To predict cervical lymph node tuberculosis based on clinical-multimodal ultrasound radiomics model

Abstract

Objectives:

In practical clinical work, when sonographers extract image features, there may be large intra-observer and inter-observer variability in subjective description and visual evaluation. Lymph node tuberculosis is often confused with other diseases of lymphadenopathy. To avoid the shortcomings of ultrasound such as strong subjectivity and low repeatability, we discussed the clinical value of imaging models based on B-mode ultrasound (B-US), elastic ultrasound (EUS) and contrast-enhanced ultrasound (CEUS) images in predicting cervical lymph node tuberculosis (CLNT).

Methods:

Herein, 215 patients with cervical lymph node enlargement confirmed via international diagnostic criteria at our hospital between January 2018 and May 2023 were included. Patients were randomly divided into training (n = 151) and validation (n = 64) sets in a 7:3 ratio. Thereafter, 42 patients with cervical lymphadenopathy who underwent ultrasound-guided lymph node puncture from March 2023 to September 2023 were considered as a prospective internal validation set. Three models (radiomics model, clinical model and clinical-radiomics model) were established. Receiver operating characteristic curves (ROCs) of different models were drawn, and the area under the curve (AUC),were compared among them. Finally, the visual color band nomogram was established.

Results:

The AUC of the clinical-radiomics model in the training dataset, validation dataset and prospective validation dataset reached 0.959, 0.906 and 0.865, respectively. The clinical-radiomics model has good diagnostic efficacy in predicting CLNT.

Conclusions:

The Multimodal ultrasound radiomics combined with clinical manifestations and imaging features, showed good judgment in identifying CLNT ability and good stability.

Keywords

cervical lymph node tuberculosis B-mode ultrasound contrast-enhanced ultrasound radiomics nomogram

Background

Tuberculosis (TB) is a chronic infectious disease caused by the bacterium Mycobacterium tuberculosis, which has existed for millennia and remains a major global health concern.¹ Approximately 10 million people develop TB each year, and this disease is one of the major causes of death worldwide.² TB can infect every organ and system in the body. If tuberculosis develops outside the lung tissue, it is called extrapulmonary tuberculosis (EPTB). The treatment of EPTB takes a long time and the cure rate is also unsatisfactory compared with that of pulmonary tuberculosis.^3–5 In 2022, Kang et al.⁶ conducted a large-scale multicenter observational survey on EPTB, including 208214 patients with EPTB who were treated in 21 chest hospitals in 15 provinces of China from January 2011 to December 2017, and found that the number of CLNT cases was as high as 15282. It ranked first in EPTB. It can be seen that CLNT is an important part of the tuberculosis epidemic, which brings challenges to clinicians in the diagnosis and treatment of EPTB. It is critical to strengthen clinical intervention and ensure timely and effective diagnosis and treatment of TB patients to control the progression of the disease.

Ultrasonography which has the advantages of non-exposure to radiation and being real-time and dynamic is the preferred diagnostic imaging modality for CLNT. It can provide information relevant to the diagnosis of lymph node tuberculosis by evaluating B-mode ultrasound (B-US) signs of size, morphology, hilum, internal echo, and surrounding tissues. Also, ultrasonography can detect microcirculatory disorders or microvascularization in tumors.^7,8 However, the ultrasonic characteristics of lymph node tuberculosis overlapped with those of other lymph node diseases, which could not be effectively distinguished by gross observation only.^9–11 Simultaneously, the limitations of ultrasonography in the evaluation of lymph node diseases are inevitable since the technique relies heavily on operational experience and subjective judgment.

Medical imaging is a large-scale extraction of imaging features that can hardly be identified by the naked eye, and quantitative analyses are carried out with the aid of machines to bring out the hidden high-dimensional spatial information in the images and correlate it with clinical diseases to facilitate the early diagnosis of these diseases.^12–14 Present studies have shown that ultrasonography can predict cervical lymph node metastases of thyroid cancer and breast cancer,^15,16 and most studies are based on a single ultrasonic mode, and CLNT is rarely studied.

The purpose of this study was to conduct multimodal imaging analysis based on static B-US, EUS and CEUS images, explore the prediction efficacy of the combined clinical and radiomics model for CLNT by combining ultrasonic characteristics, and construct a visual nomogram model to visualize the results. Clinical decision curves (DCA) were drawn to evaluate its clinical application value. By comparing the diagnostic value of different prediction models, to explore whether the combined clinical and imaging model can improve the evaluation efficiency of CLNT, better realize the early accurate prediction of CLNT, and provide more valuable reference for early individualized treatment.

