CT pipeline.

2 Literature review

2.1 The clinical challenge of underdiagnosed vertebral fractures

Osteoporotic vertebral compression fractures (VCFs) represent one of the most prevalent yet critically underdiagnosed skeletal complications in aging populations [1, 9]. Despite their significant impact on morbidity, mortality, and quality of life [10], VCFs frequently remain undetected in clinical practice. This diagnostic gap stems from multiple factors: traditional osteoporosis screening methods are substantially underutilized due to high costs, limited accessibility, and the logistical burden of scheduling separate imaging appointments [11]. Moreover, vertebral fractures are often asymptomatic or present with non-specific symptoms, leading to their oversight during routine clinical evaluations.

The emergence of opportunistic screening—the practice of extracting additional diagnostic information from imaging studies performed for other clinical indications—has gained considerable attention as a pragmatic solution to this healthcare challenge [4, 12]. Chest, abdominal, and pelvic CT scans, routinely performed for non-spinal conditions, incidentally capture vertebral structures and thus present an untapped opportunity for identifying undiagnosed fractures. Recent studies have demonstrated that systematic implementation of opportunistic screening could substantially reduce the care gap in osteoporosis diagnosis and fracture risk stratification [13].

2.2 Deep learning approaches for automated vertebral fracture detection

The application of artificial intelligence, particularly deep learning, to automated vertebral fracture detection has evolved rapidly over the past decade. Early implementations focused primarily on binary fracture detection, with Bhatnagar et al. [14] providing a comprehensive overview of the field’s progression. Contemporary approaches predominantly employ convolutional neural networks (CNNs), which have demonstrated remarkable capability in learning hierarchical feature representations from medical imaging data [15].

U-Net and its variants have emerged as the architecture of choice for vertebral segmentation tasks, owing to their encoder-decoder structure that preserves both high-level semantic information and fine-grained spatial details [16]. Building upon these foundations, several research groups have developed specialized pipelines for vertebral fracture detection. Zhang et al. [17] achieved notable success with a multicenter cohort study demonstrating high-accuracy fracture localization from spinal radiographs, while Kim et al. [18] compared the efficacy of anteroposterior vs. lateral x-ray projections, revealing projection-dependent performance variations.

For radiograph-based analysis, Paik et al. [19] implemented a Mask R-CNN approach trained on 487 lateral thoracolumbar radiographs containing 598 fractures, achieving 79.8% sensitivity and 85.7% accuracy in vertebral fracture detection. For CT-based analysis, Glessgen et al. [20] introduced a comprehensive deep learning pipeline for systematic vertebral fracture reporting, demonstrating the feasibility of integrated clinical deployment. External validation studies have further substantiated these findings, with Nicolaes et al. [21] reporting a sensitivity of 94% and specificity of 93% in identifying scans with vertebral fractures across 4,810 CT scans from Chinese patients, and Guenoun et al. [22] providing systematic performance evaluation of automated fracture quantification on opportunistic CT scans.

2.3 Clinical translation and commercial implementation

The transition from research prototypes to clinically deployable systems has begun to materialize, though with mixed results. Bendtsen and Hitz [23] evaluated a commercial algorithm (HealthVCF, Zebra Medical Imaging) in a real-world clinical setting, reporting 68% sensitivity (CI 0.581–0.776) and 91% specificity (CI 0.89–0.928) based on a performance evaluation of 1,000 CT scans out of approximately 10,000 processed. While these metrics demonstrate clinical utility, they also reveal the persistent challenge of balancing sensitivity and specificity in diverse clinical populations. Similarly, Page et al. [13] reported that machine learning algorithms could identify the majority of vertebral compression fractures in opportunistic screening scenarios, though performance varied with fracture severity and imaging quality.

Kim et al. [3] extended this work by focusing specifically on acute vertebral fractures in routine abdominal and chest CT scans, demonstrating that automated detection could facilitate earlier clinical intervention. These real-world implementations have highlighted both the promise and the limitations of current AI approaches, particularly regarding their handling of subtle or borderline cases.

2.4 Advanced techniques for model robustness and generalization

Recognizing the limitations of models trained exclusively on available clinical data, researchers have increasingly explored data augmentation strategies to enhance model robustness and generalization. El Kojok et al. [24] investigated the use of variational autoencoders (VAEs), generative adversarial networks (GANs), and transfer learning to augment spine CT scan datasets, demonstrating improved detection rates for vertebral compression fractures. Their work aligns with broader trends in medical imaging, where generative models have shown promise in addressing data scarcity challenges [25, 26].

