Clinical-Radiomics Model Enhancing Prediction of Occult Nodal Metastasis in cT1a-bN0M0-stage Lung Adenocarcinoma: A Multi-center Study

doi:10.21203/rs.3.rs-5025314/v1

Download PDF

Research Article

Clinical-Radiomics Model Enhancing Prediction of Occult Nodal Metastasis in cT1a-bN0M0-stage Lung Adenocarcinoma: A Multi-center Study

https://doi.org/10.21203/rs.3.rs-5025314/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Purpose

To construct radiomics models for predicting occult nodal metastasis (ONM) in cT1a-bN0M0-stage lung adenocarcinoma (LUAD) and evaluate the multi-center diagnostic performance of models.

Methods

1672 patients from six hospitals were collected including training set (n = 687), test set (n = 297) and external validation set (n = 688). Generalized linear model (GLM), support vector machine (SVM), random forest (RF), gradient boosting machine (GBM) and the Clinical-Radiomics (Clinic-Rad) models were constructed and validated to predict ONM. Diagnostic performance was quantified by the area under receiver operative characteristic curve (AUC), and compared using De-Long test. Correlations of radiomics features with pathological characteristics were evaluated by Mantel-test.

Results

Compared to GLM-, SVM-, RF- and GBM-models, the Clinic-Rad model integrating clinical predictors and Radscore received superior diagnostic efficacy in validation set (0.813 ± 0.019 versus 0.790 ± 0.021, 0.761 ± 0.023, 0.708 ± 0.026, 0.769 ± 0.022; all P < 0.001), although no statistical differences in test set (0.834 ± 0.023 versus 0.827 ± 0.024, 0.829 ± 0.025, 0.838 ± 0.023, 0.826 ± 0.024; all P > 0.05). The pooled sensitivity, specificity, accuracy of the Clinic-Rad model was 77.2–75.8%, 72.0–72.7%, 72.7–74.4%. Besides, it was well predictive in solid- and subsolid-appearance LUAD respectively, with pooled AUC values of 0.802–0.820 and 0.797–0.917. Furthermore, radiomics models significantly outperformed clinical predictors including solid-component diameter, consolidation-to-tumor ratio, CEA level and the combined diagnosis (AUC values: versus 0.669–0.678, 0.542–0.600, 0.571–0.613 and 0.683–0.724; all P < 0.001). The Mantel-test demonstrated 88.9%(n = 16/18) of selected radiomics features, Radscore and predicted ONM possibilities were correlated with poorly-differentiated, lymph-vessel invasion, visceral pleura invasion.

Conclusions

Radiomics features are useful to predict ONM in cT1a-bN0M0-stage LUAD and the Clinic-Rad model shows the best diagnostic performance.

Occult nodal metastasis

Lung adenocarcinoma

Radiomics

Prediction

Lung cancer comes the top lethal malignancy worldwide¹. Nodal metastasis (NM) is one of the keys to advanced diseases and uptick mortality. The five-year overall survival of stage-Ia lung cancer is 92%, significantly reducing to 53% and 36% in advanced stage-IIb and -IIIa, respectively². Conventionally, the lobectomy with systematic node dissection is standard treatment for early-stage non-small cell lung cancer (NSCLC); whereas the sub-lobectomy and selective node sampling is optimal for N0 diseases, benefiting from less trauma, more reserved lung function and less postoperative complications. However, the underestimated N stage would lead to a poorer prognosis due to the incomplete excision³.

There was approximately 10–21.7%^4–6 of NM that was unrevealed on preoperative examinations, but instead recognized after radical surgery, which is defined as occult nodal metastasis (ONM). CT scan had limited ability to reveal mediastinal NM by measuring nodal short diameter, with sensitivity and specificity of 55% and 81%⁷. As for PET/CT scan, approximately 7.6–13% and 23–24% of hilar and mediastinal NM in NSCLC were down-staged respectively^8–10. Furthermore, invasive endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA) showed a sensitivity of 35–60%^11,12 and false negative rate (FNR) of 5–30%⁷ in detecting mediastinal NM, whereas video-assisted thoracic surgery (VATS) offered a sensitivity of 73%¹³. Despite great efforts on N stage, it still remains difficult to accurately uncover ONM prior to surgery.

Studies have further evaluated correlations of tumor imaging findings with NM^14,15. Hayashi, et al.,¹⁴ found that solid-component diameter of primary tumor on CT images could predict NM in lung adenocarcinoma (LUAD). He, et al.,¹⁵ showed that morphological characteristics including pleural indentation, vascular convergence sign and preoperative carcinoembryonic antigen (CEA) level were useful to indicate metastasis in NSCLC. In addition, NM-associated pathological risk factors in lung cancer have been reported that presence of poor differentiation, lymph-vessel invasion (LVI), visceral pleura invasion (VPI) more frequently tend to metastasize^16,17. Besides, even small proportion of solid or micro-papillary type would increase chances of NM¹⁸. However, such pathological indicators are unavailable unless invasive examinations or surgery.

Radiomics can decode the heterogeneity of diseases from raw medical images and serves as imaging biomarkers for diagnosing and monitoring diseases¹⁹. Studies have shown that radiomics features could be used for identifying invasiveness, pathological grades in lung cancer^20–22. In addition, in a meta-analysis study²³, six studies using radiomics features showed a relatively high-heterogeneous diagnostic performance for NM prediction in lung cancer, yielding a pooled AUC value of 0.74, and pooled sensitivity and specificity of 0.661 and 0.598 respectively. The variances of models’ performance might be affected by scanner types and scanning parameters²⁴. Besides, most of studies were single-center and small samples^25,26, the diagnostic efficacy might be overestimated in internal validation. Hence, the extrapolation validation is necessary to evaluate the generalization of the model.

In this retrospective study, we investigated several radiomics models to predict ONM in cT1a-bN0M0-stage LUAD, and further validated the multi-center diagnostic performance of the models.