Methods

Patient

A total of 784 cases with suspicious ultrasonographic signs of cervical lymph nodes and the results of ultrasound-guided lymph node biopsy were collected in this study. Herein, patients with cervical lymph node enlargement via international diagnostic criteria standard^17,18 came to our hospital from January 2018 to May 2023. Then, 42 patients with cervical lymphadenopathy who underwent ultrasound-guided lymph node puncture from March 2023 to September 2023 were considered as a prospective internal validation set. The patient's enrollment is shown in Figure 1. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.

Figure 1.

The procedure for patient screening and enrollment is shown in the figure.

Because part of the data in this study were retrospective, informed patient consent was not required. Considering age, gender, diet, sleep habits, body weight, medication and other factors between the prospectively collected cases, the subjects were informed of the study process, asked in detail about their medical history. Thereafter, written informed consent was obtained before the examination. This study was approved by the Medical Ethics Committee of Hangzhou Red Cross Hospital (number of Ethic Committee: [2022] Review No. (44)). It was in accordance with the ethical principles of the Declaration of Helsinki.

Our inclusion criteria were as follows: (1) Patients who underwent B-US, EUS and CEUS examinations for cervical lymph node enlargement; (2) Ultrasound-guided puncture biopsy was performed to obtain specimens for histopathology;(3) Complete clinical data and original DICOM imaging data can be retrieved from the hospital's PACS system.

And our exclusion criteria were as follows: (1) Patients with HIV; (2) the lesion was not confirmed by international diagnostic criteria; and (3) the quality of the B-US images and CEUS video was poor.

Multi-mode ultrasonic image acquisition, segmentation and feature extraction

GUS and EUS image acquisition: the patient fully exposed the lesion site, the target lymph nodes were examined by B-US and EUS. The frozen images of the section with clear image, the largest lesion and no large blood vessels were selected, and the images were stored. The procedure was performed by doctors with 10 years or more experience in the diagnosis of superficial organ ultrasound.

Key frame capture of CEUS: the patient was placed in the supine position, and the lesion site was fully exposed, adjusted to the clearest scanning plane of the lesion, and the focus and gain of ultrasound development were set to an ideal state. The patient was allowed to breathe steadily, and a 2.4-ml intravenous bolus of SonoVue™ developer (Bracco, Italy) was administered through the anesthetized elbow, followed rapidly by a 5-ml bolus of saline solution. A timer was started at the beginning of the injection of contrast agent, and the development of lymph nodes, the development pattern and the phenomenon of insufficient blood supply were closely monitored and recorded. At the same time, about 60 s of dynamic imaging video data were saved.

Multi-mode ultrasonic images segmentation and feature extraction

Then，the region of interest (ROI) was delineated by two sonographers (M.X. and W.Y. with 9 and 6 years of clinical experience respectively) with extensive experience in diagnosing lymph node diseases using the open-source software (3D Slicer. v5.2.2) to manually outline ROI along lymph node boundaries blindly, and the information features of each picture were extracted by the software.

Thirty cases were randomly selected from the ROIs outlined by the two physicians for intraclass correlation coefficient (ICC) analyses to evaluate the repeatability and stability of the segmentation images. ICC value of >0.75 is considered to have high interobserver agreement, biological range, and good reproducibility.^19,20

Establishment of multi-modality ultrasound model

The enrolled participants were divided into a training set and a verification set in a 7: 3 ratio by R software. After z-score normalization, LASSO (least absolute shrinkage and selection operator) algorithm, and five-fold cross-validation to further select the best representative features. Logistic regression was used to construct the prediction model based on the selected key features, and the selected radiomics eigenvalues and their corresponding non-zero regression coefficients were used to construct the Rad-score for each model. The final B-US + EUS + CEUS combined radiomics model will be obtained. The operation flow of radiomics is shown in Figure 2.

Figure 2.

Working flow chart of the radiomics prediction model for CLNT.

Analysis of conventional ultrasonic images and the establishment of clinical model

The enrolled lymph node patients were divided into a tuberculosis group and a non-tuberculosis group according to the international diagnostic criteria, and the differences in the features of conventional ultrasound images in the training set and validation set were analyzed. Two sonographers (M.X. and Z.Y. with 9 and 16 years of clinical experience respectively), empirically analyzed and recorded 14 important features of lymph nodes according to the stored ultrasound images and clinical baseline information, including (a) Long diameter; (b) Short diameter (c) Ratio of long to short diameter; (d) Boundary; (e) Internal echo; (f) Lymphatic portal condition; (g) Annular low echo; (h) Adjacent lymph node fusion; (i) Surrounding soft tissue; (j) gender; (k) age; (l) history of tuberculosis; (m) fatigue status; (n) weight change.

