Development and validation of an MRI-based clinical-radiomics nomogram for predicting lymph node metastasis in non-small cell lung cancer

Abstract

Background

To develop and validate a nomogram that integrates clinical factors and multiparametric MRI-based radiomics features for the preoperative prediction of lymph node metastasis (LNM) in non-small cell lung cancer (NSCLC).

Methods

This retrospective diagnostic accuracy study enrolled 220 patients with pathologically confirmed NSCLC (142 males; 60.77 ± 8.70 years) between September 2021 and October 2024. Patients were randomly divided into training and validation sets. A clinical model was constructed using independent predictors identified by univariable and multivariable logistic regression analysis. A radiomics signature was developed from T1WI, T2WI, and T1 mapping sequences using the least absolute shrinkage and selection operator logistic regression algorithm. A nomogram was developed by integrating the clinical model and the radiomics signature. Diagnostic performance was assessed by receiver operating characteristic analysis, calibration, and decision curve analysis. Two radiologists independently assessed LNM status in the validation set for comparison.

Results

Tumor maximum diameter and carcinoembryonic antigen level were identified as independent predictors for the clinical model. In the validation set, the nomogram achieved an area under the curve (AUC) of 0.847, significantly greater than the clinical model (AUC = 0.710, p = 0.033) and the radiomics signature alone (AUC = 0.802, p = 0.033). The AUC of the nomogram was significantly higher than two radiologists (0.847 vs. 0.682, p = 0.022; 0.847 vs. 0.698, p = 0.041, respectively).

Conclusion

The nomogram could serve as a noninvasive tool for preoperative prediction of LNM in NSCLC, thereby aiding in clinical decision-making.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-026-02357-5.

Background

Lung cancer is the most lethal form of cancer globally, accounting for approximately 18.7% of all cancer deaths [1]. Non-small cell lung cancer (NSCLC), which accounts for about 85% of all lung cancer cases, is one of the main histological subtypes [2]. Lymph node metastasis (LNM) status directly influences the NSCLC patients’ TNM staging, the selection of treatment options, and the prognosis. NSCLC patients without LNM have a significantly better prognosis than those with metastasis [3, 4]. Consequently, non-invasively and accurately predicting LNM status is crucial for NSCLC.

Conventional imaging examination methods primarily rely on the size and morphology of lymph nodes to assess metastatic status, presenting certain limitations [5]. Radiomics has been explored for its potential in tumor diagnosis, treatment, and prognosis by extracting vast image features and converting them into high-dimensional, quantitative data. These features can characterize tumor heterogeneity, provide non-invasive biomarkers, and improve the accuracy of diagnosis and prognosis [6]. Currently, CT and PET/CT radiomics have been used in several studies to predict LNM in NSCLC [7, 8]; however, both modalities involve exposure to ionizing radiation. In contrast, as a radiation-free technique, MRI can comprehensively characterize tumors by providing morphological, functional, and metabolic information through its multiparametric sequences. Although the exploration of MRI radiomics is increasing in cancers such as breast and colorectal cancer [9, 10], its application in pulmonary oncology is still emerging, with preliminary investigations primarily dedicated to lung cancer histologic subtyping and the differentiation of benign and malignant nodules [11, 12]. Moreover, few studies have employed MRI radiomics, integrating both conventional and functional sequences, for predicting LNM in NSCLC.

Therefore, this study aimed to develop and validate a nomogram that combines clinical factors and multiparametric MRI radiomics features for the preoperative prediction of LNM in NSCLC.

Methods

Patients

This retrospective study was approved by the institutional review board. Patient informed consent was waived. The patients with suspected lung cancer based on CT were collected at Shandong Provincial Hospital between September 2021 and October 2024. The inclusion criteria were as follows: (1) MRI within one week before surgery; (2) No prior procedures or therapy; (3) Lymph node dissection during surgery; (4) Pathologically confirmed NSCLC and LNM status. Among the 245 patients who met the initial inclusion criteria, 25 patients were excluded for the following reasons: poor image quality (n = 12, 4.9%) and tumors unsuitable for delineation (n = 13, 5.3%).

