Deep Learning for Discrimination of Early Spinal Tuberculosis from Acute Osteoporotic Vertebral Fracture on CT

Introduction

Spinal tuberculosis (STB) and acute osteoporotic vertebral compression fracture (OVCF) are common conditions in spinal surgery, being characterized by symptoms such as pain, inability to stand, and even spinal kyphotic deformity and spinal cord compression, potentially leading to neurological deficits and paralysis.¹ With the intensification of population aging, osteoporosis-induced OVCF has emerged as a significant global health concern,² with OVCF becoming the third most common fracture worldwide, with an estimated annual incidence of 1.4 million new OVCF cases. Tuberculosis, a severe infectious disease detrimental to health, is caused by the hematogenous spread of Mycobacterium tuberculosis.³ STB is the most common form of extrapulmonary secondary tuberculosis⁴ and is a frequent condition in spinal surgery;⁵ worldwide, the annual incidence of STB exceeds 100 000. STB is more prevalent in developing countries,⁶ where early detection and treatment can lead to favorable prognoses.⁷

Early-stage STB and vertebral compression fractures often manifest with indistinguishable imaging features and clinical presentations.^8–10 In both cases, patients typically present with primary symptoms of lower back pain accompanied by restricted mobility, and one or multiple vertebrae may exhibit damage or alterations in signal intensity on imaging,^11,12 making early differentiation challenging, especially in elderly patients without a history of trauma.¹³ The similarities between the two conditions make their differential diagnosis difficult, hindering early clinical diagnosis. Uncertainty in early diagnosis may lead to inappropriate treatment strategies that exacerbate the patient’s condition.¹⁴ With advances in spinal surgery, percutaneous balloon kyphoplasty (PKP) is often the preferred treatment for OVCF;^15,16 however, spinal infectious diseases such as STB are contraindications for PKP, because PKP may lead to the spread of infection and difficulty in eradicating it,¹⁷ and therefore the misdiagnosis of STB as OVCF can have catastrophic consequences.¹⁸ Patients with STB require early and regular anti-tuberculosis drug treatment,¹⁹ regardless of surgical intervention. Given the stark differences in treatment approaches required for STB and OVCF, accurate early diagnosis is crucial for selecting the correct treatment method.

Clinically, the reference standards for diagnosing STB and OVCF are Mycobacterium culture²⁰ and magnetic resonance imaging (MRI), respectively. However, in many countries MRI equipment is not widely available,²¹ and even when available, various factors may preclude certain patients from undergoing MRI examinations. These factors include metallic implants or devices, pregnancy (particularly the early stages), claustrophobia, excessive body weight, and inability to remain still.²² Furthermore, the biopsy techniques used for diagnosing STB are invasive procedures that require skilled spinal surgeons, and many risks are associated with the procedure, including patient discomfort, further spread of the bacteria, and nerve root damage.²³ Additionally, the culture of Mycobacterium tuberculosis, a critical step for confirmation, is time-consuming because of the bacterium’s slow growth rate.²⁴ This extended culture time hinders early diagnosis of patients, thereby impeding early treatment.^25,26 Based on the aforementioned reasons, existing diagnostic methods (biopsy, bacterial culture, and MRI examination) fail to meet the widespread demands for differential diagnosis of these two diseases in many countries and populations. Therefore, the development and refinement of a new non-invasive early diagnostic technique is crucial for the treatment and management of these diseases.

Computed tomography (CT) is a routinely utilized diagnostic modality in orthopedics, often being employed to further investigate suspicious features observed on X-ray images.²⁷ CT has high spatial resolution, enabling differentiation between osteoporotic fractures and pathological fractures on the basis of morphological characteristics such as bone integrity and fracture edge. However, CT is not sensitive to edema or hematoma caused by tuberculosis infection or fresh fractures. Although CT can display certain features of typical STB and OVCF, it cannot differentiate between these two diseases in their early stages, especially when STB coexists with osteoporosis.

Traditional image analysis methods such as threshold-based segmentation, edge detection, and region growing struggle with complex CT images because of their dependence on manual feature extraction, which is inefficient and inflexible. These methods are also prone to image noise, frequently necessitating human intervention, and thereby reducing processing efficiency.^28,29 Furthermore, their effectiveness and adaptability falter with high-dimensional and large-scale CT data, making methods based on complex algorithms challenging to interpret.