2.5 Persistent gaps in current approaches

Despite these advances, several critical limitations remain unaddressed in the current literature. First, the vast majority of studies focus solely on binary fracture detection, neglecting fracture severity grading despite its central importance in clinical decision-making. The Genant semi-quantitative method [8], which classifies vertebral deformities as mild, moderate, or severe based on vertebral height loss, remains the clinical gold standard but is rarely integrated into automated AI pipelines [27].

Second, co-occurring vertebral pathologies such as Schmorl’s nodes—disc material herniations into vertebral endplates associated with degeneration and trauma [2]—are typically excluded from analysis despite their diagnostic and prognostic significance. The simultaneous assessment of multiple vertebral abnormalities would provide clinicians with a more comprehensive evaluation but requires multi-label classification frameworks that are uncommon in current systems.

Third, many existing approaches treat vertebral segmentation, anatomical labeling, and pathology classification as separate, sequential tasks requiring manual quality control between stages. This fragmentation reduces efficiency and impedes seamless integration into clinical workflows. The absence of end-to-end pipelines that unify these components represents a substantial barrier to clinical adoption.

Finally, while several studies have developed promising algorithms, few have progressed to actual clinical deployment with continuous performance monitoring and iterative improvement based on radiologist feedback. The gap between research validation and sustained clinical implementation remains wide.

2.6 Study objectives and novel contributions

The present study addresses these limitations through several key innovations. We developed a fully automated, end-to-end pipeline that integrates vertebral segmentation, anatomical labeling, multi-class fracture detection, Genant-based severity grading, and Schmorl’s node identification. By leveraging the YOLOv11 architecture—selected for its computational efficiency and strong performance in object detection and classification tasks—we created a unified framework that eliminates manual intervention between processing stages. Furthermore, we incorporated a continuous learning mechanism whereby radiologist corrections are systematically collected and used for bi-weekly model retraining, ensuring ongoing performance optimization in real-world deployment. To facilitate clinical adoption, we developed a comprehensive web-based platform that enables real-time image processing, interactive result visualization, and confidence score evaluation. This work thus represents a significant advance toward practical, deployable AI solutions for opportunistic vertebral health assessment in routine clinical care.

3 Datasets

This study utilizes a combination of publicly available datasets and a proprietary institutional dataset to develop and assess the segmentation and classification modules of the proposed vertebral analysis pipeline. Segmentation is trained on a diverse set of open-source CT spine scans, while classification relies on a clinically curated dataset from AUBMC, annotated with a custom-designed vertebral labeling tool.

3.1 Public datasets for segmentation

To train a robust and generalizable vertebra segmentation model, we combined three publicly available datasets: All datasets were harmonized through isotropic resampling and standardized anatomical labeling. This aggregation strategy enabled the YOLOv11-based segmentation model to generalize across different imaging protocols and patient anatomies.

• CTSpine1K: A collection of 1,005 CT volumes with over 11,100 labeled vertebrae that provide comprehensive annotations for vertebral levels. The dataset spans a broad demographic range and includes various spine pathologies.
• VerSe2019: Released for the MICCAI 2019 Vertebrae Segmentation Challenge, this dataset includes 160 CT scans with expert-labeled vertebrae for cervical, thoracic, and lumbar regions. It is widely adopted as a benchmarking standard.
• VerSe2020: An extension of VerSe2019, this dataset contains 319 CT image series from 300 subjects, offering more diverse anatomical coverage and pathological variations. It is designed to promote clinically relevant segmentation performance.

3.2 Institutional dataset for vertebra classification

This study protocol was reviewed and approved by the Institutional Review Board (IRB) at American University of Beirut Medical Center (AUBMC). Given the retrospective design and the exclusive use of de-identified data, the requirement for individual written informed consent was waived by the IRB. To support vertebra classification tasks—including fracture detection, severity grading, and Schmorl’s node identification—we assembled a large-scale institutional dataset comprising 5,000 CT scans obtained from the AUBMC. The scans were retrospectively collected between 2016 and 2020 from patients who underwent abdominal, pelvic, or lumbar CT imaging for non-spinal clinical indications. An initial list generated in 2021 included 24,095 studies from patients aged 45 years and older. This list was later refined in 2022 to 5,000 studies by restricting the cohort to patients aged 65 years and older, yielding the final dataset used for training, validation and testing. The decision to restrict the cohort to patients aged 65 and older was motivated by the substantially higher prevalence of osteoporotic vertebral compression fractures in this age group, which provided greater disease enrichment and a more clinically relevant population for training the classification models. Accordingly, the intended target population of the deployed system is patients aged 65 years and older undergoing routine CT imaging. Generalizability to younger populations (ages 45–64) has not been evaluated in this study and remains a direction for future work. This opportunistic sampling strategy enabled population-level screening for vertebral compression fractures (VCFs), helping to address the care gap in osteoporosis diagnosis. A representative example from the dataset is shown in Figure 1.

