All ultrasound examinations were performed by board-certified pediatric radiologists using high-frequency linear transducers (7–15 MHz) on contemporary scanners. Standardized planes (longitudinal and transverse) were obtained. For color Doppler, the pulse repetition frequency was set at a low range (≈0.8–1.2 kHz) with a low wall filter to detect slow flow, and gain was optimized to avoid blooming artifacts. The following features were prospectively or retrospectively recorded: nodal size (short and long axes), short-to-long axis ratio, cortical thickness (pathological if >3 mm), echotexture (homogeneous vs. heterogeneous), margin regularity, presence or absence of an echogenic fatty hilum, and vascular pattern (hilar vs. peripheral/mixed). Nodes were categorized as reactive, suspicious, or malignancy-suspected based on the pre-specified criteria detailed above. Images were reviewed by two pediatric radiologists in consensus, and any disagreements were resolved by joint reading. Elastography was not available during the study period and therefore was not performed.

Classification as suspicious or malignancy-suspected on ultrasonography was based on the presence of one or more abnormal features, including loss of the fatty hilum, rounded shape (short-to-long axis ratio >0.5), irregular margins, heterogeneous echotexture, cortical thickening (>3 mm), and/or abnormal hilar or peripheral vascularity on Doppler.

Biopsy was considered for persistent nodes (>6–8 weeks), mediastinal location, supraclavicular location (if available), B symptoms (fever, weight loss, night sweats), rapidly enlarging nodes, or suspicious US features as defined above.

Lymph node biopsy was performed when indicated, and the pathological results were categorized as reactive, granulomatous, or malignant. Additionally, bone marrow flow cytometry results were reviewed. The final diagnosis was determined as reactive lymphadenopathy, infectious lymphadenopathy, hematological malignancy (leukemia/lymphoma), or solid tumor metastasis. Follow-up duration was recorded as 0–6 months, 6–12 months, or over 12 months, and the patient's final status was classified as follow-up, complete recovery, or disease progression.

Numbers of procedures were specified as follows: fine-needle aspiration cytology (FNAC) was performed in 12 cases, excisional/incisional histopathology in 47 cases, and immunohistochemistry (IHC) in 28 cases; immunocytochemistry was not performed. Representative histopathology images are presented in Figure 1.

The relationship between lymph node size, the presence of multiple lymph nodes, lymph node location, and the final diagnosis was analyzed using statistical methods. Parametric and non-parametric tests were applied where appropriate. Data analysis was conducted using SPSS version 26.0 (IBM Corp., Armonk, NY, USA) or similar statistical software.

Categorical variables were presented as counts and percentages, while continuous variables were expressed as mean ± standard deviation or median (range), depending on distribution. The normality of distribution was assessed using the Kolmogorov–Smirnov test. For comparisons, the Student's t-test was applied to normally distributed continuous variables, and the Mann–Whitney U test was used for non-normally distributed variables. The chi-square test was used to compare categorical variables, and the Kruskal–Wallis test was employed for comparisons among three or more independent groups when data were not normally distributed. Correlation analysis was conducted using Pearson's test for normally distributed variables and Spearman's test for non-normally distributed variables.

The statistical significance level was set at two-sided p ≤ 0.05. In addition, multivariable binary logistic regression analysis was performed to determine independent predictors of malignancy, and results were reported as adjusted odds ratios (OR) with 95% confidence intervals (CI). Where applicable, relative risks (RR) with 95% CI were calculated for two-by-two comparisons. Model calibration and discrimination were evaluated using the Hosmer–Lemeshow goodness-of-fit test and the area under the receiver operating characteristic (ROC) curve, with the area under the curve (AUC) reported.

The binary outcome (malignancy) was coded 1 = present and 0 = absent; reference categories were reactive US (US groups), cervical region (location), size <2 cm (size), and single node (count).

The a priori sample size calculation indicated that approximately 288 patients in total (144 in each comparison group) would have been sufficient to detect the observed difference in malignancy rates with 80% power at a significance level of α = 0.05. However, due to the retrospective design of the study, all eligible patients during the study period (n = 500) were included, thereby further enhancing statistical power. This ensured that the study was adequately powered for detecting the observed associations.

In the multivariable logistic regression analysis, ultrasonographic findings were the only independent predictor of malignancy. Nodes categorized as suspicious on ultrasound were associated with ∼56-fold higher odds of malignancy (adjusted OR ≈ 55.6; 95% CI: 14.3–200; p < 0.001), whereas lymph node size and anatomical region did not retain statistical significance (p > 0.05). Representative ultrasonographic appearances of reactive, suspicious, and high-suspicion nodes are shown (Figure 3). The overall model was statistically significant (Omnibus χ² = 133.9, df = 4, p < 0.001), explained ∼67% of the variance (Nagelkerke R² = 0.670), and showed good calibration (Hosmer–Lemeshow χ² = 1.96, df = 2, p = 0.376). Classification accuracy was 97.4%, with very high specificity (99.6%) but limited sensitivity (60.7%) (Table 4). Estimates for the “malignancy-suspected” stratum were unstable due to sparse data (quasi-complete separation). Additional binary analysis combining suspicious and malignancy-suspected categories is provided in Supplementary Table 1, confirming the robustness of the findings.

Reactive lymphadenopathy was most frequently observed in the cervical region (93.6%, n = 409). Fewer cases were detected in the axillary (1.4%, n = 6), abdominal (2.5%, n = 11), and inguinal (2.5%, n = 11) regions; none were identified in the mediastinal region. Infectious lymphadenopathy was also predominantly cervical (91.2%, n = 31). Hematologic malignancies were most commonly cervical (81.8%, n = 18), with additional cases in abdominal (9.1%, n = 2) and mediastinal (4.5%, n = 1) regions. Solid tumor metastases were identified in cervical (42.9%, n = 3) and abdominal (42.9%, n = 3) regions, with one mediastinal case (14.3%). Overall differences by region were significant (p < 0.001) (Table 5).

