Fusing Radiomics and Ultrasonic Features: A Superior Nomogram for Diagnosing Cervical Tuberculous Lymphadenitis

Abstract

Objective

This study aims to develop and validate a radiomics-based nomogram derived from ultrasound (US) images for differentiating cervical tuberculous lymphadenitis (CTL) from non-CTL in the pre-biopsy setting.

Methods

A retrospective analysis was conducted on 287 patients with biopsy-confirmed cervical lymphadenopathy (116 CTL and 171 non-CTL), all of whom underwent conventional US examinations. The US images were randomly split into training and validation cohorts at a 7:3 ratio. Radiomics features were extracted from manually delineated regions of interest (ROIs) using PyRadiomics. Features with intra- and inter-observer ICC > 0.75 were retained. Significant features were selected via t-test and LASSO regression. A radiomics score (radscore) was calculated, and a nomogram was constructed by integrating the radscore with US features (necrosis, heterogeneous, and posterior enhancement). Diagnostic performance was evaluated using ROC curves, calibration curves, decision curve analysis (DCA), and the Hosmer–Lemeshow goodness-of-fit test.

Results

The nomogram demonstrated discriminative performance, with AUCs of 0.798 (95% CI: 0.738–0.858) in the training cohort and 0.756 (95% CI: 0.654–0.858) in the validation cohort. Its AUC was higher than that of the clinical US model (AUC: 0.764 (95% CI: 0.700–0.829) and 0.744 (95% CI: 0.640–0.847), respectively) and the radiomics-only model (AUC: 0.693 (95% CI: 0.621–0.765) and 0.611 (95% CI: 0.492–0.730), respectively). The AUC difference between the nomogram and the clinical US model was not statistically significant, but calibration and DCA confirmed its favorable clinical applicability and goodness of fit.

Conclusion

The US-based radiomics nomogram may help differentiate CTL from non-CTL and serve as an adjunct to pre-biopsy assessment. This approach shows promise for improving the pre-biopsy diagnostic workup of tuberculous lymphadenitis, although formal subgroup validation in pathologically early-stage cases is warranted in future studies. Unlike “black-box” machine learning models, this nomogram provides a clear display of each predictor’s contribution, which makes the model easier to interpret.

Introduction

Tuberculosis (TB) remains one of the top 10 causes of mortality worldwide and is the leading cause of death attributable to a single infectious agent.^1,2 Extrapulmonary tuberculosis constitutes a significant proportion of tuberculosis,³ accounting for approximately 20% of all TB infections. Among extrapulmonary manifestations, the lymphatic system is the most frequently involved.⁴ Culture testing, the tuberculin skin test (TST), and fine-needle aspiration (FNA) cytology are commonly employed to achieve an early diagnosis of cervical tuberculous lymphadenitis (CTL).⁵ However, each of these methods has inherent limitations. Culture testing requires a prolonged incubation period (6–8 weeks) and cannot provide a rapid diagnosis.⁶ TST is a valuable but non-specific evaluation method.⁷ It relies on cytological alterations rather than direct bacterial detection, and its specificity is confounded by inflammatory processes unrelated to tuberculosis.⁸

Compared with other diagnostic approaches, imaging techniques are non-invasive, rapid, and convenient. Nowadays, various imaging modalities are available for cervical lymphadenopathy evaluation. CT and MRI provide detailed anatomical information for deep or extensive lesions, while PET/CT offers functional metabolic data to help differentiate benign from malignant disease.⁹ However, these modalities are costly and less accessible—with CT and PET/CT, in particular, involving exposure to ionizing radiation. In contrast, ultrasound is non-invasive, real-time, widely accessible, cost-effective, and radiation-free, making it the preferred first-line modality for superficial lymph node assessment.¹⁰ With growing public health awareness and the widespread availability of various diagnostic methods, lymph node tuberculosis is often discovered in the early stages of the disease, prior to the formation of typical tuberculous granulomas.^11,12 Under such circumstances, characteristic findings such as necrosis and calcification may be rare or absent on US, posing diagnostic challenges for conventional sonography in distinguishing tuberculous from non-tuberculous lymph node lesions.¹³ In addition, conventional US interpretation depends on operator experience and is inherently subjective.¹⁴ Therefore, conventional US alone may be insufficient for accurate discrimination of CTL from non-CTL.¹⁵

Radiomics enables high-throughput quantitative analysis of medical imaging data through automatic or semi-automatic software, providing more comprehensive information than visual interpretation by physicians and thereby improving image analysis.¹⁶ This method was first demonstrated to be useful in analyzing CT or MRI images on diagnosis of clinical disease.¹⁷ More recently, US-based radiomics has also been applied to diagnostic tasks.¹⁸ In addition, deep learning approaches have also been shown to exhibit promising performance in medical image analysis, but their “black-box” nature limits clinical adoption. In contrast, radiomics-based nomograms provide a transparent and visual scoring system for clinical disease diagnosis and prognosis, which allows clinicians to directly visualize the contribution of each predictor and calculate individualized risk probabilities at the point of care. This interpretability is particularly valuable in routine clinical practice.