Among the 5,000 CT scans analyzed, 1,500 contained one or more vertebral fractures, while the remaining 3,500 were fracture-free. The annotation process was carried out by two experienced radiologists from the Department of Diagnostic Radiology at AUBMC. In the initial phase, 500 CT scans were independently annotated by both radiologists using the Genant semi-quantitative grading method. All disagreements between the two annotators were reviewed and resolved through consensus discussion to establish a consistent annotation standard. Following this calibration phase, the remaining scans were divided between the two radiologists, with each annotating independently based on the agreed-upon criteria established during the initial consensus process.

To ensure standardized and efficient annotation of vertebrae, we developed a custom-built annotation tool tailored for DICOM image navigation and structured metadata entry. This interface allowed radiologists to scroll through sagittal slices and annotate each vertebra by selecting its anatomical level, classification (Normal, Mild, Moderate, Severe), and indicating the presence of Schmorl’s nodes and any additional abnormalities such as sclerosis or lysis. A screenshot of the annotation interface is shown in Figure 2.

Fractures were categorized using the Genant semi-quantitative method, which defines three severity grades based on the degree of vertebral height reduction. Table 1 summarizes the grading criteria, while Figure 3 visually illustrates the differences between severity levels.

Table 1. Genant fracture classification.

Following segmentation, each CT scan was decomposed into individual vertebra-level crops. Each vertebra was labeled with multiple attributes, including classification (Normal, Mild, Moderate, Severe) and Schmorl’s node presence (Yes/No). The full dataset was randomly partitioned into training (80%), validation (10%), and testing (10%) sets, ensuring balanced representation across all severity categories and pathological labels.

4 Methodology

This section outlines the complete methodology adopted in developing the end-to-end vertebra classification pipeline, including segmentation and classification models, and the final deployment strategy. Each component is described in detail to emphasize reproducibility and clinical relevance.

4.1 YOLOv11 architecture and model selection

The YOLO (You Only Look Once) family has evolved through multiple generations, with YOLOv11 representing a recent advancement in real-time computer vision. YOLOv11 supports a wide range of tasks, including object detection, instance segmentation, image classification, pose estimation, and oriented object detection, making it a unified framework for multi-task visual understanding.

YOLOv11 provides five scaled model variants of increasing capacity: nano (n), small (s), medium (m), large (l), and extra-large (x). These variants are designed to accommodate diverse deployment scenarios ranging from resource-constrained edge devices to high-performance computing environments. All variants share a common architectural design with compound scaling of depth and width, while incorporating modern building blocks such as C3k2 modules, Spatial Pyramid Pooling–Fast (SPPF), and attention-enhanced components (C2PSA), which improve feature extraction and efficiency.

For the classification models, official Ultralytics configurations report the following parameter counts and computational complexities: YOLOv11n-cls (1.63M parameters, 0.5 GFLOPs), YOLOv11s-cls (5.55M, 1.6 GFLOPs), YOLOv11m-cls (10.46M, 5.0 GFLOPs), YOLOv11l-cls (12.94M, 6.2 GFLOPs), and YOLOv11x-cls (28.46M, 13.7 GFLOPs). These values are directly obtained from the official model configuration files and provide a reliable reference for model capacity and computational cost.

For the segmentation task, we employed the YOLOv11l-seg variant. Vertebral segmentation and anatomical labeling across 26 levels (C1–S1) constitute a fine-grained and high-complexity problem that requires strong spatial reasoning and detailed feature representation. Larger YOLOv11 variants provide increased representational capacity and have been shown to improve accuracy in complex visual tasks, at the cost of higher computational demand. The large variant therefore provides a suitable trade-off between accuracy and efficiency for clinical deployment scenarios involving volumetric CT data.