Then the preliminary univariate analysis was carried out, followed by the comprehensive multivariate analysis. The general clinical data of the patients were sorted out, and the statistically significant data (p < 0.05) were analyzed in depth to construct a prediction model for clinical use. AUC and other relevant statistics were calculated to assess diagnostic power.

Clinical-radiomics model establishment

Finally, the GUS + EUS + CEUS radiomics model (Rad-score) was combined with clinical imaging features with p < 0.05, and a multi-modality fusion model combining clinical imaging information and radiomics features was constructed by logistic regression algorithm.

Statistical analysis

All statistical analyses of this study design were processed by R-studio software, SPSS and MedCalc statistical software. Clinical variable characteristic count data were expressed as “ %”. We downloaded and used “glmnet”, “pROC”, “rms” and “rmda” software packages in R language to screen radiomics features, reduce dimensionality, construct and validate each model. The measurement data were tested for homogeneity of variance and normality, represented by mean ± standard deviation or median, and the differences between groups were compared by independent sample t test or Mann Whitney U nonparametric test. And, the Hosmer-Lemeshow test was used to evaluate the fit of each model. The Delong test of Medcalc software was used to evaluate and compare the AUC, accuracy, sensitivity, specificity, NPV and PPV among the models. P < 0.05 was considered statistically significant.

Results

Patient clinical information

A total of 257 patients with cervical lymphadenopathy of different causes were included in this study, which were divided into training set (n = 151), validation set (n = 64), and prospective validation set (n = 42). There were 120 male and 137 female patients, with an average age of 48 ± 40 years. Among them, 135 cases were diagnosed as CLNT by the comprehensive diagnostic reference standard, and 122 cases were non-CLNT(31 cases were pathologically confirmed as lymphoma, 36 cases were pathologically confirmed as reactive hyperplasia, and 55 cases were pathologically confirmed as metastatic lymph nodes).

Construction of the clinical prediction model

Table 1 shows the ultrasound features and clinical baseline information of the images were analyzed by univariate and multivariate analysis. After univariate analysis, six significant features were identified, and the study finally concluded that gender, age, history of tuberculosis diagnosis, annular hypoechoic halo, and adjacent soft tissue echo enhancement showed significant statistical differences (p < 0.05). The five features were used as independent factors to predict CLNT to construct a clinical prediction model. The AUC of the model in the training set was 0.889 (95%CI: 0.836–0.943), the sensitivity and specificity were 79.8% and 89.6%, respectively. The AUC of the model in the test set was 0.905 (95%CI: 0.834–0.977), the sensitivity and specificity were 74.3% and 96.6%, respectively. In the prospective internal validation set, the AUC was 0.668 (95%CI: 0.505–0.832), and the sensitivity and specificity were 100% and 46.2%, respectively (see Table 2).

Table 1.

Clinical ultrasound imaging characteristics of single factor and multiple factors logistic regression analysis.

	Single factor analysis			Multivariate regression analysis
Clinical imaging parameters	B	P	OR (95%CI)	B	P	OR (95%CI)
History of pulmonary tuberculosis	−2.164	<0.001*	0.115 (0.052–0.252)	2.232	<0.001*	9.322 (3.64–23.877)
Annular hypoechoic	−1.903	<0.001*	0.149 (0.064–0.350)	1.834	0.001*	6.261 (2.044–19.181)
Age	−0.044	<0.001*	0.957 (0.938–0.976)	−0.034	0.005*	0.967 (0.944–0.99)
Surrounding soft tissue of the echo	−1.825	<0.001*	0.161 (0.058–0.446)	1.497	0.038*	4.467 (1.088–18.328)
Sex	−1.149	<0.001*	0.317 (0.152–0.619)	1.194	0.009*	3.3 (1.352–8.057)
Hilum	−0.966	0.022	0.381 (0.166–0.870)	0.699	0.257	2.012 (0.6–6.74)
Short diameter (S)	−0.426	0.133	0.653 (0.375–1.138)
Fatigue	−0.486	0.141	0.615 (0.322–1.175)
Internal echo	−0.45	0.216	0.638 (0.315–1.292)
Boundary	−0.464	0.266	0.629 (0.277–1.425)
Adjacent lymph node fusion	−0.301	0.42	0.74 (0.356–1.538)
Long diameter (L)	−0.099	0.55	0.905 (0.654–1.254)
Ratio of long to short diameter (L/S)	0.078	0.766	1.081 (0.647–1.807)
weight change	0.075	0.834	1.077 (0.537–2.162)

OR:odds ratio.

Table 2.

Comparison of the diagnostic performance of the three prediction models in the train set, validation set and in the prospective internal test set.