Ultimately, a total of 220 patients with NSCLC were included in the final analysis. Patients were randomly assigned to a training set (154 cases) and a validation set (66 cases) in a 7:3 ratio. The patient selection process is summarized in Fig. 1.

Fig. 1.

The patient selection process. NSCLC non-small cell lung cancer, LNM lymph node metastasis

The following clinical characteristics were collected: age, sex, tumor maximum diameter, lesion location, smoking history (defined as consuming at least one cigarette daily for a minimum of one year), and carcinoembryonic antigen (CEA) level.

MRI acquisition

The examinations were performed on a Siemens 3.0 T Prisma MRI scanner (MAGNETOM Prisma; Siemens Healthcare, Erlangen, Germany) equipped with an 18-channel phased-array body coil. All patients were scanned using identical and strictly standardized MRI protocols (detailed in Table 1) to minimize potential batch effects or variations induced by scanner parameters. Patients were positioned supine in a head-first orientation. The scanning range covered from the lung apices to the inferior borders. Each patient underwent T1WI, T2WI, and T1 mapping.

Table 1.

Detailed MRI protocols

Sequence	Repetition Time(ms)	Echo Time(ms)	Slice Thickness(mm)	Space Between Slices (mm)	Field of View(mm)	Scanning time (s)
T2WI	2890	94	4.0	4.8	256 × 256	18
T1WI	3.3	1.3	2.5	2.5	320 × 220	13
T1 mapping	5.0	2.3	2.5	2.5	380 × 265	14

Image segmentation

Patient DICOM images were imported into the Deepwise multimodal research platform (https://keyan.deepwise.com, Beijing Deepwise & League of PHD Technology Co., Ltd, Beijing, China). Image segmentation was performed by a radiologist with over ten years of clinical experience, who was blinded to all clinical and pathological data. The radiologist manually delineated the regions of interests (ROIs) layer by layer along the primary tumor boundary, ensuring the resulting 3D volume encompassed the entire tumor while excluding adjacent vascular structures, bronchi, and necrotic areas. To assess the reliability of the radiomic feature selection, a random subset of 30 patients was selected for repeat segmentation by a second clinician with 10 years of experience. The interobserver agreement between the two sets of extracted features was then evaluated using the intraclass correlation coefficient (ICC).

Feature extraction and model establishment

The original images were subjected to filtering transformations using nine methods: wavelet, gradient, local binary pattern (2D and 3D), Laplacian of Gaussian, square root, square, logarithm, and exponential. A total of 1,598 radiomics features were extracted from each sequence (T1WI, T2WI, and T1 mapping). The extracted features included first-order statistics, shape-based features, and texture features derived from the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), gray-level dependence matrix (GLDM), and neighboring gray-tone difference matrix (NGTDM).

Radiomics features were standardized using the Z-score method. In the training set, univariate feature selection was performed with analysis of variance to assess the significance of the association between each feature and the label, retaining those with p < 0.05. To ensure the robustness and stability of the extracted radiomics features (including Haralick texture features) against segmentation variations, the interobserver agreement was evaluated using the ICC in a subset of 30 patients. Only robust features with an ICC > 0.75 were retained for subsequent analysis. Subsequently, to effectively avoid overfitting, features with high collinearity (Spearman’s correlation coefficient r ≥ 0.9) were removed to reduce redundancy. The least absolute shrinkage and selection operator (LASSO) regression algorithm, with the optimal penalty parameter (λ) optimized through 10-fold cross-validation, was then applied for further dimensionality reduction and optimal radiomics signature selection. These signatures were utilized to calculate the radiomics score (Rad-score) and construct the radiomics model.

Development of clinical model and nomogram

A clinical model was built by identifying independent risk factors for LNM through univariable and multivariable logistic regression in the training set. A nomogram incorporating clinical independent risk factors and the radiomic signature has been developed to predict LNM status in NSCLC. The nomogram operates by assigning a points value to each predictive factor, and these points are summed to calculate the total points score, which corresponds to the predicted risk of LNM in NSCLC.

