The value of intratumoral and peritumoral ultrasound radiomics in predicting lateral cervical lymph node metastasis in papillary thyroid carcinoma: a multicenter study of five hospitals

Abstract

Background

Accurate preoperative prediction of lateral cervical lymph node metastasis (LLNM) in papillary thyroid carcinoma (PTC) is crucial for surgical planning but remains challenging with conventional ultrasound. This study aimed to develop and validate a predictive model combining intratumoral and peritumoral ultrasound radiomics with clinical features for assessing LLNM.

Methods

Ultrasound images of 328 surgically treated patients with PTC across five medical centers were retrospectively collected and randomly divided into a training set (n=229) and a test set (n=99) at a 7:3 ratio. All patients underwent preoperative ultrasound examinations, and corresponding pathological results were obtained. The region of interest (ROI) within each PTC lesion was defined as the intratumoral region, while a 3 mm extension beyond this ROI was defined as the peritumoral region for radiomics feature extraction. Four clinical variables, including age, nodule length, ultrasound elasticity score, and nodule orientation, were also included in the analysis. Five predictive models were constructed: intratumoral, 3 mm peritumoral, intratumoral combined peritumoral, clinical, and clinical combined 3 mm peritumoral models. The model with the best predictive performance was used to construct a nomogram. Model performance was assessed using receiver operating characteristic (ROC) curves, DeLong tests, and decision curve analysis (DCA).

Results

The 3 mm peritumoral model demonstrated superior diagnostic performance compared to the intratumoral and combined intratumoral + 3 mm peritumoral models for predicting LLNM. The areas under the ROC curve (AUCs) for the training set were 0.909 (3 mm peritumoral), 0.891 (intratumoral), and 0.866 (combined intratumoral and 3 mm peritumoral), respectively. The AUCs for the clinical combined 3 mm peritumoral model, 3 mm peritumoral model, and clinical model alone were 0.928, 0.909, and 0.843, respectively (P<0.001). Calibration curves demonstrated good agreement between predicted and observed outcomes. DCA indicated that the model had clinical utility when the threshold probability exceeded 10%.

Conclusions

The clinical variables combined with a 3 mm peritumoral model can effectively predict the risk of LLNM in patients with PTC prior to surgery and may provide a reliable imaging-based foundation for personalized surgical planning.

Highlight box.

Key findings

• The 3 mm peritumoral ultrasound radiomics model demonstrated superior predictive performance for lateral cervical lymph node metastasis (LLNM) in papillary thyroid carcinoma (PTC), with an area under the receiver operating characteristic curve (AUC) of 0.909. Combined with clinical variables (age, nodule length, elasticity score, blood flow, and orientation), the integrated model achieved an AUC of 0.928, offering a reliable tool for preoperative risk stratification.

What is known and what is new?

• Ultrasound has limited sensitivity in detecting cervical lymph node metastasis. Radiomics based on intratumoral features has been used to assess thyroid nodules, but its value in predicting LLNM remains constrained.

• This study highlights the superior predictive value of peritumoral (3 mm) ultrasound radiomics over intratumoral features for LLNM. It further develops a clinically integrated nomogram that combines peritumoral radiomics with key clinical factors, providing a novel, imaging-based strategy for personalized surgical planning in PTC.

What is the implication, and what should change now?

• The proposed model can assist clinicians in preoperatively identifying PTC patients at high risk of LLNM, supporting more tailored surgical decisions and potentially reducing unnecessary prophylactic lymph node dissections. Future clinical assessments may prioritize peritumoral radiomics analysis to streamline workflow and enhance prediction accuracy without relying heavily on intratumoral features extraction.