Indicators	Train set			Test set			Prospective internal test set
Indicators	Clinical models	Radiomics model	Multimodal fusion model	Clinical models	Radiomics model	Multimodal fusion model	Clinical models	Radiomics model	Multimodal fusion model
AUC (95%CI)	0.889 (0.836–0.943)	0.917 (0.836–0.943)	0.959 (0.932–0.986)	0.905 (0.834–0.977)	0.812 (0.704–0.92)	0.906 (0.835–0.977)	0.668 (0.505–0.832)	0.887 (0.771–1)	0.865 (0.746–0.984)
ACC	84.1%	90%	90.1%	84.4%	76.6%	84.4%	66.7%	83.3%	83.3%
SEN	79.8%	81%	90.5%	74.3%	77.1%	85.7%	100%	93.8%	87.5%
SPEC	89.6%	81%	89.6%	96.6%	75.9%	82.8%	46.2%	76.9%	80.8%
PPV	90.5%	91.9%	91.6%	96.3%	79.4%	85.7%	53.3	71.4%	71.4%
NPV	77.9%	79.2%	88.2%	75.7%	73.3%	82.8%	100%	65.2%	91.3%

AUC: area under the curve; ACC: accuracy; SEN: sensitivity; SPEC: specificity; NPV: negative predictive value; PPV: positive predictive value; CI: confidence interval.

Establishment of the B-US + EUS + CEUS radiomics model

ICC calculated according to the features extracted by the segmentation images ROI of two senior sonographers is 0.928–0.982, and the extracted features have good repeatability.

3D Slicer was used for feature extraction of ultrasound images, a total of 837 features were extracted from each modality image. There were 14 gray-level dependence matrices (GLDM), 16 gray-level area-size zone matrices (GLSZM), 16 gray-level run-length matrices (GLRLM), 24 gray-level co-occurrence matrices (GLCM), 5 neighborhood gray-tone difference matrices and 744 the small Porter signs (wavelet). According to the dimensionality reduction by the LASSO algorithm, to determine the best number of parameters and characteristics, five-fold cross-validation was applied to the training set. Finally, two stable features of US images were selected, the US radiomics model was constructed by logistic regression, and the radiomics score was calculated. The Rad-score formula is constructed from the final radiomics eigenvalues and their corresponding regression coefficients. Rad-score = −0.1071+−1.1913*wavelet-HLH_glcm_ClusterTendency_2D + −0.2388*wavelet-HLH_glcm_Imc1_2D + −0.8484*wavelet-LHH_gldm_DependenceVariance_Elast + 0.2379*wavelet-LHL_glcm_Imc1_CEUS + −0.4064*wavelet-LHL_glszm_SizeZoneNonUniformityNormalized_CEUS + 0.6764*wavelet-LHL_glszm_ZoneEntropy_CEUS + −1.0751*wavelet-HHH_gldm_DependenceNonUniformityNormalized_CEUS.

Comparison of the clinical model, radiomics model, and combined model

ROC curves of models are shown in Figure 3. In the training and two test sets, the sensitivity, specificity, accuracy, and other diagnostic performance indices of the clinical, radiomics, and hybrid models are compared in Table 2. The calibration curve shows that the prediction of CLNT in the three sets of the clinical-radiomics model is in good agreement with the real situation, as shown in Figure 4. To further evaluate the clinical utility of the hybrid model, DCA curves were used to quantify the net benefit at different threshold probabilities (Figure 5). The DCA results showed that in the threshold probability range of 0–0.9, all models could obtain more benefits than all or no intervention in the evaluation of CLNT, and within this threshold, the clinical-radiomics model could provide the largest net benefit for the evaluation of CLNT.

Figure 3.

ROC diagram of the training set (A), test set (B) and prospective internal test set (C) of three radiomics models.

Figure 4.

Calibration curves of the training set (A), test set (B) and prospective internal test set (C) of the models. Note: Ideal represents the ideal prediction curve, Apparent represents the actual prediction curve, and Bias-corrected represents the calibrated prediction curve.

Figure 5.

DCA curves of the three predictive models.

Analysis of the clinical decision curve

The independent risk factors with P < 0.05 in clinical imaging information and Rad score were included in the production of Nomogram, as shown in Figure 6. By adding the Points of each variable in the figure, the Total Points can be used to predict the risk probability of CLNT.

Figure 6.

The visualized color strip Nomogram is shown in the figure. Note: The final CLNT score is reflected by “Total Points”, and the predicted value is reflected by “Risk”.