With the development of deep learning and advanced computer vision technologies, artificial intelligence models can automatically learn and extract features, learn complex patterns and features, and capture comprehensive structural information from images, effectively handling tasks primarily focused on image recognition. The application of artificial intelligence in medical image analysis continuous to increase, with studies demonstrating that deep learning can serve as an effective technique for fracture diagnosis.³⁰ Its effectiveness for diagnosing fractures in multiple body parts has been validated,³¹ with deep learning having been successfully applied in automatic vertebral segmentation³² and opportunistic osteoporosis screening.³³ All of this indicates that deep learning models are a promising approach for orthopedic medical image analysis.

Our study aimed to develop deep learning models using sagittal CT images of patients with STB or OVCF, and to evaluate the models’ abilities and feasibility for distinguishing between acute OVCF and early-stage STB, thereby guiding early clinical differential diagnosis.

Methods

Patient Datasets

Patient images were sourced from the orthopedic CT imaging databases of two affiliated medical university hospitals (Figure 1), with informed consent for the imaging being obtained at admission. Data were collected from individuals diagnosed with acute vertebral compression fracture or STB at these two institutions over a ten-year period. The inclusion criteria were: (1) early-stage STB without severe spinal deformity; (2) STB confirmed by pathological diagnosis; and (3) complete clinical data and preoperative CT. The exclusion criteria included: (1) concurrent severe infectious disease or other types of infectious spondylitis; (2) history of spinal trauma, spinal surgery, or internal implants; (3) absence of pathological diagnosis; and (4) paravertebral abscess or burst fractures. The age, gender, and vertebral lesion location of each patient were obtained from electronic medical records. A total of 373 patients were included in this study, with 302 being from hospital 1 (average age 62.02 years, age range 19–98 years). This group consisted of 149 patients with acute OVCF (average age 72.60 years, age range 36–98 years) and 153 patients with STB (average age 51.61 years, age range 19–96 years). Additionally, 71 patients were recruited from another affiliated hospital (hospital 2; average age 62.03 years, age range 37–94 years), with these consisting of 32 patients with acute OVCF (average age 61.63 years, age range 37–94 years) and 39 patients with STB (average age 62.36 years, age range 48–82 years). The STB cases were confirmed by biopsy results as tuberculosis. All cases of acute OVCF had no known history of tuberculosis and were diagnosed with osteoporosis or severe osteoporosis based on bone density results. The CT images were reformatted into sagittal views from the original transverse section images.

This study is reported in line with the STROCSS criteria.³⁴

Figure 1 Patient selection and grouping flowchart.

Deep Learning Models

A senior orthopedic surgeon with 20 years of experience collaborated with a junior orthopedic doctor to analyze the patient images. We specifically selected the most abnormal vertebral segments for deep learning analysis. Regions of interest (ROIs) were manually delineated on the sagittal slices, then a computer program generated rectangles that encompassed the entirety of abnormal areas, which were then input into the deep learning models. For this study, we employed four widely used neural network architectures: ResNet50, ResNet101, Multiscale Vision Transformers (MVITV2), and EfficientNet_B5. These four models were trained and their effectiveness for differentiating between STB and OVCF was assessed.

ResNet50 is a Residual Network variant with 50 layers, known for its “residual learning” feature for solving deep network training degradation by using skip connections for smoother training. ResNet101 is an extension of ResNet with 101 layers, offering deeper and more complex feature learning than ResNet50 through similar residual learning techniques. MVITV2 is designed for image processing, with the model using vision transformers to analyze images at multiple scales, capturing a wide range of features for complex analysis. EfficientNet_B5 is known for its efficient scaling of network depth, width, and resolution, and this model from the EfficientNet series achieves high accuracy with fewer parameters, optimizing three key network dimensions.

Our objective was to assess the models’ effectiveness in differentiating between STB and OVCF. The model inputs included the target slice and the four adjacent sagittal slices. Images were resized to a fixed 145×210 matrix and the pixel intensities were normalized to a mean of 0 and standard deviation of 1. The training parameters consisted of 100 training epochs, an AdamW optimizer, a learning rate (ResNet101, 0.00001; MVITV2, 0.000005; EfficientNet_B5, 0.0001; ResNet50, 0.00005), and a batch size of 32. The patients were randomly split into training and validation sets in a 4:1 ratio. To enhance the dataset, we performed image augmentation through random affine transformations, including translation, rotation, flipping, and scaling, thereby expanding the dataset by 20 times. The diagnostic performance of the models was analyzed and compared using receiver operating characteristics (ROC) curves, area under the curve (AUC), accuracy, precision, sensitivity, F1 score, and confusion matrices. Importantly, our image processing did not involve the inclusion of patient clinical features, such as medical history and laboratory test results.

