MRI radiomics-based evaluation of tuberculous and brucella spondylitis

Abstract

Objectives

We analyzed magnetic resonance imaging (MRI) and radiomics labels from tuberculous spondylitis (TBS) and brucella spondylitis (BS) to build machine learning models that differentiate TBS from BS and culture-positive TBS (TBS(+)) from culture-negative TBS (TBS(−).

Methods

This retrospective study included 56 patients with BS, 63 patients with TBS(+) and 71 patients with TBS(−). Radiomics labels were extracted from T2-weighted fat-suppression images. MRI labels were analyzed via logistic regression (LR); radiomics labels were analyzed by t-tests, SelectKBest, and least absolute shrinkage and selection operator (LASSO). Random forest (RF) and support vector machine (SVM) models were established using radiomics or joint (radiomics+MRI) labels. Models were evaluated by receiver operating characteristic curves, areas under the curve (AUCs), decision curve analysis (DCA), and Hosmer–Lemeshow tests.

Results

When joint-label models were used to compare BS vs TBS(+) and BS vs TBS(−) groups, SVM AUCs were 0.904 and 0.944, respectively, whereas RF AUCs were 0.950 and 0.947, respectively; these were higher than the AUCs of the MRI label-based LR model. DCA showed that radiomics-based machine learning models had a greater net benefit; Hosmer–Lemeshow tests demonstrated good prediction consistency for all models.

Conclusions

Radiomics can help distinguish TBS from BS and TBS(+) from TBS(−).

Introduction

Human brucellosis is the most common zoonotic infectious disease worldwide, with approximately 500,000 new cases annually; more than 50% of affected individuals have spinal involvement, which manifests as brucella spondylitis (BS).^1–4 In recent decades, the global epidemiologic characteristics of BS have substantially shifted, and the overall incidence has begun decreasing. Most developed countries report a small number of BS cases every year, and some countries in the European Union (e.g., France, Germany, Sweden, Denmark, and the United Kingdom) have been officially declared brucellosis-free. However, new endemic areas continue to emerge in other regions, including Asia, the Middle East, and Africa.^3,5

Tuberculosis, a highly lethal infectious disease, is endemic worldwide. Approximately 2 billion people (one-quarter of the global population) are infected with Mycobacterium tuberculosis. ⁶ Although M. tuberculosis primarily affects the lungs, it theoretically can colonize any organ. Typically, 16% of patients develop extrapulmonary tuberculosis, but this proportion approaches 24% in the Mediterranean region. ⁷ M. tuberculosis infection in the osteoarticular system is most likely to involve the spine; 50% of patients with such involvement develop tuberculous spondylitis (TBS). ⁸

Common radiographic manifestations of BS include varying degrees of bone destruction at the vertebral margins, indistinct or mild wedge-shaped changes in vertebral shape, mild disc destruction, and paraspinal abscess formation. BS and TBS have similar clinical manifestations and imaging features; thus, they are difficult to distinguish. When a patient exhibits early signs of BS or TBS, neither disease has formed an abscess; both can cause inflammatory vertebral edema or paravertebral abscesses. Treatment for BS comprises doxycycline, rifampicin, streptomycin, and quinolones ⁹ ; treatment for TBS consists of isoniazid, rifampicin, pyrazinamide, and ethambutol. ⁵ Because the treatment approaches considerably differ, a clear diagnosis can guide clinical management; ultimately, drug therapy should be selected on the basis of microbial culture results.

This study was performed to explore pathogen-related heterogeneity in bone microenvironments. Quantitative analysis of high-dimensional data from medical images provides a non-invasive approach to address the limitations of conventional morphology-based imaging to distinguish between similar diseases, as well as the limitations of unaided visual observation. This approach is expected to assist clinicians in rapid diagnosis of pathogenic bacteria and improve treatment outcomes.