Discussion

Tuberculosis (TB) is one of the world's oldest and deadliest infectious diseases. Despite all efforts to contain the disease, tuberculosis remains a prominent public health problem, especially in developing countries. Every year, the World Health Organization analyzes data from all over the world in the Global TB Report. However, due to lack of facilities in some areas and the impact of COVID-19 in previous years, many patients are still not counted and treated in a timely manner.^4,5 Clinicians often delay its diagnosis, which may lead to an increase in resistant strains or mortality.²¹ Thus, while risk factors for extrapulmonary TB have been investigated in previous studies, newer techniques have facilitated faster detection of EPTB, especially in complex cases. Among the many imaging methods, ultrasound is considered to be a convenient and effective method for the evaluation of cervical lymph nodes, which has excellent performance in the diagnosis of metastatic lymph nodes and other superficial lymph node lesions.^9,10,16

Gray-scale ultrasonography is often used as the preferred technique to evaluate lymph node abnormalities. It can observe the size, morphology, hilar structure, internal sonographic features, calcium deposition and adjacent tissues of lymph nodes, so as to provide a valuable basis for the diagnosis of CLNT.²² In clinical work, we often measure the maximum long diameter (L), maximum short diameter (S) and the ratio of length to short diameter (L/S) of lymph nodes to judge the pathological changes of lymph nodes. Lymph node tuberculosis can cause destruction of hilar structures without visualization. However, in our study, the maximum long diameter (L), maximum short diameter (S) and L/S ratio of lymph nodes were not statistically significant for the diagnosis of CLNT(p > 0.05), and due to the influence of anatomical relationship, the lymphatic hilum of deep benign lymph nodes may not be clearly displayed. In this study, a total of gray-scale ultrasound features were screened, both of which were GLCM, which is an important texture analysis feature and represents the respective distance between the intensity distribution feature and the intensity level in the original image.²³

Meanwhile, CEUS can more accurately reflect and observe the microcirculation perfusion of tissues.^24,25 Cui studied the blood flow injection mode and time-signal intensity curve (TIC) of CEUS, and observed that tuberculous lymph nodes showed obvious high enhancement in the arterial phase, and the descending segment of the TIC curve was extremely steep, accompanied by a significant depression, which was different from metastatic lymph nodes, and the differences were statistically significant (p<0.05).²⁶ However, the effective diagnosis of this detection method still depends on the doctor's diagnostic experience and dynamic monitoring. Some studies have shown that real-time high-definition images obtained by contrast-enhanced ultrasound (CEUS) video can provide supplementary information, which is comparable to images obtained by CT or MRI.²⁷ In this study, a total of 7 radiomics signature features were extracted, of which 4 were CEUS features, indicating that CEUS can provide more information than other modalities.

Rajendra et al.²⁸ conducted a prospective study on the differentiation of benign and malignant lymph node diseases in 50 patients, and compared the differences in sensitivity, specificity and diagnostic accuracy between ultrasound elastography and conventional gray-scale ultrasound. The results showed that ultrasound elastography was superior to traditional gray-scale ultrasound in diagnostic performance. However, he noted that chronic granulomatous lymphadenopathy, such as tuberculosis, has both inflammation and fibrosis, which can result in variable stiffness, making elastography less specific. In this study, an elastography feature was extracted and it was statistically significant (P = 0.032), indicating that elastography plays a certain role in the differential diagnosis of lymph node tuberculosis.

Chu et al.⁹ also explored the value of multimodal ultrasound images including gray-scale ultrasound, color Doppler ultrasound, elastic ultrasound and CEUS in the identification of lymph node tuberculosis and other non-lymph node tuberculosis lymphadenitis. The study showed that the imaging features of lymph node tuberculosis included ring hypoechoic lymph nodes, enhanced echo of surrounding soft tissues, and enhanced echo of lymph nodes. This is similar to the results of the present study. In this study, we conducted univariate and multivariate statistical analysis of ultrasound gray-scale features, and concluded that whether annular hypoechogenicity around lymph nodes and echo enhancement of adjacent soft tissues were statistically significant (P < 0.05), as shown in Table 1. However, based on these features alone, it is still difficult to differentiate lymph node tuberculosis from metastatic lymph nodes, necrotizing lymphadenitis, or lymphoma with compression of surrounding soft tissues.^6,9,11

In previous studies, “gender” and “age” were considered to be common clinical predictors for the diagnosis of EPTB. Qian et al.²⁹ analyzed the data of EPTB in Texas, USA, and found that the risk of EPTB in women was significantly increased. The reason for the higher incidence of EPTB in women may be related to the variation of endocrine hormones and immune-related factors.³⁰ Similarly, Pang et al. also found an association between EPTB and being female and being young.³¹ This finding is similar to our data (gender, age) in the training set, validation set, and prospective internal validation set, which also showed that younger women were more likely to be diagnosed with lymph node tuberculosis.