2.1 Patients

This multi-center study in accordance with the Declaration of Helsinki was approved by each hospital’s ethic review board and the informed content was obtained to publish the information in an online open-access publication. All patient information was anonymous. Patients diagnosed as stage cT1a-bN0M0 LUAD on serial thin-section CT scan before surgery within one month were collected from six hospitals in this study. And those with one of following criteria were excluded: a. non-adenocarcinoma, adenocarcinoma in situ (AIS) or minimally-invasive adenocarcinoma (MIA); b. with any tumor treatment history; c. pulmonary multiple malignancy; d. removal LN stations less than six; e. incomplete clinical information. f. unsatisfactory CT images; g. NM diagnosed by contemporaneous PET/CT (if any). The workflow of patients’ enrollment was shown in Fig. 1.

2.2 CT scans

All patients underwent chest serial thin-section CT scans covering the area from lung apex to base. The details of machine manufacture and scanning parameters were presented in Table S1 (seen in Supplement). Contrast-enhanced chest CT scans were performed at delayed 30–35 s following the intravenous administration of 70 mL of iodinated contrast agent at flow rate of 3.5–4.0 mL/s.

2.3 Histopathology

The histopathology was evaluated using hematoxylin-eosin staining (H&E), elastic stains and immunohistochemistry. Pathological characteristics comprised of N stage, poor differentiation, LVI, VPI were defined according to the international association for the study of lung cancer (IASLC)^27,28.

2.4 Imaging processing and Nodular segmentation

CT images were imported from picture archiving and communication systems (PACS) to a commercial software uAI research portal (uRP, Shanghai United Imaging Intelligence, Co., Ltd, China)^29,30. The three-dimensional volume of interest (VOI) on the whole tumor was automatically segmented at lung window (window width = 1500 HU, window level=-600 HU), then was reviewed and modified by two radiologists with over ten-years of experiences. Any discrepancy was determined by the third senior radiologists with over 20-years of experiences. In addition, CT images from 100 patients were randomly selected to be segmented again after three months for assessing the repeatability of radiomics features. CT images and VOIs were normalized by isotropically resampling to a voxel size of 1 × 1 × 1 mm³. Then, radiomics features were extracted from original images and images processing by filters including boxmean, additivegaussiannoise, binomalblurimage, curvatureflow, boxsigmaimage, LoG, wavelet, normalize, laplacianSharpening, discreteGaussian, mean, specklenoise, recursivegaussian, and shortnoise using an open-source software of Pyradiomics (https://pypi.org/project/pyradiomics/). Finally, 2264 radiomics features were obtained from each VOI.

2.5 Features selection and Models construction

To construct models, the model cohort was randomly classified as training and test sets respectively according to a ratio of 7 to 3. Radiomics features were normalized by Z-score. To obtain reproducible radiomics features, intra-class correlation coefficients (ICCs) were calculated for radiomics features from twice segmentations on 100 VOIs and radiomics features with ICC ≥ 0.7 were remained. Then, radiomics features with Pearson’s correlation coefficients ≥ 0.8 were further removed to avoid over-fitting. Next, a Boruta algorithm was used for selecting optimal radiomics features. Radiomics classifiers including generalize linear model (GLM), support vector machine (SVM), random forest (RF), and gradient boosting machine (GBM) were used to construct models for ONM prediction. Furthermore, the Clinic-Rad model was developed integrating the Radscore and selected clinical predictors using the logistic regression algorithm. The Radscore was calculated using the Formula as follows:

Radscore = β₀ + β₁ × F₁ + β₂ × F₂+ … + β_n × F_n. β_n represents the coefficient of feature F_n.

All statistics were analyzed using R studio software (version 4.1.3). The continuous and categorized variables were presented as mean ± standard deviation (SD), and numbers with frequency respectively. The Kruskal Wallis test was used for comparisons of patients’ characteristics among cohorts. Differences of clinical characteristics were compared using univariate logistic regression. And then the clinical predictors were selected using stepwise regression analysis with minimal Akaike information criterion (AIC). The importance of the selected radiomics features was assessed using SHapley additive explanation (SHAP). The Clinic-Rad model was visualized using the nomogram and evaluated by calibration curves. The diagnostic performance was evaluated and quantified by decision curve analysis (DCA) and the area under receiver operative characteristic (ROC) curve (AUC), with accuracy, sensitivity, specificity, positive- and negative-predictive value (PPV and NPV). The best cutoff value was determined by the maximum Youden’s index in training set. The correlations of radiomics features, Radscore and model’s predicted possibilities with pathological characteristics were evaluated using Mantel test. The De-Long test was used for comparing the difference in diagnostic efficacy between models. P value (two-sides) less than 0.05 was statistical significance.

4.1 Clinical characteristics of enrolled patients

As shown in Table 1, a total of 1672 patients (male, n = 725; female, n = 947; age: 60.6 ± 9.7/years) were enrolled in this study, including training set (n = 687), test set (n = 297) and external validation set (n = 688). The ONM incidence of 18.6% (n = 128/688) in validation set was inferior to that of 39.2% (n = 269/687) and 38.4% (n = 114/297) in training and test sets (P < 0.001). There were no statistical differences in sex, age, nodular location, clinical T stage and nodular diameter (all P > 0.05). In validation set, patients had less proportion of adenocarcinoma appearing as solid nodules (68.9% versus 78.9% versus 78.5%; P < 0.001) and preoperative elevated CEA level (6.1% versus 17.5% versus 17.5%; P < 0.001) and smaller solid-component diameter (1.4 ± 0.4/cm versus 1.6 ± 0.4/cm versus 1.5 ± 0.4/cm; P < 0.001), compared to those in training and test sets.