Introduction

In recent years, the detection rate of papillary thyroid carcinoma (PTC) has increased steadily, making it the most common endocrine malignancy worldwide (1). PTC accounts for approximately 94% of all thyroid cancers. Although it is generally considered to have low malignancy (2,3), 31% of PTC patients present with extra thyroidal invasion and cervical lymph node metastasis (LNM), and approximately 20% exhibit lateral cervical LNM (LLNM) at initial diagnosis. Several studies have reported that patients with LLNM are more likely to experience recurrent LNM and distant metastasis during follow-up (4,5).

To reduce the risk of poor outcomes and improve patient quality of life, clinicians have increasingly adopted prophylactic central neck lymph node dissection during surgery, while targeted lateral neck lymph node dissection is performed for high-risk patients (6). However, cervical lymph node dissection is associated with complications such as recurrent laryngeal nerve injury and parathyroid damage (7,8). Therefore, prophylactic dissection is not performed unless LLNM can be reliably identified preoperatively. Accurate preoperative identification of LLNM is essential to guide clinical decision-making and support the development of individualized surgical plans for PTC patients.

Previous studies have demonstrated that ultrasound has some value in detecting central cervical LNM (CLNM) and LLNM, although its sensitivity remains relatively low. Computed tomography (CT) was used to assess cervical LNM, but its limited spatial resolution may result in missed detection of small metastatic lymph nodes (9). Ultrasound radiomics, a machine learning technique based on ultrasound images, has recently emerged as a tool for extracting high-throughput, abstract features from medical images (10). Research has shown that radiomics based on the primary PTC lesion can distinguish between benign and malignant nodules and has some utility in diagnosing cervical LNM (11). Since the peritumoral microenvironment also contains information relevant to tumor progression (12), recent studies have started incorporating features from areas surrounding the tumor to improve assessments of invasiveness. However, studies using peritumoral ultrasound imaging models to evaluate LLNM remain limited.

Therefore, this study aims to extract imaging features from both the intratumoral region and a 3 mm peritumoral area of the primary PTC lesion to investigate whether ultrasound-based radiomics can support the development of a clinically applicable model for preoperative prediction of LLNM in PTC patients, thus providing imaging evidence for personalized surgical plan. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-1-572/rc).

Methods

General characteristics

This study included ultrasound image data from five medical centers: The Second Affiliated Hospital of Anhui Medical University, The First Affiliated Hospital of University of Science and Technology of China, North Area of The First Affiliated Hospital of Anhui Medical University, Lu’an People’s Hospital, and Wuhu Second People’s Hospital, collected from February 2021 to October 2022. It was a prospective multicenter study.

Inclusion criteria: (I) a complete ultrasound examination was performed within 2 weeks before surgery, and the ultrasound images were clear and intact; (II) PTC was confirmed pathologically after surgery, with pathological results based on paraffin sections; (III) if multiple malignant lesions were present, the most suspicious nodule was selected for inclusion in the study; (IV) the lymph nodes in the central neck region and clinically suspicious lymph nodes in the lateral neck region were routinely dissected during surgery, and complete surgical records were maintained; (V) clinical information, pathological data, and ultrasound reports were complete; and (VI) informed consent was signed prior to participation.

Exclusion criteria: according to the inclusion and exclusion criteria, a total of 328 surgically treated PTC patients were finally included in the study, as shown in Figure 1.

Figure 1.

Flow chart in this study.

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The Second Affiliated Hospital of Anhui Medical University (approval No. 2020078), and informed consent was obtained from all individual participants. The other medical centers were also informed and agreed to the study.

Apparatus

All ultrasound images were obtained using the Siemens ACUSON Sequoia Diagnostic Ultrasound System, equipped with a 10L4 linear array probe operating at a frequency of 3.5 MHz.

Multimodal ultrasound examination

Ultrasound images were acquired by senior sonographers with more than 10 years of experience in superficial ultrasound. During the examination, the patient was placed in a supine position. The sonographer continuously scanned the thyroid and surrounding areas in multiple directions and recorded and stored the original image data.

