This study establishes and validates the predictive value of radiomics models based on preoperative contrast-enhanced CT images for laryngeal recurrent occult lymph node metastasis in esophageal squamous carcinoma. Additionally, comprehensive nomograms based on rad score and clinical predictors (pathological N stage, peripheral nerve infiltration, length of esophagectomy, and degree of differentiation) were most predictive.

Detecting lymph node metastasis in routine CT examination is extremely challenging. Particularly metastatic lymph nodes less than 1 cm in diameter [13, 14]. Although routine CT examination is convenient and popular, the diagnostic performance of lymph node metastasis is relatively low. Previous studies [15, 16] have shown that imaging based on CT, MRI, PET-CT, etc., as a noninvasive and quantifiable method, not only allows observation of the anatomical structure of the tumor, but also reflects tumor heterogeneity and has good predictive accuracy in multiple tumor nodal status predictions. In this study, we found that in addition to texture characteristics, first-order characteristics are also of great value in the optimal model, with a total of 6 so-called first-order characteristics, which refer to the differences and patterns in the distribution of pixel gray intensity in raw data images directly based on CT scans, used to describe the distribution of signal intensity worthy of various voxels, which can be reflected by the distribution characteristics of histograms [17, 18]. Thus, the difference in gray-value intensity caused in the CT images within the tumor tumor was of great significance for the model predicting lymph node metastasis with para-laryngeal nerve occultness. In addition, we found that the correlation coefficients of GLCM, GLDM, GLRLM, and GLSZM imaging characteristics ranked fourth. Among these selected 3D imaging characteristics, Cluster Shade based on the gray-level symbiotic matrix of 3D images was the most valuable parameter for predicting recurrent laryngeal nerve lymph node metastasis in esophageal squamous carcinoma. The cluster shading reflects the uniformity and equilibrium of the gray matter value of the image, i.e., the larger the local change in the image texture, the larger the texture heterogeneity, the larger the cluster shading value.

This study is based on manually sketching the CT image of the area of interest, which reflects the heterogeneity of the tumor in the microstructure and improves the predictive performance of the model. This study uses ITK-SNAP software to extract the histogram parameters, absolute gradient model, grayscale travel matrix and grayscale symbiotic matrix characteristics of the three-dimensional space of the whole tumor, which can reflect the heterogeneity of the whole tumor. Select the imaging omics characteristics with the greatest weight of differential diagnosis. The constructed Radscore has high predictive value. In the training and test set, Radscore in the positive paralaryngeal lymphatic metastasis is higher than the negative lymph node metastasis group, indicating that the esophageal cancer lymph node metastasis group and the non-metastatic group have no For the heterogeneity of the same tumor, the AUC, which uses the KNN-based radiomics model alone, is 0.881 and 0.741 in the training and test set, which has high efficiency, indicating that the radiomics feature predicts that the case of lymph node metastasis that is negative by CT images suggests has potential value, high Rad-scores suggests that the possibility of lymph node metastasis is significantly increased.

This study also compared several machine learning approaches to differentiate preoperative recurrent laryngeal nerve lymph node status in patients with esophageal squamous cell carcinoma and found that the KNN machine learning model had the best predictive efficacy, with AUC values > or = 0.80 in both the training and test sets, and other machine learning models, but the limitations of the algorithm and the loss of important clinical observational characteristics prevented it from comparing with the KNN model. Therefore, the KNN machine learning model constructed in this study has high practicality and reliability. The reason why KNN machine learning models achieved the expected diagnostic efficiency may be as follows: first, the theoretical maturity and simplicity of the KNN algorithm can be used to construct regression as well as linear, nonlinear classification models. Second, compared to machine algorithms such as Basque Bayesian, the KNN algorithm has low complexity, high accuracy, and is insensitive to abnormalities. Third, as the KNN approach relies mainly on the surrounding limited adjacent samples rather than the discriminant-domain approach to determine the class to which it belongs, the KNN algorithm is more appropriate than other approaches for sets of unclassified samples with more crossovers or duplications of the class. Forth, the KNN algorithm compares automatically classifications applicable to categories with larger sample sizes, whereas those with smaller sample sizes tend to be subject to misclassification with this algorithm. Also the reason why other machine learning models failed to achieve the expected diagnostic efficiency may be as follows: first, the small sample size of the present study and the modeling of other machine learning models was simpler than that of the KNN machine learning model. Therefore, they are not applicable to small data volumes. Second, the relationship of CT imaging histological features of esophageal squamous carcinoma to recurrent laryngeal nerve paralymph node status is unclear and is likely to be nonlinear. The KNN machine learning model is more explanatory and suitable for solving a series of complex nonlinear linear problems.

