Radiomics-based prediction model for lymph node metastasis in thyroid microcarcinoma

Abstract

Objective

To develop and validate a predictive model for lymph node metastasis (LNM) in thyroid microcarcinoma (TMC) based on clinical, ultrasonographic, and radiomic features, providing a basis for accurate preoperative risk assessment and individualized surgical planning.

Methods

A total of 426 TMC patients treated at our institution from June 2022 to December 2024 were retrospectively enrolled and randomly divided into a training set (n = 300) and a validation set (n = 126) at a 7:3 ratio. Demographic characteristics (age, gender), preoperative clinical and ultrasonographic features (tumor size measured by ultrasound, multifocality assessed by ultrasound, capsular invasion on ultrasound, etc.), ultrasound features (lymph node size, sphericity, etc.), laboratory indicators (preoperative TSH level), and radiomic parameters (3D tumor volume, surface area, etc.) were collected. In the training set, univariate analysis was performed to screen LNM-associated factors, followed by LASSO regression for variable selection. Multivariate logistic regression was used to identify independent predictors. Random forest (RF), K-nearest neighbors (KNN), and gradient boosting (GB) models were constructed. The model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA) in the training and internal validation sets.

Results

No significant differences in baseline characteristics were observed between the training and validation sets (P > 0.05). Univariate analysis revealed that tumor size, lymph node size, TSH level, central lymph node metastasis, 3D tumor volume, and sphericity were associated with LNM (P < 0.05). Multivariate logistic regression identified tumor size, lymph node size, TSH level, central lymph node metastasis, and 3D tumor volume as independent risk factors for LNM (P < 0.05), while sphericity was an independent protective factor (P < 0.05). The RF model exhibited superior performance (training AUC: 0.838, internal validation AUC: 0.815) compared to KNN (0.815, 0.792) and GB (0.787, 0.763), demonstrating good calibration and stable clinical net benefit.

Conclusion

The RF model constructed using multidimensional features effectively predicts LNM in TMC, with TSH level, tumor size, and sphericity as key predictors, demonstrating high clinical utility.

Introduction

Thyroid microcarcinoma (TMC) refers to thyroid malignancies with a maximum diameter ≤ 10 mm. Although its overall prognosis is favorable, the incidence has shown a steady increase in recent years, with lymph node metastasis (LNM) remaining a critical factor influencing postoperative recurrence risk and treatment strategy selection [1, 2]. In clinical practice, TMC patients with central or lateral cervical LNM require lymph node dissection, whereas excessive central lymph node dissection in patients without LNM may lead to typical complications such as recurrent laryngeal nerve injury and hypoparathyroidism, and lateral lymph node dissection is associated with other adverse effects such as neck numbness and shoulder dysfunction; notably, prophylactic lymph node dissection is never a routine clinical indication for patients with suspected thyroid microcarcinoma. Therefore, precise preoperative prediction of LNM status is essential for balancing therapeutic benefits and risks [3–4].

Traditional prediction methods primarily rely on ultrasonographic features (e.g., tumor taller-than-wide shape, calcification) and Preoperative clinical characteristics.

However, single indicators exhibit limited predictive efficacy and are subject to operator-dependent variability, failing to meet the demands of individualized treatment [5–6]. Recent advances in radiomics—quantifying three-dimensional tumor features such as volume and sphericity—have provided novel insights into assessing tumor biological behavior. Meanwhile, machine learning algorithms demonstrate superior capability in integrating multi-source data for disease risk prediction.

Existing studies confirm that tumor size and TSH levels correlate strongly with TMC LNM, and radiomic parameters may supplement the predictive value of conventional ultrasonography. However, most prior studies focus on isolated indicators without systematically integrating clinical, ultrasonographic, radiomic, and laboratory variables, limiting model generalizability and clinical utility [7–8]. To address this gap, this study aims to combine multi-dimensional predictors, apply LASSO regression for dimensionality reduction, and construct a machine learning-based predictive model for TMC LNM. The findings will identify independent risk factors and key predictors, offering an objective basis for preoperative risk stratification and surgical planning.

Materials and methods

Study population

The sample size was calculated based on the expected LNM incidence (25%–30%, derived from prior radiomics literature, pilot data, and clinical TMC characteristics).

Using SPSS 26.0 (sample size module) and R 4.2.1 (“pwr” and “epiR” packages), we set a significance level (α = 0.05, two-tailed) and a statistical power (1 − β) of 80%, which is the conventional standard for clinical prediction model development. Under these parameters, the minimum required sample size for the training set was 290 patients. Our actual training set included 300 patients, exceeding this requirement. In addition, the training set contained 80 LNM events and 6 candidate predictors, yielding an events-per-variable (EPV) ratio of 13.3, which exceeds the classic EPV ≥ 10 criterion, confirming adequate sample size for model development and internal validation. A total of 426 TMC patients were retrospectively enrolled and randomly divided into a training set (n = 300) and a validation set (n = 126) at a 7:3 ratio.