Two-dimensional (2D) ultrasound: ultrasound images of the thyroid nodules were captured in longitudinal and transverse sections, recorded the location and size of the nodules. The long diameter of a nodule was defined as the maximum diameter measured in the longitudinal section. The nodules were categorized as either parallel or vertical based on their aspect ratio. Microcalcification was recorded when calcifications smaller than 1 mm in diameter were present within the nodule. Calcifications equal to or greater than 1 mm were not classified as microcalcifications.

Color Doppler flow imaging (CDFI): the types of blood flow signals within the nodules were recorded. Based on the distribution of blood flow, nodules were classified into four types: no blood flow, central blood flow, peripheral blood flow, and mixed blood flow.

Elastography: during ultrasound elastography (UE) scanning, patients were instructed to breathe calmly. The probe was kept perpendicular to the skin and gently placed on the neck. The UE was observed in real time, and the elasticity of the nodule was scored using the UE scoring standard (Itoh 5-point method), with scores ranging from 0 to 5 points. The sonographers performing these examinations and assigning scores (e.g., elasticity score) were not provided with information about the patient’s other clinical predictors or eventual surgical pathology outcomes during the image acquisition and feature recording process.

Ultrasound image segmentation

Feature extraction was performed on ultrasound images of PTC lesions to ensure that each image fully contained the lesion. The open-source software ITK-SNAP version 3.9.0 (https://www.itksnap.org/pmwiki/pmwiki.php) was used. Two senior sonographers, who were blinded to the final pathological results of LNM, manually segmented the intratumoral region of interest (ROI) on the ultrasound images. A 3 mm peritumoral region around the tumor was then automatically generated by the ITK-SNAP software (see Figure 2).

Figure 2.

Illustration of three ROIs in thyroid cancer ultrasound imaging. (A) The original ultrasound image. (B) Intratumoral. (C) Intratumoral (the red region) + 3 mm peritumoral (the green circle). ROI, region of interest.

Prediction model establishment with image feature screening and radiomics

All ultrasound images were preprocessed, and image-diomic features were extracted using the open-source Python software package Pyradiomics (version 2.1.0; https://pyradiomics.readthedocs.io/). Preprocessing mainly includes: (I) image normalization, ensuring that features have the same scale and range, avoiding the influence of factor or proportion differences between different data on feature extraction; and (II) pre-cropping, clipping the original image to the ROI region size, reducing the computational load of feature extraction, and excluding the influence of adjacent regions of the tumor.

A total of 1,561 imaging features were extracted from each ROI, including shape-based features (SBF), first-order statistical features (FOSF), gray level co-occurrence matrix (GLCM) features, gray level run length matrix (GLRLM) features, gray level size zone matrix (GLSZM) features, gray level dependency matrix (GLDM) features, and neighboring gray tone difference matrix (NGTDM) features.

To ensure methodological rigor and prevent information leakage, the entire dataset was first randomly partitioned into a training set and a completely independent test set before any feature analysis. Subsequently, feature selection and dimensionality reduction were performed using a multi-step pipeline strictly within the training set. First, the extracted features were initially screened using the Mann-Whitney U test to identify features with significant differences between groups. Next, highly correlated features (Pearson correlation coefficient >0.9) were removed to reduce redundancy. Finally, the least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was applied to select the most predictive, non-redundant features and construct the radiomics signature. The final radiomics score (Rad-score) for each patient was calculated based on the selected features and their LASSO coefficients derived from the training set. This finalized signature was then applied in a single step to the held-out test set for validation, ensuring the test data played no role in the feature selection or model tuning process.

Following this, a Rad-score was calculated from the selected features and LASSO coefficients. A K-nearest neighbors (KNN) classifier was trained on this score to predict LLNM. Model optimization was conducted within the training set via grid search and five-fold cross-validation, with key parameters (K, distance metric) tuned to obtain the final model.

Finally, three image-histologic models, including intratumoral, peritumoral, and intratumoral + peritumoral models, were constructed through image segmentation and feature screening.