For the classification tasks (fracture detection, Schmorl’s node identification, and severity grading), we employed the YOLOv11m-cls variant (10.46M parameters, 5.0 GFLOPs). Since classification operates on cropped vertebral regions rather than full-resolution CT slices, the spatial complexity is significantly reduced. The medium variant thus provides sufficient representational capacity while maintaining lower computational overhead, enabling efficient and scalable per-vertebra inference during deployment.

4.2 Vertebra segmentation model

To extract, segment, and label individual vertebrae from volumetric CT scans, we trained a vertebra detection and segmentation model based on the YOLOv11l-seg architecture appropriately. YOLOv11 was selected for its computational efficiency, real-time inference capability, and strong performance in both object detection and instance segmentation tasks. Compared to U-Net and nnU-Net, which are primarily designed for semantic segmentation, YOLOv11 natively supports simultaneous object detection, segmentation, and classification within a unified framework, making it well-suited for the joint vertebral localization and labeling task required by our pipeline. Furthermore, its single-pass architecture enables substantially faster inference compared to transformer-based models, which is critical for clinical deployment where processing speed directly impacts workflow integration. The model was trained on a unified dataset combining CTSpine1K, VerSe2019, and VerSe2020, which provide both vertebral masks and corresponding vertebral level annotations. These datasets were harmonized through isotropic resampling and consistent anatomical labeling, enabling the model to segment each vertebra and assign its anatomical level (e.g., C1, T1, L1). The dataset was randomly split into training (80%), validation (10%), and testing (10%) subsets, ensuring balanced representation across fracture severity classes. To improve generalization, the training process incorporated data augmentation techniques such as random rotations, scaling, brightness normalization, and left–right flipping.

To handle cases where the field of view did not fully capture the upper cervical or lower sacral vertebrae, we adopted a Top-N prediction strategy. For each vertebra, the model produced multiple candidate detections ranked by class probability. We retained the top three predictions, allowing the system to account for alternative possibilities when vertebral boundaries were incomplete or ambiguous. Segmentation accuracy was then evaluated using both Top-1 (highest-ranked prediction only) and Top-3 (any of the top three predictions) metrics, providing a more robust assessment of performance under challenging imaging conditions. It is important to note that the Top-N accuracy metric refers specifically to vertebral anatomical labeling accuracy (i.e., whether the correct vertebral level was assigned), while segmentation mask quality was separately assessed using mean Intersection over Union (mIoU).

4.3 Multi-stage vertebra classification

Following segmentation, each vertebra is cropped and passed through a multi-stage classification pipeline composed of three models: All classifiers were implemented using the YOLOv11 framework, fine-tuned separately for each task. Models were trained with cross-entropy loss and evaluated using standard classification metrics, including precision, recall (Sensitivity), Specificity, F1-score, and confusion matrices. Early stopping and learning rate decay were used to prevent overfitting. Given the class imbalance in the institutional dataset (3,500 non-fracture vs. 1,500 fracture cases at scan level), the training data was balanced at the vertebra level through stratified sampling to ensure adequate representation of all classes during training.

• Primary Classification: This stage assigns each vertebra to one of three mutually exclusive classes: normal, fractured, or unknown. The “unknown” class accommodates ambiguous cases where image quality or anatomical distortions prevent confident classification. Specifically, a vertebra is assigned to the “unknown” category when factors such as motion artifacts, partial field-of-view coverage, severe degenerative changes, or post-surgical hardware obscure the vertebral body to the extent that a reliable fracture assessment cannot be made. In clinical practice, vertebrae flagged as “unknown” are routed for mandatory radiologist review, ensuring that no actionable findings are missed.
• Schmorl’s Node Detection: For vertebrae labeled as either normal or fractured, a binary classifier is applied to determine the presence of Schmorl’s nodes. This model was trained on annotated data with consensus labels provided by the radiologists.
• Fracture Severity Grading: Vertebrae classified as fractured are routed to a third classifier that assigns one of three grades: mild, moderate, or severe.

4.4 Pipeline architecture and visualization

The complete pipeline integrates segmentation and classification in a sequential and modular framework. Figure 4 illustrates the pipeline architecture. A CT-scan is first processed by the segmentation model, producing masks for each vertebra. These vertebrae are cropped and forwarded to the primary classifier. Depending on the output class, each vertebra proceeds through one or both of the secondary classifiers (Schmorl and severity). Final outputs are aggregated per vertebra, supporting multi-label outputs that capture both structural abnormalities and severity assessments.