Inclusion criteria

Patients were included if they met the following criteria: first, the diagnosis of papillary thyroid microcarcinoma was confirmed in accordance with the WHO Classification of Thyroid Tumors (2022) [6] and the Chinese Expert Consensus on Thyroid Microcarcinoma (2023) [7], with the maximum diameter of the tumor ≤ 10 mm verified by postoperative pathological examination; second, all patients completed preoperative thyroid color Doppler ultrasonography and serum thyroid stimulating hormone (TSH) detection, and a subset of patients additionally underwent contrast-enhanced computed tomography (CT) of the neck; third, complete clinical data (demographics, medical history), pathological data (including postoperative lymph node metastasis status), imaging data (ultrasound/CT images available for radiomic feature extraction) and follow-up data were collected and stored in the electronic medical record system; and fourth, there was no history of prior thyroid-related therapeutic interventions, including radioiodine therapy, thyroid surgery, targeted therapy and endocrine therapy.

Exclusion criteria

Other thyroid malignancies (e.g., follicular, medullary, anaplastic carcinoma) or extra-thyroidal malignancies.

Severe cardiac/hepatic/renal/coagulation disorders or immunological diseases affecting surgery/data collection.

Poor-quality imaging (e.g., blurred ultrasound, CT artifacts) precluding radiomic analysis.

Missing critical data (e.g., unconfirmed LNM status, unreported lab results).

Pregnancy/lactation (hormonal influences on TSH/tumor biology).

Data collection

Demographics (age, gender), preoperative clinical and imaging features (tumor size on ultrasound, multifocality on ultrasound, capsular/extrathyroidal invasion on ultrasound), ultrasonographic characteristics (margin regularity, calcification, echogenicity, aspect ratio, lymph node size/morphology/hilar structure/echo pattern/vascularity/elasticity score), radiomic parameters (3D volume, surface area, sphericity [0–1, higher=more spherical and regular tumor shape], short-/long-run emphasis, size-zone variability) extracted mainly from preoperative thyroid contrast-enhanced CT images (partial morphological parameters were verified by high-resolution thyroid ultrasound images), and laboratory indices (TSH, chemiluminescence immunoassay) were retrieved from electronic medical records, PACS, and LIS. Notably, histological characteristics such as lymphatic invasion, surgical margin status and aggressive variants of papillary thyroid carcinoma were not included in the analysis, as these indicators can only be obtained by postoperative pathological examination and are not available for preoperative risk assessment. In multifocal cases, the largest tumor focus was used for radiomic feature extraction. The most suspicious lymph node on preoperative ultrasound was selected for lymph node feature extraction.

Outcome definition

LNM was defined per reference standards [6, 7]: LNM group: Pathologically confirmed central/lateral cervical LNM (including micrometastases ≤ 2 mm). Preoperative ultrasound-guided fine-needle aspiration confirming metastasis. Intraoperative frozen section positivity. Non-LNM group: Pathologically negative nodes without preoperative/intraoperative metastasis evidence.

Statistical analysis

Statistical analyses were performed using SPSS 26.0 (IBM Corp., Armonk, NY, USA) and R 4.2.1. Normally distributed continuous variables were expressed as mean ± standard deviation (x̄±s) and compared between the training and validation sets or between metastatic and non-metastatic groups using independent samples t -tests.