Methods

Patients

We reviewed the clinical data, imaging, laboratory, and pathological findings of all patients who underwent spinal magnetic resonance (MR) examination at the Second Hospital of Shanxi Medical University between January 2016 and January 2022. Patients with TBS and BS were identified using the diagnostic and inclusion/exclusion criteria described below. Data were recorded concerning patient age, sex, affected vertebrae, intervertebral disc involvement, and paravertebral abscess size (no abscess, no obvious abscess signal on magnetic resonance imaging [MRI]; small abscess, largest abscess smaller than adjacent vertebral body; large abscess, largest abscess larger than adjacent vertebrae). This study protocol was approved by the Ethics Review Committee of Shanxi Medical University (approval no. 2023 (YX) NO: 115), which waived the requirement for informed consent. This study was conducted in accordance with the Declaration of Helsinki. All patient details were deidentified prior to analysis. The reporting of this study conforms to the STARD guidelines.^10–12

Diagnostic criteria

Patients were diagnosed with TBS on the basis of the following criteria: ¹³ 1) common clinical manifestations of TBS and positive findings on spinal imaging examination; 2) current symptoms of tuberculosis or a history of tuberculosis; 3) detection of ongoing tuberculosis or old tuberculosis lesions on computed tomography (CT); 4) positive result in at least one purified protein derivative (PPD) test or T-SPOT test; 5) imaging-confirmed efficacy of regular anti-tuberculosis treatment; 6) positive bacterial culture findings on analysis of specimens collected via puncture or surgical excision; 7) negative bacterial culture findings, despite lesion pathological characteristics consistent with tuberculosis. A diagnosis of TBS(+) was made when conditions 1, any of 2 through 5, and 6 were met; a diagnosis of TBS(−) was made when conditions 1, any of 2 through 5, and 7 were met.

Patients were diagnosed with BS on the basis of the following criteria: 1) common clinical manifestations of BS and positive findings on spinal imaging examination; 2) an epidemiological history or any possible history of contact (e.g., animal husbandry, work in the livestock industry, or contact with individuals engaged in those activities; or consumption of immature beef, mutton, pork, or milk and goat milk products; or a history of rural travel); 3) positive Rose Bengal test results (i.e., slide agglutination); 4) standard tube agglutination test (SAT) titer ≥1:160; 5) positive bacterial culture findings on analysis of specimens collected via puncture or surgical excision. A diagnosis of BS was made when conditions 1, either 2 or 3, and either 4 or 5 were met.

Inclusion and exclusion criteria

Inclusion criteria were as follows: complete clinical, imaging, etiological, and pathological data; initial diagnosis of TBS or BS made in our hospital; and good spinal MR image quality without obvious artifacts.

Exclusion criteria were as follows: antibiotic therapy and puncture biopsy performed within 3 months before MR examination; history of spinal surgery; history of hematological disease; history of metabolic abnormalities affecting bone, calcium, and/or phosphorus; history of malignant tumor or radiotherapy; spinal trauma within the preceding 3 months; involvement of spinal attachment sites only, without abnormal vertebral body signals; and/or severe vertebral bone destruction or spinal deformity preventing vertebral body segmentation in the region of interest.

Imaging protocol

A General Electric MR scanner was used for imaging in this study, with spinal or body coils. The following imaging sequences were collected: sagittal T1-weighted imaging (repetition time [TR], 450–640 ms; echo time [TE], 8–11 ms), T2-weighted imaging (T2WI) (TR, 2400–2800 ms; TE, 100–130 ms), and T2WI–fat-suppression (TR, 1700–2300 ms; TE, 50–60 ms). All sagittal images had the following parameters: slice thickness, 3.5 to 4.5 mm; slice spacing, 4 to 4.5 mm; field of view, 320 mm  × 320 mm; and matrix, 320 × 320. Axial T2WI was performed with the following parameters: TR, 4500 to 5500 ms; TE, 100 to 130 ms; slice thickness, 5 mm; slice spacing, 6 mm; field of view, 200 mm  × 200 mm; and matrix, 512 × 512. Two 1.5T devices (Signa HDxt and Signa EXCITE) were used to suppress fat signals by short time inversion recovery (STIR) technology (inversion time [TI], 150 ms); two 3.0T devices (Discovery 750w and Pioneer) were used to suppress fat signals via water-lipid separation (i.e., Dixon technique).