Statistical Analysis

Python 3.8 was used for coding and developing the deep learning networks with Pytorch and TensorFlow. The Scikit-learn toolkit was used to implement model comparisons and perform related analyses, such as drawing ROC curves and calculating AUC. A p value of less than 0.05 was defined as statistically significant.

Ethical Approval

This study was approved by the Ethics Committee of the First Affiliated Hospital of Chongqing Medical University (Approval No. K2024-028-01) and the Ethics Committee of the Army Specialty Medical Center of the Chinese People’s Liberation Army (Approval No. [2024] No. 29). All procedures performed in this study were in accordance with the ethical standards of the institutional and national research committees and with the 1964 Declaration of Helsinki and its later amendments.

Results

Study Participants

This study included a total of 373 patients (Table 1). Among the 302 patients collected from hospital 1, 153 patients had STB (male/female: 77/76, average age: 51.71 ± 16.52 years, 358 affected vertebrae) and 149 patients had OVCF (male/female: 35/114, average age: 72.24 ± 11.38 years, 175 affected vertebrae). We randomly assigned 80% of the patients (123 patients with STB and 120 patients with OVCF) to the training set, with the remaining 20% (30 STB patients and 29 OVCF patients) assigned to the internal validation set. From hospital 2, we collected data from 71 patients, including 39 with STB (male/female: 17/22, average age: 62.36 ± 10.65 years, 45 affected vertebrae) and 32 with OVCF (male/female: 16/16, average age: 61.63 ± 13.32 years, 41 affected vertebrae). As shown in Table 1, there were significant differences in age and gender ratios between the STB and OVCF groups (p < 0.05). The distributions of lesions across different spinal regions were as follows: in the training set and internal validation set, the STB group consisted of 199 lesions in the thoracic spine, 143 in the lumbar spine, and 16 in the sacral spine, while the OVCF group consisted of 89 lesions in the thoracic spine, 86 in the lumbar spine, and none in the sacral spine. There were no cervical spine lesions in either group. In the external validation set, the STB group consisted of 4 lesions in the cervical spine, 24 in the thoracic spine, and 27 in the lumbar spine, while the OVCF group consisted of 25 lesions in the thoracic spine, 16 in the lumbar spine, and none in the cervical spine. There were no sacral spine lesions in either group.

Table 1 Patient Characteristics Comparison Between the STB and OVCF Groups

In the comparative analysis presented in Table 2 and illustrated in Figure 2, the MVITV2 framework demonstrated superior performance to its counterparts, achieving an accuracy rate of 98.98%, precision rate of 100% (indicating that all predictions labeled as fractures were accurately validated, with no false positives), sensitivity of 97.97%, F1 score of 98.98%, and AUC of 0.997. Following closely, the ResNet101 architecture achieved accuracy of 96.95%, precision of 95.86%, sensitivity of 97.89%, F1 score of 96.86%, and AUC of 0.995. The ResNet50 model achieved accuracy of 92.20%, precision of 93.79%, sensitivity of 90.67%, F1 score of 92.20%, and AUC of 0.968. The EfficientNet_B5 model achieved accuracy of 86.78%, precision of 91.72%, sensitivity of 83.12%, F1 score of 87.21%, and AUC of 0.948.

Table 2 Comparison of Diagnostic Accuracy Across Different Deep Learning Models and Experienced Surgeons

Figure 2 Curves for all models on the validation set, in the following order: ResNet50, ResNet101, MVITV2, EfficientNet_B5.

On the external test set, the MVITV2 model demonstrated excellent performance, with accuracy of 83.77%, precision of 84.84%, sensitivity of 83.09%, F1 score of 83.38%, and AUC of 0.879, as shown in Table 2 and Figure 3. The ResNet101 model achieved accuracy of 72.75%, precision of 72.61%, sensitivity of 72.57%, F1 score of 72.59%, and AUC of 0.751. The ResNet50 model showed slightly poorer performance on the external validation set, with accuracy of 60.00%, precision of 74.05%, sensitivity of 62.58%, F1 score of 55.74%, and AUC of 0.686. The EfficientNet_B5 model performed well, with accuracy of 76.81%, precision of 77.69%, sensitivity of 77.41%, F1 score of 76.80%, and AUC of 0.844.