Non-normally distributed continuous variables (e.g., short-run emphasis and long-run emphasis in radiomics texture parameters) were presented as median (interquartile range) [M (Q1, Q3)] and compared using the Mann-Whitney U test. Categorical variables were expressed as number (percentage) [n (%)] and compared using the chi-square (χ²) test. Fisher’s exact test was applied when the theoretical frequency (T) was < 5.Univariate analysis was first conducted in the training set, and variables with P < 0.05 were included in subsequent analyses. To avoid multicollinearity, least absolute shrinkage and selection operator (LASSO) regression analysis was performed using the “glmnet” package in R, with the optimal λ value determined by 10-fold cross-validation, to compress redundant variables and screen for core predictive indicators. The LASSO-selected variables were then incorporated into multivariate logistic regression analysis to identify independent predictors of lymph node metastasis (LNM) in thyroid microcarcinoma (TMC).Sample size calculation followed the classic principles for predictive model construction. Based on the expected incidence of the primary outcome (LNM in TMC, 25%–30%), a significance level (α) of 0.05 (two-tailed), a statistical power (1 − β) of 85%, and an anticipated 15% loss to follow-up rate, the minimum required sample size was verified using both the “Sample Size Calculation” module in SPSS 26.0 and the “pwr” package in R. Before machine learning modeling, z-score standardization was performed on all continuous variables (tumor size, lymph node size, TSH level, 3D tumor volume, sphericity) to eliminate the influence of dimensional differences, and the standardized data met the standard normal distribution (mean = 0, standard deviation = 1). The predictive model was established based on independent risk factors: The random forest (RF) model was constructed using the “randomForest” package in R. K-nearest neighbors (KNN), gradient boosting (GB) model, and other machine learning algorithms were implemented via the “caret” package. Receiver operating characteristic (ROC) curves were plotted using the “pROC” package, and the area under the curve (AUC) with 95% confidence intervals (CI) was calculated to compare the discriminative performance of different models. Calibration curves were generated using the “briq” package, and model calibration was assessed by calibration curves and the Brier score (the Hosmer-Lemeshow test is not applicable to machine learning models). Decision curve analysis (DCA) was performed using the “rmda” package to evaluate the clinical net benefit across different risk thresholds. Additionally, SHAP analysis was performed using the “shap” package (version 0.41.0) in Python 3.9.7 to interpret the Random Forest model. Figure 5A was generated using R 4.2.1 with the ggplot2 and corrplot packages, and Fig. 5B was generated using the SHAP package in Python. The complete code is provided in the supplementary materials. Internal validation was conducted for the training set using the bootstrap method (1,000 resamples) to evaluate model stability; the test set (validation set) was used for independent validation without bootstrap. SHAP values were computed using the ‘shap’ package in Python to interpret feature importance of the RF model. 3D tumor volume was automatically calculated by CT imaging software. Missing values were handled by multiple imputation, outliers were removed if Z-score > 3, and all continuous variables were standardized by z-score. Machine learning modeling parameters are listed in the supplementary materials. The area under the curve (AUC) was calculated to assess model discrimination, with values closer to 1 indicating better performance. The model satisfies the classic “10 events per variable (EPV)” rule. Six core predictive variables were included, with 80 metastatic events in the training set (EPV ≥ 10), ensuring model stability. The variance inflation factor (VIF) was computed via the “car” package to assess multicollinearity. A two-tailed P < 0.05 was considered statistically significant for all analyses.

Fig. 5.

SHAP summary plot of the RF model showing feature importance and impact on lymph node metastasis prediction (A: Correlation trends, B: Dominant predictors)

Machine learning model parameters setting

All machine learning models were constructed based on R 4.2.1 and Python 3.9.7, and the hyperparameters were optimized by 10-fold cross-validation to avoid overfitting. The specific parameter settings are as follows: (1) Random Forest (RF) model: The number of decision trees (ntree) was set to 500, the number of variables tried at each split (mtry) was 3 (determined by cross-validation), the minimum number of observations at the terminal node (nodesize) was 5, and the sampling method was bootstrap with replacement. (2) K-nearest neighbors (KNN) model: The number of nearest neighbors (k) was 7 (optimized by cross-validation), the distance metric was Euclidean distance, the weight was set to “uniform”, and the data was standardized before modeling for distance calculation. (3) Gradient Boosting (GB) model: The number of base learners (n_estimators) was 300, the learning rate (learning_rate) was 0.01, the maximum depth of decision trees (max_depth) was 3, the minimum number of samples required to split an internal node (min_samples_split) was 2, and the loss function was “log_loss” for binary classification. SHAP values for model interpretability were computed using the ‘shap’ package in Python 3.9.7, and the summary plot (Fig. 5B) was generated with the same package.

Results

Baseline characteristics

The training set (n = 300) included 80 LNM cases (26.67%) and the validation set (n = 126) had 35 LNM cases (27.78%). No significant intergroup differences were observed (P > 0.05) (Table 1).

Table 1.