The pipeline was implemented in Python using PyTorch and YOLO for model deployment. FastAPI was used to expose the APIs, enabling efficient communication between modules. Each component was encapsulated as a containerized service to support modular deployment. The web interface was developed using ReactJs web technology, providing an accessible and interactive user experience.

4.5 Web-based platform

To facilitate real-time clinical applicability and integration into radiological workflows, we developed a comprehensive web-based platform encompassing the full vertebra segmentation and classification pipeline. This system consists of three primary modules: (1) DICOM image upload and preprocessing, (2) vertebral segmentation and anatomical labeling, and (3) classification analysis and visualization. The interface, implemented using ReactJs, HTML, CSS, and JavaScript, ensures cross-platform compatibility and interactive user experience.

4.5.1 Initial view of the web portal

Upon accessing the portal, users are presented with an intuitive file drop zone where they can upload CT-scan images. The landing view is designed to guide users through the process with minimal instruction, ensuring ease of use for clinicians with varying levels of technical expertise.

4.5.2 Segmentation interface

Upon uploading a CT-scan, the system initiates automatic preprocessing followed by vertebra segmentation using our deep learning model. The model outputs include binary vertebral masks and anatomical vertebral labels (e.g., C1, C2, T1, L5). These are overlaid on the sagittal CT slices to provide a clear interpretable visualization of the segmentation results.

Figure 5 presents the segmentation interface, where each vertebral body is highlighted with a distinct colored mask and annotated with its corresponding vertebral level. A dynamic legend facilitates visual identification of each vertebra.

4.5.3 Classification interface

Following successful segmentation, the system automatically extracts each vertebral region of interest and routes it through the vertebra classification pipeline. For every segmented vertebra, the classification module outputs the following: These results are visualized using vertebra-level cropped images, with class labels and confidence scores overlaid for each case. This enables clinicians to review automated decisions and assess model reliability on a per-vertebra basis. An example of the classification results interface is illustrated in Figure 6.

• Predicted class label: Normal, Fractured, or Unknown
• Presence of Schmorl’s nodes: Yes or No
• Fracture severity grade: Mild, Moderate, or Severe (if applicable)

4.5.4 Deployment and radiologist feedback loop

The pipeline is currently deployed at AUBMC and is actively integrated into routine radiology workflows. The system processes CT scans that contain sagittal reconstructions, as sagittal slices are required for reliable vertebral visualization and analysis. For each eligible scan, the model automatically generates per-vertebra predictions—covering fracture classification, Schmorl’s node detection, and severity grading—and presents them directly to the reviewing radiologist through the web interface.

To support continuous model improvement, the platform incorporates a built-in correction mechanism that allows radiologists to review and override model predictions at the per-vertebra level. Any correction made by a radiologist is automatically logged and stored alongside the original model output, creating a continuously growing pool of expert-validated annotations grounded in real clinical data. These corrections are collected over time and used to perform scheduled model retraining at regular intervals. Each retraining cycle enforces strict separation between training and evaluation data to prevent data leakage, and model performance is systematically benchmarked before and after retraining to verify consistent improvement prior to redeployment. This feedback-driven loop ensures that the system progressively adapts to the institutional data distribution and maintains high predictive reliability in a live clinical environment.

5 Experiments and evaluation

To assess the effectiveness of our proposed vertebra analysis system, we designed a three-stage evaluation protocol. Each stage corresponds to a critical component of the pipeline: segmentation, classification, and the integrated end-to-end system. The experiments were conducted using public segmentation datasets and our institutional classification dataset, as previously described. The evaluation focuses on standard classification and detection metrics, as well as qualitative assessment for clinical usability.

5.1 Experiment 1: vertebra segmentation accuracy

This experiment aimed to assess the effectiveness of the YOLOv11-based segmentation model in detecting, extracting, and labeling vertebrae from CT-scans. A total of 1,479 CT volumes were used for training and evaluation, drawn from the CTSpine1K, VerSe2019, and VerSe2020 datasets. The model was trained to segment and identify vertebral levels spanning from C1 to S1.

To account for anatomical variability and differences in scan coverage, model performance was assessed under two conditions: (1) scans that included at least one end of the spine (cervical or sacral), and (2) partial scans that lacked both spinal extremities.

