Preoperative diagnosis of lateral cervical lymph node metastasis in papillary thyroid carcinoma: a multimodal ultrasound model integrating super-resolution ultrasound

Abstract

Background

Accurate preoperative diagnosis of lateral lymph node metastasis (LLNM) in papillary thyroid carcinoma (PTC) is critical but remains challenging.

Objective

To evaluate the diagnostic value of a multimodal model integrating super-resolution ultrasound (SRUS) for the preoperative diagnosis of suspicious lateral cervical lymph nodes (LNs) in PTC.

Methods

This prospective study enrolled consecutive participants with PTC and suspicious lateral LNs from January to November 2025. All LNs underwent US, color Doppler flow imaging (CDFI), contrast-enhanced ultrasound (CEUS), SRUS, fine needle aspiration cytology (FNAC) and fine needle aspiration thyroglobulin (FNA-Tg), with surgical pathology as the reference standard. Three generalized estimating equations models were constructed: US+CDFI, US+CEUS, and US+CEUS+SRUS. Receiver operating characteristic curves were plotted for these models, FNAC and FNA-Tg; and sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were calculated. The nomogram based on the optimal model was constructed and assessed.

Results

A total of 149 lateral LNs (74 metastatic, 75 non-metastatic) from 112 participants (median age, 46 years [IQR, 35–54 years]; 72 female) were analyzed. The multimodal US+CEUS+SRUS model achieved an area under the curve (AUC) of 0.961(95% CI: 0.933, 0.988), superior to both the US+CDFI (AUC, 0.851; 95% CI: 0.791, 0.911; P < 0.001) and US+CEUS (AUC, 0.921; 95% CI: 0.878, 0.964; P = 0.040) models. It outperformed FNAC (AUC, 0.858; 95% CI: 0.806, 0.910; P < 0.001) and demonstrated comparable to FNA-Tg (AUC, 0.961; 95% CI: 0.928, 0.994; P = 0.989). Five independent predictors were identified: microcalcifications, abnormal hyperechogenicity, centripetal/synchronous enhancement, microvascular complexity level > 1.64, and mean density > 9.82. The resultant nomogram showed excellent discrimination, calibration and clinical utility.

Conclusion

The multimodal US+CEUS+SRUS model provides an accurate, non-invasive strategy for preoperative diagnosis LLNM in PTC.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40644-026-01026-8.

Introduction

Papillary thyroid carcinoma (PTC) is the most common thyroid cancer [1]. Although most patients have a favorable prognosis, lateral cervical lymph node metastasis (LLNM) occurs in 20%-50% of cases [2, 3]. LLNM significantly impacts management strategies, surgical planning and prognosis of PTC patients. Guidelines recommend therapeutic lateral lymph node dissection for lateral lymph node (LN) with clinically/pathologically confirmed metastasis preoperatively or intraoperatively [4]. US remains the first-line imaging tool, but its sensitivity for LLNM is limited, approximately 60%-73% [3–5]. Furthermore, US features overlap between benign and malignant LNs, and early or focal metastatic signs are often non-specific, making accurate preoperative assessment a significant clinical challenge.

According to guidelines [4], suspicious LNs may undergo US-guided fine needle aspiration cytology (FNAC). However, its diagnostic sensitivity remains limited [6], as cytological interpretation is operator- and pathologist-dependent. Alternatively, fine needle aspiration thyroglobulin (FNA-Tg) measurement offers higher diagnostic efficacy [6–8]. Nevertheless, FNA-Tg is not universally available, and a standardized diagnostic cutoff for LLNM is lacking due to variability in assays and patient populations. Moreover, the rationale for performing invasive procedures on LNs that appear relatively clear on US is questionable. Therefore, an effective, non-invasive imaging method is needed to accurately assess LLNM.

Tumor development is characterized by early microvascular alterations, with progressive increases in vessel number, density, diameter, and tortuosity from early to advanced stages [9, 10]. Furthermore, study indicates that during early stages of lymphatic metastasis, changes in vascular volume and density precede alterations in LN size [11]. While color Doppler flow imaging (CDFI) cannot effectively detect microvessels. Contrast-enhanced ultrasound (CEUS) allows real-time evaluation of microcirculatory perfusion by observing intravascular contrast microbubbles. Super-resolution ultrasound (SRUS) overcomes the diffraction limit of conventional US, enabling microvascular imaging at micron-scale resolution [12–14]. SRUS provides clearer delineation of tumor microvasculature and permits quantitative assessment of microvascular morphology and hemodynamics, thereby offering a potential tool for detecting early microvascular alterations in disease. To date, SRUS has shown considerable promise in preclinical models and preliminary clinical studies, emerging as a research hotspot [15–17].