Comparison of baseline characteristics between training and validation sets

Variables	Training set (n = 300)	Validation set (n = 126)	t/χ²	P
Age (years)	45.62 ± 9.38	44.89 ± 8.75	0.748	0.455
Gender (male/female)	108(36.0%)/192(64.0%)	51(40.5%)/75(59.5%)	0.760	0.383
Tumor size (mm)	6.85 ± 2.13	6.62 ± 1.98	1.038	0.299
Multifocality (yes/no)	89(29.7%)/211(70.3%)	37(29.4%)/89(70.6%)	0.003	0.950
Capsular invasion (yes/no)	65(21.7%)/235(78.3%)	27(21.4%)/99(78.6%)	0.003	0.956
Extrathyroidal extension (yes/no)	42(14.0%)/258(86.0%)	18(14.3%)/108(85.7%)	0.006	0.938
Ultrasound margin features (regular/irregular)	130(43.3%)/170(56.7%)	66(52.4%)/60(47.6%)	2.924	0.087
Ultrasound calcification type (present/absent)	143(47.7%)/157(52.3%)	60(47.6%)/66(52.4%)	0.001	0.992
Ultrasound echogenicity (high/medium/low)	35(11.7%)/98(32.7%)/167(55.7%)	15(11.9)/41(32.5%)/70(55.6%)	0.328	0.849
Ultrasound taller-than-wide shape (≥ 1/<1)	125(41.7%)/175(58.3%)	52(41.3)/74(58.7%)	0.005	0.939
Lymph node size (mm)	8.23 ± 3.15	7.98 ± 2.87	0.767	0.443
Lymph node morphology (regular/irregular)	201(67.0%)/99(33.0%)	84(66.7%)/42(33.3%)	0.004	0.946
Hilum status (preserved/obscured/absent)	185(61.7%)/72(24.0%)/43(14.3%)	78(62.0%)/30(23.8%)/18(14.3%)	0.412	0.814
Lymph node internal echotexture (homogeneous/heterogeneous)	168(56.0%)/132(44.0%)	70(55.6%)/56(44.4%)	0.007	0.932
Lymph node vascularity (high/moderate/low)	62(20.7%)/145(48.3%)/93(31.0%)	26(20.6%)/61(48.4%)/39(30.9%)	0.285	0.867
TSH level (mIU/L)	2.35 ± 1.08	2.28 ± 1.01	0.622	0.534
Central compartment lymph node metastasis (yes/no)	78(26.0%)/222(74.0%)	33(26.2%)/93(73.8%)	0.001	0.967
Lateral cervical lymph node metastasis (yes/no)	45(15.0%)/255(85.0%)	19(15.1%)/107(84.9%)	0.001	0.984
Tumor 3D volume (mm³)	128.65 ± 45.32	123.98 ± 42.17	0.990	0.322
Tumor surface area (mm²)	89.42 ± 28.67	86.75 ± 26.93	0.893	0.372
Sphericity	0.68 ± 0.15	0.66 ± 0.14	1.281	0.201
Ultrasound elasticity score	2.85 ± 1.03	2.79 ± 0.98	0.557	0.578
Short-run emphasis	15.32 ± 4.87	14.98 ± 4.62	0.668	0.505
Long-run emphasis	9.65 ± 3.21	9.51 ± 3.19	0.412	0.681
Regional size non-uniformity	0.38 ± 0.12	0.37 ± 0.11	0.804	0.422

Univariate analysis of factors influencing lymph node metastasis in thyroid microcarcinoma

Univariate analysis revealed statistically significant differences (P < 0.05) between the metastatic (n = 80) and non-metastatic (n = 220) groups in the training cohort regarding tumor size, lymph node size, TSH level, central lymph node metastasis, 3D tumor volume, and sphericity index (Table 2).

Table 2.

Univariate analysis of factors associated with lymph node metastasis in thyroid microcarcinoma