Table 1

Characteristics of enrolled patients
	Training set (n = 687)	Test set (n = 297)	Validation set (n = 688)	P-value
Sex, M	294 (42.8)	144 (48.5)	287 (41.7)	0.134
Age / y, mean (SD)	60.4 (9.6)	60.7 (9.3)	59.4 (10.3)	0.111
Nodular location Central zone Peripheral zone	54 (7.9) 633 (92.1)	17 (5.7) 280 (94.3)	39 (5.7) 649 (94.3)	0.211
Clinical T stage T1a T1b	69 (10.0) 618 (90.0)	31 (10.4) 266 (89.6)	90 (13.1) 598 (86.9)	0.178
Pathological N stage N0 N1 N2	418 (60.8) 97 (14.2) 172 (25.0)	183 (61.6) 44 (14.8) 70 (23.6)	560 (81.4) 64 (9.3) 64 (9.3)	< 0.001
Nodular types Solid nodules Subsolid nodules	541 (78.7) 146 (21.3)	230 (77.4) 67 (22.6)	470 (68.3) 218 (31.7)	< 0.001
Nodular type subgroups Solid nodules LN positives LN negatives Subsolid nodules LN positives LN negatives	226 (41.8) 315 (58.2) 37 (25.3) 109 (74.7)	93 (40.4) 137 (59.6) 21 (31.3) 46 (68.7)	107 (22.8) 363 (77.2) 21 (9.6) 197 (90.4)	< 0.001 < 0.001
Solid-component diameter / cm, mean (SD)	1.6 (0.4)	1.5 (0.4)	1.4 (0.4)	< 0.001
Nodular diameter / cm, mean (SD)	1.7 (0.5)	1.7 (0.5)	1.7 (0.5)	0.078
Preoperative CEA level,µg/L > 5 ≤ 5	120 (17.5) 567 (82.5)	52 (17.5) 245 (82.5)	46 (6.1) 642 (93.3)	< 0.001
Note. SD, standard deviation. CEA, carcinoembryonic antigen. LN, lymph node.

4.2 Clinical predictors of ONM

Compared to the non-ONM disease, LUAD with ONM had lager solid-component diameter (1.7 ± 0.3/cm versus 1.5 ± 0.4/cm, P < 0.001) and CTR value (0.96 ± 0.1 versus 0.89 ± 0.2, P < 0.001) (Table 2). Besides, stratification analysis showed nodules with solid-component diameter > 1.0cm or CTR value > 0.75 indicated increased chances of metastasis. However, the nodular diameter between ONM and non-ONM disease had no difference (1.8 ± 0.4/cm versus 1.7 ± 0.5/cm, P = 0.273). Additionally, ONM disease had more proportion of elevated CEA level than non-ONM disease (30.8% versus 9.2%; P < 0.001).

Table 2

Differences of clinical characteristics between ONM and non-ONM disease
	ONM (n = 263)	non-ONM (n = 424)	P-value
Solid-component diameter /cm, mean (SD)	1.7 (0.3)	1.5 (0.4)	< 0.001
Solid-component diameter subgroups /cm (0, 1.0] (1.0, 1.5] (1.5, 2.0]	11 (4.2) 69 (26.2) 183 (69.6)	57 (13.4) 166 (39.2) 201 (47.4)	reference 0.033 < 0.001
CTR, mean (SD)	0.96 (0.1)	0.89 (0.2)	< 0.001
CTR subgroups (0.75, 1.0] (0.5, 0.75] (0, 0.5]	243 (92.4) 19 (7.2) 1 (0.4)	327 (77.1) 64 (15.1) 33 (7.8)	reference < 0.001 0.002
Nodular diameter /cm, mean (SD)	1.8 (0.4)	1.7 (0.5)	0.273
CEA level, > 5µg/L	81 (30.8)	39 (9.2)	< 0.001
Note. ONM, occult nodal metastasis. CTR, consolidation-to-tumor ratio. SD, standard deviation. CEA, carcinoembryonic antigen.

The stepwise regression analysis indicated CTR value (P = 0.006), solid-component diameter (P < 0.001) and preoperative elevated CEA level (P < 0.001) were optimal predictors of ONM. The combined diagnosis (Cutoff value = 0.35) performed significantly better diagnostic performance with AUC values of 0.683 ± 0.031 and 0.724 ± 0.025 in test and validation sets respectively (Table 3), compared to CTR value (versus 0.542 ± 0.024, 0.600 ± 0.019; both P < 0.001. Cutoff value = 0.80), solid-component diameter (versus 0.678 ± 0.031, 0.669 ± 0.026; P = 0.855, P = 0.004. Cutoff value = 1.55) and CEA level (versus 0.571 ± 0.033, 0.613 ± 0.022; both P < 0.001). The pooled accuracy, sensitivity, specificity, PPV and NPV of the combined diagnosis were 64.6–84.0%, 25.0–75.4%, 57.9–97.5%, 52.8–69.5% and 79.1–85.0% in test and validation sets.

Table 3

Diagnostic performance of clinical predictors to predict ONM
	Cutoff	AUC (SD)	Accuracy%	Sensitivity%	Specificity%	PPV%	NPV%
Test set
Solid-component diameter	1.55 cm	0.678 (0.031)	65.0	71.9	60.7	53.2	77.6
CTR	0.80	0.542 (0.024)	48.5	87.7	24.0	41.8	75.9
CEA level	increased	0.571 (0.033)	64.3	26.3	88.0	57.7	65.7
Combined diagnosis	0.35	0.683 (0.031)	64.6	75.4	57.9	52.8	79.1
Validation set
Solid-component diameter	1.55 cm	0.669 (0.026)	67.9	59.4	69.8	31.0	88.3
CTR	0.80	0.600 (0.019)	40.4	87.5	29.6	22.1	91.2
CEA level	increased	0.613 (0.022)	84.0	25.0	97.5	69.6	85.0
Combined diagnosis	0.35	0.724 (0.025)	84.0	25.0	97.5	69.5	85.0
Note. ONM, occult nodal metastasis. SCD, solid-component diameter. CTR, consolidation-to-tumor ratio. ND, nodular diameter. SD, standard deviation. CEA, carcinoembryonic antigen. AUC, area under the curve. PPV, positive predictive value. NPV, negative predictive value.