We hypothesized that SRUS can sensitively detect abnormal microvascular perfusion within metastatic LNs. This study aimed to prospectively evaluate suspicious lateral neck LNs in patients with PTC using US, CEUS, and SRUS quantitative parameters. By developing and comparing multimodal diagnostic models with the established standards of FNAC and FNA-Tg, we sought to identify an optimal model for preoperative diagnosis of LLNM, thereby reducing unnecessary invasive biopsies.

Materials and methods

This prospective study was approved by the Ethics Committee of our hospital (approval no. KYLL-2025-268), and written informed consent was obtained from all participants.

Participants

Consecutive participants with suspected thyroid cancer and potential LLNM were enrolled at our hospital from January to November 2025. Inclusion criteria were: (a) age > 18 years; (b) preoperative US showing at least one suspicious feature in a lateral cervical LN according to the 2025 American Thyroid Association guidelines [4], including loss of fatty hilum, round shape, hyperechogenicity, cystic change, calcifications, or peripheral/disorganized vascularity; (c) scheduled to undergo both FNAC and FNA-Tg; (d) surgery planned within two weeks after CEUS and SRUS examinations. Exclusion criteria included: (a) known or suspected hypersensitivity to sulfur hexafluoride lipid microsphere or its components; (b) prior radioactive iodine 131 therapy; (c) obvious cystic change; (d) poor CEUS/SRUS image quality; (e) received ablation therapy after diagnosis; (f) surgery without lateral neck dissection; (g) final pathological diagnosis other than PTC; (h) loss to follow-up. The flowchart of study is shown in Fig. 1.

Fig. 1.

The flowchart of study

US examination

Participants were positioned supine with the neck fully exposed. A comprehensive scan of the thyroid and cervical LNs was performed. Gray-scale US documented the location, size, shape, echogenicity, calcifications, cystic change, abnormal hyperechogenicity, and lymphatic hilum. CDFI was used to assess vascularity, with particular attention to peripheral vascularity.

CEUS examination and SRUS images acquisition

ULTIMUS 9E (Vinno, Suzhou, China) system equipped with a U5-15 L linear transducer was used. The system was configured for ultrafast contrast-enhanced imaging and ultra-resolution microscopy.

The contrast agent SonoVue (Bracco, Milan, Italy) was diluted in 5.0 mL of 0.9% sodium chloride, shaken thoroughly, and administered as a 1.5 mL intravenous bolus followed by a 5.0 mL saline flush. The target LN was observed in real-time without interruption for up to 60 s. Upon visual confirmation of microbubbles within the target LN, the ultra-resolution microscopy mode was activated for a 10-second acquisition. Participants were instructed to breathe quietly, with breath-holding if necessary, while the transducer was held steady to minimize motion artifacts. Imaging settings were standardized: mechanical index (0.073), acoustic power (9%), and a contrast-specific preset (“Low”).

Qualitative CEUS features were assessed, including: (a) enhancement intensity (hypo-, iso-, or hyper-enhancement); (b) enhancement pattern (centrifugal, centripetal or synchronous); (c) perfusion defect (no or yes); (d) enhancement homogeneity (homogeneous or heterogeneous).

SRUS microvascular maps were reconstructed from the dynamic CEUS sequences using the built-in ultra-resolution microscopy quantification toolkit. The ULTIMUS 9E system is equipped with a built-in motion compensation function with adjustable levels (0, 1, and 2). If needed, level 2 motion compensation was uniformly applied to correct residual motion artifacts. This algorithm filters out motion-corrupted data through inter-frame correlation analysis and subsequently performs rapid and precise motion compensation on stable data segments. The target lesions were manually contoured on the SRUS images with reference to corresponding 2D images. Subsequently, the software automatically generated microvascular quantitative parameters, including density-velocity analysis, displaying a comprehensive analysis of microvascular distribution across the entire LN, and density fit, showing vascular distribution, spacing, and diameter along the nodal maximum diameter. The US, CEUS features and SRUS quantitative parameters evaluated in this study are detailed in Supplementary Material 1.

All US, CEUS and SRUS examinations were performed by two sonographers with 7 and 28 years of thyroid imaging experience. Image analysis was conducted independently by two physicians (with 7–10 years of experience) blinded to final histopathology. Discrepancies were resolved by consensus, with the senior sonographer (28 years) serving as arbitrator when necessary.

FNAC and FNA-Tg

Written informed consent was obtained from all participants prior to the procedure. Examinations were performed using one of two US systems: a Canon Aplio i800 (Canon Medical Systems Corporation, Tochigi, Japan) equipped with an i18LX5 wide-band linear probe, or a Resona R9 Pro (Mindray, Shenzhen, China) equipped with an L15-3WU wide-band linear probe. With the patient in a supine position and the neck fully exposed, standard aseptic preparation was applied and local anesthesia (2% lidocaine) was administered. Under real-time US guidance, a 23-gauge (50-mm) needle was used to perform repeated effective suction-lifting maneuvers on the target LN until visible tissue filled the needle lumen. This procedure was repeated five times per target lesion. The tissue samples from the first two passes were rinsed with 1.0 mL of normal saline, and the resulting eluate was collected in a sterile tube for FNA-Tg measurement. The samples from the last three passes were placed into a ThinPrep^® cytologic solution container for cytological examination. All aspirations were performed by one of two interventional radiologists, each with 5–8 years of experience in aspiration.

