Differentiation of Benign and Malignant Lymph Nodes using Ultrasound-based Radiomics and Machine Learning

Abstract

Background

The evaluation of lymph node characteristics is crucial for tumor staging and patient prognosis assessment, but cytological and histopathological examinations of lymph nodes are invasive and costly. This study aims to develop machine learning models for differentiating benign and malignant lymph nodes based on radiomics features of grayscale ultrasound images and patients‘ clinical characteristics.

Methods

Between 2021 and 2023, a total of 285 ultrasound images of lymph nodes were collected from 88 patients. The diagnosis of lymph nodes was confirmed by pathological examination. The image feature reduction process was done by student's t-test, Pearson correlation analysis, and Random Forest feature importance selection. Six well-established machine learning models, including Support Vector Machines (SVM), Stochastic Gradient Descent (SGD), k-nearest Neighbors (KNN), Random Forest, XGBoost, and LightGBM, were developed using a combination of patient's clinical features and radiomics features of ultrasound images. The cases were randomly divided into training and test sets in an 8:2 ratio, and the area under the receiver operating characteristic curve (AUC) was adopted to evaluate model performance.

Results

There were 135 malignant and 150 benign cases in this study, including neck and axillary lymph nodes. A total of 11 radiomics features and one clinical feature were generated after the selection process, and they were used to build the final model. The AUC values of the SGD, SVM, KNN, Random Forest, XGBoost, and LightGBM in differentiating benign and malignant lymph nodes were 0.817, 0.765, 0.746, 0.816, 0.766, and 0.747, respectively.

Conclusion

By utilizing machine learning models, particularly the SGD and Random Forest, it is possible for radiomics features from ultrasound images to effectively classify benign and malignant lymph nodes, thereby improving diagnostic efficiency.