In this study, the clinical model was established by analyzing the basic clinical and ultrasound image information of the patients. Gender, age, history of tuberculosis, and ultrasound image features (ring hypoechoic halo and echo enhancement of adjacent soft tissue) were statistically significant, and the statistically significant features were used to establish the model. The AUCs of the clinical model in the training set, validation set and internal validation set were 0.889,0.905 and 0.668, respectively. It was seen that the clinical model had a certain predictive ability in the diagnosis of CLNT, but in the prospective validation set, the AUC of the clinical model was only 0.668. According to Delong's test results, the AUC value of the clinical model in the validation set was 0.905, which was higher than that of the radiomics model (0.812). However, the difference of AUC values between the two methods did not reach the degree of statistical difference either in the training set or the test set (P > 0.05). Calibration curve analysis also reflected the instability of the clinical model.

As an emerging area of multidisciplinary research, radiomics can extract high-throughput image features and perform quantitative analysis on big data in order to compensate for the subjective defects of observers.^23,32 Stable radiomics features significantly related to CLNT, such as order features, GLCM, GLSZM, GLRLM, and the wavelet feature, are the key to diagnosis.³³ The radiomics features finally screened in this study are based on the wavelet feature, which indicates the important role of the wavelet feature in ultrasonography diagnostic. Recently, radiomics has been increasingly used in the diagnosis and differential diagnosis of lymph node diseases. In the study of Chen et al.,³⁴ they extracted the dual-modality ultrasound image data set of gray-scale ultrasound and elastic ultrasound, and used radiomics analysis to diagnose benign lymph node lesions, lymphoma and metastatic lymph nodes. The AUC of benign and lymphoma, benign and metastatic, lymphoma and metastatic, and benign and malignant in the validation set were 0.960, 0.716, 0.933, and 0.856, respectively. In 2022, Lin et al.³⁵ used gray-scale ultrasound radiomics technology to perform a diagnostic study on cervical lymph node metastasis in patients with nasopharyngeal carcinoma. They established three models, the AUC of the clinical model was 0.838, the AUC of the radiomics model was 0.810, and the AUC of the clinical and radiomics combined model was 0.880. Radiomics has potential application value in screening meaningful ultrasound features and improving the diagnostic efficiency of ultrasound for Lymph node metastasis in patients with nasopharyngeal carcinoma. A large number of literatures have shown that radiomics can improve the accuracy of lymph node disease diagnosis, and a large number of studies have shown that multi-modality image fusion can promote the improvement of diagnostic accuracy.³⁶ However, at present, there are few studies on ultrasound radiomics methods combining gray-scale, elasticity, CEUS and other modalities to predict CLNT, and the clinical value is not clear. In order to adapt to the development trend of multimodal information integration, we integrated the clinical information and GUS, EUS and CEUS image information, constructed and compared three prediction models (including clinical model, radiomics model and clinical-radiomics model), and differentiated CLNT and other lymphadenitis based on clinical and radiomics features.

In this study, a total of 2511 features were extracted from the three modality ultrasound images in the training set, and 1368 features were obtained after univariate analysis. The least absolute shrinkage and selection operator (LASSO) regression algorithm was used to reduce the dimension of feature data, and 7 optimal features were finally obtained to construct the radiomics score (Rad score). Its AUC in the training set was 0.917 (95CI: 0.87–0.964). The model was then validated in the validation set and the prospective internal validation set, with AUC of 0.812 (95CI: 0.704–0.92) and 0.887 (95CI: 0.771-1), respectively. Liu et al.¹⁰ used radiomics analysis of gray-scale ultrasound to diagnose several common lymph node diseases in a multicenter study, and the AUC of CLNT diagnosis in the validation set was 0.626. Yang et al.³⁷ explored the predictive value of gray-scale ultrasound and elastic ultrasound radiomics features for CLNT, and the results showed that the AUC of single gray-scale and elastic ultrasound radiomics models in the validation set were 0.772 and 0.822, respectively. It can be seen that the single-modality radiomics has a certain potential in the diagnosis and prediction of CLNT. In this study, the AUC of combined GUS, EUS and CEUS radiomics in the diagnosis of CLNT was 0.887 in the prospective validation set, which was higher than the diagnostic efficacy of single-modality ultrasound for CLNT in previous literature.