Cytological result was diagnosed as positive if malignant cells were identified, and negative otherwise. FNA-Tg levels were quantified using a fully automated electrochemiluminescence immunoassay system (Cobas 8000, Roche, Germany), with a detection range of 0.04–500 ng/mL. The optimal diagnostic cutoff value for FNA-Tg was determined from the study cohort using receiver operating characteristic (ROC) analysis. Measurements ≥ the cutoff were considered positive, while < the cutoff were considered negative.

Preoperative localization of targeted LN

A solution of Mitoxantrone Hydrochloride (Shanghai Chuangnuo Pharmaceutical Co., Ltd.) was injected percutaneously into the suspicious LN under real-time US guidance preoperatively. Using a 1 mL syringe with a 23-gauge (50 mm) needle, a volume of 0.15 mL was administered. The overlying skin surface was marked with a surgical pen to guide subsequent intraoperative identification and targeted resection.

Statistical analysis

SPSS (version 26.0; IBM) and R (version 4.5.1, RStudio) were used for statistical analyses. Surgical pathology served as the gold standard. Continuous variables are presented as mean ± standard deviation (SD) or median with interquartile range (IQR; P25-P75) as appropriate. Categorical variables are summarized as frequencies (percentages). A two-sided P < 0.05 was considered statistically significant.

Generalized estimating equations (GEE) with an exchangeable correlation structure were employed to account for clustering of multiple LNs within the same participant (112 participants with 149 LNs). Variables with P < 0.05 in univariable GEE analyses were entered into multivariable GEE models. Three predictive models were constructed using GEE: US + CDFI, US + CEUS, and US + CEUS + SRUS.

ROC curves were plotted for these three models as well as for FNAC and FNA Tg. The area under the curve (AUC) with 95% confidence interval (CI) was calculated for each method. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated based on the optimal cutoff determined by the maximum Youden index. Pairwise comparisons of AUCs were performed using the DeLong test, with significance set at an adjusted P < 0.05 (Benjamini-Hochberg false discovery rate procedure). Agreement with pathology was assessed with the Kappa statistic: > 0.75 indicated good, 0.40–0.75 moderate, and ≤ 0.40 poor agreement.

Independent predictors identified in the optimal GEE model are presented in a forest plot with odds ratios (ORs) and 95% CIs. Specifically, the complexity level was standardized (z-score transformation) to ensure scale stability. A nomogram was developed based on the optimal GEE model coefficients to assess the individualized risk of LLNM. Model performance was evaluated in terms of discrimination, calibration and clinical utility.

Results

Clinical and pathologic characteristics of participants and lateral LNs

A total of 149 lateral LNs from 112 participants (40 male, 72 female) were included, with 74 metastatic and 75 non-metastatic. The non-metastatic group comprised 71 with reactive hyperplasia and 4 with tuberculous lymphadenitis. The median age of the cohort was 46 years (IQR, 35–54 years; range, 19–77 years), with a mean age of 45 ± 12 years. Most LNs (140 of 149; 94.0%) were ipsilateral to the primary PTC, and 9 (6.0%) were contralateral. The distribution across cervical level was as follows: level II, 15 (10.1%); level III, 65 (43.6%); level IV, 67 (45.0%); and level V, 2 (1.3%). The median long- axis and short-axis diameters were 1.18 cm (IQR, 0.86–1.61 cm) and 0.50 cm (IQR, 0.36–0.60 cm), respectively. Basic characteristics of participants and lateral cervical LNs are presented in Table 1.

Table 1.

Basic characteristics of participants and lateral cervical lymph nodes

Characteristic	Value
Participants (n = 112)
Lymph nodes (n = 149)
Sex
Female	72 (64.3)
Male	40 (35.7)
Age (y)	46 (35–54)
Location
Same side	140 (94.0)
Opposite side	9 (6.0)
Level
II	15 (10.1)
III	65 (43.6)
IV	67 (45.0)
V	2 (1.3)
Size (cm)
Long axis	1.18 (0.86–1.61)
Short axis	0.50 (0.36–0.60)
Surgical pathology
Metastatic	74 (49.7)
Non-metastatic	75 (50.3)

Univariable analyses of US, CEUS, and SRUS quantitative parameters for metastatic lateral LNs

GEE univariable analyses of US and CEUS features for lateral LNs are shown in Table 2. GEE univariable analyses of SRUS quantitative parameters for lateral LNs are demonstrated in Table 3 (limited to parameters with P < 0.05), with the complete data available in Supplementary Material 2.

Table 2.