Construction of clinical model and clinical-imaging model

Univariate analysis was performed on the ultrasound image characteristics and clinical indicators of all patients included in the training set, including age, sex, nodule length, microcalcification, nodule orientation, elasticity score, and blood flow signal. Independent sample t-test or χ² test was used for statistical analysis. P<0.05 was considered statistically significant. Indicators associated with cervical LNM were identified, and independent risk factors were selected using stepwise multivariate logistic regression analysis. Clinical logistic predictive models were built based on these factors, and model performance was tested in the validation set. Finally, the clinical predictive index and the optimal imaging group score were combined to establish the ultrasound-combined clinical model and draw the nomogram predictive model.

Model validation

The receiver operating characteristic (ROC) curve was used to calculate the area under the ROC curve (AUC) to evaluate the predictive value of the intratumoral, peritumoral 3 mm, clinical, intratumoral + peritumoral 3mm, and clinical + peritumoral 3mm models. Decision curve analysis (DCA) was used to assess the clinical applicability of the clinical-radiomics nomogram model.

Statistical analysis

Statistical data analysis was performed using R version 4.2.2. Continuous variables were tested for normality using the Shapiro-Wilk test in univariate analysis. For normally distributed data, the independent sample t-test was used, and results were expressed as mean ± standard deviation; for non-normally distributed data, the Mann-Whitney U test was used. Categorical data were compared using the χ² test or Fisher’s exact probability method. For comparisons among more than two groups, analysis of variance (ANOVA) and the Kruskal-Wallis H test were applied. ROC curves of the predictive models were compared using the DeLong test. P<0.05 was considered statistically significant.

Results

General data

A total of 328 patients with PTC were divided into a training set (n=229) and a test set (n=99) in a 7:3 ratio. Clinical and imaging characteristics were collected, including age, gender, nodule length, nodule orientation, presence of microcalcifications, blood flow, and elasticity score. The results showed no statistically significant differences between the training and test sets (P>0.05, as shown in Table 1).

Table 1. Characteristics of patients in the training and test cohorts.

Clinical features	Training set (n=229)					Test set (n=99)
	All	LLNM− (n=160)	LLNM+ (n=69)	P value		All	LLNM− (n=63)	LLNM+ (n=36)	P value
Age (years)	41.94±11.85	42.54±11.82	40.54±11.86	0.15		41.70±13.62	41.81±12.71	41.50±15.27	0.81
Tumor size (mm)	10.96±6.60	9.68±4.65	13.91±9.10	<0.001		11.39±7.29	10.13±5.99	13.59±8.80	0.02
Gender				0.80					0.88
Male	157 (68.56)	111 (69.38)	46 (66.67)			72 (72.73)	45 (71.43)	27 (75.00)
Female	72 (31.44)	49 (30.63)	23 (33.33)			27 (27.27)	18 (28.57)	9 (25.00)
Orientation				0.22					0.03
Parallel	148 (64.63)	108 (67.50)	40 (57.97)			57 (57.58)	42 (66.67)	15 (41.67)
Non-parallel	81 (35.37)	52 (32.50)	29 (42.03)			42 (42.42)	21 (33.33)	21 (58.33)
Microcalcifications (mm)				0.049					0.005
<1	114 (49.78)	87 (54.37)	27 (39.13)			53 (53.54)	41 (65.08)	12 (33.33)
≥1	115 (50.22)	73 (45.62)	42 (60.87)			46 (46.46)	22 (34.92)	24 (66.67)
CDFI				0.002					0.41
No blood flow	18 (7.86)	18 (11.25)	–			4 (4.04)	2 (3.17)	2 (5.56)
Central blood flow	24 (10.48)	19 (11.88)	5 (7.25)			10 (10.10)	8 (12.70)	2 (5.56)
Peripheral blood flow	80 (34.93)	59 (36.88)	21 (30.43)			44 (44.44)	30 (47.62)	14 (38.89)
Mixed blood flow	107 (46.72)	64 (40.00)	43 (62.32)			41 (41.41)	23 (36.51)	18 (50.00)
Elastic score				<0.001					0.07
2	7 (3.06)	3 (1.88)	4 (5.80)			2 (2.02)	1 (1.59)	1 (2.78)
3	142 (62.01)	112 (70.00)	30 (43.48)			69 (69.70)	48 (76.19)	21 (58.33)
4	76 (33.19)	41 (25.62)	35 (50.72)			25 (25.25)	14 (22.22)	11 (30.56)
5	4 (1.75)	4 (2.50)	–			3 (3.03)	–	3 (8.33)