Accurate discrimination between CTL and non-CTL is clinically important because CTL can be treated with antituberculous therapy.^19,20 In this study, radiomics features were extracted from tuberculous and non-tuberculous lymph nodes, and an integrated US-radiomics nomogram was developed. The diagnostic performance of the model was then evaluated for differentiating CTL from non-CTL.

Materials and Methods

Study Cohorts

A total of 116 patients with CTL and 171 patients with non-CTL lesions, all biopsy-confirmed, who had undergone conventional US examinations prior to biopsy or surgery between January 2019 and February 2023, were retrospectively reviewed. Patient enrollment was consecutive; all eligible patients who met the inclusion criteria during the study period were included, provided none of the exclusion criteria applied. The non-CTL group comprises reactive hyperplasia (n = 96, 56.1%), lymphoma (n = 41, 24.0%), and metastatic lymph nodes (n = 34, 19.9%), thereby reflecting the diverse differential diagnoses encountered in routine clinical practice. The images were randomly divided into training and validation cohorts at a ratio of 7:3. The requirement of informed consent was waived due to the retrospective nature of the study.

The exclusion criteria were as follows: (1) poor image quality or incomplete data; (2) indistinct or unidentifiable lesion boundaries in imaging; (3) previous history of anti-tuberculosis therapy; (4) concurrent HIV infection. The study workflow and the flowchart of patient inclusion are shown in Figures 1 and 2, respectively. This study was reported in accordance with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement.

Figure 1.

Flowchart of patient inclusion and radiomics feature extraction process. A total of 287 patients with cervical lymphadenopathy were retrospectively enrolled, including 116 with cervical tuberculous lymphadenitis (CTL) and 171 with non-CTL lesions. Ultrasound images were divided into training and validation sets in a 7:3 ratio. Regions of interest (ROIs) were manually delineated, and radiomics features were extracted using PyRadiomics. Features with intra- and inter-observer ICC > 0.75 were retained for further analysis.

Figure 2.

Flowchart of patient selection and cohort division. A total of 650 patients with cervical lymphadenopathy were initially enrolled. After applying exclusion criteria (poor image quality, indistinct lesion boundaries, previous anti-tuberculosis therapy, and HIV co-infection), 287 patients with biopsy-confirmed diagnoses were included in the final study cohort. These patients were then randomly divided into a training cohort (n=200) and a validation cohort (n=87) at a ratio of approximately 7:3.

Abbreviations: CTL, cervical tuberculous lymphadenitis; non-CTL, non-tuberculous lymphadenitis.

US Data Acquisition

Conventional US examinations were performed using a GE LOGIQ E9 system with an ML6-5 linear-array transducer operating at 6–15 MHz. Patients were examined in the supine or lateral position, as appropriate for the lymph-node location. Experienced radiologists (more than 10 years of US experience) assessed echogenicity, lymph-node hilum, heterogeneity, margin, internal calcification, necrosis, and posterior acoustic enhancement.

Image Region of Interest Segmentation

The grayscale US image showing the largest cross-sectional area of the target lymph node was selected. The image was imported into the professional radiomics open-source software ITK-SNAP, where a senior attending physician manually delineated the boundaries of the lymph node to define the Region of Interest (ROI), ensuring that the delineated ROI tightly covers the entire target lymph node. For intra-observer reproducibility, the same radiologist repeated the delineation in 30 randomly selected images after one week. For inter-observer reproducibility, a second senior radiologist independently delineated the same 30 images while blinded to the first radiologist’s results. Intra- and inter-observer reproducibility of the extracted radiomics features was assessed using the intraclass correlation coefficient (ICC).