Some scholars have also found [19] that the size of the primary tumour is not a determinant of lymph node metastasis for lymph node metastasis. Similar to this conclusion, in this study we found that none of the final model-selected features were shaped. This suggests that three-dimensional morphological and size characteristics of esophageal squamous carcinoma tumors are not decisive factors in the prediction of lymph node status, and that esophageal squamous carcinoma recurrent para-neural lymph node status may be more dependent on the degree of differentiation, pathological type, and progression [20] of the tumor.

In recent years, more and more studies have developed nomograms to help clinical decision-making processes intuitively, making treatment strategies more precise and personalized for patients with cancer. In previous studies of other tumors, it was found that the imaging histological characteristics of the primary tumor could evaluate and predict lymph node metastasis of gastric adenocarcinoma, lung cancer, bladder cancer and other tumors. At the time of the deadline, the authors searched Pubmed with the MeSH subject-matter ‘recurrent paranodal lymph node metastases in esophageal cancer’ and the keyword ‘Radiomics’, but did not retrieve relevant literature that predicted recurrent paranodal lymph node metastases by radiomics features of primary esophageal cancer. However, metastatic involvement of lymph nodes adjacent to the recurrent laryngeal nerve in esophageal cancer has some commonalities with axillary lymph nodes in breast cancer, i.e., no accurate localization of metastatic lymph nodes can be achieved in both cases.

Therefore, the related study of using imaging histology to predict axillary lymph node metastasis in breast cancer is of great reference significance for laryngeal recurrent nerve paralymph node metastasis in esophageal cancer. Yu et al. [21] developed a nomogram with imaging histological features and clinical features to provide individualized prediction of the risk of axillary lymph node metastasis and disease recurrence in patients with early breast cancer. Tan et al. [22] established nomograms (AUC = 0.805) containing clinical-pathological features of radiohistology based on T2-FS images using linear regression models. Shan et al. [23] validated a nomogram model for invasive detection of axillary lymph node metastasis in patients with breast cancer by combining a kinetic curve model and extraction of imaging histology features from DCE-MRI. These nomograms are based on the analysis of imaging characteristics of breast tumors, and although they are significant in predicting axillary lymph node metastasis, they do not accurately localize metastatic lymph nodes. This is similar to the subject matter of the present study. The nomogram developed in this study was composed of imaging histological features, with Rad-score being the most significant independent influencing factor in differentiating axillary lymph node metastatic status (OR = 7.86, P < 0.001). After adding Rad-score to the prediction model for imaging features, we found significant improvements in the diagnostic efficacy of nomograms compared with imaging or imaging histology alone, with internal and external validations demonstrating good discrimination and calibration, and decision curve analysis demonstrating clinical utility. In summary, the use of clinical-imaging histogram nomograms to predict small volumes of recurrent laryngeal nerve paralymph node metastasis in esophageal cancer can improve the accuracy of prediction by conventional CT techniques and help clinical decision making.

The nomogram in this study is a linear model based on the principle of logistic regression. Finally, this study includes independent risk factors such as rad-score, differentiation degree, N-staging and peripheral nerve infiltration. Using the joint prediction model, the diagnostic efficiency of the training and test set is higher. The KNN-based radiomics model shows that disease assessment needs to integrate different information such as clinical, pathology, imaging, etc. This study integrates multiple-dimensional prediction factors to build a visual line chart model. The calibration curve display model has good fitting advantages, and the model established in this study has good stability and extrapourability. The AUC of the nomogram were 0.97, 0.86 and 0.63 in the training set, internal test group and external test set respectively.

There are several limitations in this study. First, the number of patients included is limited, and the application of machine learning models to big data sets yields more stable results. Several imaging models, including KNN, MLP, and SVM, were included in this study and are a subset of machine learning models with a high ability to simulate nonlinear characteristic data. However, they did not exhibit the expected predictive power in this study, possibly because variable characteristics were not efficiently extracted and the data volume was small. Therefore, in subsequent studies, more multicenter data can be added for training and external validation, resulting in more reliable prediction models. Secondly, the development and validation of this study using retrospective data should be preceded by a prospective validation study to confirm the reliability of the model before formal clinical practice.