Figure 3 Comparison of results between DL models and surgeons using confusion matrix.

In the independent and blind evaluation of the training set and internal test set, the two spine surgeons achieved accuracy rates of 77.38% (junior doctor) and 93.56% (senior surgeon). A confusion matrix (Figure 3) indicates that the neural network models surpassed the level of the junior resident doctor and matched or exceeded the performance of the experienced spinal surgeon. Several applications of the models are visually illustrated in Figure 4, which showcases four true positive cases. The CT sagittal images of these four patients were analyzed using the MVITV2 model, and all four patients were not included in the training dataset. The numbers on the images represent the model’s probability estimates for the corresponding diseases. The vertebrae with the most evident lesions, denoted by green boxes, indicate the ROIs that the model focused on.

Figure 4 Visual examples demonstrate four True Positive (TP) cases, all examples of patients not included in the training dataset: The first case involves osteoporosis and vertebral compression changes due to a fall, while the second case, without trauma or osteoporosis, highlighting that spinal tuberculosis can have different appearances on CT scans. The final two cases, both with osteoporosis but no trauma, showcase the model’s ability to differentiate osteoporotic compression fractures in CT sagittal images.

Discussion

In many countries, STB and OVCF are common, and both require early intervention and customized treatment. In early-stage STB and OVCF, the clinical symptoms may not facilitate early differentiation, with the patient often presenting with back pain, emphasizing the importance of imaging or laboratory tests in the differential diagnosis.^35,36 However, various limitations may restrict the use of percutaneous biopsy, mycobacterial culture, and MRI examination. In comparison, CT machines are already widely available in most hospitals, and even in primary healthcare centers. Therefore, the application of deep learning models to sagittal CT images allows the models to be more easily applied in clinical practice. Our study aimed to utilize artificial intelligence models trained on reconstructed CT images to differentiate between OVCF and STB on CT images, addressing the current clinical challenges and aiding in early and improved diagnosis and prognostication for patients. Our results showed that all models performed robustly on the internal dataset, with the MVITV2 model achieving 98.98% accuracy and AUC of 0.997. The models also yielded notable outcomes on the external validation dataset, aside from the slightly lower performance of ResNet50. The MVITV2 model achieved 83.77% accuracy and AUC of 0.879. When these values were compared with those of the spinal surgeons of different seniority, the deep learning model matched the diagnostic abilities of the experienced surgeon, offering a non-invasive tool for junior clinicians to accurately and swiftly differentiate between STB and OVCF, and providing early diagnostic insights for senior surgeons. This demonstrates that the artificial intelligence models (especially the MVITV2 model) developed in this study could effectively differentiate between STB and OVCF on CT images.

Artificial intelligence, particularly deep learning and transfer learning, is increasingly being utilized in medical image analysis for lesion localization, disease diagnosis, tumor classification, and progression prediction.^37,38 Deep learning surpasses traditional machine learning by extracting complex features without predefined parameters.³⁸ Duan et al³⁹ used MRI-based deep learning models to distinguish spinal tumors from tuberculosis. Ke Liu et al⁴⁰ applied an MRI-based ResNet to identify the origins of spinal metastases, which included lung, kidney, prostate, thyroid, and breast cancer. Yuan Li et al⁴¹ demonstrated that ResNet50 could differentiate benign from malignant vertebral fractures on CT images. Chen et al⁴² showed that an attention-based model could separate multiple myeloma from spinal metastases. Kim et al⁴³ retrained Inception v3 with wrist X-rays for fracture classification. Cheng et al⁴⁴ employed Dense-U-Net for segmentation of vertebrae on CT, achieving full spine segmentation. Such neural network models convert medical images into high-dimensional features for tasks like lesion classification and segmentation,^45–48 offering applications in tumor detection, disease differentiation, and pathological analysis.^49–51

In recent years, the application of artificial intelligence, particularly deep learning, to medical imaging has seen remarkable progress. Previous studies have demonstrated the potential of deep learning models to differentiate between diseases with similar radiologic features. For example, Wang et al² successfully employed convolutional neural networks (CNNs) to distinguish between malignant and benign spinal lesions on MRI, while our study extends this approach to the differential diagnosis of STB and OVCF using CT images. Deep learning models offer the advantage of automatically extracting complex features from imaging data, often with better sensitivity and specificity than traditional methods (Zhao et al⁵²).

