Differentiation of Benign and Malignant Cervical Lymph Nodes Using a Multi-Modal Ultrasound-Based Machine Learning Model with SHAP Interpretability.

Authors

Abstract

To evaluate the diagnostic value of a machine learning (ML) model based on multi-modal ultrasound features in differentiating benign from malignant cervical lymph nodes, and to provide a visual interpretation of model decisions using shapley additive explanations (SHAP). This retrospective study included 190 patients with suspected cervical lymph node lesions who obtained a pathological result at the Fourth Affiliated Hospital of Harbin Medical University between August 2022 and January 2025. All patients underwent 2D ultrasound, color Doppler flow imaging (CDFI), microvascular flow imaging (MVFI), and contrast-enhanced ultrasound (CEUS). Clinical data (age and sex) and multi-modal ultrasound features were collected. Univariate analysis was used to identify variables significantly associated with lymph node malignancy. Ten ML algorithms were developed and compared to construct predictive models. Model performance was evaluated using multiple metrics such as the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. The SHAP framework was applied to interpret the decision-making process of the optimal model. Univariate analysis identified 18 features significantly associated with malignancy, including age, short axis, L/S ratio, morphology, cortical thickness, cortical echogenicity, calcification, cystic change, hyperechoic foci, reticular/cord-like echogenicity, hilum visibility, CDFI vascular pattern, pulsatility index (PI), MVFI vascular pattern, vascularity index (VI), CEUS enhancement pattern, enhancement uniformity, and necrotic regions. Among the 10 ML algorithms, the gradient boosting machine (GBM) model achieved the best diagnostic performance, with an AUC of 0.987 (95% CI: 0.967-1.000), accuracy of 0.929, sensitivity of 0.879, specificity of 0.965, and F1 score of 0.935 on the test set. The GBM model significantly outperformed an experienced ultrasound physician (AUC = 0.904, p = .03). SHAP interpretation revealed that the most influential features for prediction included CEUS enhancement pattern, L/S ratio, age, PI, and VI. The case-based analysis further demonstrated that malignant lymph nodes were commonly associated with a combination of centripetal heterogeneous enhancement on CEUS, L/S <2, presence of hyperechoic foci, mixed-type vascularity on MVFI, and elevated VI values. The GBM model based on multi-modal ultrasound features enables accurate differentiation between benign and malignant cervical lymph nodes. SHAP provides a transparent, visual interpretation of model decisions, demonstrating a novel methodological framework that integrates multi-modal data with explainable AI, supporting its potential as a reliable tool for clinical decision-making.

Tags

Topics