Radiomics Feature Extraction and Selection

The ROI images undergo radiomics feature extraction using the open-source Python package Pyradiomics (http://pyradiomics.readthedocs.io). Features were extracted from both the original grayscale images and wavelet-filtered images (using the PyWavelets package, which yields eight decompositions per level corresponding to all possible combinations of high- and low-pass filters in each of the two dimensions). Laplacian of Gaussian (LoG) filtering was not applied, as the PyRadiomics LoG implementation requires three-dimensional input and our analysis was based on 2D ultrasound images. The extracted features were primarily classified into three categories: First Order Features, Texture Features and Shape Features.

Development of Radscore and Nomogram Models

Development of Radscore

Radiomics features with intra- and inter-ICC values exceeding 0.75 were retained, and the data were then split into training and validation cohorts (7:3), followed by Z-score normalization. First, independent-samples t tests were applied to the radiomics features, and features with P < 0.05 were retained for further analysis. Second, Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression with 5-fold cross-validation was then used to select an optimized feature subset. The optimal tuning parameter λ was determined as the value that minimized the mean squared error (MSE) across the cross-validation folds. Finally, a linear combination of the selected features and coefficient vector was utilized to compute the radiomics signature (radscore) for each individual lymph node.

Development and Validation of Nomogram Model

Chi-square tests (or Fisher’s exact test when expected frequencies were < 5) were performed on US-based characteristics, and variables with P < 0.05 were selected. A radiomic nomogram was then created through the integration of both the Rad score and the Ultrasound characteristics. Receiver operating characteristic (ROC) curves were plotted to evaluate the diagnostic efficacy of the radiomic nomogram. Calibration curves were generated to assess the calibration efficiency of the nomogram, and the Hosmer-Lemeshow test was employed to evaluate its calibration capability. Decision curve analysis (DCA) was utilized to evaluate the clinical utility of the predictive models. These analyses were performed in both the training and validation cohorts.

Statistical Analysis

Statistical analyses were performed using SPSS 23.0 software. Categorical variables were expressed as proportions and compared using the chi-square test, or Fisher’s exact test when expected frequencies were < 5. Histopathological results were used as the diagnostic gold standard. ROC curves were plotted, and the area under the curve (AUC), specificity, sensitivity, accuracy, positive predictive value, and negative predictive value were calculated to evaluate model performance. Pairwise comparisons of correlated AUCs were performed using the DeLong test. All tests were two-sided, with P < 0.05 considered statistically significant.

Results

Ultrasound Characteristics

Table 1 presents the comparison of US-based characteristics. The training cohort included 85 (42.5%) CTL and 115 (57.5%) non-CTL, the validation cohort included 31 (35.6%) CTL and 56 (64.4%) non-CTL.

Table 1.

Baseline Sonographic Characteristics of Tuberculous and Non-Tuberculous Cervical Lymphadenopathy

Feature	Overall (n=287)	Non-CTL Group (n=171)	CTL Group (n=116)	P-value
Echo
-Abnormal	5(1.7%)	3(1.8%)	2(1.7%)	1.000†
-Normal	282(98.3%)	168(98.2%)	114(98.3%)
Lymph node hilum				0.119
-Absent	70(24.4%)	36(21.1%)	34(29.3%)
-Present	217(75.6%)	135(78.9%)	82(70.7%)
Heterogeneous				0.246
-Yes	117(40.8%)	65(38.0%)	52(44.8%)
-No	170(59.2%)	106(62.0%)	64(55.2%)
Margin				<0.001
-Irregular	138(48.1%)	66(38.6%)	72(62.1%)
-Well-defined	149(51.9%)	105(61.4%)	44(37.9%)
Internal calcification				0.542†
-Present	11(3.8%)	8(4.7%)	3(2.6%)
-Absent	276(96.2%)	163(95.3%)	113(97.4%)
Necrosis				<0.001
-Present	56(19.5%)	16(9.4%)	40(34.5%)
-Absent	231(80.5%)	155(90.6%)	76(65.5%)
Posterior enhancement				<0.001
-Present	88(30.7%)	37(21.6%)	51(44.0%)
-Absent	199(69.3%)	134(78.4%)	65(56.0%)

Note: Data are presented as number of cases (percentage). P-values were calculated using the chi-square test, or ^†Fisher’s exact test when expected frequencies were < 5.

Abbreviations: CTL, cervical tuberculous lymphadenitis; non-CTL, non-tuberculous lymphadenitis.