Model performance was evaluated using Top-1 and Top-3 accuracy alongside mean Intersection over Union (mIoU). Of the evaluation set, 81% of cases had at least one spine endpoint (cervical or sacral) visible, while the remaining 19% lacked both endpoints. For scans containing spinal extremities, the model achieved a Top-1 accuracy of 96%, compared to 86% for partial scans. Remarkably, Top-3 accuracy reached 100% in both cases, underscoring the model’s resilience in handling incomplete or anatomically ambiguous inputs. The mean IoU averaged 0.94, indicating high spatial concordance between predicted and ground truth segmentations.

These results confirm that the segmentation model consistently localizes and labels individual vertebrae with high precision across diverse imaging conditions, establishing a reliable foundation for subsequent classification tasks.

5.2 Experiment 2: multi-stage classification performance

This experiment assesses the performance of a multi-stage vertebra-level classification pipeline using vertebrae extracted from the AUBMC dataset. The ground truth labels were confirmed by consensus between two experienced radiologists. The pipeline comprises three consecutive classification tasks: (1) fracture detection, (2) Schmorl’s node identification, and (3) fracture severity grading.

Stage 1: Fracture Detection. The initial classification model categorizes each vertebra into three classes: Normal, Fractured, and Unknown. In the validation and test datasets, the Normal and Fractured categories each represent approximately 40% of vertebra-level instances, while the remaining 20% correspond to the Unknown class. This distribution allows the model to be evaluated across clearly defined cases as well as instances where the vertebral condition is ambiguous or difficult to determine.

As illustrated in Figure 7, the training process demonstrates stable convergence across the full training duration of 100 epochs. Both training and validation loss curves decrease progressively, indicating that the model is effectively minimizing prediction error while maintaining consistent generalization performance. At the same time, the top-1 accuracy steadily increases for both datasets, reflecting improved capability in correctly identifying the most probable class label for each vertebra.

The top-5 accuracy remains consistently high throughout training, suggesting that even when the top prediction is incorrect, the correct class is frequently ranked among the model’s highest-confidence predictions. This behavior indicates strong probabilistic ranking of class candidates and reliable internal feature representations learned by the model.

Additional insight is provided by the normalized confusion matrix shown in Figure 8. Most predictions are concentrated along the diagonal, demonstrating that the majority of vertebrae are correctly classified. Misclassifications occur only in a limited number of cases and are mainly observed between categories with subtle visual differences, particularly those involving the Unknown class, which inherently represents uncertain or borderline cases.

The model achieves a vertebra-level classification accuracy of 92.7%. A detailed class-wise breakdown of performance metrics, including precision, recall, and F1-score, is presented in Table 2.

Table 2. Performance metrics for fracture classification across three categories.

The Fractured class achieved a precision of 0.907, recall of 0.926, specificity of 0.958, and an F1-score of 0.916. The Normal class was classified with a precision of 0.930, recall of 0.909, specificity of 0.955, and F1-score of 0.919. These results reflect the model’s high discriminative power and robustness across all three categories.

Stage 2: Schmorl’s Node Detection. The second model performs binary classification to detect the presence or absence of Schmorl’s nodes. Figure 9 shows the training and validation performance across epochs, demonstrating stable convergence. The confusion matrix in Figure 10 confirms the model’s effectiveness in distinguishing between the two classes.

As shown in Table 3, the model achieved an accuracy of 91.9%, a macro precision of 92.7%, a macro recall of 90.6%, and a macro F1-score of 91.4%. At the class level, the Schmorl class had a precision of 90.0%, recall of 97.3%, specificity of 84.0%, and an F1-score of 93.5%, whereas the Not-Schmorl class achieved a precision of 95.4%, recall of 84.0%, specificity of 97.3%, and an F1-score of 89.3%. These results confirm the model’s robustness in identifying Schmorl’s nodes in vertebrae.

Table 3. Performance metrics for Schmorl’s node detection.

Stage 3: Fracture Severity Grading. The final classification stage categorizes vertebral fractures into three severity levels: Mild, Moderate, and Severe. As shown in Figure 11, the training and validation curves exhibit stable and gradual convergence, indicating good learning behavior. Figure 12 presents the confusion matrix, which shows that most misclassifications occur between the Mild and Moderate categories, an expected pattern due to overlapping visual characteristics in these two classes. Despite this, the model performs strongly across all categories.

As summarized in Table 4, the model achieved an overall accuracy of 85%, a macro precision of 79.9%, a macro recall of 79.2%, and a macro F1-score of 79.5%. Specifically, the Severe class had the best performance, with a precision of 91.2%, recall of 93.9%, and F1-score of 92.5%. The Mild class achieved balanced metrics of 87.5% across precision, recall, and F1-score. The Moderate class was the most challenging to classify, with lower precision (61.0%) and recall (56.3%). This difficulty is partly due to the opportunistic nature of the CT-scans, which were not specifically acquired using a bone window protocol, resulting in reduced resolution of the vertebrae and making it harder to accurately differentiate moderate severity fractures.