Variables	Metastatic group (n = 80)	Non-metastatic group (n = 220)	t/χ²	P
Age (years)	46.20 ± 9.38	45.01 ± 9.02	0.999	0.318
Gender (male/female)	32(40.0%)/48(60.0%)	80(36.4%)/148(63.6%)	0.617	0.432
Tumor size (mm)	7.85 ± 2.15	6.35 ± 2.11	5.417	0.001
Multifocality (yes/no)	29(36.3%)/51(63.7%)	60(27.3%)/160(72.7%)	1.989	0.158
Capsular invasion (yes/no)	20(25.0%)/60(75.0%)	45(20.5%)/175(79.5%)	0.714	0.398
Extrathyroidal extension (yes/no)	15(18.8%)/65(81.2%)	27(12.3%)/193(87.7%)	2.044	0.153
Ultrasound margin features (regular/irregular)	38(47.5%)/42(52.5%)	100(45.5%)/120(54.5%)	0.098	0.753
Ultrasound calcification type (present/absent)	33(41.3%)/47(58.7%)	110(50.0%)/110(50.0%)	1.801	0.179
Ultrasound echogenicity (high/medium/low)	11(13.8%)/28(35.0%)/41(51.2%)	24(10.9%)/70(31.8%)/126(57.3%)	0.402	0.621
Ultrasound taller-than-wide shape (≥ 1/<1)	35(43.8%)/45(56.2%)	90(40.9%)/130(59.1%)	0.195	0.659
Lymph node size (mm)	9.21 ± 3.36	8.01 ± 3.02	2.952	0.003
Lymph node morphology (regular/irregular)	50(62.5%)/30(37.5%)	151(68.6%)/79(31.4%)	0.260	0.611
Hilum status (preserved/obscured/absent)	45(56.2%)/22(27.5%)/13(16.3%)	140(63.6%)/50(22.7%)/30(13.6%)	0.502	0.322
Lymph node internal echotexture (homogeneous/heterogeneous)	50(62.5%)/30(37.5%)	118(53.6%)/102(46.4%)	1.871	0.171
Lymph node vascularity (high/moderate/low)	22(27.5%)/35(43.8%)/23(28.7%)	40(18.2%)/110(50.0%)/70(31.8%)	2.301	0.851
TSH level (mIU/L)	3.25 ± 1.08	2.01 ± 1.01	3.587	0.001
Central compartment lymph node metastasis (yes/no)	30(37.5%)/50(62.5%)	48(21.8%)/172(78.2%)	7.499	0.006
Lateral cervical lymph node metastasis (yes/no)	15(18.8%)/65(81.2%)	30(13.6%)/190(86.4%)	1.203	0.273
Tumor 3D volume (mm³)	148.65 ± 45.02	125.65 ± 44.32	4.819	0.001
Tumor surface area (mm²)	89.52 ± 28.67	89.12 ± 28.67	0.106	0.915
Sphericity	0.54 ± 0.15	0.69 ± 0.16	7.299	0.001
Ultrasound elasticity score	2.86 ± 1.03	2.84 ± 1.02	0.149	0.881
Short-run emphasis	15.52 ± 4.87	15.31 ± 4.83	0.332	0.739
Long-run emphasis	9.68 ± 3.21	9.61 ± 3.16	0.169	0.865
Regional size non-uniformity	0.39 ± 0.13	0.38 ± 0.12	0.624	0.533

Multivariate logistic regression analysis of lymph node metastasis in thyroid microcarcinoma

Lymph node metastasis status (non-metastatic = 0, metastatic = 1) was set as the dependent variable. Six statistically significant variables from univariate analysis were included. LASSO regression was first applied for feature selection (variable assignments in Table 3), followed by multivariate logistic regression (Fig. 1). Results identified tumor size, lymph node size, TSH level, central lymph node metastasis (yes/no), and 3D tumor volume as independent risk factors, while sphericity was a protective factor (P < 0.05) (Table 4).

Table 3.

Variable assignments

Variable	Meaning	Assignment
X1	Tumor size	Continuous
X2	Lymph node size	Continuous
X3	TSH level	Continuous
X4	Central lymph node metastasis	Categorical (no = 0/yes = 1)
X5	3D tumor volume	Continuous
X6	Sphericity	Continuous
Y	Metastasis status	Non-metastatic = 0, Metastatic = 1

Fig. 1.

LASSO feature selection

Table 4.

Multivariate logistic regression analysis of influencing factors

Factor	β	SE	Wald	P	OR	95%CI
Tumor size	0.449	0.095	22.106	0.001	1.566	1.299–1.889
Lymph node size	0.119	0.059	4.053	0.044	1.126	1.003–1.264
TSH level	1.404	0.227	38.094	0.001	4.071	2.607–6.358
Central lymph node metastasis	1.281	0.420	9.291	0.002	3.599	1.508–8.201
3D tumor volume	0.020	0.005	14.876	0.001	1.021	1.010–1.031
Sphericity	-6.485	1.241	27.308	0.001	0.002	0.001–0.017
Constant	-8.182	1.657	24.383	0.001	0.001

Performance evaluation of machine learning models

Key predictors from multivariate analysis were used to construct machine learning models. All models demonstrated robust performance in training and validation sets (Fig. 2, ROC curves). The RF model achieved the highest AUC (0.815) in the validation set, with a sensitivity of 74.3% and specificity of 79.1%, outperforming the KNN and GB models across all evaluated metrics; detailed performance indicators including accuracy, precision, recall, F1 score, positive predictive value (PPV), and negative predictive value (NPV) are presented in Table 5. Figure 3 (calibration curves) confirmed high agreement between predicted probabilities and observed risks.

Fig. 2.

ROC curves (A: training set. B: validation set)

Table 5.

Comprehensive performance metrics of machine learning models in training and internal validation sets

Model	Dataset	Accuracy	Sensitivity (Recall)	Specificity	PPV	NPV	F1
RF	Training	0.813	0.788	0.823	0.650	0.891	0.712
	Validation	0.778	0.743	0.791	0.583	0.872	0.654
KNN	Training	0.777	0.750	0.786	0.591	0.864	0.661
	Validation	0.746	0.714	0.756	0.526	0.849	0.606
GB	Training	0.737	0.700	0.750	0.538	0.836	0.608
	Validation	0.714	0.686	0.721	0.482	0.826	0.566

Fig. 3.