Table 4

Diagnostic performance of radiomics models for ONM prediction
	Cutoff	AUC (SD)	Accuracy%	Sensitivity%	Specificity%	PPV%	NPV%
Test set
GLM	0.443	0.827 (0.024)	75.1	67.5	79.8	67.5	76.1
SVM	0.272	0.829 (0.025)	74.7	79.8	71.6	63.6	85.1
RF	0.515	0.838 (0.023)	76.1	63.2	84.2	71.3	78.6
GBM	0.389	0.826 (0.024)	74.1	70.2	76.5	65.0	80.5
Clinic-Rad model	0.383	0.834 (0.023)	74.4	77.2	72.7	63.8	83.6
Validation set
GLM	0.443	0.790 (0.021) ^*	71.2	68.8	71.8	35.8	91.0
SVM	0.272	0.761 (0.023) ^*#	68.5	71.9	67.7	33.7	91.3
RF	0.515	0.708 (0.026) ^*#	72.5	51.6	77.3	34.2	87.5
GBM	0.389	0.769 (0.022) ^*	71.5	64.1	73.2	35.3	89.9
Clinic-Rad model	0.383	0.813 (0.019)	72.7	75.8	72.0	38.2	92.9
Note. ONM, occult nodal metastasis. AUC, area under the curve. SD, standard deviation. PPV, positive predictive value. NPV, negative predictive value. GLM, generalized linear model. SVM, support vector machine. RF, random forest. GBM, gradient boosting machine (GBM). Clinic-Rad model, the model integrating clinical predictors and Radscore. * The AUC value of the Clinic-Rad model is statistically higher than that of GLM model, SVM model, RF model and GBM model. # The AUC value of the GLM model is significantly better than SVM model and RF model, but comparable to GBM model.

4.3 Radiomics models for predicting ONM

18 radiomics features were finally selected for constructing radiomics models (Fig. 2a). Figure 2b and 2c showed the importance of selected radiomics features and the impact of each radiomics feature on the model output respectively. And the Radscore was obtained according to the Formula (seen in the Table S2). The Clinic-Rad model was developed using the Radscore combined with the clinical predictors including solid-component diameter, CTR value and elevated CEA level, as presented in the nomogram (Fig. 3a).

GLM model, SVM model, RF model, GBM model and the Clinic-Rad model all received good performance for identifying ONM, with AUC values of 0.827 ± 0.024, 0.829 ± 0.025, 0.838 ± 0.023, 0.826 ± 0.024 and 0.834 ± 0.023 in test set, and 0.790 ± 0.021, 0.761 ± 0.023, 0.708 ± 0.026, 0.769 ± 0.022 and 0.813 ± 0.019 in validation set separately (Fig. 4a and 4b). Besides, compared to those clinical predictors, GLM model, SVM model, RF model, GBM model and the Clinic-Rad model all had remarkably better diagnostic performance in test and validation sets (all P < 0.001).

In test set, there were no differences in diagnostic efficacy among radiomics models (all P > 0.05). However, in validation set, GLM model obviously outperformed SVM model (P = 0.021), RF model (P < 0.001) and was comparable to GBM model (P = 0.072). In addition, through further combining with clinical predictors, the Clinic-Rad model was significantly superior to GLM model, SVM model, RF model, and GBM model (all P < 0.001). The pooled accuracy, sensitivity, specificity, PPV and NPV of the Clinic-Rad model were 72.7–74.4%, 75.8–77.2%, 72.0–72.7%, 38.2–63.8% and 83.6–92.9%. Furthermore, the calibration curves (Fig. 3b) demonstrated the Clinic-Rad model performed good prediction of ONM in test and validation set separately. DCA curves (Fig. 4b and 4d) showed that GLM model, GBM model and the Clinic-Rad model had larger clinical net benefit than treat none or treat all when occurrence rate of ONM beyond approximately 5%, and the Clinic-Rad model had better net benefit when the predicted-possibilities cutoff value ranges from approximately 5–38% in validation set, compared to GLM model, SVM model, RF model, and GBM model.

4.4 Diagnostic performance of the subgroups

4.4.1 Identifying ONM in LUAD presenting as solid nodules

In test set, GLM model, SVM model, RF model, and GBM model and the Clinic-Rad model had equal diagnostic performance for identifying ONM, yielding AUC values of 0.808 ± 0.028, 0.817 ± 0.029, 0.829 ± 0.028, 0.815 ± 0.029 and 0.820 ± 0.027 (all P > 0.05). However, in validation set, the Clinic-Rad model (AUC value: 0.802 ± 0.022) was better than GLM (versus 0.777 ± 0.024, P = 0.001), SVM model (versus 0.745 ± 0.025, P = 0.002), RF model (versus 0.708 ± 0.028, P < 0.001) and GBM model (versus 0.750 ± 0.026, P = 0.002). The pooled accuracy, sensitivity, specificity, PPV and NPV of the Clinic-Rad model were 70.6–74.3%, 69.9–75.7%, 69.1–77.4%, 42.0–67.7% and 79.1–90.6% (Table 5).