Reactive lymphadenopathy occurred more often with nodes <2 cm (46.9%, n = 205), though 53.1% (n = 232) of reactive cases had nodes ≥2 cm. Infectious lymphadenopathy clustered in the ≥2 cm group (76.5%, n = 26). Hematologic malignancies were strongly linked to ≥2 cm nodes (86.4%, n = 19), as were solid tumor metastases (85.7%, n = 6). The association between size category and final diagnosis was significant (p < 0.001) (Table 6).

There was a significant difference between single and multiple node involvement (p = 0.001). Multiple nodes predominated among hematologic malignancies (95.5%, 21/22) and infectious lymphadenopathy (94.1%, 32/34), whereas solitary nodes were relatively more frequent in solid tumor metastasis (42.9%, 3/7) (Table 7).

Final diagnoses differed by biopsy status (p < 0.001). Biopsy was performed in 47 patients (9.4%). Reactive lymphadenopathy was predominantly diagnosed without biopsy (96.6%, 422/437), as was infectious lymphadenopathy (82.4%, 28/34). All hematologic malignancies underwent biopsy (100%, 22/22). Among solid tumor metastases, 57.1% (4/7) had biopsy (Table 8).

Pre-biopsy classification integrated clinical, US, and laboratory findings into three categories: reactive, suspicious, and malignancy-suspected. Reactive nodes typically showed smooth borders, homogeneous echotexture, preserved fatty hilum, and normal Doppler vascularity. Suspicious nodes exhibited one or more concerning features (rounded shape, irregular margins, heterogeneity, cortical thickening >3 mm, abnormal vascularity). Malignancy-suspected nodes combined multiple abnormal US features and/or clinical red flags (rapid growth, B symptoms) and were prioritized for histopathology.

Definitive diagnoses differed significantly across categories (p < 0.001). The reactive category was largely benign (reactive 93.0%; infectious 6.1%; hematologic malignancy 0.7%). The suspicious group showed higher proportions of infectious (1.3%) and hematologic malignancy (1.8%), with some solid metastasis (0.9%). The malignancy-suspected group was predominantly malignant (hematologic 2.4%; solid metastasis 0.4%) (Table 9).

In the multivariable logistic regression analysis, ultrasonographic findings emerged as the only independent predictor of malignancy. Nodes classified as suspicious on ultrasound were associated with approximately 56-fold higher odds of malignancy (adjusted OR ≈ 55.6; 95% CI: 14.3–200; p < 0.001), whereas lymph node size and anatomical region did not retain statistical significance (p > 0.05). The model was statistically significant (Omnibus χ² = 133.9, df = 4, p < 0.001), explained ∼67% of the variance in malignancy (Nagelkerke R² = 0.670), and showed good calibration (Hosmer–Lemeshow χ² = 1.96, df = 2, p = 0.376). The overall classification accuracy was 97.4%, with very high specificity (99.6%) but limited sensitivity (60.7%) (Table 4).

We then evaluated model discrimination. ROC analysis demonstrated that lymph node size, enlarged lymph node region, and number of lymph nodes had limited discriminatory power for malignancy (AUC = 0.581, 0.545, and 0.520, respectively) (Figure 4). In contrast, the multivariable logistic regression model achieved excellent discrimination, with an AUC of 0.92 (95% CI: 0.88–0.96), indicating strong ability to distinguish malignant from benign lymphadenopathy (Figure 5).

Our findings emphasize the central role of ultrasonography in evaluating childhood lymphadenopathy. Based on these results, we propose a preliminary diagnostic approach where ultrasonographic features serve as the primary determinant for biopsy referral. Specifically, lymph nodes showing suspicious or malignancy-suspected US features should prompt tissue sampling, while isolated clinical factors such as size ≥2 cm or multiplicity, in the absence of abnormal US findings, may be managed conservatively with close follow-up. This approach may help reduce unnecessary invasive procedures while ensuring timely identification of malignant cases. Future multicenter, prospective studies are needed to validate and refine this algorithm for broader clinical use.

One of the main strengths of this study is the large sample size. The actual number of patients included (n = 500) exceeded the minimum required according to a priori power analysis (288 patients), thereby providing adequate statistical power for the primary comparisons and enhancing the robustness of the results. In addition, the study comprehensively evaluated multiple clinical, ultrasonographic, and pathological parameters, and representative histopathological images were included to illustrate the diagnostic spectrum. The use of multivariable logistic regression further strengthened the analysis by identifying independent predictors of malignancy.

Nonetheless, several limitations should be acknowledged. First, the retrospective and single-center design introduces the possibility of referral and selection biases, and the results may not be fully generalizable to different populations. Second, although the sample size was large overall, some subgroups (e.g., solitary lymph node cases) were relatively small, which reduced statistical power for those specific comparisons. In particular, the small number of malignancy-suspected cases led to unstable regression estimates. To address this limitation, we also performed a binary analysis combining suspicious and malignancy-suspected categories, which yielded consistent results (Supplementary Table 1). Third, ultrasound examinations were performed without elastography, which may limit generalizability to centers routinely using this technique. Finally, potential long-term outcomes could not be evaluated due to the retrospective nature of the study.

Taken together, while these limitations should be considered when interpreting the results, the strengths of the study—particularly the large cohort size, systematic evaluation of diagnostic parameters, and robust statistical analysis—enhance the credibility of the findings.