Radiologists’ assessment

Two radiologists (Radiologist 1, P.Z., with 20 years of experience; Radiologist 2, S.Z., with 10 years of experience, both in cardiothoracic imaging), blinded to pathological results, independently evaluated the images in the validation set. In the absence of standardized criteria, they assessed LNM in NSCLC based on their individual experience.

Statistical analysis

SPSS (version 26.0, IBM) and R (version 3.4.3) were employed. All statistical comparisons were two-sided. Comparisons were considered statistically significant at p < 0.05. Univariable logistic regression was first used to screen candidate clinical factors. Those with p < 0.05 were then entered into a multivariable logistic regression model to identify independent risk factors, with results expressed as odds ratios (OR) and 95% confidence intervals (CI). To assess multicollinearity among the independent variables in the multivariable logistic regression model, the variance inflation factor (VIF) was calculated. A VIF value greater than 5 was considered indicative of significant multicollinearity. The diagnostic performance of the predictive models was evaluated using the area under the receiver operating characteristic curve (AUC). The DeLong test was employed to compare the AUC values between different models. Calibration was assessed with the Hosmer-Lemeshow (H-L) test and visualized through calibration curves. To further quantify calibration and evaluate the confidence intervals of the model’s probabilistic predictions, the Brier score was calculated using a bootstrapping method with 1000 resamples. Additionally, when comparing the AUCs between the nomogram and the two radiologists using the DeLong test, the Bonferroni correction was applied to adjust for multiple comparisons. The clinical utility of the models was determined by decision curve analysis (DCA), which calculated the net benefit across various threshold probabilities. The intra-observer agreement in segmentation and radiomics feature extraction was assessed using ICC, with values > 0.75, 0.50–0.75, and < 0.50 indicating excellent, moderate, and poor agreement, respectively. Subgroup analysis was then conducted for the predominant adenocarcinoma cohort.

Results

Patient characteristics

A total of 220 NSCLC patients were included in this study (142 males and 78 females; average age, 60.77 ± 8.70 years), which was stratified into 119 patients without LNM [LNM (-)] and 101 patients with LNM [LNM (+)]. The clinical characteristics of the patients are summarized in Table 2.

Table 2.

Clinical characteristics of patients included in this study

	Training set(n = 154)			p value	Validation set(n = 66)		p value
		LNM(+) (n = 70)	LNM(-)(n = 84)		LNM (+) (n = 31)	LNM (-) (n = 35)
Age(y)		60.81 ± 9.35	59.80 ± 8.24	0.475	61.19 ± 8.40	62.66 ± 8.69	0.490
Sex				0.030			0.081
Male		51 (72.9%)	47 (56.0%)		24 (77.4%)	20 (57.1%)
Female		19 (27.1%)	37 (44.0%)		7 (22.6%)	15 (42.9%)
Tumor maximum diameter	4.280 ± 0.989		2.835 ± 0.759	<0.001	4.181 ± 1.023	2.906 ± 0.872	<0.001
Lesion location				0.406			0.851
Right lung		37 (52.9%)	50 (59.5%)		17 (54.8%)	20 (57.1%)
Left lung		33 (47.1%)	34 (40.5%)		14 (45.2%)	15 (42.9%)
Smoking history				0.007			0.828
Present		28 (40.0%)	52 (61.9%)		16 (51.6%)	19 (54.3%)
Absent		42 (60.0%)	32 (38.1%)		15 (48.4%)	16 (45.7%)
CEA level				<0.001			<0.001
≤ 5ng/ml		24 (34.3%)	61 (72.6%)		10 (32.3%)	29 (82.9%)
> 5ng/ml		46 (65.7%)	23 (27.4%)		21 (67.7%)	6 (17.1%)

Note: Continuous variables are presented as mean ± SD, and categorical variables are expressed as n (%). CEA carcinoembryonic antigen, LNM lymph node metastasis

In the training set, statistically significant differences were observed between LNM (+) and LNM (-) groups in terms of gender, tumor maximum diameter, smoking history, and carcinoembryonic antigen levels (p < 0.05). No statistically significant differences were found for the remaining factors (p > 0.05).