Table 5

Diagnostic performance of radiomics models in LUAD subgroups
	Cutoff	AUC ± SD	Accuracy%	Sensitivity%	Specificity%	PPV%	NPV%
LUAD presenting as solid nodules
Test set
GLM	0.443	0.808 (0.028)	73.5	72.0	74.5	65.7	79.7
SVM	0.272	0.817 (0.029)	72.2	82.8	65.0	61.6	84.8
RF	0.515	0.829 (0.028)	75.2	68.8	79.6	69.6	79.0
GBM	0.472	0.815 (0.029)	73.9	69.9	76.6	67.0	78.9
Clinic-Rad model	0.505	0.820 (0.027)	74.3	69.9	77.4	67.7	79.1
Validation set
GLM	0.443	0.777 (0.024) ^*	63.6	79.4	59.0	36.3	90.7
SVM	0.272	0.745 (0.025) ^*	62.3	76.6	58.1	35.0	89.4
RF	0.515	0.708 (0.028) ^*	66.4	57.9	68.9	35.4	84.7
GBM	0.472	0.750 (0.026) ^*	69.6	67.3	70.2	40.0	87.9
Clinic-Rad model	0.505	0.802 (0.022)	70.6	75.7	69.1	42.0	90.6
LUAD presenting as subsolid nodules
Test set
GLM	0.268	0.922 (0.035)	85.1	76.2	89.1	76.2	89.1
SVM	0.202	0.906 (0.039)	86.6	81.0	89.1	77.3	91.1
RF	0.448	0.903 (0.040)	83.6	52.4	97.8	91.7	81.8
GBM	0.219	0.909 (0.035)	85.1	85.7	84.8	72.0	92.9
Clinic-Rad model	0.251	0.917 (0.033)	83.6	76.2	87.0	72.7	88.9
Validation set
GLM	0.268	0.775 (0.053) ^#	84.4	42.9	88.8	29.0	93.6
SVM	0.202	0.685 (0.062) ^*	75.7	42.9	79.2	18.0	92.9
RF	0.448	0.547 (0.077) ^*	80.7	23.8	86.8	16.1	91.4
GBM	0.219	0.725 (0.059) ^#	76.6	47.6	79.7	20.0	93.5
Clinic-Rad model	0.251	0.797 (0.050)	83.5	41.9	87.8	27.3	93.5
Note. LUAD, lung adenocarcinoma. AUC, area under the curve. SD, standard deviation. PPV, positive predictive value. NPV, negative predictive value. GLM, generalized linear model. SVM, support vector machine. RF, random forest. GBM, gradient boosting machine (GBM). Clinic-Rad model, the model integrating clinical predictors and Radscore. * The AUC value of the Clinic-Rad model is statistically higher than that of GLM model, SVM model, RF model and GBM model in solid-appearance LUAD, as well as is higher than that of SVM model and RF model in subsolid-appearance LUAD. # The AUC value of the Clinic-Rad model is comparable to that of GLM model and GBM model in subsolid-appearance LUAD.

4.4.2 Identifying ONM in LUAD presenting as subsolid nodules

For LUAD presenting as subsolid nodules, GLM model, SVM model, RF model, and GBM model and the Clinic-Rad model received AUC values of 0.922 ± 0.035, 0.906 ± 0.039, 0.903 ± 0.040, 0.909 ± 0.035 and 0.917 ± 0.033 in test set, with no statistical differences among models (all P > 0.05). In validation set, the Clinic-Rad model (AUC value: 0.797 ± 0.050) had comparable diagnostic performance to GLM model (versus 0.775 ± 0.053, P = 0.291) and GBM model (versus 0.725 ± 0.059, P = 0.066), but was significantly better than SVM model (versus 0.685 ± 0.062, P = 0.025) and RF model (versus 0.547 ± 0.077, P < 0.001). The pooled accuracy, sensitivity, specificity, PPV and NPV of the Clinic-Rad model were 83.5–83.6%, 41.9–76.2%, 87.0–87.8%, 27.3–72.7% and 88.9–93.5% (Table 5).

4.5 Pathological relevance

As shown in Fig. 5 and Table S3, in validation set, 16 of selected radiomics features (88.9%, n = 16/18) showed correlations with pathological characteristics, whereas “wavelet_firstorder_wavelet.HHL.MeanAbsoluteDeviation” and “wavelet_glcm_wavelet.HHL.Contrast” had no relevance with LVI, VPI and poor differentiation (all P > 0.05). Specifically, “boxsigmaimage_glcm_InverseVariance”, “laplaciansharpening_firstorder_Median”, “laplaciansharpening_gldm_LargeDependenceHighGrayLevelEmphasis”, “laplaciansharpening_glrlm_LongRunHighGrayLevelEmphasis”, “specklenoise_gldm_LargeDependenceHighGrayLevelEmphasis”, “specklenoise_glrlm_LongRunEmphasis”, “wavelet_gldm_wavelet.HHL.LargeDependenceEmphasis”, “wavelet_gldm_wavelet.LLL.LargeDependenceHighGrayLevelEmphasis”, “wavelet_glszm_wavelet.LLL.ZoneVariance” had correlations with LVI (r = 0.04–0.23, all P < 0.05), VPI (r = 0.05–0.15, all P < 0.01) and poor differentiation (r = 0.09–0.22, all P < 0.01). “boxsigmaimage_ngtdm_Contrast” was associated with LVI (r = 0.12, P < 0.01), VPI (r = 0.07, P < 0.01) and poor differentiation (r = 0.18, P < 0.01). “wavelet_glcm_wavelet.HHL.Idm”, “wavelet_firstorder_wavelet.HHL.InterquartileRange”, “wavelet_firstorder_wavelet.HHH.InterquartileRange”, “wavelet_firstorder_wavelet.HHH.MeanAbsoluteDeviation”, “wavelet_glcm_wavelet.HHL.JointEntropy” and “wavelet_glcm_wavelet.HHL.Id” were correlated with poor differentiation (r = 0.02–0.04, all P < 0.05). In addition, the Radscore showed relevance with LVI (r = 0.18, P < 0.01), VPI (r = 0.13, P < 0.01) and poor differentiation (r = 0.22, P < 0.01). Furthermore, the predicted ONM possibilities on the Clinic-Rad model had positive correlations with LVI (r = 0.17, P < 0.01), VPI (r = 0.12, P < 0.01) and poor differentiation (r = 0.27, P < 0.01).

In this retrospective study, we constructed radiomics models for identifying ONM in stage cT1a-bN0M0 LUAD, and validated the multi-center diagnostic performance of models. The results showed that the Clinic-Rad model integrating the Radscore, solid-component diameter, preoperative elevated CEA level and CTR value performed superior diagnostic efficacy, compared to alone GLM model, SVM model, RF model and GBM model. Besides, the Clinic-Rad model also showed good prediction of ONM in solid- and subsolid-appearance LUAD respectively.

