A machine learning model based on high-frequency ultrasound for differentiating benign and malignant skin tumors.

Abstract

Listen

Translate article

This study aims to explore the potential of machine learning as a non-invasive automated tool for skin tumor differentiation. Data were included from 156 lesions, collected retrospectively from September 2021 toFebruary 2024. Univariate and multivariate analyses of traditional clinical features were performed to establish a logistic regression model. Ultrasound-based radiomics features are extracted from grayscale images after delineating regions of interest(ROIs). Independent samples t-tests, Mann-Whitney U tests, and Least Absolute Shrinkage and Selection Operator (LASSO)regression were employed to select ultrasound-based radiomics features. Subsequently, five machine learning methods wereused to construct radiomics models based on the selected features. Model performance was evaluated using receiver operatingcharacteristic (ROC) curves and the Delong test. Age, poorly defined margins, and irregular shape were identified asindependent risk factors for malignant skin tumors. The multilayer perception (MLP) model achieved the best performance,with area under the curve (AUC) values of 0.963 and 0.912, respectively. The results of DeLong's test revealed a statisticallysignificant discrepancy in efficacy between the MLP and clinical models (Z=2.611, p=0.009). Machine learning based skin tumor models may serve as a potential non-invasive method to improve diagnostic efficiency.