Development of clinical model, radiomics signature, and nomogram

In multivariable logistic regression analysis, tumor maximum diameter (OR = 12.056, 95% CI: 4.507–32.248) and carcinoembryonic antigen level (OR = 4.147, 95% CI: 1.219–14.103) were identified as independent risk factors for predicting LNM in NSCLC (Table 3). The VIF was calculated to evaluate multicollinearity between the independent predictors. The VIF values were 1.163 for tumor maximum diameter and 1.122 for CEA level, indicating no significant multicollinearity. These factors were subsequently incorporated into the construction of a clinical model.

Table 3.

Univariable and multivariable analysis of predictors for lymph node metastasis in non-small cell lung cancer

	Univariable analysis			p value		Multivariable analysis		p value
	OR	OR 95% CI				OR	OR 95% CI
Age(y)	1.014		0.977–1.051	0.472		-	-	-
Sex	2.113		1.070–4.173	0.031		1.629	0.307–8.654	0.567
Tumor maximum diameter	9.475		4.517–19.876	< 0.001		12.056	4.507–32.248	< 0.001
Lesion location	1.312		0.691–2.488	0.406		-	-	-
Smoking history	2.437		1.272–4.669	0.007		2.294	0.539–9.772	0.261
CEA level	5.083		2.554–10.116	< 0.001		4.147	1.219–14.103	0.023
Rad-score	2.953		2.083–4.186	< 0.001		3.005	1.810–4.989	< 0.001

Note: OR odds ratios. CI confidence intervals. CEA carcinoembryonic antigen. Dash (-) indicates not calculated

The radiomic features demonstrated high inter-observer reproducibility, with ICC values exceeding 0.75 for both radiologists, indicating excellent consistency. A radiomics signature was constructed using logistic regression with 14 features selected by the LASSO method. The corresponding Rad-score was calculated as follows:

Rad-score=-0.283 + 0.370*Map_lbp-3D-k_glszm_SmallAreaHighGrayLevelEmphasis + 0.421*t2_log-sigma-1-mm-3D_firstorder_Skewness-0.358*t2_wavelet-HLH_glszm_LowGrayLevelZoneEmphasis + 0.207*t2_exponential_gldm_LargeDependenceHighGrayLevelEmphasis + 0.237*t1_lbp-3D-k_glszm_ZoneEntropy + 0.186*t1_wavelet-LHH_glszm_SmallAreaHighGrayLevelEmphasis + 0.110*Map_lbp-3D-k_glszm_SizeZoneNonUniformityNormalized + 0.282*t2_wavelet-HHL_glcm_Contrast-0.293*t2_square_glrlm_ShortRunLowGrayLevelEmphasis + 0.442*Map_wavelet-LLH_glcm_Correlation + 1.253*t1_wavelet-HLL_glszm_LargeAreaLowGrayLevelEmphasis + 0.434*t2_lbp-3D-k_firstorder_Entropy + 0.269*t2_log-sigma-1-mm-3D_firstorder_90Percentile + 0.062*t1_square_glrlm_LongRunHighGrayLevelEmphasis

In the formula, Map denotes the T1 mapping sequence, while t1 and t2 denote features from T1WI and T2WI, respectively; lbp-3D-k refers to local binary pattern (3D) with kurtosis. glszm is the gray-level size zone matrix; glcm is the gray-level co-occurrence matrix; gldm is the gray-level dependence matrix; glrlm is the gray-level run-length matrix; firstorder denotes first-order statistics features.

A nomogram was constructed by integrating the Rad-score with the independent clinical predictors of LNM in NSCLC patients, providing a visual representation of the contribution of each factor (Fig. 2a).

Fig. 2.