Features Selection and Model Construction

A total of 108 radiomics features were initially extracted, of which 14 were retained after t-testing. Non-zero coefficients were subsequently identified using the LASSO logistic regression model (Figure 3). Multiple models were evaluated in both cohorts for diagnostic performance, including Logistic Regression, Decision Tree, Random Forest, Naive Bayes, and Support Vector Machine (SVM) classifiers (Figure 4). Within the training cohort, the combined model achieved an AUC of 0.798 (95% CI: 0.738–0.858). Within the validation cohort, the Unite model achieved an AUC of 0.756 (95% CI: 0.654–0.858). Based on the confusion matrix comparison (Table 2), the Logistic Regression model was selected for nomogram construction due to its superior and stable performance across both cohorts, with the smallest performance degradation between training and validation.

Figure 3.

Feature selection using Least Absolute Shrinkage and Selection Operator (LASSO) regression. (A) Mean squared error (MSE) versus log(λ) curve with 5-fold cross-validation. The dashed vertical line indicates the optimal λ value selected based on minimum MSE. (B) Coefficient profiles of radiomics features across different λ values. The vertical line corresponds to the optimal λ where features with non-zero coefficients are selected. (C) Weight distribution of the selected features used to compute the radiomics score (radscore).

Figure 4.

Receiver operating characteristic (ROC) curves of five machine learning models in the training and validation cohorts. (A) Training cohort: Logistic Regression (AUC = 0.798), Decision Tree (AUC = 0.652), Random Forest (AUC = 0.738), Naive Bayes (AUC = 0.774), and Support Vector Machine (SVM, AUC = 0.784). (B) Validation cohort: Logistic Regression (AUC = 0.756), Decision Tree (AUC = 0.624), Random Forest (AUC = 0.609), Naive Bayes (AUC = 0.732), and SVM (AUC = 0.716). The Logistic Regression model was selected for nomogram construction due to its superior and stable performance.

Table 2.

Confusion Matrices of Five Machine Learning Classifiers in Training and Validation Cohorts

Model	Cohort	TP	FP	FN	TN
Logistic Regression	Training	62	32	23	83
	Validation	21	17	10	39
Decision Tree	Training	61	12	24	103
	Validation	17	14	14	42
Random Forest	Training	57	33	28	82
	Validation	21	17	10	39
Naive Bayes	Training	73	48	12	67
	Validation	28	30	3	26
SVM	Training	72	28	13	87
	Validation	24	23	7	33

Note: Data are presented as numbers of cases.

Abbreviations: TP, true positive; FP, false positive; FN, false negative; TN, true negative. The Logistic Regression model was selected for nomogram construction due to its stable performance across both cohorts.

Development and Validation of the US Radiomics Nomogram

A nomogram combining the clinical US and radiomics signatures was constructed. The combined model achieved AUCs of 0.798 (95% CI: 0.738–0.858) in the training cohort and 0.756 (95% CI: 0.654–0.858) in the validation cohort. The corresponding AUCs were 0.764 (95% CI: 0.700–0.829) and 0.744 (95% CI: 0.640–0.847) for the clinical US model, and 0.693 (95% CI: 0.621–0.765) and 0.611 (95% CI: 0.492–0.730) for the radiomics-only model. DeLong tests showed significant differences between the combined and radiomics-only models, but not between the combined and clinical US models (Table 3). Calibration and DCA results are shown in Figures 5 and 6.

Table 3.

Pairwise Comparisons of AUCs Using DeLong Tests

Cohort	Comparison	ΔAUC	95% CI	Z	P
Training	Combined vs Clinical	+0.034	–0.003 to 0.070	1.819	0.069
	Combined vs Rad	+0.105	0.042–0.168	3.272	0.001
	Clinical vs Rad	+0.072	–0.017 to 0.160	1.584	0.113
Validation	Combined vs Clinical	+0.012	–0.032 to 0.056	0.537	0.592
	Combined vs Rad	+0.145	0.034–0.257	2.552	0.011
	Clinical vs Rad	+0.133	–0.007 to 0.273	1.862	0.063

Note: DeLong tests were performed for paired ROC curves. Bold P values indicate statistical significance (P < 0.05).

Abbreviations: AUC, area under the curve; ΔAUC, difference in AUC; CI, confidence interval.

Figure 5.

Development and validation of the radiomics nomogram. (A) The nomogram integrates the radscore and significant ultrasound features (heterogeneity, necrosis, and posterior enhancement) to predict the probability of CTL. (B) ROC curves of the radiomics model (Rad), clinical US model (Clinic), and combined model (Unite) in the validation cohort. (C) ROC curves of the three models in the training cohort. The combined nomogram (Unite) demonstrated the highest AUC in both cohorts.

Figure 6.