On CT images, a mediastinal LN with short diameter of equal to or beyond 10mm is considered enlarged, indicating a higher possibility of metastasis. However, small-size LN (short diameter < 10mm) is still possibly involved, even with no uptake of tracer on PET/CT scan^8–10. Radiomics is far more efficient than human eyes to capture imaging biomarkers of tumor on medical images¹⁹. Prior study²⁵ demonstrated that texture features in combination with CEA level could predict ONM in stage cT1N0M0 LUAD, achieving AUC values of 0.88 and 0.83 in training and test sets respectively. Similarly, Liu, et al.,²⁶ found that a logistic regression model combing with radiomics and semantic features reached an AUC value of 0.76 for identifying ONM in peripheral LUAD. However, these studies were conducted within a single center, neglecting the evaluation of models’ stability and generalization across various institutions, scanners and scanning parameters.

In this multi-center study, the Clinic-Rad model had significantly superior diagnostic performance and stability for ONM prediction in cT1a-bN0M0 LUAD, receiving AUC values of 0.834 and 0.813 in test and external validation sets respectively; whereas RF model showed inferior diagnostic efficacy (AUC value: 0.708) and poor robustness in validation set, which might be attributed to variances in nodular types, nodular solid-component diameter and frequency of elevated CEA level among cohorts. Besides, the subgroup analysis showed that the Clinic-Rad model had good prediction of ONM in solid- and subsolid-appearance LUAD respectively, with pooled AUC values of 0.802–0.820 and 0.797–0.917 respectively. For subsolid-appearance LUAD, the model showed high-specificity and -NPV of 87.0–87.8% and 88.9–93.5%, which was favorable for the diagnosis of pathological N0 disease. Likewise, a high-NPV of 90.6% in validation set was also helpful to exclude ONM in solid-appearance LUAD that was recognized with potentially higher metastasis risk. Compared to previous PET/CT scan and invasive EBUS examination, CT-based radiomics models may be safer, simpler and less costly to indicate NM in cT1a-bN0M0-stage LUAD.

In addition, radiomics models in this study provided significantly better diagnostic performance than the clinical predictors to predict ONM in stage cT1a-bN0M0 LUAD. The CTR value that calculates the ratio of maximum consolidation size to the maximum tumor size at lung window has been used for determining surgical modalities and prognostic evaluation in NSCLC, because ground-glass opacity of adenocarcinoma on serial thin-section CT images usually corresponds to the lepetid pattern, indicating the non-invasive growth. However, there still remains controversial about the cutoff value of CTR including 0.25, 0.5, 0.75, 0.8, 0.85 and continuous CTR values³¹. Iwamoto, et al.,³² reported a CTR cutoff of 0.8 could predict recurrence with highest sensitivity and specificity. Besides, Shao, et al.,⁵ recommend the nodular maximum diameter on the mediastinal window larger or equal to 11.8mm and CTR vale larger or equal to 79.5% to indicate NM in stage T1N0M0 LUAD. Similarly, in this study, a CTR cutoff of 0.8 was preferred to reveal ONM in stage cT1a-b LUAD.

Although CTR value was an enough sensitive indicator, there was most likely a false positive diagnosis due to the low specificity. Contrarily, the elevated CEA level provided the pooled high-specificity of 88.0–97.5% to avoid over-diagnosis of ONM, but it was unfavorable to screen out high-risk patients with very low sensitivity. In addition, compared to CTR value and CEA level, the solid-component diameter greater or equal to 1.55cm had slightly better performance to indicate the likelihood of ONM. Similarly, Deng, et al.,³³ also demonstrated that NSCLC greater than 1.5cm but less than or equal to 2cm had relatively high rates of NM, and recommended for lobectomy and an extensive resection of intrapulmonary LNs. Moreover, tumor disappearance ratio less than or equal to 0.75 on the mediastinal window was found to be associated with high metastasis rates of 31.0% in LUAD³⁴, however, this was not measured in this study.

On the other hand, from a standpoint of biological background, the predicted ONM possibilities on the Clinic-Rad model in validation set had correlations with pathological characteristics including LVI, VPI and poor differentiation which were confirmed as unfavorable factors of NM^16–18. Moreover, pathological relevance analysis further demonstrated that 16 of selected radiomics features (88.9%, n = 16/18) and the Radscore that was calculated based on radiomics features were associated with those pathological predictors, which provided histopathological interpretability and validation for the model output. Among those pathological predictors, the radiomics biomarker associated with poor differentiation contributed mostly to the ONM prediction, followed by LVI, and then VPI. However, other pathological factors for instance, high-grade patterns of solid or micro-papillary types¹⁸ were not involved in this study.

The variable importance plot and SHAP summary plot showed that GLCM (“Id”, “Idm”), GLDM (LargeDependenceEmphasis), firstorder (MeanAbsoluteDeviation) contributes the most to the prediction of ONM. “Id” and “Idm” measures the local homogeneity of an image^35, and larger overall value indicates a more uniform local density and texture. A study³⁶ using gene-expression profiling and immunohistocheistry demonstrated that the “Id” was positively associated with hypoxia-related carbonic anhydrase 9, and higher “Id” that indicates poor survival was found in dense and uniform lesions. In addition, “Id” and “Idm” were also found to be useful to predict synchronous liver metastasis in colorectal cancer³⁷, and distinguish solitary brain metastasis from glioblastoma³⁸. In present study, the higher “Id” and lower “Idm” had positive contributions to ONM prediction. Besides, consistently with previous studies, our study manifested that “Id” and “Idm” were associated with poorly-differentiated adenocarcinoma. In addition, the higher “LargeDependenceEmphasis” indicative of larger dependence and more homogeneous texture and higher “MeanAbsoluteDeviation” was useful for N categorization. Moreover, the “LargeDependenceEmphasis” was also found to be a predictor of overall survival in resectable NSCLC³⁹.

There were some limitations in present study. Firstly, the clinical utility of the Clinic-Rad model still needs further prospective validation and long-time follow-up of patients’ prognosis. Additionally, only radiomics models were investigated for ONM prediction, other methods, for instance, deep learning features and gene expression profiles were not conducted in this study, which were expected to further improve the diagnostic performance. Thirdly, the inclusion of two types of images (non-contrast enhanced and contrast-enhanced CT images) rather than a single type might have influenced the stability of the radiomics features, potentially affecting the results, although the model based on a hybrid of image types performed good prediction of ONM in this study. Moreover, the diagnosis of cN0 diseases in present study mainly depended on CT scan. And only a minority of patients had both CT and PET/CT scan before surgery due to high radiation exposure and high cost of PET/CT, therefore, there was absence of the comparison of diagnostic performance between PET/CT and radiomics models, which would be further investigated in our prospective study.