Nomogram developed in training set, which assigns corresponding scores to the maximum diameter, carcinoembryonic antigen level, and Rad-score (a). Calibration curves for the clinical model, radiomics signature, and nomogram in training (b), validation (c) sets. CEA carcinoembryonic antigen level≤5ng/ml(0); >5ng/ml(1)

Comparison of clinical model, radiomics signature, and nomogram

The diagnostic performance of the clinical model, radiomics signature, and nomogram is presented in Table 4. In the training set, the AUC values for the clinical model, radiomics signature, and nomogram were 0.742 (95% CI: 0.665–0.809), 0.895 (95% CI: 0.836–0.939), and 0.922 (95% CI: 0.868–0.959) (Fig. 3a). In the validation set, the AUC values for the clinical model, radiomics signature, and nomogram were 0.710 (95% CI: 0.585–0.815), 0.802 (95% CI: 0.686–0.890), and 0.847 (95% CI: 0.737–0.924) (Fig. 3b). The optimal cut-off value for predicting LNM was determined to be 0.52 in the training set (Youden’s index = 0.654). At this threshold, the nomogram achieved a positive predictive value (PPV) of 77.9% and a negative predictive value (NPV) of 87.0% in the training set. In the validation set, the PPV was 69.2% and the NPV was 85.2%.

Table 4.

Diagnostic performance of clinical model, radiomics signature, and nomogram

	Clinical model (1)	Radiomics signature (2)	Nomogram (3)	p value
				1v2	1v3	2v3
Training set
AUC (95% CI)	0.742(0.665–0.809)	0.895(0.836–0.939)	0.922(0.868–0.959)	<0.001	<0.001	0.039
Accuracy	68.2%	82.3%	82.5%	-	-	-
Sensitivity	72.5%	81.2%	85.3%	-	-	-
Specificity	64.7%	83.5%	80.1%	-	-	-
PPV	62.5%	80.0%	77.9%	-	-	-
NPV	74.3%	84.5%	87.0%	-	-	-
Validation set
AUC (95% CI)	0.710(0.585–0.815)	0.802(0.686–0.890)	0.847(0.737–0.924)	0.121	0.033	0.033
Accuracy	71.2%	74.2%	75.8%	-	-	-
Sensitivity	71.0%	64.5%	87.1%	-	-	-
Specificity	71.4%	82.9%	65.7%	-	-	-
PPV	68.8%	76.9%	69.2%	-	-	-
NPV	73.5%	72.5%	85.2%	-	-	-

Note: AUC the area under the receiver operating characteristic curve. CI confidence intervals. PPV positive predictive value. NPV negative predictive value. Accuracy, sensitivity, specificity, PPV and NPV reported as percentages. Dash (-) indicates not calculated

Fig. 3.

The Receiver operating characteristic curves for differentiation of lymph node metastasis in non-small cell lung cancer. Training set: clinical model, radiomics signature, and nomogram (a). Validation set: clinical model, radiomics signature, and nomogram (b). Validation set: nomogram versus two radiologists (c)

Compared with the clinical model (AUC = 0.742), both the radiomics signature and the nomogram yielded significantly higher AUCs for predicting LNM in the training set. The AUC for the radiomics signature was 0.895 (p < 0.001), and that for the nomogram was 0.922 (p < 0.001). Furthermore, the nomogram exhibited a statistically significant improvement over the radiomics signature alone (0.922 vs. 0.895, p = 0.039). In the validation set, the nomogram continued to demonstrate a superior AUC of 0.847. This was significantly higher than that of the clinical model (AUC = 0.710; p = 0.033) and the radiomics signature alone (AUC = 0.802; p = 0.033).

The calibration curves of the radiomics signature and the nomogram showed good agreement between the mean predicted value and the observed fraction of positives in both the training (Fig. 2b) and validation sets (Fig. 2c). The H-L test indicated good calibration for the nomogram in both the training (p = 0.149) and validation (p = 0.310) sets. This reliable probabilistic calibration was further confirmed by the Brier scores, which were 0.104 (95% CI: 0.078–0.130, evaluated via 1000-iteration bootstrapping) in the training set and 0.139 (95% CI: 0.091–0.190) in the validation set. Decision curve analysis further demonstrated that both the radiomics signature and the nomogram provided a greater overall net benefit than the clinical model for predicting LNM, with the nomogram achieving the highest net benefit (Fig. 4). In the validation set, within the clinically relevant threshold probability range of 10% to 90%, the nomogram achieved net benefit values ranging from 0.258 to 0.411, consistently exceeding those of the clinical model and the radiomics signature alone.