Radiomics

The image biomarker standardization initiative ¹⁴ and radiomics quality score ¹⁵ were used to ensure reproducibility. Images were segmented using 3D Slicer (https://www.Slicer.org, version 4.11); image preprocessing, radiomics feature extraction, and feature selection were conducted in Python software (https://www.python.org, version 3.8).

For image preprocessing and segmentation, deidentified sagittal T2WI–fat-suppression images of the spine were converted to DICOM format, then converted to NIfTI format in Python software. N4 bias field correction was performed using the Simple ITK package in Python software to eliminate local differences in image grayscale values related to magnetic field inhomogeneity. Two physicians (each with >10 years of experience in bone and joint MR examinations) independently used 3D Slicer software to segment the involved vertebrae. Beginning at the second layer of the vertebral body, the involved vertebral body was delineated along a 2-mm area outside of the bone cortex until the first layer where the vertebral body disappeared; this process yielded a volume of interest (VOI), as shown in Figure 1. Both physicians were blinded to the final diagnosis of each image; disagreements about vertebral body involvement were resolved by discussion. Each segmented original image and mask were resampled; voxels were adjusted to a consistent size (1 mm × 1 mm × 1 mm) to ensure uniformity in all directions and to reduce the effects of slice thickness and spacing related to analysis of images from different devices.

Figure 1.

Schematic representation of image segmentation. (a) The patient's original T2-WI–fat-suppression images showed diffuse hyperintensities in the third and fourth vertebral bodies of the lower back, with an SAT titer of 1:320; the patient was diagnosed with BS. (b) Stratification of affected vertebral bodies and (c) VOI generation for each affected vertebral body.

BS, brucella spondylitis; SAT, standard tube agglutination test; T2-WI, T2 weighted imaging; VOI, voxel of interest.

For feature extraction, the PyRadiomics package in Python software was used to extract radiomics features from resampled VOIs. In total, 1409 labels were obtained from seven basic features and nine higher-order transformation features.

For feature selection, samples were divided into three comparison groups: BS vs TBS(+), BS vs TBS(−), and TBS(+) vs TBS(−); separate radiomics features were selected for each comparison. In patients with multiple vertebral lesions, data in each feature label column were sequentially averaged according to the number of lesioned vertebrae ( $\overline{x}=\frac{{\sum }_{i=1}^n{\chi }_i}{n}$ , where n indicates number of lesioned vertebrae); subsequently, the mean was used for feature identification. The Epps–Pulley test was used to analyze data normality under each label; the Leven test was used to analyze homogeneity of variance. Features with P-values <0.05 were initially selected using Student's t-test, Welch's t-test, or the Mann–Whitney U test. Next, the linear correlation between each feature and the classification result was calculated and transformed into an F_score using the SelectKBest function; features were selected according to F_score, yielding features strongly correlated with classification results. Finally, the least absolute shrinkage and selection operator (LASSO) method was used to select the top 100 features most strongly correlated with the classification results; non-zero coefficient features with optimal λ values were identified by cross-validation (CV), as shown in Figure 2.

Figure 2.

Schematic representation of radiomics feature identification process in BS vs TBS(+) comparison integrated by CV and LASSO; the same processing method was used for the other two comparisons. (a) Diagram of process to select optimal λ parameter in LASSO equation by CV: y-axis represents the mean square error, x-axis represents log (λ) value retained, according to number of features. Points indicate corresponding log (λ) values of mean square error (with 95% confidence intervals). Dotted lines indicate unique log (λ) values, where values on the left side represent minimum mean square error and suggest 16 radiomics features, while values on the right represent one standard deviation increase in mean square error of the equation and suggest 14 radiomics features; optimal λ values were chosen from both left and right sides. Use of values on the right led to retention of fewer features and prevention of model overfitting. (b) LASSO equations in which each coefficient changed as a function of λ: x-axis represents log (λ) and y-axis shows the labels of all characteristic coefficients as λ increases; each tag coefficient gradually approaches 0 (the coefficient of 0 is excluded).