In conclusion, this retrospectively multi-center study exhibits the Clinic-Rad model that shows superior predictability and robustness to identify ONM in stage cT1a-bN0M0 LUAD before surgery, and is expected to provide evidences for individualized treatment.

Ethics approval:

This study was in accordance with the Declaration of Helsinki and approved by each center’s ethic review board. The patient informed content was obtained.

Funding:

This project was approved by the National Natural Science Foundation of China (General Program) [grant number 82172030], the three-year Action Plan (2020-2022) of Shanghai Shen Kang Hospital Development Center to Promote Clinical Skills and Innovation [grant number SHDC2020CR3080B].

Conflicts of Interest:

The authors have no relevant financial or non-financial interests to disclose.

Data and/or Code availability

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

Authors’ contribution statements

QQ. Y. wrote the original draft and performed formal analysis, FH. Y. reviewed and edited the manuscript, and performed formal analysis. SP. W., ZZ. J., F. F., S. Y., ZH. C. and ZY. Z. helped data curation. F. S. provided funding acquisition and performed formal analysis.

Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin Jan-Feb. 2024;74(1):12–49.
Goldstraw P, Chansky K, Crowley J, et al. The IASLC Lung Cancer Staging Project: Proposals for Revision of the TNM Stage Groupings in the Forthcoming (Eighth) Edition of the TNM Classification for Lung Cancer. J Thorac Oncol Jan. 2016;11(1):39–51.
Mimae T, Saji H, Nakamura H, et al. Survival of Octogenarians with Early-Stage Non-small Cell Lung Cancer is Comparable Between Wedge Resection and Lobectomy/Segmentectomy: JACS1303. Ann Surg Oncol Nov. 2021;28(12):7219–27.
Yu Y, Jian H, Shen L, Zhu L, Lu S. Lymph node involvement influenced by lung adenocarcinoma subtypes in tumor size ≤ 3 cm disease: A study of 2268 cases. European Journal of Surgical Oncology (EJSO). 2016/11/01/ 2016;42(11):1714–1719.
Shao W, Zhang Z, Liu Z, et al. The value of pulmonary nodule diameter and consolidation/tumor rate in the prediction of lymph node metastasis in early-stage (cT1N0M0) lung adenocarcinoma. Transl Cancer Res Jan. 2021;10(1):38–46.
Zhang D, Chen X, Zhu D, et al. Intrapulmonary lymph node metastasis is common in clinically staged IA adenocarcinoma of the lung. Thorac Cancer Feb. 2019;10(2):123–7.
Silvestri GA, Gonzalez AV, Jantz MA, et al. Methods for staging non-small cell lung cancer: Diagnosis and management of lung cancer, 3rd ed: American College of Chest Physicians evidence-based clinical practice guidelines. Chest May. 2013;143(5 Suppl):eS211–50.
Dooms C, Tournoy KG, Schuurbiers O, et al. Endosonography for mediastinal nodal staging of clinical N1 non-small cell lung cancer: a prospective multicenter study. Chest Jan. 2015;147(1):209–15.
Giron J, Lacout A, Marcy PY. Accuracy of positron emission tomography may be improved when combined with postcontrast high-resolution computed tomography scanIn Regard to Pepek et al. Pract Radiat Oncol Sep-Oct. 2015;5(5):e549–50.
Collin-Castonguay MM, Guinde J, Laflamme L, Marcoux S, Fortin M. Added value of invasive needle techniques in mediastinal and hilar nodal staging of clinical N0-N1 non-small cell lung cancer after positron emission tomography. Clin Transl Radiat Oncol Sep. 2020;24:49–51.
Ong P, Grosu H, Eapen GA, et al. Endobronchial ultrasound-guided transbronchial needle aspiration for systematic nodal staging of lung cancer in patients with N0 disease by computed tomography and integrated positron emission tomography-computed tomography. Ann Am Thorac Soc Mar. 2015;12(3):415–9.
Shingyoji M, Nakajima T, Yoshino M, et al. Endobronchial ultrasonography for positron emission tomography and computed tomography-negative lymph node staging in non-small cell lung cancer. Ann Thorac Surg Nov. 2014;98(5):1762–7.
Decaluwé H, Dooms C, D'Journo XB et al. Mediastinal staging by videomediastinoscopy in clinical N1 non-small cell lung cancer: a prospective multicentre study. Eur Respir J Dec 2017;50(6).
Hideyuki H, Kazuto A, Yukihiro O, et al. Comparison between solid component size on thin-section CT and pathologic lymph node metastasis and local invasion in T1 lung adenocarcinoma. Jpn J Radiol. 2017;01:23.
He XQ, Luo TY, Li X, Huo JW, Gong JW, Li Q. Clinicopathological and computed tomographic features associated with occult lymph node metastasis in patients with peripheral solid non-small cell lung cancer. Eur J Radiol Nov 2021;144.
Ding N, Mao Y, Gao S, et al. Predictors of lymph node metastasis and possible selective lymph node dissection in clinical stage IA non-small cell lung cancer. J Thorac Dis Jul. 2018;10(7):4061–8.
Chen B, Xia W, Wang Z, et al. Risk analyses of N2 lymph-node metastases in patients with T1 non-small cell lung cancer: a multi-center real-world observational study in China. J Cancer Res Clin Oncol Nov. 2019;145(11):2771–7.
Zhao Y, Wang R, Shen XX, et al. Minor Components of Micropapillary and Solid Subtypes in Lung Adenocarcinoma are Predictors of Lymph Node Metastasis and Poor Prognosis. Ann Surg Oncol Jun. 2016;23(6):2099–105.
van Timmeren JE, Cester D, Tanadini-Lang S, Alkadhi H, Baessler B. Radiomics in medical imaging-how-to guide and critical reflection. Insights Imaging Aug. 2020;12(1):91.
Cheng DO, Khaw CR, McCabe J et al. Predicting histopathological features of aggressiveness in lung cancer using CT radiomics: a systematic review. Clin Radiol May/17 2024.
Zebin Y, Hao D, Chunlong F et al. A nomogram based on CT intratumoral and peritumoral radiomics features preoperatively predicts poorly differentiated invasive pulmonary adenocarcinoma manifesting as subsolid or solid lesions: a double-center study. Front Oncol 0205 2024.
Yong L, Jieke L, Xi Y, et al. Radiomic and quantitative-semantic models of low-dose computed tomography for predicting the poorly differentiated invasive non-mucinous pulmonary adenocarcinoma. Radiol Med. 2023;01:14.
Xiushan Z, Bo H, Yunhai H, et al. Diagnostic Accuracy of Deep Learning and Radiomics in Lung Cancer Staging: A Systematic Review and Meta-Analysis. Front Public Health. 2022;08:05.
Peng XQ, Yang SY, Zhou LX, et al. Repeatability and Reproducibility of Computed Tomography Radiomics for Pulmonary Nodules A Multicenter Phantom Study. Invest Radiol Apr. 2022;57(4):242–53.
Gu Y, She Y, Xie D, et al. A Texture Analysis-Based Prediction Model for Lymph Node Metastasis in Stage IA Lung Adenocarcinoma. Ann Thorac Surg Jul. 2018;106(1):214–20.
Liu Y, Kim J, Balagurunathan Y, et al. Prediction of pathological nodal involvement by CT-based Radiomic features of the primary tumor in patients with clinically node-negative peripheral lung adenocarcinomas. Med Phys Jun. 2018;45(6):2518–26.
Travis WD, Brambilla E, Noguchi M, et al. International association for the study of lung cancer/american thoracic society/european respiratory society international multidisciplinary classification of lung adenocarcinoma. J Thorac Oncol Feb. 2011;6(2):244–85.
Rusch VW, Crowley J, Giroux DJ et al. The IASLC Lung Cancer Staging Project: proposals for the revision of the N descriptors in the forthcoming seventh edition of the TNM classification for lung cancer. J Thorac Oncol Jul 2007.
Mu G, Chen Y, Wu D, Zhan Y, Zhou XS, Gao Y. Relu Cascade of Feature Pyramid Networks for CT Pulmonary Nodule Detection. Paper presented at: Machine Learning in Medical Imaging; 2019//, 2019; Cham.
Wu J, Xia Y, Wang X, et al. uRP: An integrated research platform for one-stop analysis of medical images. Front Radiol. 2023;3:1153784.
Wu Y, Song W, Wang D, et al. Prognostic value of consolidation-to-tumor ratio on computed tomography in NSCLC: a meta-analysis. World J Surg Oncol Jun. 2023;22(1):190.
Iwamoto R, Tanoue S, Nagata S, et al. T1 invasive lung adenocarcinoma: Thin-section CT solid score and histological periostin expression predict tumor recurrence. Mol Clin Oncol Nov. 2021;15(5):228.
Deng HY, Zhou J, Wang RL, et al. Surgical Choice for Clinical Stage IA Non-Small Cell Lung Cancer: View From Regional Lymph Node Metastasis. Ann Thorac Surg Apr. 2020;109(4):1079–85.
Wu JJ, Wu CY, Wu CY, et al. Impact of tumor disappearance ratio on the prognosis of lung adenocarcinoma ≤ 2 cm in size: A retrospective cohort study. J Formos Med Assoc Feb. 2021;120(2):874–82.
Scapicchio C, Gabelloni M, Barucci A, Cioni D, Saba L, Neri E. A deep look into radiomics. Radiol Med Oct 2021.
Tunali I, Tan Y, Gray JE et al. Hypoxia-Related Radiomics and Immunotherapy Response: A Multicohort Study of Non-Small Cell Lung Cancer. JNCI Cancer Spectr Aug 2021;5(4).
Yang KF, Li SJ, Xu J, Zheng YB. Machine learning prediction model for gray-level co-occurrence matrix features of synchronous liver metastasis in colorectal cancer. World J Gastrointest Surg Jun. 2024;27(6):1571–81.
Zhang G, Chen X, Zhang S, et al. Discrimination Between Solitary Brain Metastasis and Glioblastoma Multiforme by Using ADC-Based Texture Analysis: A Comparison of Two Different ROI Placements. Acad Radiol Nov. 2019;26(11):1466–72.
Peng B, Wang K, Xu R, et al. Preoperative computed tomography-based tumoral radiomic features prediction for overall survival in resectable non-small cell lung cancer. Front Oncol. 2023;13:1131816.