Fig. 4.

Curves from decision curve analysis show net benefit of clinical model, radiomics signature, and nomogram in training (a), validation (b) sets

In the adenocarcinoma subgroup (training set: n = 87; validation set: n = 45), the nomogram achieved an AUC of 0.948 in the training set and 0.867 (95% CI: 0.762–0.972) in the validation set. The H–L test yielded p = 0.441 and the Brier score was 0.138 in the validation set. The ROC curves and calibration curve for this subgroup are presented in Additional file 1.

Incremental value of T1 mapping

To quantify the incremental value of T1 mapping, a baseline model was constructed incorporating clinical factors and radiomics features from T1WI and T2WI sequences. In the validation set, the baseline model achieved an AUC of 0.812 (95% CI: 0.715–0.909). After incorporating T1 mapping features, the full model achieved an AUC of 0.847 (95% CI: 0.737–0.924), with a DeLong p value of 0.482.

The continuous net reclassification improvement (NRI) was 0.948 (95% CI: 0.528–1.367, p < 0.001) and the integrated discrimination improvement (IDI) was 0.099 (95% CI: 0.008–0.189, p = 0.033).

Comparison of nomogram and radiologists in validation set

In the validation set, radiologist 1 demonstrated an accuracy of 68.2%, a sensitivity of 67.7%, and a specificity of 68.6%, with an AUC of 0.682 (95% CI: 0.555–0.791). Similarly, radiologist 2 demonstrated an accuracy of 69.7%, a sensitivity of 71.0%, and a specificity of 68.6%, with an AUC of 0.698 (95% CI: 0.572–0.805). Before adjusting for multiple comparisons, the AUC of the nomogram was significantly higher than those of both radiologist 1 and radiologist 2 (0.682 vs. 0.847, raw p = 0.022; 0.698 vs. 0.847, raw p = 0.041, respectively). After applying the rigorous Bonferroni correction for the two tests, the nomogram’s performance remained significantly superior to radiologist 1 (adjusted p = 0.044) and demonstrated a marginally superior trend compared to radiologist 2 (adjusted p = 0.082). The corresponding ROC curves in the validation set are displayed in Fig. 3c.

Two representative cases from the validation set are shown in Fig. 5 to illustrate the nomogram’s predictive performance.

Fig. 5.

Application of nomogram to predict probability of lymph node metastasis in non-small cell lung cancer patients. (a) A patient with a Rad-score of -3.2 (black solid arrow, score 28.1), tumor maximum diameter = 1.5 cm (blue solid arrow, score 8.2), and carcinoembryonic antigen level > 5 ng/ml (green dotted arrow, score 22.8). The predicted probability of lymph node metastasis was < 0.010 (red solid arrows) (b). Pathological examination confirmed the absence of lymph node metastasis. (c) Another patient with a Rad-score of 0.7 (score 57.2), tumor maximum diameter = 6.0 cm (score 82.8), and carcinoembryonic antigen level ≤ 5 ng/ml (score 0). The predicted probability of lymph node metastasis was > 0.990 (d). Pathological examination confirmed the presence of lymph node metastasis

Discussion

Surgical resection remains the primary treatment for patients with stage I and II NSCLC [13]. However, on the basis of standardized implementation of systematic lymph node sampling, mediastinal lymph node dissection has not significantly improved survival rates in NSCLC patients without LNM [14]. Therefore, accurate preoperative assessment of lymph node staging is of paramount importance for formulating treatment strategies and improving prognosis in NSCLC. In this study, we developed and validated a nomogram that integrated a radiomics signature with key clinical predictors. The nomogram demonstrated good diagnostic performance, with AUCs of 0.922 and 0.847 in the training and validation sets. These findings suggest that the nomogram could serve as a noninvasive and reliable tool for preoperative prediction of LNM in NSCLC, thereby aiding in clinical decision-making and personalized treatment planning.