The application of deep learning, convolutional neural networks (CNNs), and Transformer models has significantly impacted medical imaging analysis. These algorithms are designed to autonomously learn from training datasets, enabling them to identify key features.^50,51 In this learning process, they adjust their internal parameters to appropriately weight these features, thereby creating models that can be tested with new datasets. In this study, we trained and evaluated three CNN models (ResNet50, ResNet101, and EfficientNet_B5) and a representative Transformer model (MVITV2).

The pathogenesis of spinal tuberculosis (STB) involves the direct infection of the vertebrae by Mycobacterium tuberculosis, leading to localized granulomatous inflammation and bone destruction. In the early stages, imaging primarily shows inflammatory changes, with minimal abnormal bone density. In contrast, osteoporotic vertebral compression fractures (OVCF) are mainly caused by a decrease in bone density due to osteoporosis. The occurrence of OVCF results from a reduction in the mechanical strength of the vertebrae, leading to compression fractures under minimal external forces. Early imaging may show a combination of hemorrhage, edema, and osteoporotic changes. Based on these mechanisms, our artificial intelligence-based image recognition model can accurately differentiate between these two diseases with high accuracy.

These models collectively represent the forefront of deep learning applications in the field of medical imaging, offering new avenues for accurate and efficient analysis. Our experiments demonstrated that the aforementioned deep learning models are feasible for the differentiation of fresh vertebral compression fractures and STB on CT images.

Although previous studies have highlighted the benefits of incorporating clinical attributes into diagnostic models,⁵³ our focus was on using CT imaging to differentiate STB from OVCF, especially among elderly women.⁵⁴ Given the unreliability of the patients’ tuberculosis history, with many cases lacking evidence of pulmonary tuberculosis or disease at other primary sites, and the commonness of osteoporosis in such patients, our research emphasized radiological over clinical data. Our findings suggest that deep learning models, without incorporating extensive medical history data or laboratory results, can match the diagnostic accuracy of experienced spinal surgeons. It is worth noting that the deep learning models developed in this study are still an auxiliary diagnostic tool. When their diagnostic results conflict with medical opinion, the final diagnosis should be comprehensively determined by doctors.

This study has several limitations. First, the relatively small sample size of the dataset may limit the generalization ability of the deep learning network models. Further training and validation with larger sample sizes from more centers are needed. Second, the neural network models developed in this study did not perform lesion localization and automatic segmentation, but were based on manual segmentation and correction by physicians. Third, the retrospective design of the study may introduce case selection bias, potentially affecting the credibility of the results. Our goal for future research is to continuously incorporate new cases, expand the sample size to enhance the efficiency and generalization ability of the model, and introduce vertebral automatic segmentation and localization functions into the model to avoid bias introduced by manual segmentation and improve segmentation efficiency.

In conclusion, this study explored the application of deep learning methods in differentiating spinal tuberculosis (STB) and acute osteoporotic vertebral compression fractures (OVCF) using CT images. Four models (ResNet50, ResNet101, EfficientNet_B5, and MVITV2) all demonstrated high accuracy, with the MVITV2 model performing exceptionally well on the internal dataset, and showing strong robustness in external validation. However, to further improve the generalizability and diagnostic accuracy of these models, continued training with larger, multi-center datasets is necessary.

From a clinical perspective, the implementation of these deep learning models provides clinicians with a rapid, efficient, and non-invasive diagnostic tool. Particularly in primary care settings where MRI may not be widely available but CT machines are more commonly used, the CT image analysis model developed in this study holds significant value in resource-limited environments. These models can reduce dependence on MRI, lower the need for invasive biopsy procedures and their associated risks, and assist clinicians in identifying lesions more quickly and accurately, thus improving the diagnostic pathway for patients.

Furthermore, the performance of these deep learning models approaches or even surpasses the diagnostic capabilities of some experienced spinal surgeons, providing strong technical support for less experienced clinicians or those in primary care. In challenging cases, these models can also offer supplementary information to senior clinicians, enhancing diagnostic confidence. Particularly for elderly female patients, where medical history may be unreliable and osteoporosis is highly prevalent, this study demonstrates that relying solely on CT imaging data, rather than patient history or laboratory results, can still achieve a high level of diagnostic accuracy. This is crucial for optimizing clinical decision-making processes and reducing diagnostic time.