Data are presented as mean ± standard deviation or n (%). −, negative; +, positive. CDFI, color Doppler flow imaging; LLNM, lateral cervical lymph node metastasis.

Univariate and multivariate analyses were performed on patients in the training set. Univariate analysis indicated that gender, microcalcification, nodule orientation, blood flow, elasticity score, tumor length, and age were significantly associated with LLNM (P<0.05). Multivariate logistic regression analysis further identified age, nodule length, elasticity score, blood flow, and nodule orientation as independent predictors of LLNM (P<0.05, as shown in Table 2). The AUC of the clinical model based on multivariate analysis was 0.843 in the training set and 0.720 in the test set.

Table 2. Univariable and multivariate analyses of clinical characteristics in the training set (n=229).

Clinical features	Univariable analysis			Multivariable analysis
	OR	P value		OR	P value
Gender	0.469	0.003		1.141	0.69
Orientation	0.558	0.01		1.915	0.049
Microcalcifications	0.575	0.004		1.289	0.41
CDFI	0.784	<0.001		1.472	0.03
Elastic score	0.793	<0.001		0.608	0.02
Tumor size	0.967	0.003		1.085	0.003
Age	0.98	<0.001		0.968	0.008

CDFI, color Doppler flow imaging; OR, odds ratio.

Extraction and screening of intratumoral, peritumoral radiomics, and combined histological features

A total of 1,561 ultrasound features were extracted from the 2D ultrasound images of 229 cases in the training set. Initially, 367 intratumoral features were selected using the Mann-Whitney U test, followed by 42 features using Pearson correlation analysis. Finally, the feature selection process using the LASSO method is visualized in Figure 3. This led to the identification of 21, 18, and 8 features most strongly associated with LLNM in the intratumoral, 3 mm peritumoral, and combined intratumoral + 3 mm peritumoral regions, respectively (Figure 4). Based on the analysis of radiomics feature weights in the peritumoral (3 mm) region, features with higher positive coefficients were mainly associated with small-area homogeneity and gray-level distribution, such as “peri_wavelet_HHH_glsm_SmallAreaEmphasis” (coefficient =0.060) and “peri_wavelet_HHH_glcm_Imc2” (coefficient =0.050). Among features with negative weights, several were related to texture coarseness, intensity variability, and low-gray-level zone distribution, including “peri_exponential_ngtdm_Busyness” (coefficient =−0.040) and “peri_wavelet_HLL_ngtdm_Coarseness” (coefficient =−0.050). Ultrasound radiomics prediction models were then constructed for each group. Results showed that the 3 mm peritumoral image-omics model demonstrated the highest predictive performance in the test set, with AUCs of 0.778, 0.771, and 0.768 for the peritumoral, intratumoral, and combined models, respectively (Figure 5).

Figure 3.

Radiomic feature selection was performed using a parametric method and the LASSO logistic regression. (A,D) Intratumoral; (B,E) 3 mm peritumoral; (C,F) intratumoral + 3 mm peritumoral. LASSO, least absolute shrinkage and selection operator; MSE, mean squared error.

Figure 4.

