Advancing Spine Fracture Detection: The Role of Artificial Intelligence in Clinical Practice

Authors: Seonghoon Jeong, Byung-Jou Lee

Abstract

Vertebral fractures are prevalent skeletal injuries commonly associated with osteoporosis, trauma, and degenerative diseases. Early and accurate diagnosis is crucial to prevent complications such as chronic pain and progressive spinal deformities. In recent years, artificial intelligence (AI) has emerged as a powerful tool in medical imaging to support automatic detection and classification of vertebral fractures. This review provides an overview of AI-based approaches for spinal fracture diagnosis and summarizes recent advances in deep learning (DL) and machine learning (ML) models. The performance of AI models, mainly evaluated by sensitivity, specificity, and accuracy metrics, varies with imaging modality and dataset size, with computed tomography-based models demonstrating superior diagnostic accuracy. In addition, AI-assisted workflows have been shown to improve diagnostic efficiency, reducing the time required for fracture detection. Despite these advances, challenges remain, such as dataset variability, the need for large-scale annotated datasets, and standardization of evaluation metrics. Future research should focus on improving model generalization, integrating multimodal imaging data, and validating AI applications in real-world clinical settings to further improve vertebral fracture diagnosis and patient management.

INTRODUCTION

Vertebral fractures are one of the most common skeletal injuries and are often caused by osteoporosis, trauma, or degenerative changes.35924) These fractures frequently occur due to mechanical stresses that exceed the structural integrity of the vertebral bodies, especially in individuals with reduced bone density. Diagnosis of vertebral fractures relies primarily on imaging modalities such as radiographs, computed tomography (CT), and magnetic resonance imaging (MRI).4152025272932) Radiographs are commonly used as an initial screening tool, but have limited sensitivity, especially for subtle fractures. This limitation is particularly evident in cervical spine fractures, where complex anatomy and subtle fracture patterns make radiological evaluation even more challenging. CT scans are effective in detecting compression fractures and vertebral deformities by providing detailed visualization of bone structures. MRI is particularly useful in differentiating acute from chronic fractures by assessing bone marrow edema, which indicates recent injury. It is important to diagnose vertebral fractures accurately and in a timely manner because spinal deformities may gradually worsen, causing chronic pain and increasing the risk of subsequent fractures, if the diagnosis is delayed.1213) Early intervention, including pharmacologic treatment, vertebroplasty, and surgical stabilization, can significantly improve patient outcomes. Given the high prevalence and clinical significance of vertebral fractures, there is an increasing need for more efficient and reliable diagnostic tools to assist clinicians identify these fractures at an early stage.

Artificial intelligence (AI) has been increasingly integrated into various aspects of medical imaging to assist in the detection, classification, and prognosis of various diseases.1271114171923) AI algorithms, including deep learning (DL) and machine learning (ML) approaches, are being applied in fields such as oncology for tumor detection, cardiology for automated electrocardiogram interpretation, and neurology for stroke assessment using imaging data. In musculoskeletal radiology, AI models have been developed for tasks such as bone age estimation, fracture detection, and osteoporosis assessment.1018) For vertebral fractures, researchers have explored the application of AI models in detecting and classifying fractures from medical imaging scans. These models use automated feature extraction and classification techniques to assist radiologists in identifying fractures with greater accuracy and efficiency. While AI-based vertebral fracture diagnosis is still evolving, research efforts are ongoing to continue improving these models with the aim of increasing generalizability and clinical applicability. Although various datasets and methodologies have been explored to improve AI performance in spinal fracture detection, challenges such as variability in imaging protocols and the need for large, annotated datasets remain areas of ongoing investigation.

Given these developments, this narrative review aims to provide an overview of AI-based vertebral fracture diagnosis by summarizing recent researches. FIGURE 1 illustrates the role of AI in this process, highlighting how AI models analyze radiographs and CT data to detect vertebral fractures and classify fractures. We will review various AI approaches used for vertebral fracture detection, discuss their advantages and limitations, and explore the potential clinical implications of these technologies. By reviewing the existing literature, we aim to highlight current trends in AI-based spinal fracture diagnosis and identify areas for future research and improvement.