BS, brucella spondylitis; CV, cross-validation; LASSO, least absolute shrinkage and selection operator; (+), culture-positive.

Statistical analysis

Statistical analyses in this study were conducted in Python and R (version 4.0.5, https://www.rproject.org) softwares. Two-sided P-values <0.05 were considered statistically significant. Continuous data with a normal distribution were expressed as $\overline{x}\hspace{0.17em}\pm \hspace{0.17em}\text{s}$ . Differences in MRI labels (age, sex, affected vertebrae, intervertebral disc involvement, and paravertebral abscess size) were analyzed via multifactorial logistic regression (LR). The Dice coefficient was used to evaluate similarity between VOIs outlined by the two physicians, $\mathit{Dice}=\frac{2|A\cap B|}{\left|A\right|+\left|B\right|}$ , where A and B represent VOIs outlined by physicians 1 and 2, respectively. The Dice coefficient takes a value between 0 and 1; values closer to 1 indicate greater VOI similarity and smaller error. Intra- and interobserver correlation coefficients (ICCs) were used to measure intra- and interobserver agreement concerning radiomic feature extraction. ICCs >0.75 indicated good agreement.

The LR model was built using MRI labels; the machine learning model was built with radiomics labels and radiomics labels plus MRI labels (i.e., joint labels) using random forest (RF) and support vector machines (SVM) algorithms. A CV method was used to divide the data into a training set and a test set. Parameters were optimized and each model was trained using the training set. The radial basis function kernel constituted the core function of the SVM; c and γ were determined by CV. In this study, n_estimators was set to 500, with the goal of balancing accuracy and computational burden; the remaining parameters were set to default values in the sklearn package in Python software.

Receiver operating characteristic (ROC) curves were used to evaluate the classification ability of each model. ¹⁶ Secondary indicators were 95% confidence intervals (CIs), as well as specificity and sensitivity values. Decision curve analysis (DCA) was used to evaluate the net benefits of each model; the Hosmer–Lemeshow test was used to establish calibration curves that could assess agreement between predicted and actual probabilities generated by the models.

Results

Sample composition

In total, 190 patients were included in this study: 56 in the BS group (29 men and 27 women, mean age 56.25 ± 12.89 [21–74] years), 63 in the TBS(+) group (34 men and 29 women, mean age 55.24 ± 14.83 [23–81] years), and 71 in the TBS(−) group (39 men and 32 women, mean age 55.03 ± 13.26 [23–81] years) (Table 1). There were significant differences (all P < 0.05) in intervertebral disc involvement and paravertebral abscess size between BS and TBS(+) and between BS and TBS(−), but the etiologies of TBS could not be predicted.

Table 1.

Clinical characteristics and MRI features of patients with BS, TBS(+) and TBS(−).

Label	BS (n = 56)	TBS (+) (n = 63)	TBS (−) (n = 71)	P-value
Age	56.25 ± 12.89	55.24 ± 12.89	55.03 ± 12.89	0.990 a
				0.314 b
				0.949 c
Sex
Male	29 (51.8)	34 (54.0)	39 (54.9)	0.724 a
Female	27 (48.2)	29 (46.0)	32 (45.1)	0.346 b
				0.970 c
Location
Neck	0 (0)	0 (0)	2 (2.8)	0.112 a
Neck–thoracic	0 (0)	0 (0)	0 (0)	0.573 b
Thoracic	17 (30.3)	16 (25.4)	17 (23.9)	0.873 c
Thoracic–lumbar	8 (14.3)	21 (33.3)	19 (26.8)
Lumbar	31 (55.4)	26 (41.3)	33 (46.5)
Disc involvement
Yes	19 (33.9)	42 (66.7)	47 (66.2)	0.002 a
No	37 (66.1)	21 (33.3)	24 (33.8)	0.001 b
				0.957 c
Paravertebral abscess
No	14 (25.0)	6 (9.5)	7 (9.9)	0.009 a
Small	31 (55.4)	26 (41.3)	28 (39.4)	0.009 b
Large	11 (19.6)	31 (49.2)	36 (50.7)	0.983 c

Note: Binary logistic regression was used for pairwise comparison to identify relevant clinical characteristics and MRI features.