Table 4. Performance metrics for fracture severity grading.

Finally, the three-stage classification pipeline shows high effectiveness in detecting and grading vertebral abnormalities, achieving consistent and interpretable results across binary and multi-class classification tasks.

5.3 Experiment 3: end-to-end pipeline evaluation

This final experiment evaluates the performance of the complete pipeline—from raw CT-scan input through segmentation and classification, including web-based deployment—using the testing CT-scans dataset from AUBMC that were held out from the training and validation phases. Ground truth per-vertebra annotations were available for all cases.

The system achieved an overall vertebra segmentation and classification pipeline success rate of 85%, defined as the proportion of cases where all vertebrae were correctly segmented and classified. Multi-label consistency, representing agreement between model predictions and radiologist consensus for both fracture and Schmorl’s node labels, reached 90%.

Finally, this experiment confirms the clinical relevance, robustness, and deployment readiness of the proposed pipeline for real-world opportunistic spinal health assessment.

5.4 Comparison with state-of-the-art methods

To contextualize our results, we compare the proposed pipeline against recent state-of-the-art (SOTA) methods for automated vertebral fracture detection and severity grading from CT imaging. Table 5 summarizes the key performance metrics reported in the literature alongside those achieved by our system.

Table 5. Comparison with state-of-the-art methods for vertebral fracture detection and severity grading from CT scans.

As shown in Table 5, our fracture detection model achieves a sensitivity of 92.6% and F1-score of 91.6%, surpassing or matching most existing approaches. Notably, the commercially deployed Zebra Medical system [23] reported lower sensitivity (68.0%) in a real-world evaluation, highlighting the challenge of translating research-grade performance to clinical deployment. Our pipeline achieves competitive performance while operating on a larger and more heterogeneous dataset (5,000 CT scans) that reflects real-world clinical variability, unlike many prior studies that used curated or single-center datasets.

For severity grading, direct comparisons are limited since few studies report Genant-based multi-class severity classification from CT imaging. Our system achieves 85.0% accuracy on a large and heterogeneous dataset, demonstrating competitive performance despite the additional challenge of operating on opportunistic (non-bone-window) CT acquisitions.

To our knowledge, no prior study has reported automated Schmorl’s node detection performance from CT imaging, making our result of 91.9% accuracy a novel benchmark in this domain. The integration of Schmorl’s node detection alongside fracture classification and severity grading within a single end-to-end pipeline represents a unique contribution that is not addressed by any of the compared methods.

5.5 Model interpretability via gradient-weighted class activation mapping

To enhance interpretability and provide insight into the regions driving model predictions, we applied Eigen-CAM [28] to visualize class activation heatmaps from the classification models. Eigen-CAM computes the principal components of the feature maps from the final convolutional layer, producing spatially resolved heatmaps that highlight the most discriminative image regions without requiring gradient computation. This approach is particularly suitable for YOLO-based architectures where standard Grad-CAM may not be directly applicable due to the detection head structure.

Figure 13 presents representative Eigen-CAM heatmaps for each class addressed by the pipeline: Normal, Mild, Moderate, Severe, and Schmorl. The first four columns visualize activations from the vertebra classification models, while the rightmost column shows activations from the Schmorl’s node detection model. Several clinically meaningful observations can be drawn from these heatmaps.

The model demonstrates severity-dependent activation patterns. Normal vertebrae show compact, central activations, while mild fractures shift attention toward the superior endplate, capturing early deformities. Moderate fractures exhibit diffuse, widespread activations across the vertebral body, reflecting broader structural compromise and increased classification difficulty. Severe fractures produce strong, highly localized activations at the site of maximal collapse. In contrast, Schmorl’s nodes are identified by sharply focused, circular activations at the endplate, highlighting their distinct focal morphology.

These visualizations confirm that both the severity classification and Schmorl’s node detection models attend to anatomically relevant vertebral structures rather than spurious background features. The progressive change in activation patterns—from compact and centered in normal vertebrae to increasingly diffuse and endplate-oriented in fractured vertebrae—mirrors the clinical progression of vertebral compression fractures as defined by the Genant grading system. This alignment between model attention and established clinical criteria supports the trustworthiness, interpretability, and clinical validity of the proposed system.