Calibration curves (A: training set. B: validation set)

Decision curve analysis (Fig. 4) showed the RF model’s net benefit surpassed “None” and “All” thresholds across clinically relevant risk ranges, indicating superior clinical utility.In summary, the RF model exhibited optimal discriminative power, calibration, and net benefit for predicting lymph node metastasis in thyroid microcarcinoma. In the validation set, the RF model showed the optimal clinical performance with a sensitivity of 74.29%, a specificity of 79.07%, a PPV of 58.33%, a NPV of 87.23% and an AUC (95%CI) of 0.815 (0.741–0.926), which were all higher than those of the KNN model (71.43%, 75.58%, 52.63%, 84.88%, 0.792 (0.651–0.867)) and the GB model (68.57%, 72.09%, 48.15%, 82.56%, 0.763 (0.676–0.886)). A similar trend was observed in the training set, where the RF model achieved a sensitivity of 78.75%, a specificity of 82.27%, a PPV of 65.00%, a NPV of 89.09% and an AUC (95%CI) of 0.838 (0.779–0.897), outperforming the KNN model (75.00%, 78.64%, 59.09%, 86.36%, 0.815 (0.754–0.876)) and the GB model (70.00%, 75.00%, 53.85%, 83.64%, 0.787 (0.719–0.854)). These results indicated that the RF model has a good ability to identify metastatic and non-metastatic patients in clinical practice. We further quantified the calibration accuracy and clinical net benefit of each model. The Brier scores of the RF model were 0.185 in the training set and 0.203 in the validation set, which were lower than those of the KNN model (0.212 in the training set, 0.235 in the validation set) and the GB model (0.238 in the training set, 0.257 in the validation set), indicating the optimal calibration performance of the RF model. Decision curve analysis (DCA) showed that the net benefit of the RF model was consistently higher than that of the KNN and GB models in the clinical risk threshold range of 0.1–0.8; the maximum net benefit of the RF model in the validation set was 0.27 at the risk threshold of 0.3, which was 17.4% and 28.6% higher than that of the KNN (0.23) and GB (0.21) models, respectively.

Fig. 4.

Decision curves (A: training set. B: validation set)

Interpretability of model predictions

Six core features (X1–X6) were analyzed via logistic regression and correlation analyses (Fig. 5 A–B). TSH level (X3), tumor size (X1), and sphericity (X6) were top contributors. Elevated TSH and tumor size increased metastasis risk, while higher sphericity was protective. (A) Correlation trends: Tumor size (X1), lymph node size (X2), TSH (X3), central metastasis (X4), and 3D volume (X5) were risk factors. sphericity (X6) was protective. (B) Dominant predictors: TSH (X3) had the strongest positive impact on metastasis prediction, followed by tumor size (X1) and reduced sphericity(X6).

Discussion

Accurate prediction of lymph node metastasis (LNM) in thyroid microcarcinoma (TMC) is critical for surgical decision-making (e.g., prophylactic lymph node dissection) and prognosis assessment. The integration of multidimensional indicators to construct high-performance predictive models is essential for reducing overtreatment or missed diagnoses, thereby improving patient outcomes [8, 9]. In this study, clinical, ultrasonographic, and radiomics data from 426 TMC patients (training set: n = 300, validation set: n = 126) were analyzed. Six differential indicators were initially screened via univariate analysis, followed by LASSO regression for dimensionality reduction. Multivariate logistic regression identified tumor size, lymph node size, TSH level, central compartment LNM, and tumor 3D volume as independent risk factors for LNM, while sphericity was a protective factor. The variance inflation factor (VIF) for all variables was < 2, indicating low multicollinearity and confirming model stability. A RF model demonstrated optimal performance (AUC = 0.838 in the training set, AUC = 0.815 in the validation set).

Notably, TSH level, tumor size, and sphericity ranked highest in feature importance, suggesting that thyroid function indices, morphological characteristics, and radiomics parameters are pivotal predictors of LNM in TMC, providing a practical basis for individualized surgical planning.