^aBS vs TBS(+).

^bBS vs TBS(−).

^cTBS(+) vs TBS(−).

BS, brucella spondylitis; MRI, magnetic resonance imaging; TBS, tuberculous spondylitis; (+), culture-positive; (−), culture-negative.

Radiomics labeling

The mean Dice coefficient of the VOIs outlined by the two physicians was 0.983, indicating good agreement. Fourteen, 12, and 10 radiomics labels were identified based on BS vs TBS(+), BS vs TBS(−), and TBS(+) vs TBS(−) comparisons, respectively. The ICCs of all retained features were >0.75, indicating good measurement consistency. As shown in Figure S1, the three most relevant features for classification in the BS vs TBS(+) comparison were size zone non-uniformity normalization (SZNN) under gray level size-zone matrix (GLSZM), run entropy under gray level run-length matrix (GLRLM), and dependence non-uniformity under gray level dependence matrix (GLDM). In the other two comparisons, the top three most relevant features were shape flatness, maximum two-dimensional diameter column on the coronal plane, contrast under neighboring gray tone difference matrix (NGTDM), shape sphericity, dependent variance under gray level dependence matrix (GLDM), and informational measure of correlation (IMC) under gray level co-occurrence matrix (GLCM).

Model evaluation

In the BS vs TBS(+), BS vs TBS(−), and TBS(+) vs TBS(−) comparisons, the areas under the curve (AUCs) of the radiomics label-based SVM model were 0.877 (95% CI 0.753–1.000), 0.923 (95% CI 0.843–1.000), and 0.857 (95% CI 0.739–0.957), as shown in Table S1. For the radiomics label-based RF model, these values were 0.914 (95% CI 0.824–1.000), 0.924 (95% CI 0.846–1.000), and 0.929 (95% CI 0.854–1.000), as shown in Table S2. The AUCs of the joint label-based SVM model in the BS vs TBS(+) and BS vs TBS(−) comparisons were 0.904 (95% CI 0.799–1.000) and 0.944 (95% CI 0.863–1.000), respectively; these values for the joint label-based RF model were 0.950 (95% CI 0.886–1.000) and 0.947 (95% CI 0.884–1.000), respectively (Table S3). Notably, for the MRI label-based LR model, these values were 0.769 (95% CI 0.683–0.856), and 0.776 (95% CI 0.967–0.858), respectively, as shown in Table S3. Because MRI labeling could not assess the etiologies of TBS, joint label-based SVM and RF models and an MRI label-based LR model could not be constructed for the TBS(+) vs TBS(−) comparison. The specificity and sensitivity parameters of each model are detailed in Tables S1 to S3.

The ROC curves of each model are shown in Figure S5. Z tests of the ROC curves showed that the classification abilities of the radiomics label-based RF model and joint label-based RF model were better than the classification ability of the MRI label-based LR model only in the BS vs TBS(+) comparison (both P < 0.05). In the BS vs TBS(−) comparison, the SVM and RF models outperformed the LR model (both P < 0.05), as shown in Table 2. DCA of each model (Figure 3) revealed that the net benefit was better for the joint label-based RF model than for the other models in the BS vs TBS(+) comparison when the risk threshold was >0.36; the net benefit of the MRI label-based LR model was lowest among the three models. Similar results were observed in the BS vs TBS (−) comparison. Calibration curves derived from Hosmer–Lemeshow tests demonstrated that there was no statistically significant difference between the predicted and actual probabilities generated by the three models. Notably, the curves were not significantly biased before or after calibration; the mean prediction errors of the models are depicted in Figures S2 to S4.

Table 2.

ROC curve Z tests of various models.

