Evaluating the Risk of Inguinal Lymph Node Metastases before Surgery Using the Morphonode Predictive Model: A Prospective Diagnostic Study in Vulvar Cancer Patients

Abstract

Ultrasound examination is an accurate method in the preoperative evaluation of the inguinofemoral lymph nodes when performed by experienced operators. The purpose of the study was to build a robust, multi-modular model based on machine learning to discriminate between metastatic and non-metastatic inguinal lymph nodes in patients with vulvar cancer. One hundred and twenty-seven women were selected at our center from March 2017 to April 2020, and 237 inguinal regions were analyzed (75 were metastatic and 162 were non-metastatic at histology). Ultrasound was performed before surgery by experienced examiners. Ultrasound features were defined according to previous studies and collected prospectively. Fourteen informative features were used to train and test the machine to obtain a diagnostic model (Morphonode Predictive Model). The following data classifiers were integrated: (I) random forest classifiers (RCF), (II) regression binomial model (RBM), (III) decisional tree (DT), and (IV) similarity profiling (SP). RFC predicted metastatic/non-metastatic lymph nodes with an accuracy of 93.3% and a negative predictive value of 97.1%. DT identified four specific signatures correlated with the risk of metastases and the point risk of each signature was 100%, 81%, 16% and 4%, respectively. The Morphonode Predictive Model could be easily integrated into the clinical routine for preoperative stratification of vulvar cancer patients.

Keywords: lymph nodes; machine learning; ultrasound; vulvar cancer.

Figure 6

Morphonode Predictive Model output scheme.

Figure 1

Ultrasound features registration form. All features are described according to VITA nomenclature. The corresponding terms previously used in the Morphonode Study are reported in brackets.

Figure 2

Predictive performances of ultrasound features, subjective assessment and Morphonode–RFC. Ultrasound variables are grouped in four blocks according to the Area Under the ROC Curve: AUC > 75% (first), AUC > 70% (second), AUC > 65% (third), and AUC <= 65% (fourth). The optimal cut-point value (threshold) was defined by a max-Sensitivity/max-Specificity criterion. AUCs, performance indices, and their 95% confidence intervals are reported. Performance indices acronyms for ultrasound variables, subjective assessment, and Morphonode–RFC are the following: TP, True Positives; TN, True Negatives; FP, False Positives; FN, False Negatives; Se, Sensitivity; Sp, Specificity; PPV, Positive Predictive Value (Precision); NPV, Negative Predictive Value; LR+, Positive Likelihood Ratio, LR−, Negative Likelihood Ratio; F1, F1 score; Acc, Accuracy. The F1 score is defined as the harmonic mean between precision (PPV) and sensitivity (Se), thus penalizing huge differences between sensitivity and specificity. This provides a more reliable predictive accuracy measure than Acc = (TP + TN)/(TP + TN + FP + FN).

Figure 3

Flow chart: patients’ selection.

Figure 4

Decision tree flow chart: risk signatures.

Figure 5

Metastasis risk signatures (MRSs) and the number of positive nodes.

Figure 7

Morphonode Predictive Model workflow.

Figure 8

Morphonode Predictive Model workflow.