CT radiomics features to predict lymph node metastasis in advanced esophageal squamous cell carcinoma and to discriminate between regional and non-regional lymph node metastasis: a case control study

Abstract

Background: Prediction of lymph node status in esophageal squamous cell carcinoma (ESCC) is critical for clinical decision making. In clinical practice, computed tomography (CT) has been frequently used to assist in the preoperative staging of ESCC. Texture analysis can provide more information to reflect potential biological heterogeneity based on CT. A nomogram for the preoperative diagnosis of lymph node metastasis in patients with resectable ESCC has been previously developed. However, to the best of our knowledge, no reports focus on developing CT radiomics features to discriminate ESCC patients with regional lymph node metastasis (RLNM) and non-regional lymph node metastasis (NRLNM). We, therefore, aimed to develop CT radiomics models to predict lymph node metastasis (LNM) in advanced ESCC and to discriminate ESCC between RLNM and NRLNM.

Methods: This study enrolled 334 patients with pathologically confirmed advanced ESCC, including 152 patients without LNM and 182 patients with LNM, and 103 patients with RLNM and 79 patients NRLNM. Radiomics features were extracted from CT data for each patient. The least absolute shrinkage and selection operator (LASSO) model and independent samples t-tests or Mann-Whitney U tests were exploited for dimension reduction and selection of radiomics features. Optimal radiomics features were chosen using multivariable logistic regression analysis. The discriminating performance was assessed by area under the receiver operating characteristic curve (AUC) and accuracy.

Results: The radiomics features were developed based on multivariable logistic regression and were significantly associated with LNM status in both the training and validation cohorts (P<0.001). The radiomics models could differentiate between patients with and without LNM (AUC =0.79 and 0.75, and accuracy =0.75 and 0.71 in the training and validation cohorts, respectively). In patients with LNM, the radiomics features could effectively differentiate between RLNM and NRLNM (AUC =0.98 and 0.95, and accuracy =0.94 and 0.83 in the training and validation cohorts, respectively).

Conclusions: CT radiomics features could help predict the LNM status of advanced ESCC patients and effectively discriminate ESCC between RLNM and NRLNM.

Keywords: Esophageal neoplasms; X-ray; computed tomography (CT); lymphatic metastasis; squamous cell carcinoma.

Conflict of interest statement

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at http://dx.doi.org/10.21037/qims-20-241). The authors have no conflicts of interest to declare.

2021 Quantitative Imaging in Medicine and Surgery. All rights reserved.

Figures

Figure 1

Flow chart of patient recruitment in this study. LN+, lymph node metastasis positive; LN−, lymph node metastasis negative; RLNM, regional lymph node metastasis; NRLNM, non-regional lymph node metastasis.

Figure 2

In the examples of esophageal squamous cell carcinoma (ESCC) without lymph node metastasis (LNM) (A) and with regional (B) and non-regional (C) LNM, the tumor contours were segmented manually on contrast-enhanced computed tomography (CT) images.

Figure 3

Evaluation of feature stability with inter- and intra-observer agreements based on the interclass correlation coefficient (ICC). In the 152 patients without lymph node metastasis (LNM) and 182 patients with LNM, features had good inter-observer (A) and intra-observer (B) agreements with ICCs more than 0.75 (above the red cutoff line). After this assessment, 496 features were selected from the 556 features.

Figure 4

Of the ICC analysis of 182 patients with lymph node metastasis (LNM), features had good inter-observer (A) and intra-observer (B) agreements with ICCs more than 0.75. After this assessment, 319 features were selected from the 352 features. ICC, intra-class correlation coefficient

Figure 5

The least absolute shrinkage and selection operator (LASSO) binary logistic regression model used to select radiomics features. (A) Tuning parameter (λ) selection in the LASSO model used 10-fold cross-validation via minimum criteria. The area under the receiver operating characteristic curve (AUC) is plotted versus log(λ). Dotted vertical lines are drawn at the optimal values by using the minimum criteria and the 1 standard error of the minimum criteria (the 1-SE criteria). (B) LASSO coefficient profiles of the 386 radiomics features. A coefficient profile plot was produced against the log(λ) sequence. As a result, 11 non-zero coefficients were chosen.

Figure 6

Feature selection using least absolute shrinkage and selection operator (LASSO) algorithm. (A) Tuning parameter (λ) selection in the LASSO model used 10-fold cross-validation via minimum criteria. The binomial deviance is plotted versus log(λ). Dotted vertical lines are drawn at the optimal values by using the minimum criteria and the 1-SE criteria. (B) LASSO coefficient profiles of the 279 radiomics features. A coefficient profile plot was produced against the log(λ) sequence. As a result, 13 non-zero coefficients were chosen.

Figure 7

The receiver operating characteristic (ROC) curves of the multivariable logistic regression model for the prediction of lymph node metastasis (LNM) in esophageal squamous cell carcinoma (ESCC) in the training cohort (A) and the validation cohort (B). Notes: Values outside parentheses represent cut-off values; and the first and second values in parentheses indicate sensitivity and specificity, respectively. AUC indicates area under the receiver operating characteristic curve.

Figure 8

The receiver operating characteristic (ROC) curves of the multivariable logistic regression model for the discrimination of patients with regional lymph node metastasis (RLMN) and with non-regional lymph node metastasis (NRLNM) in the training cohort (A) and the validation cohort (B). Values outside parentheses represent cut-off values; and the first and second values in parentheses indicate sensitivity and specificity, respectively. AUC indicates area under the receiver operating characteristic curve.