Model	BS vs TBS (+)		BS vs TBS (−)		TBS (+) vs TBS (−)
	Z-value	P-value	Z-value	P-value	Z-value	P-value
SVM a vs RF a	−0.481	0.63	−0.017	0.986	−1.007	0.314
SVM b vs RF b	−0.731	0.465	−0.052	0.959
SVM a vs LR	1.393	0.164	2.512	0.012
SVM b vs LR	1.937	0.053	2.882	0.004
RF a vs LR	2.275	0.022	0.564	0.01
RF b vs LR	3.301	<0.001	3.249	0.001

^aModel constructed based on radiomics tags.

^bModel constructed based on joint tags; LR classification is solely based on MRI labels.

BS, brucella spondylitis; LR, logistic regression; MRI, magnetic resonance imaging; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; TBS, tuberculous spondylitis; (+), culture-positive; (−), culture-negative.

Figure 3.

DCA curves of label-based models in each comparison.

ALL, all patients are culture-positive and receive treatment (negative net benefit); BS, brucella spondylitis; DCA, decision curve analysis; LR, logistic regression; MRI, magnetic resonance imaging; NO, all patients are culture-negative and receive no treatment (no net benefit); RF, random forest; SVM, support vector machine; TBS, tuberculous spondylitis; (+), culture-positive; (−), culture-negative ALL and NO are the two extremes of the DCA curve.

Figure 4.

(a) Magnetic resonance image of a patient with diffuse edema in the thoracic 11th and 12th vertebrae, wedge-shaped transformation of thoracic 12th vertebrae, stenosis of thoracic 11 and 12th vertebrae, and kyphosis. Arrow indicates infected vertebrae and (b) Signs of a small paravertebral abscess, supporting an imaging diagnosis of “infectious lesion of spine with high possibility of tuberculosis.” Arrows indicate abscesses around infected vertebrae. An in vitro agglutination test yielded a titer of 1:320 and brucella culture results were positive; the patient was diagnosed with BS. The images in this case were classified as BS according to SVM and RF models, consistent with the results of clinical examinations.

BS, brucella spondylitis; RF, random forest; SVM, support vector machine.

Discussion

BS and TBS are infectious bone marrow lesions caused by different pathogenic bacteria. After bacterial colonization of bone marrow, distinct inflammation and swelling manifestations occur. These pathological changes, which produce heterogeneity within bone marrow lesions, provide a theoretical basis for pathophysiological analyses. The SLC11A1 gene is reportedly associated with susceptibility to TBS, indirectly indicating that the bone marrow microenvironment differs between TBS and BS. ¹⁷ Extraction of ¹⁸F-fluorodeoxyglucose positron emission tomography (PET)/CT radiomics features from examinations of thoracic 5–7 and lumbar 2–4 vertebrae in patients with multiple myeloma revealed that low gray-level zone emphasis (LGLZE) is an important marker for multiple myeloma progression. ¹⁷ In contrast to GLSZM, both GLDM and GLCM are matrices with directional properties. GLDM indicates the number of identical grayscale voxels within a specific distance from the central voxel, representing textural homogeneity with dependent variance; thus, it constitutes an imaging biomarker that reflects heterogeneity within the cellular microenvironment. ¹⁸ GLCM represents the number of times that a combination of two (similar or different) grayscale voxels occurs at a particular distance and angle. The IMC is a measure of the proportion of the two voxels in a particular combination that quantify texture complexity; Ekert et al. reported that the IMC is an important radiomics label for evaluating multiple myeloma prognosis. ¹⁹

In addition to the three radiomics labels mentioned above, which were present in all groups, skewness under first-order statistical features is strongly correlated with bone density ²⁰ ; coarseness and contrast under NGTDM can help to distinguish osteogenic metastases from sclerotic bone lesions, on the basis of mineral content in bone. ²¹ Furthermore, we found that radiomics labels in the TBS(+) vs TBS(−) comparison were mainly wavelet and Lorentzian over Gaussian (LoG) features. Wavelets have been widely used in computer graphics in recent years because they overcome the inability of conventional Fourier transform to analyze the time domain of a signal, and they can arbitrarily focus the time-frequency information of a particular signal. ²² The LoG approach integrates Gaussian and Laplace transformations, applying them secondarily to the spatial parameters of the derivative image after smoothing to eliminate noise; thus, it can identify voxels in images with substantial changes in grayscale. ²³ Correlation studies have shown that wavelet and LoG transformations reflect lesion heterogeneity.^15,24 The pathogenic bacteria causing TBS(+) and TBS(−) are identical, and there is minimal heterogeneity between these diseases at the cellular level. The difference presumably arises from the amount of bacteria involved, and thus more sensitive imaging biomarkers are needed to clearly distinguish these diseases; the relevant radiomics labels mostly constitute higher-order transformed features.

