Development, Validation, and Comparison of 2 Ultrasound Feature–Guided Machine Learning Models to Distinguish Cervical Lymphadenopathy

Abstract

Abstract The objective of this study is to develop and validate the performance of 2 ultrasound (US) feature–guided machine learning models in distinguishing cervical lymphadenopathy. We enrolled 705 patients whose US characteristics of lymph nodes were collected at our hospital. B-mode US and color Doppler US features of cervical lymph nodes in both cohorts were analyzed by 2 radiologists. The decision tree and back propagation (BP) neural network were developed by combining clinical data (age, sex, and history of tumor) and US features. The performance of the 2 models was evaluated by calculating the area under the receiver operating characteristics curve (AUC), accuracy value, precision value, recall value, and balanced F score (F1 score). The AUC of the decision tree and BP model in the modeling cohort were 0.796 (0.757, 0.835) and 0.854 (0.756, 0.952), respectively. The AUC, accuracy value, precision value, recall value, and F1 score of the decision tree in the validation cohort were all higher than those of the BP model: 0.817 (0.786, 0.848) vs 0.674 (0.601, 0.747), 0.774 (0.737, 0.811) vs 0.702 (0.629, 0.775), 0.786 (0.739, 0.833) vs 0.644 (0.568, 0.720), 0.733 (0.694, 0.772) vs 0.630 (0.542, 0.718), and 0.750 (0.705, 0.795) vs 0.627 (0.541, 0.713), respectively. The US feature–guided decision tree model was more efficient in the diagnosis of cervical lymphadenopathy than the BP model.