In this study, both tumor maximum diameter and carcinoembryonic antigen level were identified as independent risk factors for LNM in NSCLC, in agreement with previous reports [15, 16]. However, the clinical model based on these factors demonstrated only modest performance (AUC = 0.710) in the validation set. The two radiologists also achieved modest AUCs of 0.682 and 0.698, respectively. These limitations may be attributed to the inherent constraints of conventional MRI, which is susceptible to respiratory artifacts and limited spatial resolution. Furthermore, clinical indicators primarily provide macroscopic anatomical information, which struggles to reflect tumor internal heterogeneity comprehensively. Therefore, our study incorporated radiomics features to complement the clinical model. The resulting integrated nomogram achieved significantly improved diagnostic performance, with an AUC of 0.847 in the validation set, demonstrating the value of combining clinical and radiomic data for enhanced LNM prediction in NSCLC. Clinically, the high NPV of 85.2% in our validation set provides a crucial theoretical basis for personalized surgical planning. For NSCLC patients preoperatively stratified into the low-risk category by our nomogram, multidisciplinary teams could theoretically consider less invasive nodal evaluation strategies, potentially sparing these patients from the increased morbidity, prolonged operative time, and postoperative complications associated with systematic lymph node dissection. However, this theoretical clinical benefit must be strictly weighed against the risks of false-negative predictions (14.8% in our validation set). Missed occult metastases could lead to postoperative understaging, inadequate adjuvant therapy, and ultimately a detrimental impact on patient survival. Therefore, our nomogram is designed to act as a robust noninvasive auxiliary tool to complement, rather than completely replace, comprehensive clinical judgment and intraoperative pathological assessment.

Radiomics extracts a large number of quantitative features from medical images, providing deeper insights into the tumor’s microstructure and functional characteristics, thereby compensating for the limitations of conventional indicators. Currently, CT and PET/CT are commonly used imaging modalities for preoperative lymph node staging in NSCLC. Several studies have developed radiomics models based on CT and PET/CT to predict lymph node metastasis in NSCLC, demonstrating considerable diagnostic potential [17, 18]. A meta-analysis by Yu et al. [19] confirmed that ¹⁸F-FDG PET/CT and PET/MRI have comparable diagnostic efficacy for nodal staging in NSCLC (sensitivity: 0.82 vs. 0.86; specificity: 0.88 vs. 0.90). Furthermore, a recent advanced study by Duan et al. [20] developed a combined model integrating clinical factors, radiomics, and deep learning features from ¹⁸F-FDG PET/CT to predict LNM. Their combined model achieved an AUC of 0.869 in an external validation set, which is remarkably consistent with the performance of our nomogram (AUC = 0.847). In contrast, MRI offers distinct advantages, including the absence of ionizing radiation and no requirement for iodinated contrast agents, making it particularly suitable for patients with iodine allergies or renal impairment. Previous studies have demonstrated that certain quantitative MRI parameters hold potential value in the preoperative prediction of LNM in NSCLC [21, 22]. However, the application of multiparametric MRI-based radiomics in the context of lung cancer remains relatively limited. Previous studies [23, 24] demonstrated that models integrating features from multiple sequences achieve superior performance compared to those relying on a single sequence. Different MRI sequences reflect distinct tumor characteristics, such as tissue cellular density and microvascular density. Therefore, this study utilized multiparametric MRI radiomics combined with clinical factors to comprehensively characterize tumor profiles and enhance the predictive accuracy for LNM in NSCLC.

In recent years, T1 mapping has been applied to the study of tumor intrinsic characteristics [25, 26], yielding improvements in the assessment of pulmonary lesions. For instance, Li et al. [27] demonstrated that quantitative parameters derived from T1 mapping hold value in distinguishing pathological subtypes of lung cancer, while Yan et al. [28] developed radiomics models derived from T1 mapping to differentiate benign from malignant pulmonary nodules. These advantages may be attributed to the ability of T1 mapping to quantitatively reflect rich tumor information at a microscopic level while producing fewer artifacts, distortions, and positional deviations, thereby enhancing the stability and reproducibility of radiomics features.