BASIC CHARACTERISTICS OF CASES

This review includes studies published between 2021 and 2024 that utilized CT or radiography to diagnose spinal fractures using AI models, as summarized in TABLES 1 and 2. These studies were selected based on their relevance to AI-based diagnostic performance, and a minimal number of searches were performed in PubMed and Google Scholar using keywords such as spinal fracture, artificial intelligence, deep learning, radiograph, and CT.

Diverse patient populations with spinal fractures were examined (TABLE 1). Sample sizes varied considerably, from approximately 229 to 4,050 cases, depending on study design and data availability. It is noteworthy that some studies used the number of vertebral bodies as the unit of analysis rather than individual patients, which resulted in larger sample sizes in certain studies.28) Patients in these studies were aged 18 years or older, and many studies have focused specifically on the elderly population because of the high prevalence of osteoporotic fractures. However, some studies did not report the exact age range of participants.

The vertebral fractures examined in these studies were primarily investigated as osteoporotic compression fractures, traumatic fractures, and non-pathological fractures. Compression fractures, especially those seen in elderly individuals with low bone mineral density, were commonly reported.16263335) While some studies also considered traumatic fractures, more complex fracture patterns, such as burst fractures, were not included.

Regarding the anatomical distribution, studies primarily focused on fractures of the cervical spine, though some also investigated fractures in the thoracic, lumbar, or thoracolumbar regions. A subset of studies specifically analyzed fractures in the thoracolumbar or cervicothoracic-lumbar regions, while others included fractures across the whole spine.

Regarding imaging modality, radiographs and CT scans were predominantly used to detect spine fractures. Although radiographs have limited sensitivity, especially for subtle fractures, it is still the initial imaging modality of choice due to its accessibility and cost-effectiveness. Therefore, this widespread use of radiography has led to a significant amount of research into the use of AI for detecting spinal fractures. CT scans, on the other hand, were preferred for their superior ability to detect fine bone details and characterize fractures more accurately. Although MRI is known to be useful in diagnosing vertebral fractures, none of the reviewed studies utilized MRI for fracture detection.

Overall, the patient population in this study reflected a wide clinical spectrum of vertebral fractures, from routine osteoporotic cases to more complex traumatic injuries, which were analyzed using radiographs and CT imaging modalities.

AI MODELS FOR VERTEBRAL FRACTURE DIAGNOSIS

Various AI models were used for diagnosing spinal fractures, ranging from commercial products to research-based models. Among the AI models used, BoneView (Gleamer, Paris, France) and the Aidoc AI algorithm (AIDOC Medical, Tel Aviv, Israel) are commercial products.163135) Both tools are integrated with the picture archiving and communication system (PACS) of the institution, allowing for efficient image transfer and analysis. Once images are uploaded to the PACS, the AI systems process the radiographs or CT scans and return results within minutes, enabling timely diagnostic support. These systems are designed to assist radiologists by automatically detecting and flagging potential fractures, enhancing clinical decision-making and improving workflow efficiency.

On the research side, various deep learning models have been used, mainly convolutional neural network (CNN)-based architectures.212830) Popular CNN architectures such as AlexNet and GoogleNet have been applied to analyze spinal fractures in medical images.26) Although these models are not commercially available, they have been shown to be effective at identifying fractures by learning complex patterns from large data sets. You Only Look Once (YOLO) v4, a real-time object detection model, was also used for spinal fracture detection, providing the advantage of quickly detecting and locating fractures in images.6) Vision Transformer (ViT), a recent approach that applies transformer models to image analysis, has also shown promising results in fracture detection by taking advantage of its ability to capture long-range dependencies in images.8)

From commercially available products to research models, various AI models highlights the growing versatility and potential of AI in improving spinal fracture diagnosis, providing a range of tools to suit a variety of clinical needs.

PERFORMANCE OF AI MODELS

The performance of AI models for diagnosing spinal fractures was evaluated using several key metrics, including sensitivity, specificity, and accuracy. These metrics were reported in most studies and were used as the main indicators of model effectiveness. Sensitivity, which measures the ability of a model to correctly identify fractures, varied across studies and reported values ranging from approximately 49.0% to 100.0% (TABLE 2). Specificity, which measures the ability of a model to correctly classify non-fracture cases, reported values ranging from 54.9% to 100.0%. Accuracy, which provides an overall measure of performance, was generally reported to be greater than 75%. Some models showed near-perfect sensitivity, specificity, and accuracy, while others had lower performance due to differences in dataset composition and annotation criteria.

