Ultrasound-based radiomics for phyllodes tumor diagnosis: A comprehensive single-center study with extensive patient cohort

Background Accurate differentiation of phyllodes tumor (PT) from breast cancer (BC) and fibroadenoma (FA) is crucial for appropriate clinical management; however, this remains challenging owing to overlapping imaging features. This study aimed to evaluate the potential of ultrasound-based radiomics analysis to distinguish PT from BC and FA. Methods A retrospective analysis was conducted on the ultrasound images of 651 patients, including 251 with PT, 200 with BC, and 200 with FA. Radiomics features were extracted from the ultrasound images, and machine learning models were developed to differentiate PT from BC and FA. The dataset was divided into training and validation sets, and feature selection was performed using least absolute shrinkage and selection operator (LASSO) algorithms. The performance of the model was assessed using receiver operating characteristic (ROC) curve analysis, with the area under the curve (AUC) as the primary metric. Results The radiomics models demonstrated significant discriminative power. For PT vs BC differentiation, Logistic Regression achieved the highest validation performance (AUC = 0.917, accuracy = 86.3%), with shape-based features showing highest importance. For PT vs FA differentiation, seven models achieved excellent performance with AUCs >0.90, led by Multilayer Perceptron (AUC = 0.964, accuracy = 93.2%). Texture-based features were most discriminative for PT vs FA differentiation. Conclusion In this single-center internal-validation study, ultrasound-based radiomics differentiated PT from BC and FA, particularly excelling in PT vs FA differentiation, and may represent a promising adjunctive diagnostic tool that warrants external, multi-center validation.