No competing interests reported.

SupplementaryMaterials.docx

Download PDF

Editorial decision: Revision requested
10 Oct, 2024
Reviews received at journal
03 Oct, 2024
Reviewers agreed at journal
30 Sep, 2024
Reviewers agreed at journal
12 Sep, 2024
Reviewers invited by journal
12 Sep, 2024
Editor assigned by journal
07 Sep, 2024
Submission checks completed at journal
03 Sep, 2024
First submitted to journal
03 Sep, 2024

You are reading this latest preprint version

Clinical-Radiomics Model Enhancing Prediction of Occult Nodal Metastasis in cT1a-bN0M0-stage Lung Adenocarcinoma: A Multi-center Study

Status:

Version 1

Abstract

Purpose

Methods

Results

Conclusions

Figures

1. Introduction

2. Materials and methods

2.1 Patients

2.2 CT scans

2.3 Histopathology

2.4 Imaging processing and Nodular segmentation

2.5 Features selection and Models construction

3. Statistics

4. Results

4.1 Clinical characteristics of enrolled patients

4.2 Clinical predictors of ONM

4.3 Radiomics models for predicting ONM

4.4 Diagnostic performance of the subgroups

4.4.1 Identifying ONM in LUAD presenting as solid nodules

4.4.2 Identifying ONM in LUAD presenting as subsolid nodules

4.5 Pathological relevance

5. Discussion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1