Building upon this foundation, our study further corroborates and expands the utility of T1 mapping in pulmonary lesions by incorporating it into the multiparametric MRI radiomics framework for LNM prediction. To rigorously validate its incremental value, we compared a baseline model (incorporating clinical factors and standard T1WI/T2WI radiomics) with our full nomogram. Although the difference in AUCs (0.812 vs. 0.847) did not reach statistical significance by the DeLong test (p = 0.482), evaluations using more sensitive metrics revealed substantial improvements. Specifically, the addition of T1 mapping yielded a highly significant NRI of 0.948 (95% CI: 0.528–1.367, p < 0.001) and an IDI of 0.099 (95% CI: 0.008–0.189, p = 0.033). These metrics confirm that T1 mapping significantly enhances the model’s ability to precisely stratify individual risk. Beyond serving as quantitative imaging biomarkers, these T1 mapping-derived features carry potential biological interpretations, aligning with the fundamental premise of radiomics that medical images encapsulate underlying pathophysiological information [29]. Among the 14 features constituting the Rad-score, those derived from T1 mapping, such as SmallAreaHighGrayLevelEmphasis, may specifically reflect regions of high cellular density or microvascular proliferation within the tumor microenvironment. Such microenvironments are known to facilitate lymph node metastasis by providing routes for tumor cell invasion [30]. Although direct spatial co-registration with histopathology was not performed in this retrospective study, these features serve as valuable noninvasive surrogate biomarkers for tumor invasiveness.

This study has several limitations. First, its single-center retrospective design and relatively small sample size limit the generalizability of our findings. The limited clinical implementation of MRI for lung cancer at other institutions, particularly T1 mapping sequences, hindered the collection of an external validation set. Second, manual ROI segmentation proves time-consuming, whereas automated deep-learning segmentation offers greater efficiency for future large-scale analyses. Third, the radiologists’ assessments were primarily based on their clinical experience rather than strictly standardized morphological criteria, which might introduce subjectivity when compared to the fully quantitative nomogram. Fourth, due to the retrospective nature of this study, we were unable to assess the real-world clinical utility of the nomogram. The theoretical clinical benefits based on the current NPV require rigorous real-world validation. Accordingly, future research will focus on establishing prospective, multicenter collaborations with larger sample sizes, integrating additional functional MRI sequences (such as diffusion-weighted and dynamic contrast-enhanced imaging), and incorporating long-term patient follow-up to definitively evaluate the model’s impact on treatment decision-making and patient survival.

Conclusion

The nomogram integrating clinical and multiparametric MRI radiomics features outperformed the clinical model and two radiologists in predicting LNM in NSCLC patients, providing a scientific basis for clinicians’ treatment decisions.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUC

Area under the curve

CEA

Carcinoembryonic antigen

CI

Confidence intervals

LASSO

Least absolute shrinkage and selection operator

LNM

Lymph node metastasis

NSCLC

Non-small cell lung cancer

OR

Odds ratio

Rad-score

Radiomics score

ROC

Receiver operating characteristic

ROI

Region of interest

Author contributions

XCL conducted the research, performed data acquisition and analysis, and drafted the manuscript. SZ, NC, HH, MT and NL performed statistical analysis of the data and provided material support. XTL conceived and designed the study. PZ conceived and designed the study, and critically reviewed the manuscript for important intellectual content. All authors read and approved the final manuscript.

Funding

This study has received funding from the Natural Science Foundation of Shandong Provincial (No. ZR2024MH018).

Data availability

The data used or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This retrospective study was approved by the Institutional Review Board of Shandong Provincial Hospital (SWYX: NO.2025 − 545) and conducted in accordance with the Declaration of Helsinki. The requirement for written informed consent was waived.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.