GEE Univariable analyses of US and CEUS features of lateral cervical lymph nodes

Characteristic	Total (n = 149)	Metastatic (n = 74)	Non-metastatic(n = 75)	P
US
Round shape				< 0.001
Yes	62 (41.6)	42 (56.8)	20 (26.7)
No	87 (58.4)	32 (43.2)	55 (73.3)
Microcalcifications				< 0.001
Yes	68 (45.6)	49 (66.2)	19 (25.3)
No	81 (54.4)	25 (33.8)	56 (74.7)
Cystic change				< 0.001
Yes	15 (10.1)	15 (20.3)	0 (0.0)
No	134 (89.9)	59 (79.7)	75 (100.0)
Abnormal hyperechogenicity				< 0.001
Yes	39 (26.2)	33 (44.6)	6 (8.0)
No	110 (73.8)	41 (55.4)	69 (92.0)
Loss of fatty hilum				< 0.001
Yes	103 (69.1)	67 (90.5)	36 (48.0)
No	46 (30.9)	7 (9.5)	39 (52.0)
CDFI
Peripheral vascularity				< 0.001
Yes	37 (24.8)	33 (44.6)	4 (5.3)
No	112 (75.2)	41 (55.4)	71 (94.7)
CEUS
Enhancement intensity				< 0.001
Hypo-enhancement	14 (9.4)	4 (5.4)	10 (13.3)
Iso-enhancement	54 (36.2)	8 (10.8)	46 (61.3)
Hyper-enhancement	81 (54.4)	62 (83.8)	19 (25.3)
Enhancement pattern				< 0.001
Centrifugal	73 (49.0)	11 (14.9)	62 (82.7)
Centripetal/Synchronous	76 (51.0)	63 (85.1)	13 (17.3)
Perfusion defect				0.004
Yes	24 (16.1)	20 (27.0)	4 (5.3)
No	125 (83.9)	54 (73.0)	71 (94.7)
Enhancement homogeneity				< 0.001
Heterogeneous	80 (53.7)	54 (73.0)	26 (34.7)
Homogeneous	69 (46.3)	20 (27.0)	49 (65.3)

Table 3.

GEE Univariable analyses of SRUS quantitative parameters of lateral cervical lymph nodes

Characteristic	Total (n = 149)	Metastatic (n = 74)	Non-metastatic (n = 75)	P
Density-velocity analysis
Vessel ratio	67.64(53.43–83.46)	81.07(67.60-90.47)	58.01(37.37–68.29)	< 0.001
Complexity level	1.60 ± 0.13	1.67 ± 0.10	1.54 ± 0.12	< 0.001
Max density	22.31 ± 5.50	23.58 ± 4.95	21.06 ± 5.75	0.010
Mean density	9.52 ± 3.00	11.22 ± 2.51	7.84 ± 2.45	< 0.001
Blood volume	3.65(1.58–5.63)	4.00(2.44–7.96)	2.63 (1.21–4.99)	0.019
Max velocity	24.68 ± 2.74	25.25 ± 2.25	24.12 ± 3.06	0.016
Mean velocity	12.07(9.61–14.05)	13.96(12.54–15.05)	9.85(7.60-11.25)	< 0.001
Std velocity	4.56 ± 0.94	4.20 ± 0.96	4.92 ± 0.77	< 0.001
Perfusion index	9.59(6.40-12.47)	11.98(10.06–13.90)	6.89(4.30–8.62)	< 0.001
Density fit
Min diameter	107.40(72.39-124.66)	120.06(95.11-130.14)	87.25(65.56-111.43)	0.003
Mean diameter	294.70 ± 59.21	315.27 ± 56.55	274.40 ± 54.91	< 0.001

Significant differences between metastatic and non-metastatic LNs were observed in the following US and CEUS features (all P < 0.05): round shape, microcalcifications, cystic change, abnormal hyperechogenicity, loss of fatty hilum, peripheral vascularity on CDFI, enhancement intensity, enhancement pattern, perfusion defect, and enhancement homogeneity (Table 2). Among SRUS quantitative parameters, vessel ratio, complexity level, max density, mean density, blood volume, max velocity, mean velocity, std velocity, perfusion index, min diameter, and mean diameter were significantly different (all P < 0.05) (Table 3).

Diagnostic performance of US+CDFI, US+CEUS, US+CEUS+SRUS, FNAC, and FNA-Tg

The diagnostic performance comparison of US+CDFI, US+CEUS, US+CEUS+SRUS, FNAC, and FNA-Tg are shown in Table 4. The ROC curves for each model are shown in Fig. 2.

Table 4.