Decision curve analysis (DCA) for the radiomics nomogram, clinical model, and radiomics-only model. The DCA shows the net benefit of each model across a range of threshold probabilities. (A) Training cohort: overall DCA of the three models. (B) Validation cohort: overall DCA of the three models. (C) Calibration curve in the training cohort. (D) Calibration curve in the validation cohort. The combined nomogram (red line) showed higher net benefit than the all-or-none strategies (treat all or treat none) and, over part of the threshold range, than the clinical and radiomics-only models, as shown in the figure.

Discussion

The TB has continued to rise over the past three decades. In 2019 alone, approximately 10 million individuals were infected with TB worldwide, resulting in 1.4 million deaths.²¹ Therefore, establishing a relatively accurate diagnosis of lymph node tuberculosis (LNTB) is of critical importance.²²

Cervical tuberculous lymphadenitis remains a considerable diagnostic challenge in clinical practice, particularly in its early stages or when presenting with atypical features.^23,24 While conventional ultrasound serves as a first-line imaging modality, its diagnostic accuracy is often hampered by subjective interpretation and the absence of pathognomonic signs in incipient disease.²⁵ In this retrospective study, we developed a radiomics nomogram combining the radiomics signature and US features. The combined model performed better than the radiomics-only model, but its AUC was not significantly higher than that of the clinical US model. The nomogram may serve as a pre-screening tool to improve the diagnostic rate of CTL and reduce unnecessary biopsies.

The concept of radiomics was first introduced in 2012 by Dutch scholar Lambin and others.¹⁶ Nowadays, radiomics models built based on ultrasound can also be used to diagnose many diseases.^26,27 For LNTB, it induces specific histopathological changes—such as caseous necrosis, granulomatous inflammation, and fibrosis—that generate distinct echotexture patterns. Although these microstructural alterations may not present as overt signs like necrosis or calcification until advanced stages, radiomics features—particularly texture descriptors from GLCM and wavelet transforms—can capture these changes quantitatively. In this way, radiomics converts subjective interpretation into objective, reproducible metrics, thereby complementing conventional US and improving diagnostic accuracy, especially in cases without typical sonographic features. In this study, five models were evaluated for CTL diagnosis, and Logistic Regression was selected for constructing the final nomogram due to its interpretability.^28,29 Unlike “black-box” models, LR provides explicit coefficients for each feature, allowing clinicians to understand the contribution of each factor to the final prediction.^30,31 Although Heterogeneous was not significant on univariate analysis (P = 0.246), it was retained in the final nomogram because variable selection was guided by LASSO regression, which prioritizes overall model performance in the multivariable setting over univariate significance.

The suboptimal performance of the standalone clinical US model (AUC: 0.764 95% CI: 0.700–0.829 training, 0.744 95% CI: 0.640–0.847 validation) in our study aligns with the well-documented limitations of conventional ultrasound.³² The current study showed that certain US features are indeed associated with the diagnosis of LNTB, with significant features including calcification, necrosis, posterior enhancement, matting, and surrounding tissue changes.^33,34 However, in our multivariate regression analyses, only necrosis and posterior enhancement were associated with tubercular lymphadenitis, which differs from the results of previous conventional ultrasound diagnosis in detecting LNTB. The reason may be that LNTB in different patients might be at different pathological stages.¹ This also highlights the limitations of ultrasound diagnosis in detecting LNTB.

Although the combined nomogram showed numerically higher AUCs than the clinical model in both cohorts, DeLong tests indicated that this improvement did not reach statistical significance. We attribute this to a “ceiling effect”: our clinical model already incorporated necrosis and posterior enhancement—two of the strongest sonographic predictors for CTL—leaving limited statistical room for further AUC elevation. Importantly, statistical nonsignificance in AUC comparison does not equate to lack of clinical utility. Our Decision Curve Analysis (Figure 6) demonstrated that the combined nomogram provided higher net benefit over part of the clinically relevant threshold range, suggesting that even a modest AUC increment can positively influence biopsy decision-making. Thus, we position the nomogram as a complementary tool, particularly valuable when typical ultrasound features are equivocal or absent.