Distribution of selected features associated with LLNM. (A) Intratumoral model, (B) peritumoral 3-mm model, and (C) combined radiomics model. LLNM, lateral cervical lymph node metastasis.

Figure 5.

Comparison of ROC curves for the intratumoral, peritumoral 3-mm, and combined radiomics models. (A) ROC curves of the three models in the training set (n=229). (B) ROC curves of the three models in the independent test set (n=99). The corresponding AUC values are annotated for each model. AUC, area under the ROC curve; CI, confidence interval; ROC, receiver operating characteristic.

Establishment and validation of a clinical combined 3 mm peritumoral model

Multivariate logistic regression analysis demonstrated that age, nodule length, elasticity score, blood flow, and orientation were independent predictors of LLNM (P<0.05). Among the imaging models, the 3 mm peritumoral group showed superior predictive performance. A nomogram model was constructed by integrating clinical features with radiomics features from the 3 mm peritumoral region to predict the status of LLNM in patients with PTC.

As shown in Figure 6, the nomogram model exhibited the highest predictive accuracy, with AUCs of 0.928 in the training set and 0.816 in the test set, outperforming both the clinical model and the 3 mm peritumoral model. To comprehensively evaluate the model performance beyond the AUC, we further analyzed key classification metrics, including precision, recall (sensitivity), and the F1-score, as detailed in Table 3. Among all models, the nomogram achieved the highest F1-scores in both the training (0.817) and test (0.667) cohorts, indicating its robust and well-balanced predictive capability. In contrast, the Peri3mm model exhibited a pronounced trade-off: it achieved exceptional precision (0.952 in training and 0.500 in test) and specificity (>0.99 in training). The clinic model demonstrated consistently high recall across both cohorts (0.896 in training and 0.778 in test), confirming its sensitivity in detecting target cases, though its lower precision reflects a higher rate of false positives. Calibration curves indicated that the predicted probabilities in both the training and test sets closely approximated the ideal diagonal line, suggesting high consistency between the predicted and actual outcomes.

Figure 6.

Nomogram constructed based on the combined model. (A) A radiomic nomogram was developed in the training set incorporating peritumoral 3 mm radiomic features, age, nodule length, elasticity score, blood flow, and nodule orientation. In the training set (B,D) and test set (C,E), the nomogram, clinical model, and peritumoral 3 mm model were compared for predicting LLNM in PTC patients. (F,G) The calibration curves of the three models in the training set and test set, respectively. CDFI: 0, no blood flow; 1, central blood flow; 2, peripheral blood flow; 3, mixed blood flow. Orientation: 0, parallel; 1, non-parallel. AUC, area under the ROC curve; CDFI, color Doppler flow imaging; CI, confidence interval; DCA, decision curve analysis; LLNM, lateral cervical lymph node metastasis; PTC, papillary thyroid carcinoma; ROC, receiver operating characteristic.

Table 3. Comprehensive performance metrics of the clinic, peri-tumoral (Peri3mm), and nomogram models on the training and test sets.

Cohort	Model	Accuracy	AUC (95% CI)	Sensitivity	Specificity	Precision	Recall	F1-score
Train	Clinic	0.783	0.843 (0.780–0.905)	0.896	0.735	0.589	0.896	0.711
	Peri3mm	0.82	0.909 (0.863–0.955)	0.417	0.991	0.952	0.417	0.58
	Nomogram	0.894	0.928 (0.884–0.971)	0.792	0.938	0.844	0.792	0.817
Test	Clinic	0.656	0.720 (0.507–0.932)	0.778	0.609	0.437	0.778	0.56
	Peri3mm	0.719	0.778 (0.587–0.967)	0.444	0.826	0.5	0.444	0.471
	Nomogram	0.781	0.816 (0.646–0.986)	0.778	0.783	0.583	0.778	0.667

AUC, area under the ROC curve; CI, confidence interval; ROC, receiver operating characteristic.