In addition to these basic metrics, some studies have also reported additional performance metrics such as recall, F-score, area under the receiver operating characteristic curve (AUC), and predictability scores.62630) Recall, which is equivalent to sensitivity in binary classification tasks, showed similar trends to sensitivity values. F-score, which balances precision and recall by calculating the harmonic mean, has been reported in several studies, but not consistently across all studies. The AUC, which represents the ability of a model to distinguish between fractured and non-fractured cases at various threshold, has been reported in limited studies but often exceeds 0.7, indicating their notable classification performance. However, not all studies provide these additional metrics, making direct comparisons between models difficult.

AI models using CT scans tended to demonstrate better performance in vertebral fracture detection compared to those using radiographs.30) This observation may be due to the higher spatial resolution and enhanced bone contrast of CT, which allows for more accurate identification of fracture lines and differentiation from surrounding anatomy. Sensitivity and specificity values were generally higher in CT-based models, suggesting a potential advantage in diagnostic accuracy. However, it is important to consider that radiographs would be more appropriate for initial screening in certain clinical situations because it offers advantages such as faster acquisition times and lower radiation exposure.

Among the AI architectures, CNNs have demonstrated superior performance in vertebral fracture detection.2630) A typical CNN architecture used for spinal fracture classification is shown in FIGURE 2, which shows the sequential process of feature extraction through convolutional and pooling layers and classification using fully connected layers. CNN-based models generally achieved higher sensitivity and specificity than other AI methodologies. This superior performance can be attributed to the ability of CNNs to automatically extract hierarchical features from imaging data and capture complex spatial patterns and subtle fracture characteristics. These results suggest that CNNs play an important role in improving the diagnostic accuracy of AI models for detecting spinal fractures.

AI models trained on larger datasets tend to perform better in detecting vertebral fractures.630) Higher sensitivity, specificity, and accuracy were more commonly observed in studies using extensive training data, which may be due to improved generalization and reduced overfitting. Larger datasets have the potential to provide a more comprehensive representation of fracture patterns, anatomical variations, and imaging conditions, thereby contributing to greater model robustness. In contrast, studies using smaller sample sizes sometimes reported greater variability in performance, suggesting potential limitations in model generalizability. These results suggest that size of dataset can play a critical role in the development of reliable AI-based diagnostic tools.

DISCUSSION

This review examined several AI models for vertebral fracture detection, focusing on their performance across different imaging modalities, patient populations, and dataset sizes. Sensitivity, specificity, and accuracy varied across studies, with CT-based models generally performing better than models using radiographs.3033) This trend may be due to the higher spatial resolution and superior bone contrast of CT, which allows for more accurate identification of fractures. However, AI models for radiograph-based fracture detection remain an important area of research as radiographs remain widely used due to their accessibility and low radiation exposure. Larger datasets were associated with improved model performance due to better generalization and reduced overfitting.630) Larger datasets provide a more comprehensive representation of fracture patterns and anatomical variations, contributing to more robust AI models. Commercial AI tools such as BoneView and Aidoc have shown clinical applicability.16313335) Furthermore, research-driven deep learning models, including CNN-based architectures and ViTs, continue to evolve and provide promising advances in the diagnosis of spinal fractures.82128) However, the diversity of study designs, datasets, and evaluation metrics complicates direct comparisons between models, highlighting the need for standardized evaluation methods.

AI-based fracture detection offers several benefits that can improve clinical workflow and patient care. First, integrating AI into the diagnostic workflow can automate the detection of spine fractures, improving efficiency and reducing the time radiologists need to interpret images. This automation not only accelerates the diagnostic process, but also minimizes human error, especially in high-volume workflows where fatigue and time constraints can impact diagnostic accuracy. AI can act as an adjunct tool to help prioritize cases requiring urgent treatment and identify and address critical fractures immediately.