In this study, TSH level emerged as the strongest independent risk factor (OR = 4.071) for LNM, likely mediated using the “TSH-thyroid cancer proliferation-invasion/metastasis” axis. As a key regulator of thyroid cell growth, elevated TSH activates TSH receptors on cancer cells, accelerating the G1/S phase transition and enhancing proliferation [10, 11]. Concurrently, TSH upregulates matrix metalloproteinases (MMP-2, MMP-9), degrading the extracellular matrix and basement membrane to facilitate invasion [12–14]. Additionally, high TSH may promote VEGF secretion, increasing microvessel density and accelerating lymphatic metastasis [15]. The significantly higher TSH in the metastasis group (3.25 ± 1.08 mIU/L vs. 2.01 ± 1.01 mIU/L, P < 0.05) further supports this mechanistic link.

Tumor size (OR = 1.566) and 3D volume (OR = 1.021) directly quantify tumor burden, aligning with the clinical axiom of “tumor burden-metastatic potential” [16].

Compared to unidimensional size, 3D volume better reflects true growth status. larger volumes correlate with increased cell proliferation, capsular penetration, and lymphatic invasion [17]. Studies confirm a 2–3-fold higher LNM rate for tumors > 8 mm or volumes > 150 mm³, consistent with our findings (metastasis group: 7.85 ± 2.15 mm, 148.65 ± 45.02 mm³ vs. non-metastasis group, P < 0.05). Tumor expansion may also compress local lymphatic networks, altering drainage patterns and elevating metastasis risk.

Sphericity, the sole protective factor (OR = 0.002), reflects the association between tumor morphology and aggressiveness. Low sphericity (< 0.6) indicates irregular growth along tissue planes or vascular bundles, often correlating with higher invasiveness [18]. Mechanistically, non-spherical tumors have larger contact areas for “contact-driven invasion” and may harbor metastatic subclones [19, 20]. The metastasis group exhibited significantly lower sphericity (0.54 ± 0.15 vs. 0.69 ± 0.16, P < 0.05), underscoring its utility in preoperative risk stratification.

Lymph node size (OR = 1.126) and central LNM (OR = 3.599) underscore the “local lymph node status–overall metastasis risk” relationship. A size ≥ 9 mm is a conventional cutoff for suspected metastasis. our metastasis group showed larger nodes (9.21 ± 3.36 mm, P < 0.05), pathologically explained by cancer infiltration [21].

Central LNM, the most frequent metastatic site in TMC, signals lymphatic dissemination and predicts lateral neck involvement (35–40% incidence) [22, 23], justifying its high OR value. Lymph node size was included as a continuous predictive index in the model, and although lymph node enlargement is a clinically suspected sign of metastasis, it is not an absolute diagnostic criterion—lymph node micrometastasis can occur in normal-sized lymph nodes, and reactive hyperplasia can lead to non-metastatic lymph node enlargement [24]. The quantitative inclusion of lymph node size can effectively reflect the morphological changes of lymph nodes, and its combination with other preoperative indicators (e.g., TSH level, tumor 3D volume) can avoid the one-sidedness of a single morphological sign and improve the overall predictive accuracy of the model [25]. It is worth noting that lymph node size is not fully correlated with lymph node metastasis status, and the model does not rely on this single indicator for prediction but integrates multi-dimensional features for comprehensive judgment [26].

The RF model achieved excellent performance (AUC > 0.8) by integrating multidimensional data. Unlike logistic regression, it handles non-linear interactions (e.g., TSH-tumor volume synergy) and mitigates measurement errors (e.g., ultrasonographic variability). Feature importance scores highlighted TSH and tumor size as key indicators, aiding clinical monitoring. The model’s clinical utility lies in its noninvasive risk assessment (combining ultrasound and TSH data), guiding prophylactic dissection for high-risk patients (predicted probability > 50%) while sparing low-risk patients unnecessary interventions. The RF model showed superior performance to the KNN and GB models in this study, and the reasons are mainly as follows: First, the RF model is an integrated learning model based on multiple decision trees, which can effectively capture the non-linear interaction between multi-dimensional features (e.g., the synergistic effect of TSH level and tumor 3D volume on metastasis risk), while the KNN model is limited by its distance-based judgment logic and is difficult to fit the complex non-linear relationship in the data.

Second, the RF model uses bootstrap sampling and feature random selection, which has strong anti-overfitting ability and good generalization performance for high-dimensional radiomic data; the GB model is prone to overfitting when the learning rate is not properly adjusted, and the model performance decreases in the validation set. Third, the RF model is less sensitive to outliers and missing values in the retrospective clinical data, while the KNN model is highly sensitive to data outliers due to its reliance on distance calculation, which affects the model’s stability.

In summary, the RF model has better adaptability to the multi-dimensional, non-linear and heterogeneous characteristics of preoperative clinical, ultrasonographic and radiomic data of TMC patients, making it the optimal predictive model in this study.