Discussion

LLNM is an important prognostic factor in PTC patients (13). The surgical treatment strategy differs depending on the presence or absence of LLNM. In this study, we developed a model with PTC peritumoral 3 mm ultrasound radiomics combined with clinical features of the elasticity score, age, tumor length, and blood flow to predict LLNM in PTC patients. The model achieved an AUC of 0.928, showing that it can effectively predict the risk of LLNM in patients with PTC. This performance not only surpasses that of a prior clinical-intratumoral model, which reported an AUC of 0.843 (14), but more importantly, methodologically underscores the necessity and advantage of shifting the predictive focus from intratumoral features to the peritumoral microenvironment.

Ultrasonography is the most commonly used imaging method for PTC patients before surgery. It can evaluate the nature of nodules by observing the size, shape, blood flow, and elasticity of tumors (15). However, it has limitations in assessing the metastatic status of contralateral cervical lymph nodes (16). Some researchers have attempted to establish predictive models for distinguishing benign and malignant thyroid nodules and for evaluating LNM in PTC using radiomics (17). However, previous studies have mainly focused on intratumoral information and have lacked exploration of peritumoral features. Therefore, these studies generally demonstrate high sensitivity and specificity for differentiating benign from malignant thyroid lesions, but their sensitivity in detecting cervical LNM is limited. Recent studies have shown that perifocal sonographic features can provide additional tumor-related information, particularly in assessing local invasion and LNM of malignant lesions (18,19). The PTC microenvironment contains various immune cells, blood vessels, and extracellular matrix components (20). Changes in immune cell distribution and angiogenesis may promote tumor progression and metastasis, which indicates that peritumoral information is a crucial factor in evaluating LNM (21). Therefore, this study included the extraction of peritumoral ultrasound characteristics to improve sensitivity in assessing LLNM.

In the training set of this study, the AUC of the 3 mm peritumoral ultrasound imaging feature model was 0.909, which is higher than that of the models constructed using intratumoral ultrasound imaging features and the combined intratumoral+3 mm peritumoral features. This indicates that ultrasound imaging features around the tumor have greater predictive value for LLNM in PTC patients than features from within the tumor itself. This finding confirms the necessity of incorporating additional peritumoral imaging information when assessing cervical LNM. It was consistent with previous studies suggesting that the peritumoral microenvironment of PTC lesions better reflects tumor aggressiveness than intratumoral information, thereby providing more insight into LLNM (22). Meanwhile, the results of this study indicate that adding intratumoral ultrasound radiomics features offers limited improvement in predicting LLNM. This suggests that for future clinical evaluations focused solely on the presence or absence of LLNM in PTC, radiomics features extracted from the 3 mm peritumoral area may suffice to achieve optimal predictive accuracy, without the need to extract features from inside the tumor. This finding could significantly reduce the workload associated with extracting numerous intratumoral ultrasound features during evaluation and simplify the clinical assessment of LLNM in PTC.

Unlike previous studies that defined the peritumoral region using a percentage-based rectangular expansion from the tumor centroid (23), our model employs a fixed 3-mm annular expansion from the actual tumor contour. This provides a more anatomically faithful representation of the peritumoral region, likely reducing inclusion of unrelated distant tissue. The absolute distance also offers a standardized, size-invariant parameter, enhancing reproducibility and clinical relevance for PTC LLNM prediction.