Diagnostic performance comparison of US+CDFI, US+CEUS, US+CEUS+SRUS, FNAC, and FNA-Tg for lateral cervical lymph node metastasis

Model	AUC (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	Accuracy (95% CI)	Kappa	Cut-off value	P a	P b	P c	P d
US+CDFI	0.851(0.791,0.911)	0.851 (0.750, 0.923)	0.733 (0.619, 0.829)	0.759 (0.653, 0.846)	0.833 (0.721, 0.914)	0.792 (0.718, 0.854)	0.584	0.518
US+CEUS	0.921(0.878,0.964)	0.892 (0.798, 0.952)	0.800 (0.692, 0.884)	0.815 (0.713, 0.892)	0.882 (0.781, 0.948)	0.846 (0.777, 0.900)	0.691	0.445	0.005
US+CEUS+SRUS	0.961(0.933,0.988)	0.892 (0.798, 0.952)	0.907 (0.817, 0.962)	0.904 (0.812, 0.961)	0.895 (0.803, 0.953)	0.899 (0.839, 0.943)	0.799	0.551	< 0.001	0.040
FNAC	0.858(0.806,0.910)	0.716 (0.599, 0.815)	1.000 (0.952,1.000)	1.000 (0.933,1.000)	0.781 (0.685, 0.859)	0.859 (0.793, 0.911)	0.718	0.500	0.941	0.062	< 0.001
FNA-Tg (ng/mL)	0.961(0.928,0.994)	0.932 (0.849, 0.978)	0.973 (0.907,0.997)	0.972 (0.902,0.997)	0.936 (0.857, 0.979)	0.953 (0.906, 0.981)	0.906	6.705	0.003	0.177	0.989	< 0.001

a, P for multiple comparisons were adjusted using the Benjamini-Hochberg method to control the false discovery rate (FDR). Significance was defined by an FDR < 0.05. b, P ^a (all vs. US+CDFI), P ^b (all vs. US+CEUS), P ^c (all vs. US+CEUS+SRUS), P ^d (all vs. FNAC).

Fig. 2.

Receiver operating characteristic curves showing the diagnostic performance of US+CDFI, US+CEUS, US+CEUS+SRUS, FNAC, FNA-Tg for lateral cervical lymph nodes metastasis

The combined US + CEUS + SRUS model achieved an AUC of 0.961 (95% CI: 0.933, 0.988), superior to both the US + CDFI (AUC, 0.851; 95% CI: 0.791, 0.911; P < 0.001) and US + CEUS (AUC, 0.921; 95% CI: 0.878, 0.964; P = 0.040) models. At the optimal cutoff, the sensitivity, specificity, PPV, NPV, and accuracy were respectively 0.851, 0.733, 0.759, 0.833, and 0.792 for US + CDFI; 0.892, 0.800, 0.815, 0.882, and 0.846 for US + CEUS; and 0.892, 0.907, 0.904, 0.895, and 0.899 for US + CEUS + SRUS. The agreement with postoperative pathology, assessed by the Kappa statistic, was moderate for US + CDFI (κ = 0.584) and US + CEUS (κ = 0.691), and good for US + CEUS + SRUS (κ = 0.799).

Among the 149 lateral LNs, FNA-Tg at an optimal cutoff of 6.705 ng/mL showed excellent diagnostic performance for LLNM, with an AUC of 0.961(95% CI: 0.928, 0.994). The sensitivity, specificity, PPV, NPV, and accuracy were 0.932, 0.973, 0.972, 0.936, and 0.953, respectively. FNAC exhibited the highest specificity and PPV but lower sensitivity (0.716). Pairwise DeLong tests indicated that the US + CEUS + SRUS model performed better than FNAC (AUC, 0.858; 95% CI: 0.806, 0.910; P < 0.001), and was comparable to FNA-Tg (P = 0.989). Agreement with final pathology was good for both US+CEUS+SRUS (κ = 0.799) and FNA-Tg (κ = 0.906).

Independent predictors and performance evaluation of the multimodal US+CEUS+SRUS model

The GEE multivariable analysis of the optimal US+CEUS+SRUS model identified five independent predictors for LLNM in PTC (Fig. 3): microcalcifications (OR, 5.73; 95% CI: 1.57, 20.98; P = 0.008), abnormal hyperechogenicity (OR, 27.86; 95% CI: 4.50, 172.29; P < 0.001), centripetal/synchronous enhancement pattern (OR, 10.08; 95% CI: 3.02, 33.61; P < 0.001), complexity level (OR per SD increase, 2.43; 95% CI: 1.07, 5.53; P = 0.034), and mean density (OR, 1.53; 95% CI: 1.06, 2.22; P = 0.023). Among these individual features, enhancement pattern showed the highest individual AUC (0.839; 95% CI: 0.776, 0.895), sensitivity (0.851), NPV (0.849), and accuracy (0.839), whereas abnormal hyperechogenicity had the highest specificity (0.920) and PPV (0.846) (Table 5).

Fig. 3.

Forest plot of the GEE multivariable analysis for the US+CEUS+SRUS model. Complexity level(sta)= (complexity level - mean) / SD, and mean = 1.60, SD = 0.13, and its OR represents the change in odds associated with each SD increase

Table 5.