One study directly evaluated the time-efficiency impact of AI in the detection of spinal fractures.31) By analyzing PACS log files containing timestamps for every instances in which studies were modified or reviewed, the researchers found that AI implementation reduced the detection and notification time (DNT) by an average of 16 minutes compared to a non-AI assisted workflow. DNT was defined as the time interval between the moment the study became accessible on PACS and the moment the image was first opened by the radiologist. This time reduction will allow AI to expedite the diagnostic process, allowing for accurate fracture identification and early treatment for patients. Earlier fracture detection has the advantage of supporting decision-making and timely patient management, particularly in settings where imaging interpretation is delayed, such as emergency rooms and other busy departments.

AI can also assist in triaging patients with important clinical implications. By classifying patients based on fracture severity, AI models can help guide patients in either conservative management or those requiring surgery. For example, patients with mild osteoporotic compression fractures may be referred for medication or rehabilitation, whereas patients with unstable fractures may require additional imaging or surgery. In addition, AI-based classification could aid research by standardizing fracture assessment across large datasets, and improve consistency in patient selection in clinical trials.

In addition to the advantages mentioned above, AI has the potential to support institutions with limited resources or experienced experts. By automatically providing pre-assessments, AI can improve the quality of fracture diagnosis and management by helping less experienced clinicians make decisions. Despite these advantages, challenges remain, including regulatory considerations, data bias, and integration with clinical systems.

One study presented the results of applying AI to the Genant classification system for vertebral fracture assessment.16) The Genant classification system is a method to evaluate vertebral fractures through vertebral height loss and morphologic changes. This classification system divides vertebral fractures into three grades. Grade 1 is a minor vertebral height loss of 20%–25%, grade 2 is a moderate vertebral height loss of 25%–50%, and grade 3 is a severe vertebral height loss of more than 50%. In this study, the AI model showed almost similar performance for vertebral fractures of Genant grades 2 and 3. However, it shows insufficient accuracy for Genant grade 1, suggesting that the performance of AI models needs to be improved for subtle spinal deformities. Additionally, the authors demonstrated that the presence of foreign material within the spine has a significant impact on the performance of AI models, both in sensitivity and specificity. These results suggest that AI models should be improved, especially for subtle fracture detection or cases where foreign materials are implanted.

Although current studies have mainly focused on imaging-based approaches for vertebral fracture detection, research is being conducted using multimodal AI models in other medical diagnostic fields. These models have shown promising results in areas such as abdominal trauma assessment and ophthalmologic disease classification by integrating imaging information with non-imaging information such as patient symptoms, trauma history, and laboratory findings.223436) Integrating this clinical context can provide a more complete understanding of the patient’s condition and improve the performance and clinical applicability of AI models. However, these approaches have rarely been applied to vertebral fracture detection. Applying such methods in vertebral fracture detection would potentially improve diagnostic accuracy and support clinical decision-making more effectively.

Several challenges need to be addressed for AI models to be integrated into clinical practice. First, regulatory approval and ethical deployment are important considerations. Given the impact of AI-based diagnostics on patient care, there is a need to validate that these models are compatible with clinical standards and consistently reliable across diverse populations. Second, the lack of transparency in deep learning models presents challenges, as their decision-making processes are often difficult to interpret. Until there is proper interpretation, clinicians will have difficulty trusting AI-generated output, especially for complex cases where decisions require explanation. Finally, AI models require constant monitoring and updates to maintain accuracy, reliability, and clinical relevance. Medical imaging data, imaging protocols, and patient demographics are constantly changing and can affect the performance of AI models. Regular retraining and validation using up-to-date data are important factors to ensure that AI models are reliable tool in clinical practice.

CONCLUSION

AI-based spinal fracture detection has shown tremendous potential to improve diagnostic accuracy, efficiency, and clinical decision-making. Because the AI model is still evolving, its integration into clinical workflow would improve vertebral fracture detection accuracy, patient triage, and support for resource-limited institutions. However, challenges still remain, such as regulatory approval, improved generalization to diverse populations, clear interpretation of AI models, and their consistent updates. Future studies should focus on better detection performance for subtle fractures in AI models, studies on diverse populations, and addressing regulatory considerations. With continued advances in deep learning and imaging technologies, AI has shown potential to improve performance in diagnosing vertebral fractures and ultimately patient management.