To enhance the accuracy of the radiomics model and incorporate sonographers’ clinical experience, this study included clinically relevant parameters such as age, nodule length, elasticity score, blood flow score, and nodule orientation. The inclusion of these clinical features improved the interpretability of the final predictive model. It is worth noting that microcalcification, an important feature of thyroid malignancy, did not demonstrate strong diagnostic value in this study. Although previous research has identified microcalcification as a key marker of cervical LNM in PTC, there is no consensus on its role specifically in LLNM (24). For example, Kim et al. reported that in cases with primary lesions accompanied by microcalcification, the rate of multizone LLNM is 67%, which was significantly higher than single-zone metastasis with 33%, suggesting that microcalcification may indicate stronger tumor invasion (25). However, a recent study found that microcalcification was not an independent risk factor for lateral cervical metastasis (26). Therefore, whether microcalcification is a risk factor for lateral cervical metastasis in PTC remains debatable. We believe that microcalcification partially reflects PTC aggressiveness. Patients with microcalcification tend to have more aggressive disease, which may primarily manifest as CLNM, while its value in predicting LLNM may be limited. Future studies could further explore differences in predictive factors for CLNM versus LLNM.

The superior performance of the 3 mm peritumoral model over both the intratumoral and the combined model underscores the critical predictive value of the tumor-host interface. Biologically, the peritumoral region is a dynamic niche encompassing the invasive front, immune cell infiltration, and stromal remodeling (27). Bayarmagnai et al. confirmed in their study on the tumor microenvironment that the tumor periphery often exhibits more aligned collagen fibers, higher macrophage density, and more active vasculature—features that are strongly associated with invadopodia-driven slow locomotion and successful metastasis (28). Conversely, the observation that the combined (intratumoral and peritumoral) model underperformed the peritumoral-only model may be attributed to several factors: (I) feature redundancy and multicollinearity between highly correlated features from adjacent regions, which can destabilize model estimation; and (II) the introduction of non-prognostic noise from intratumoral areas (e.g., necrosis or hemorrhage) that dilutes the dominant predictive signal from the peritumoral microenvironment. Klyuzhin et al. (29) demonstrate that necrotic regions within tumors can introduce interference signals, which aligns with our perspective. Radiomics features from this zone, particularly those reflecting textural homogeneity and architectural complexity, may thus encode direct information about these microenvironmental interactions. Therefore, the peritumoral 3 mm region appears not only to be information-rich but also to harbor a dominant and relatively self-contained signal; simply merging it with intratumoral data without advanced fusion strategies can dilute this signal, highlighting the principle that more features do not invariably yield a better model.

This study has several limitations. First, we only extracted ultrasonic features from 2D ultrasound images and considered clinical indicators such as CDFI and elastography as clinical data. In future research, we try to explore the combined extraction of multimodal ultrasound information to evaluate whether it can further improve the specificity and sensitivity in predicting LLNM. Second, the tumor and peri-tumoral regions were manually segmented by two sonographers. While we employed a standardized protocol and achieved good inter-observer agreement [intraclass correlation coefficient (ICC) >0.85 for stable features], this manual process inherently limits reproducibility and poses a barrier to large-scale clinical deployment. And we did not implement an automated segmentation pipeline. Developing and validating such a tool—for instance, using a convolutional neural network trained on our dataset—is a critical future endeavor. Finally, regarding the validation strategy, while internal validation based on a randomly split test set from the same multicenter cohort provides a reliable performance assessment, this still represents an optimal-scenario evaluation. Therefore, the reported performance metrics should be regarded as a robust yet preliminary internal benchmark. The critical next step for future research is to conduct external validation on independent prospective multicenter cohorts, which is essential for confirming the clinical applicability of the model.

Conclusions

In this study, the peritumoral 3 mm ultrasound radiomics nomogram model demonstrated high clinical value in evaluating LLNM in PTC patients. This model can provide a reliable imaging method to assist clinicians in selecting optimal surgical strategies for PTC patients.

Supplementary

The article’s supplementary files as

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 and its subsequent amendments. The study was approved by the Ethics Committee of The Second Affiliated Hospital of Anhui Medical University (approval No. 2020078), and informed consent was obtained from all individual participants. The other medical centers were also informed and agreed to the study.

Data Sharing Statement

Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-1-572/dss

gs-15-03-68-dss.pdf ^{(77.6KB, pdf)}

DOI: 10.21037/gs-2025-1-572