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 scan8–10. 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 NM16–18. 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 image35, 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.