6 Discussion

This study intentionally deviates from the common approach found in the literature, where algorithms are typically evaluated using pristine, high-resolution, and carefully curated images. While such datasets can yield high reported accuracies, they often overestimate real-world performance and fail to capture the challenges encountered in everyday clinical practice. In contrast, our dataset, which is characterized by its heterogeneity, provides a more realistic benchmark by reflecting the full spectrum of variability and imperfection commonly seen in clinical settings. Specifically, the sources of heterogeneity in our dataset include variations in scanner vendors and models, differences in reconstruction kernels (e.g., soft tissue vs. bone), variable slice thickness, the presence or absence of intravenous contrast, and differences in patient positioning and field-of-view coverage. This includes a mix of CT scans obtained from different machines with varying acquisition parameters, as well as the involvement of two radiologists for image annotation, introducing additional inter-rater variability. The lower performance observed in moderate fractures should, therefore, not be viewed solely as a limitation of the model, but rather as a consequence of diagnostic ambiguity, the inherent variability in clinical imaging, and less-than-ideal imaging conditions.

Moderate fractures, in particular, present a diagnostic challenge as they often exhibit features that overlap with both mild deformities and severe collapses, making them less visually distinct. While mild fractures are marked by subtle yet consistent alterations, and severe fractures show pronounced structural compromise, moderate fractures occupy a diagnostically ambiguous space. Even expert radiologists may disagree on their classification, further complicated by the variability introduced by using different CT machines, which affected factors like field of view, slice thickness, reconstruction kernels, and radiation dose. This variability in imaging conditions, coupled with patient-specific anatomical differences such as vertebral dimensions, spinal alignment, and degenerative changes, makes fracture classification more complex.

Our aim was to provide a more realistic predictive model that mirrors the challenges faced in clinical imaging, rather than one that is fine-tuned to produce idealized results. While post-processing techniques like brightness and contrast normalization were applied to mitigate some of these effects, they could not fully compensate for the diagnostic details lost during image acquisition. This approach ensures that our findings represent a more authentic reflection of real-world performance, highlighting the diagnostic challenges present in routine clinical practice.

7 Conclusion

This study presents an end-to-end deep learning-based pipeline for vertebral health assessment using routinely acquired chest and abdominal CT scans. The system automates the detection, classification, and interpretation of VCFs and Schmorl’s nodes, achieving high accuracy in clinical settings. The tool integrates deep learning-driven segmentation and hierarchical classification modules, offering a clinically relevant, three-stage classification framework. This system has been deployed as a web-based platform, integrated seamlessly into radiology workflows, and is actively used at AUBMC. Continuous improvement is ensured through a feedback loop, where radiologists review and correct predictions, with bi-weekly model retraining enhancing the system’s accuracy and clinical reliability. The retraining process includes data curation and label verification by radiologists, strict separation of training and evaluation data to prevent leakage, and performance comparison between pre- and post-retraining models to ensure consistent improvement.

Our approach intentionally departs from the common practice of using pristine, high-resolution datasets, instead utilizing a heterogeneous dataset reflective of real-world clinical conditions. The dataset consists of CT scans acquired for chest and abdominal evaluation, introducing variability in imaging quality and acquisition parameters. Additionally, two radiologists annotated the images, introducing potential inter-rater variability, particularly in the classification of moderate fractures, which remain inherently ambiguous and difficult to diagnose. This realistic and unrefined approach better mirrors the challenges encountered in clinical practice, highlighting the system’s ability to handle the complexities of diverse, real-world data.

Looking ahead, we aim to expand the system’s capabilities to include other imaging modalities such as MRI and DEXA scans, further enhancing its diagnostic power. We will also explore advanced data fusion techniques to integrate clinical data, offering more personalized fracture risk assessments. Future work includes leveraging NLP and LLMs to automate the generation of structured reports, streamlining the imaging-to-reporting workflow and reducing clinician workload. The system will also be tested in broader clinical settings to improve scalability and impact, ultimately supporting more timely and data-driven clinical decisions in vertebral health management.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmscd.2026.1718482/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Sandeep Kumar Mishra, Yale University, United States

Reviewed by: Lokeswari Pinneboyana, Independent Researcher, Farmington, MI, United States

Ningtao Sun, Zhuhai College of Science and Techology, China

Abdul Basit Mughal, University of L’Aquila, Italy