Although there are differences in MRI presentation between BS and TBS, misdiagnosis often occurs because there is minimal specificity in terms of intervertebral space alteration and paravertebral abscess size.^25–27 This lack of specificity was confirmed in the present study. in the BS vs TBS(+) and BS vs TBS(−) comparisons, the MRI label-based LR model demonstrated lower AUC, upper limit of 95% CI, specificity, sensitivity, and other indicators, compared with the radiomics label-based and joint label-based RF models. Previous studies have shown that the most common types of TBS based on imaging presentation are epiphyseal and subligamentous, which can be easily distinguished from BS using the MRI labels described above. However, recent studies have shown that the number of patients with TBS involving atypical imaging features is gradually increasing because of the prevalence of drug-resistant strains and human immunodeficiency virus. ²⁸ Additionally, although the amount of necrotic bone and extent of bone hyperplasia differ between TBS and BS, TBS lesions are often detected at a later stage; BS lesions are difficult to distinguish from osteophyte activity related to spinal degeneration. ²⁹

This study also revealed that a radiomics label-based machine learning model can facilitate the assessment of bacterial culture results in patients with suspected TBS. Because TBS and BS are infectious lesions, bacterial culture remains the gold standard laboratory test to identify both diseases. Serological examination of M. tuberculosis infection is recommended in various guidelines, and the results are reliable; nevertheless, PPD and T-SPOT tests only provide reference information in most instances because of the complex immune mechanism involved, and bacterial culture remains the primary assessment method in clinical practice. Because M. tuberculosis is an aerobic bacillus with minimal colonization of deep tissues, focal specimens of TBS yield successful culture results much less frequently, compared with pulmonary tuberculosis specimens.^30–32 There is a particular lack of characteristic tuberculosis symptoms in children with TBS; thus, rapid and accurate examination techniques that can detect M. tuberculosis infection are needed to prevent sequelae. ³³ In recent years, the Xpert MTB/RIF technology promoted by the World Health Organization has partially solved the problem of extended culture to detect M. tuberculosis and guide timely treatment, but it is not suitable for extrapulmonary specimens. ³⁴ Held et al. demonstrated that the GeneXpert technique could be used to diagnose TBS, but this approach requires surgical samples; therefore, it is invasive, expensive, and impractical. ³⁵ In contrast, the present study showed that a radiomics-based machine learning model could distinguish TBS(+) and TBS(−), enabling the initial diagnosis of TBS through routine pathological observation and imaging examination. This approach provides a novel solution to the problems of extended culture time and low positivity rate: non-invasive evaluation of pathogenic lesion information using rapid and inexpensive imaging techniques.

This study had some limitations. First, it was a single-center study, and the results may not be generalizable to other institutions. Therefore, multicenter analyses should be performed to verify the findings. Second, the number of retained imaging histology labels was relatively large; model overfitting is possible, and the AUC may decrease with increasing sample size. Third, this was a retrospective study; because the incidence of BS is relatively low and SAT results have diagnostic value according to various guidelines, bacterial culture results were unavailable for patients in the BS group, which may have led to some classification bias.

Conclusion

The results of this study indicate that radiomics can aid in the differential diagnosis of TBS and BS; it also has the potential to facilitate analyses of bacterial culture results in patients with suspected TBS. Moreover, radiomics label-based RF models and joint label-based RF models displayed greater classification abilities, compared with an MRI label-based LR model. Radiomics has high applied and experimental value with respect to infectious spinal lesions.