In this study, we constructed a model to predict occult lymph node metastasis in NSCLC based on the primary tumor focus, the peri-tumor environment of 5 mm around cancer, and the mixed features of both, and our ROC results showed that the GTV model, the CTV model and the hybrid GTV+CTV model all showed good predictive efficacy (The AUCs were 0.821, 0.812, and 0.906 in the training group, and 0.821, 0.812, and 0.906 in the validation group), respectively; the P-values of the Delong test for the hybrid model and the GTV and CTV models were less than 0.05, which were statistically significant. Additionally, the decision curves demonstrated a good net benefit, and the outcomes demonstrated that the hybrid model delivered the best prediction performance (AUC of 0.869 in the training group and AUC of 0.906 in the validation group). The Delong test performed for the GTV validation group and the CTV validation group models suggested that the AUC values were not significantly different. Moreover, the Delong tests for the hybrid model and the other two models were both statistically significant.
Two types of radiomics features, GLCM (25%) and GLDM (50%) were found to occupy the major proportion in the prediction of LNM, suggesting that researchers can focus on these two types of features in the prediction of occult LNM in lung cancer. The radiomics features in this study demonstrated better predictive efficacy compared with the clinical features. The continued predictive value of these two features for other cancers needs to be confirmed. The difference between the results of this study in the two groups of GTV and CTV was not satisfactory, considering that this result was due to the small data set of the study (only 401 cases were included in our hospital, and an additional 197 cases came from elsewhere), but the mixed training model and validation model formed by mixing GTV and CTV in the training and testing groups (which is equivalent to expanding the data set and increasing the exposure rate of the effective features) AUC predicted better than training the GTV and CTV models alone, suggesting that radiomics studies may achieve better results with larger samples. A multicenter study was conducted in this study, allowing the results to exclude single-center bias.
Many studies have been published in the literature on predicting lymph node metastasis in a variety of fields, including lung cancer, bile duct cancer, gastric cancer, and colorectal cancer, thanks to the rapid development of radiomics in recent years [12, 17-19]. Some studies have demonstrated the value of the tumor microenvironment for clinical assessment of the aggressive biological behavior of tumors, while the majority of studies concentrated on feature extraction from the primary tumor focus, ignoring the role of the environment in which the tumor cells are located in distant tumor metastasis [20-22]. In contrast, radiomics can respond to macroscopic features of tumors, such as size, shape, and texture, through high-throughput, high-dimensional features [22], which can be used to quantitatively describe various aspects of tumors, especially for assessing regional heterogeneity. Previous studies have shown that quantitative imaging features based on histogram, texture, and shape in imaging omics can respond to information related to the tumor microenvironment [23, 24]. The present study precisely considered that the peritumor environment may contain some information about the tumor. First, according to the definition of CTV, there may be micro infiltrations of the tumor around the bulk of the tumor. Secondly, this study uses radiomics to describe the characteristics of the microenvironment around the tumor to determine how it relates to occult lymph node metastasis in non-small cell lung cancer and determine the best course of action for clinical patients.
Several published literatures have reported that clinical features such as tumor size, density, morphological features, and serum CEA are highly correlated with lymph node metastasis in lung cancer [25,26], and in recent years, studies of lung cancer radiomics combined with clinical features to predict lymph node metastasis have produced good predictive results [12, 27, 28]. The small number of cases and inclusion of clinical features in this study, however, was thought to be related to the clinical features' poor predictive performance, so only radiomics features were included. The predictive advantage of the radiomics model was also significant, and in comparison to the radiomics model based on the bulk tumor volume, the combined model including the CTV had a higher predictive efficacy, which was thought to be related to the increased patient population.
In this study, only non-small cell lung cancer was included, and the tumor bulk volume under enhanced CT of the cases was clearly outlined. Cases with positive lymph nodes in the imaging report were excluded for the prediction of occult lymph node metastasis. Complete data from the CT report and postoperative pathological reports were gathered, and an external test group was added for a more thorough investigation of the influencing factors. Various traits were included to create three models, which, on the one hand, can be compared to show the importance and benefits of the integrated radiomics model of tumor bulk volume and CTV for predicting occult lymph node metastasis in NSCLC were better reflected.
There are still the following shortcomings in this study: Ⅰ Although this study is a multicenter study, the data set is less studied, the number of positive cases is low, and the sample distribution is more biased. Ⅱ More clinical characteristics should be added for thorough prediction to increase the predictive efficacy since we did not combine clinical characteristics for modeling prediction. Ⅲ Previous studies on lung cancer showed that peritumoral tissue 5-20 mm away from the tumor is closely related to tumor prognosis [29,30], but we only explored the effect of features around the tumor 5 mm. Ⅳ In this study, the radiomics features were resampled and standardized, however, the variability brought on by various scanning methods and manual image segmentation could not be eliminated, which represents a significant issue in the field of radiomics today.