In this study, the AUC of combined GUS, EUS and CEUS radiomics in the diagnosis of CLNT was 0.887 in the prospective validation set, which was higher than the diagnostic efficacy of single-modality ultrasound for CLNT in previous literature. In addition, it can be seen from Table 2 and Figure 3 that the clinical-radiomics model combining clinical imaging features and radiomics features helps to improve the AUC value of CLNT and non-CLNT diagnostic tasks, and its AUC in the validation set and prospective validation set were 0.906 and 0.865, respectively. It is proved that the multimodal fusion model has good clinical application value and better stability in the diagnosis of CLNT. Moreover, from the DCA curves in Figure 5, we can see that compared with the radiomics and clinical models, the clinical-radiomics model has better clinical applicability in the three datasets.

However, Our study still has some limitations. First, this study included a certain amount of prospectively collected data. However, most of the data were collected retrospectively, and the limited data size may reduce the broad applicability of the model in different scenarios. Future prospective analyses with larger sample sizes from different medical centers are needed to demonstrate the validity of the study. Secondly, all case data in this study were obtained from specialized TB hospitals, and there may be inevitable selection bias in the selection of cases, because patients who have been diagnosed with TB prefer to visit specialized TB hospitals. In addition, since manual delineation of ROI is subjective and requires a significant amount of time, repeatability and objectivity still need to be improved. In the future, more clinical information (e.g., lymphocytes, tumor biomarkers, etc.) can be collected to evaluate the clinical validity of this prediction model and facilitate precision medicine.

Conclusions

The multimodal ultrasound radiomics combined with clinical manifestations and imaging features, showed good judgment in identifying CLNT ability and good stability, can provide noninvasive and highly efficient for medical decision making process.

Footnotes

Acknowledgments

We thank Hangzhou Red Cross Hospital for providing the cases.

Ethical statement

The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013).

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by Funding Zhejiang Provincial Natural Science Foundation of China [grant number LTGY23H180005]; Hangzhou Biomedical and Health Industry Development Support Science and Technology Project (the Sixth Phase) [grant number 2022WJC046]; Major Project of Hangzhou Health Commission (Z20230098); Zhejiang Medical and health technology project [grant number 2024KY1231].

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

World Health Organization. Global tuberculosis report 2022. Geneva: World Health Organization, 2022.

Xie

Cao

, et al. Bactericidal effects of high intensity focused ultrasound on Bacillus Calmette-Guerin in vivo and in vitro. Int J Hyperthermia 2019; 36: 885–895.

Ben Ayed

Koubaa

Marrakchi

, et al. Extrapulmonary tuberculosis: update on the epidemiology, risk factors and prevention strategies. Int J Trop Dis 2018; 1: 006.

World Health Organization. Global tuberculosis report 2023. Geneva: World Health Organization, 2023.

Brehm

Terhalle

. Extrapulmonale Tuberkulose [extrapulmonary tuberculosis]. Dtsch Med Wochenschr 2023; 148: 1242–1249.

Kang

Liu

, et al. Epidemiology of concurrent extrapulmonary tuberculosis in inpatients with extrapulmonary tuberculosis lesions in China: a large-scale observational multi-centre investigation. Int J Infect Dis 2022; 115: 79–85.

Zhang

Chen

, et al. The value of contrast-enhanced ultrasound combined with microflow imaging in predicting microvascular invasion of hepatocellular carcinoma before operation. Clin Hemorheol Microcirc 2024; 88: 33–41.

Wang

, et al. Comparison of microvascular flow imaging and contrast-enhanced ultrasound for blood flow analysis of cervical lymph node lesions. Clin Hemorheol Microcirc 2023; 85: 249–259.

Chu

Zhang

, et al. The value of multimodal ultrasonography in differential diagnosis of tuberculous and non-tuberculous superficial lymphadenitis. BMC Surg 2021; 21: 416.

10.

Liu

Chen

Zhang

, et al. Ultrasound-based radiomics can classify the etiology of cervical lymphadenopathy: a multi-center retrospective study. Front Oncol 2022; 12: 856605.

11.

Wang

Chen

, et al. Case report: primary mediastinal large B-cell lymphoma invasion of extranodal thyroid tissue mimicking tuberculosis and confounded by similar ultrasonic appearance. Front Oncol 2022; 12: 879295.

12.

Lambin

Leijenaar

RTH

Deist

, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017; 14: 749–762.

13.

van Timmeren

Cester

Tanadini-Lang

, et al. Radiomics in medical imaging-“how-to” guide and critical reflection. Insights Imaging 2020; 11: 91.

14.

Mayerhoefer

Materka

Langs

, et al. Introduction to radiomics. J Nucl Med 2020; 61: 488–495.

15.

Tagliafico

Piana

Schenone

, et al. Overview of radiomics in breast cancer diagnosis and prognostication. Breast 2020; 49: 74–80.

16.