Diagnostic performance of independent predictors identified by the multimodal US+CEUS+SRUS model for lateral cervical lymph node metastasis

Independent Predictor	AUC (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	Accuracy (95% CI)	Kappa	Cut-off value
Microcalcifications	0.704(0.638,0.776)	0.662(0.546, 0.783)	0.747(0.649, 0.837)	0.721(0.607, 0.828)	0.691(0.585, 0.790)	0.705(0.641, 0.777)	0.409	-
Abnormal hyperechogenicity	0.683(0.623,0.752)	0.446(0.327, 0.567)	0.920(0.857, 0.975)	0.846(0.734, 0.953)	0.627(0.526, 0.716)	0.685(0.607, 0.756)	0.367	-
Enhancement pattern	0.839(0.776,0.895)	0.851(0.769, 0.928)	0.827(0.735, 0.917)	0.829(0.729, 0.911)	0.849(0.775, 0.927)	0.839(0.775, 0.895)	0.678	-
Complexity level	0.799(0.727,0.862)	0.649(0.486, 0.963)	0.787(0.446, 0.965)	0.750(0.595, 0.935)	0.694(0.618, 0.930)	0.718(0.670, 0.797)	0.436	1.635
Mean density	0.831(0.765,0.904)	0.757 (0.524, 0.909)	0.773(0.641, 0.984)	0.767(0.682, 0.975)	0.763(0.644, 0.880)	0.765(0.719, 0.853)	0.530	9.815

The US+CEUS+SRUS-based nomogram for predicting individualized LLNM risk is shown in Fig. 4A. The model demonstrated excellent discrimination (AUC, 0.961; 95% CI: 0.933, 0.988) and good calibration upon bootstrap validation (mean absolute error = 0.028; Hosmer-Lemeshow test P = 0.78) (Fig. 4B). Decision curve analysis confirmed superior clinical utility, showing a higher net benefit across a wide range of threshold probabilities than “treat all” or “treat none” strategies (Fig. 4C). The clinical impact curve further validated its predictive accuracy, with predicted high-risk cases closely matching the actual metastasis incidence at higher risk thresholds (Fig. 4D).

Fig. 4.

Construction and performance evaluation of the US+CEUS+SRUS-based nomogram for predicting the individualized risk of lateral cervical lymph node metastasis. (A) Nomogram for predicting individualized LLNM risk. (B) Calibration curve. (C) Decision curve analysis. (D) Clinical impact curve. Complexity level(sta)= (complexity level - mean) / SD, and mean = 1.60, SD = 0.13; MAE, mean absolute error

Discussion

For PTC, a relatively indolent thyroid cancer, personalized precision management based on standardized protocols are of particular importance. The completeness of initial surgery is a critical determinant of patient prognosis [4]. The multimodal US+CEUS+SRUS model achieved high diagnostic performance for identifying LLNM in PTC (AUC, 0.961; 95% CI: 0.933, 0.988). Incorporating SRUS improved specificity (0.907 vs. 0.800), PPV (0.904 vs. 0.815) and accuracy (0.899 vs. 0.846) over the US+CEUS model while maintaining equivalent sensitivity and NPV. Compared with guideline-recommended methods, the multimodal US+CEUS+SRUS model superior to FNAC (AUC, 0.858; 95% CI: 0.806, 0.910; P < 0.001), and comparable to FNA-Tg (AUC, 0.961; 95% CI: 0.928,0.994; P = 0.989). Multivariable analysis identified five independent predictors of LLNM, among these, enhancement pattern showed the highest sensitivity (0.851), NPV (0.849), and accuracy (0.839); while abnormal hyperechogenicity offered the highest specificity (0.920) and PPV (0.846), albeit with limited sensitivity. These results indicated that no single imaging feature is uniformly reliable, highlighting the value of an integrated multimodal approach for precise diagnosis.

This study confirms the efficacy of SRUS in visualizing and quantifying microvascular architecture in lateral LNs. We identified two SRUS-derived parameters as independent risk factors for LLNM: complexity level (OR per SD increase, 2.43; 95% CI: 1.07, 5.53; P = 0.034) and mean density (OR, 1.53; 95% CI: 1.06, 2.22; P = 0.023). The complexity level (Supplementary Material 3) reflects the geometric intricacy of the microvascular network, incorporating information on morphology, margins, and texture. A higher value indicates greater architectural disorganization, thereby serving as a valuable indicator for differentiating between benign and malignant tumors. Mean density quantifies the average cumulative count of microbubbles passing through a specific spatial location within the acquisition time, and its primary clinical value lies in its ability to characterize angiogenesis in lesions. In our cohort of 149 LNs, both parameters were significantly higher in metastatic than in non-metastatic LNs, reflecting the more tortuous, irregular, and disorganized architecture and richer vascular perfusion of metastatic LNs. Zhu et al. [18] previously applied SRUS in a small cohort (6 metastatic, 4 benign) and observed that local microvascular flow direction irregularity was 60% higher in metastatic LNs compared with benign reactive nodes. However, while that study primarily identified group-level differences, our work further extends the clinical utility of SRUS-derived microvascular parameters, thereby providing quantitative imaging evidence for vascular architectural disorganization in tumor angiogenesis.