This study has several limitations that should be acknowledged. First, this study adopted a single-center retrospective design, and all the research population were from medical institutions in the Qingdao area. Multifocality was assessed by preoperative ultrasound, which may be inconsistent with postoperative pathological results, leading to potential bias. The relatively homogeneous demographic characteristics, iodine nutritional status, and clinical practice norms of the regional population may lead to inherent selection bias, which makes the model more suitable for local TMC patients and greatly limits the external generalizability of the model to other regions or medical centers. A test set in multi-center, prospective cohorts is necessary to confirm the model’s robustness. Second, the definition of central lymph node metastasis was based on preoperative ultrasonography, which, while practical and reflective of real-world clinical decision-making, has inherent sensitivity and specificity limitations compared to the final histopathological gold standard.

Ultrasound has limited sensitivity and specificity for detecting lymph node micrometastases (≤ 2 mm), which may lead to misclassification of partial lymph node status. In addition, the cohort may include more patients with typical malignant lymph nodes, leading to artificially inflated model performance in atypical or early metastatic cases, which limits clinical generalizability. Subsequent studies can combine preoperative minimally invasive examination methods (e.g., ultrasound-guided fine-needle aspiration) to improve the accuracy of preoperative lymph node status evaluation. Third, the impact of TSH suppression therapy on metastasis risk was not investigated, and we also did not include iodine nutritional status, thyroid autoantibodies (e.g., TgAb, TPOAb), or specific genetic markers (e.g., BRAF V600E) closely related to the invasive and metastatic potential of TMC. In addition, the potential confounding effects caused by differences in the technical levels of surgeons (in clinical index collection) and ultrasound examiners (in imaging feature interpretation) were not controlled. These omitted factors may reduce the comprehensiveness and prediction accuracy of the model. Subsequent studies can integrate the above molecular and nutritional indicators and unify the collection and interpretation standards of clinical and imaging indicators through standardized protocols to eliminate human factors and further optimize model performance. Fourth, our follow-up period was insufficient to evaluate the model’s prognostic value for long-term outcomes such as disease recurrence. The research only completed the construction and internal validation of the preoperative predictive model for TMC-LNM, without conducting long-term follow-up and outcome analysis of the included patients. Thus, we cannot assess the predictive value of the model for long-term clinical outcomes such as postoperative tumor recurrence, disease-free survival, and overall survival, which greatly limits the long-term clinical significance and application value of the model. In subsequent research, we will conduct prospective long-term follow-up of the included patients, track their long-term prognosis, and further verify the prognostic value of the model. Fifth, this study only included preoperative accessible clinical, ultrasonographic, and radiomic features, and did not incorporate postoperative histological characteristics such as lymphatic invasion, surgical margins, and aggressive variants of PTC, which may have certain impacts on the comprehensiveness of the predictive model. Subsequent studies can integrate preoperative and postoperative multi-dimensional indicators to further optimize model performance. Sixth, this study only conducted internal validation by randomly dividing the same single-center dataset into a training set and a test set, and did not perform independent multi-center external validation. Moreover, the study did not set a control group of traditional clinical judgment (i.e., LNM risk assessment of TMC patients by experienced endocrinologists and thyroid surgeons based on conventional clinical and ultrasonographic features), and failed to compare the model’s predictive performance with traditional clinical judgment methods, making it difficult to quantify the actual improvement of the model in clinical decision-making efficiency and accuracy. Subsequent research will carry out multi-center prospective test set of the model and establish traditional clinical judgment control groups to comprehensively evaluate the stability, generalizability, and clinical utility of the model, thereby promoting its translation and application in clinical practice.

Conclusion

Tumor size, lymph node size, TSH level, central LNM, 3D volume, and sphericity are core predictors of LNM in TMC. The RF model provides a robust tool for precision staging and personalized treatment, optimizing clinical decisions through preoperative multidimensional assessment. A web-based risk calculator based on the RF model will be developed for clinical bedside use, supporting precise risk stratification and individualized surgical planning.

Authors' contributions

Conception and design: YFY and CCH. Method: ALY. Data Collection: YFY and CCH. Manuscript Writing: YFY and CCH. Manuscript revision: YFY and CCH. Research supervision: YFY and CCH. All authors contributed to the article and approved the submitted version.

Funding

None.

Data availability

All data that support the findings of this study are available from the corresponding authors upon reasonable request.

Declarations

Ethics approval and consent to participate

The study was approved by the Committees for the Ethical Review of Research involving Human Subjects from Qilu Hospital (Qingdao) of Shandong University University (Approval No. QLYY-00437), and informed consent was obtained from all patients. This study was conducted in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable, as the manuscript does not contain any individual person’s data in any form.

Competing interests

The authors declare no competing interests.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.