Liu

, et al. Comparison of the application of B-mode and strain elastography ultrasound in the estimation of lymph node metastasis of papillary thyroid carcinoma based on a radiomics approach. Int J Comput Assist Radiol Surg 2018; 13: 1617–1627.

17.

WHO. WHO consolidated guidelines on tuberculosis: module 3: diagnosis–rapid diagnostics for tuberculosis detection. Geneva: World Health Organization, 2021.

18.

National Collaborating Centre for Chronic Conditions. Tuberculosis: clinical diagnosis and management of tuberculosis, and measures for its prevention and control. London: National Institute for Health and Clinical Excellence (UK), 2011, pp. 13–16.

19.

Baeßler

Weiss

Pinto dos Santos

. Robustness and reproducibility of radiomics in magnetic resonance imaging: a phantom study. Invest Radiol 2019; 54: 221–228.

20.

Adelsmayr

Janisch

, et al. CT Texture analysis reliability in pulmonary lesions: the influence of 3D vs. 2D lesion segmentation and volume definition by a Hounsfield-unit threshold. Eur Radiol 2023 May; 33: 3064–3071.

21.

Berg

Schelling

Hailu

, et al. Investigation of the high rates of extrapulmonary tuberculosis in Ethiopia reveals no single driving factor and minimal evidence for zoonotic transmission of Mycobacterium bovis infection. BMC Infect Dis 2015; 15: 112.

22.

Gupta

Rahman

Shahid

, et al. Sonographic assessment of cervical lymphadenopathy: role of high-resolution and color Doppler imaging. Head Neck 2011; 33: 297–302.

23.

Colling

Pitman

Oien

, et al. Artificial intelligence in digital pathology: a roadmap to routine use in clinical practice. J Pathol 2019; 249: 143–150.

24.

Zhang

Zheng

Zhang

, et al. Quantitativeanalysis of contrast-enhanced ultrasound and superb microvascular imaging for the evaluation of disease activity in inflammatory bowel disease. Clin Hemorheol Microcirc 2024; 87: 451–464.

25.

Jung

Stroszczynski

Jung

. Advanced multimodal imaging of solid thyroid lesions with artificial intelligence-optimized B-mode, elastography, and contrast-enhanced ultrasonography parametric and with perfusion imaging: initial results. Clin Hemorheol Microcirc 2023; 84: 227–236.

26.

Cui

Yin

Fan

, et al. Diagnostic value of contrast-enhanced ultrasonography and time-intensity curve in differential diagnosis of cervical metastatic and tuberculous lymph nodes. J Ultrasound Med 2018; 37: 83–92.

27.

Chung

Kim

. Contrast-enhanced ultrasonography: advance and current status in abdominal imaging. Ultrasonography 2015; 34: 3–18.

28.

Rajendra

Varatharajaperumal

Renganathan

, et al. Estimation of accuracy of B-mode sonography and elastography in differentiation of benign and malignant lymph nodes with cytology as reference standard: a prospective study. Cureus 2021; 13: e14147.

29.

Qian

Nguyen

Lyu

, et al. Risk factors for extrapulmonary dissemination of tuberculosis and associated mortality during treatment for extrapulmonary tuberculosis. Emerg Microbes Infect 2018; 7: 1–14.

30.

Al-Ghafli

Varghese

Enani

, et al. Demographic risk factors for extra-pulmonary tuberculosis among adolescents and adults in Saudi Arabia. PLOS ONE 2019; 14: e0213846.

31.

Pang

Shu

, et al. Epidemiology of extrapulmonary tuberculosis among inpatients, China, 2008–2017. Emerg Infect Dis 2019; 25: 457–464.

32.

Avanzo

Stancanello

, et al. Beyond imaging: the promise of radiomics. Phys Med 2017; 38: 122–139.

33.

Zwanenburg

Vallières

Abdalah

, et al. The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology 2020; 295: 328–338.

34.

Chen

Jiang

Shi

, et al. Dual-mode ultrasound radiomics and intrinsic imaging phenotypes for diagnosis of lymph node lesions. Ann Transl Med 2020; 8: 742.

35.

Lin

Tang

Cao

, et al. Using ultrasound radiomics analysis to diagnose cervical lymph node metastasis in patients with nasopharyngeal carcinoma. Eur Radiol 2023; 33: 774–783.

36.

Boehm

Khosravi

Vanguri

, et al. Harnessing multimodal data integration to advance precision oncology. Nat Rev Cancer 2022; 22: 114–126.

37.

Yang

Zhang

, et al. Exploratory study on the predictive value of ultrasound radiomics for cervical tuberculous lymphadenitis. Clin Imaging 2022; 86: 61–66.