Centripetal/synchronous enhancement pattern on CEUS was identified as an independent predictor (OR, 10.08; 95% CI: 3.02, 33.61; P < 0.001), consistent with prior studies [19–21]. This pattern may reflect the pathophysiological process in which tumor cells enter the LN via afferent lymphatic vessels, accumulate in the subcapsular region, and stimulate the release of pro-angiogenic factors, thereby promoting perinodal vascular proliferation and intranodal neovascularization [10, 22]. In contrast, the most common non-metastatic LN in PTC patients is the benign reactive node, characterized by a preserved hilum and a centrifugal enhancement pattern that progresses from the central hilum toward the periphery.

Microcalcifications and abnormal hyperechogenicity were confirmed as independent predictors of metastasis, aligning with prior literature [4, 23]. Sonographic punctate echogenic foci typically correspond to psammoma bodies histopathology [24, 25]. The number of such foci is noteworthy, while a few or isolated punctate echoes often lack diagnostic significance for metastasis, a higher count strongly suggests metastatic involvement. We propose that these foci may represent either early calcific deposits from micro-metastases or short linear echogenic strands visualized en face, rather than true microcalcifications. Notably, coarse calcifications, though less frequent, do not exclude metastasis, as evidenced by cases in our cohort where both primary tumor and metastatic LNs showed similar calcification patterns. Marked abnormal hyperechogenicity often provides a direct diagnostic clue, possibly corresponding to thyroglobulin aggregation [26]. The formation of small thyroglobulin-filled vesicles creates high-impedance acoustic interfaces, resulting in hyperechoic appearances on US. However, isolated, well-defined hyperechoic foci remain ambiguous. SRUS effectively addresses this difficulty by providing micron-scale imaging of microvascular perfusion, enabling direct visualization of tumor-associated angiogenesis within suspicious hyperechoic foci. Hyperechoic foci with marked hyperenhancement indicated metastasis (Fig. 5A and B), reflecting underlying tumor angiogenesis; in contrast, foci showing no abnormal hyperenhancement were likely benign (Fig. 5C), potentially corresponding to avascular structures such as adipocyte aggregates (Fig. 6). This underscores the value of functional microvascular assessment over morphology alone. LNs with extensive cystic change were excluded from this study to focus on SRUS-based microvascular analysis. Thus, cystic change, an acknowledged indicator of metastasis, was not included in the final model. In clinical practice, CEUS and SRUS can definitively identify small or focal cystic areas that are often missed on US, thereby providing a stronger indicator for LNM. Distinguishing metastatic from tuberculous lymphadenitis is challenging. Although microcalcifications suggest metastasis, tuberculous LNs may also show punctate hyperechoic foci (Fig. 5D). In such equivocal cases, CEUS and SRUS may aid differentiation by revealing perfusion patterns, where extensive non-enhancement was observed (Fig. 5D). When imaging remains inconclusive, adjunctive FNAC and FNA-Tg are recommended.

Fig. 5.

A-D. Images of surgical pathology confirmed lateral cervical LNs in patients with PTC. A. Metastatic LN: a 34-year-old female. (a) US: a 22 × 4 mm suspicious LN with microcalcifications and abnormal hyperechogenicity. (b) CDFI: punctate and strip-shaped blood flow signals within the hyperechoic area. (c) CEUS: the hyperechoic area shows marked enhancement. (d) SRUS (microvascular density map): twisted and disordered microvessels are displayed within the hyperechoic region. B. Metastatic LN: a 53-year-old male. (e) US: a 14 × 3 mm suspicious LN with abnormal hyperechogenicity. (f) CDFI: punctate blood flow signals within the hyperechoic area. (g) CEUS: the hyperechoic area shows marked enhancement. (h) SRUS (microvascular density map): twisted and disordered microvessels are displayed within the hyperechoic region. C. reactive LN: a 29-year-old female. (i) US: a 13 × 4 mm suspicious LN with abnormal hyperechogenicity. (j) CDFI: no obvious blood flow is detected. (k) CEUS: the hyperechoic area shows no enhancement. (l) SRUS (microvascular density map): sparse microvessels are shown within the hyperechoic area. D. Tuberculous LN: a 56-year-old female. (m) US: a 24 × 8 mm suspicious LN with microcalcifications and loss of fatty hilum. (n) CDFI: peripheral blood flow signals are visible in the LN. (o) CEUS: the LN shows heterogeneous, hyper-enhancement, with a large central area of perfusion defect. (p) SRUS (microvascular density map): twisted and disordered microvessels are displayed within the peripheral region

Fig. 6.