In this study, we built a US-based radiomics model to differentiate CTL from non-CTL. The diagnostic model LR had an area under the receiver operating characteristics curve (AUC) of 0.693 (95% CI: 0.621–0.765) in the training set, outperforming the model reported in a previous study.³⁵ Subsequently, we built a nomogram incorporating the rad signature and US features in the diagnosis of CTL. This model demonstrated favorable predictive performance, yielding AUCs of 0.798 (95% CI: 0.738–0.858) and 0.756 (95% CI: 0.654–0.858) in the training and validation sets, respectively, and outperforming the single radiomics model or the ultrasound features in terms of predictive performance. Radiomics models based on B-mode and color Doppler US images have previously been developed to predict unexplained cervical lymphadenopathy,³⁶ but the sample sizes of LNTB in these studies were small (n=83). Consequently, models trained on lymph node tuberculosis data may be particularly susceptible to overfitting. In contrast, our study demonstrates that a radiomics model based solely on B-mode US combined with US features can achieve comparable diagnostic efficacy. While previous studies³⁷ have investigated the predictive value of grayscale ultrasound (G-US) and strain elastography ultrasound (SE-US) radiomic features for cervical tuberculous lymphadenitis (CTL), our study offers two distinct advantages. Firstly, the acquisition of elastography images requires additional specialized training, whereas the methodology employed in our study demonstrates greater generalizability. Secondly, by systematically comparing multiple models, we constructed a combined diagnostic model based on the most robust and compelling candidate. In addition, to address potential overfitting, the Hosmer-Lemeshow goodness-of-fit test was employed for internal validation across all model groups. The results indicated adequate model calibration, as evidenced by non-significant P-values (P > 0.05).

Despite the promising findings, several limitations should be acknowledged. First, the retrospective design may have introduced selection bias, and the single-center nature with a relatively limited sample size restricts the generalizability of our findings. Second, although rigorous radiomic feature extraction and selection were performed, manual ROI delineation remains subject to inter-observer variability. Third, the use of uniform ultrasound equipment and imaging parameters limits the applicability of our model to other machines or scanning protocols; the inherent operator and equipment dependence of ultrasound further underscores the need for multi-center validation and standardized image acquisition protocols. Finally, while our nomogram is intended to aid in the pre-biopsy diagnostic workup, we did not perform a formal subgroup analysis specifically for pathologically early-stage (eg, non-necrotic or pre-granulomatous) CTL cases. Future prospective, multi-center studies with larger cohorts and dedicated early-stage subgroups are warranted to validate the clinical utility and stability of the model.

Conclusions

In conclusion, we developed and validated an ultrasound-based radiomics nomogram incorporating Rad-score and key sonographic features for diagnosing cervical tuberculous lymphadenitis. The nomogram achieved AUCs of 0.798 (95% CI: 0.738–0.858) and 0.756 (95% CI: 0.654–0.858) in the training and validation cohorts, respectively, and significantly outperformed the radiomics-only model (training: ΔAUC = 0.105, P = 0.001; validation: ΔAUC = 0.145, P = 0.011). Although the improvement over the clinical US model was modest and not statistically significant, the nomogram demonstrated superior net benefit in decision curve analysis, supporting its utility as a complementary, non-invasive tool in the pre-biopsy workup, particularly when typical sonographic features are equivocal. By facilitating risk stratification—prioritizing high-risk patients for biopsy and enabling conservative follow-up for low-risk individuals—this approach may reduce unnecessary invasive procedures. Prospective multi-center studies are warranted to confirm its generalizability.

Funding Statement

This project was supported by the Finance Department of Jilin Province, People’s Republic of China (No. JLSWSRCZX2020-0027).

Abbreviations

CTL, cervical tuberculous lymphadenitis; DCA, decision curve analysis; FNA, fine-needle aspiration; ICC, intraclass correlation coefficient; LASSO, Least Absolute Shrinkage and Selection Operator; LR, logistic regression; LNTB, lymph node tuberculosis; Radscore, radiomics score; ROC, receiver operating characteristic; ROI, region of interest; TB, tuberculosis; TST, tuberculin skin test; US, ultrasound.

Data Sharing Statement

The datasets generated and analyzed during the current study are available from the corresponding author, Associate Professor Lirong Zhao (zhaolir@jlu.edu.cn), on reasonable request.

Ethics Approval and Informed Consent

This single-center retrospective study was conducted in compliance with the principles of the Declaration of Helsinki. The study protocol was reviewed and approved by the Medical Ethics Committee of The First Hospital of Jilin University (approval number: 24k024-001). Due to the retrospective nature of the study, which involved only the analysis of existing imaging data and medical records without any additional intervention, the requirement for informed consent was waived by the ethics committee. All patient data were anonymized and de-identified prior to analysis to ensure privacy and confidentiality.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declare that they have no competing interests.