Workflow for pathological correlation of a target LN. (A) The target LN, stained blue with mitochloropyridinol hydrochloride for intraoperative identification, is excised and submitted separately for pathological analysis. (B) Post-excision ultrasonographic rescanning of the target LN for definitive confirmation, the black asterisk shows the abnormal hyperechogenicity observed on US. (C) Intraoperative frozen section of the target LN; the black asterisk denotes the hyperechoic area observed on US. (D) Postoperative pathological result: reactive lymphoid hyperplasia (HE staining, 10X); the black asterisk indicates adipocytes, and the white asterisk indicates lymphocytes

Among the 149 LNs analyzed, FNAC demonstrated the highest specificity and PPV, but limited sensitivity (0.716). In contrast, FNA-Tg achieved excellent diagnostic performance at an optimal cutoff of 6.705 ng/mL, yielding an AUC of 0.961 (95% CI: 0.928, 0.994), with sensitivity, specificity, PPV, NPV, and accuracy were 0.932, 0.973, 0.972, 0.936, and 0.953, respectively. A meta-analysis of 22 studies (2,670 LNs) reported pooled sensitivity and specificity of FNA-Tg at 91% and 94%, respectively, noting that lower values (e.g., 1 ng/mL) maximize sensitivity whereas higher ones (e.g., 40 ng/mL) maximize specificity [8]. Although a universal cutoff remains undefined, its high diagnostic efficacy is undeniable [6–8, 27]. Therefore, it is recommended that individual centers establish their own diagnostic thresholds based on patient population and assay platforms.

Importantly, our multimodal US+CEUS+SRUS model demonstrated diagnostic accuracy comparable to FNA-Tg. Furthermore, the US+CEUS+SRUS-based nomogram showed excellent discrimination, calibration and clinical utility. This positions SRUS as a robust, non-invasive alternative in settings where FNAC and FNA-Tg are not routinely accessible. While SRUS entails modest incremental time (about 5 min per node) and contrast agent cost (about 513 CNY) relative to conventional US, these trade-offs remain substantially lower than the time, expense, and patient discomfort associated with invasive FNAC or FNA-Tg procedures. Thus, its application could reduce unnecessary biopsies, alleviate patient anxiety, and lower healthcare costs.

A key methodological strength of this study lies in the preoperative marking of targeted LNs with Mitoxantrone Hydrochloride, which ensured precise one-to-one correspondence between imaged and resected LNs. This approach minimizes node mismatch and strengthens diagnostic validity.

Our study has several limitations. First, although GEE was used to account for within-patient clustering, 37 participants contributed two LNs; future studies with one node per participant would avoid this concern. Second, the relatively modest sample size may have limited the statistical power to detect smaller yet potentially meaningful differences. Furthermore, the single-center, prospective design may introduce selection bias. Our findings need to be confirmed in large-scale, multicenter studies to establish their generalizability.

Conclusion

The multimodal US+CEUS+SRUS model provides an effective strategy for the preoperative diagnosis of LLNM in PTC. Specifically, lateral neck LNs with microcalcifications, abnormal hyperechogenicity, centripetal/synchronous enhancement pattern, microvascular complexity level > 1.64, and mean density > 9.82 are highly predictive of metastasis. These findings confirm the clinical value of SRUS as an efficient, non-invasive tool to reduce unnecessary biopsies and guide personalized management.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

PTC

Papillary thyroid carcinoma

LLNM

Lateral lymph node metastasis

LN

Lymph node

FNAC

Fine needle aspiration cytology

FNA-Tg

Fine needle aspiration thyroglobulin

CDFI

Color Doppler flow imaging

SRUS

Super-resolution ultrasound

ROC

Receiver operating characteristic

SD

Standard deviations

GEE

Generalized estimating equations

AUC

Area under the curve

CI

Confidence intervals

PPV

Positive predictive value

NPV

Negative predictive value

OR

Odds ratio

Author contributions

WWF: Conceptualization, Investigation, Methodology, Data curation, Formal analysis, Writing-original draft, Writing-review & editing. LLS: Investigation, Methodology, Data curation. JL: Supervision, Resources. YDC: Methodology, Formal analysis. GL: Data curation, Validation. RNW: Visualization, Validation. ZLL: Visualization, Software. QZ: Visualization, Software. YFZ: Data curation. GLY: Data curation. PL: Supervision, Investigation, Conceptualization. LPL: Conceptualization, Resources, Supervision, Project administration, Funding acquisition. All authors read and approved the final manuscript.

Funding

This work was supported by Shanxi Province Health Commission Scientific Research Project (2025YD023), and Key Research and Development Program of Science and Technology Department of Shanxi Province (201903D321190).

Data availability

The data supporting this study can be provided by the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the First Hospital of Shanxi Medical University (Approval No. KYLL-2025-268). Written informed consent was obtained from all participants in this study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.