Distinction Between Benign and Borderline/Malignant Phyllodes Tumor in Breast Mammography and Ultrasound Based on...

This study aims to evaluate whether radiomics methods used on breast mammography (MG) and ultrasound (US) could distinguish between benign and borderline/malignant phyllodes tumors (PTs).

A total of 362 female patients with PTs were retrospectively evaluated for the study, including 220 benign and 142 borderline/malignant cases. US and MG examinations were performed in all cases before pretreatment between 2013 and 2024. All patients were divided into the training (253 cases) and validation (109 cases) groups at a 7:3 ratio. Age, tumor size, tumor growth speed, MG findings, and US results for patients with benign and borderline/malignant PTs were analyzed and compared. Radiomics features were extracted from the lesion and perilesional regions of interest in US images, as well as craniocaudal (CC) and mediolateral oblique (MLO) MG images. The Least Absolute Shrinkage and Selection Operator algorithm was employed for feature selection. Six machine learning classifiers, including logistic regression (LR), support vector machine (SVM), random forest (RF), extremely randomized trees (ExtraTrees), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM), were implemented in the radiomics, clinical, and imaging models (CC, MLO and US). A predictive nomogram model was developed by integrating the intratumoral and peritumoral region-based combined radiomics model with the clinical model. The receiver operating characteristic (ROC) curve was used to calculate the area under the curve (AUC), sensitivity (Sens), specificity (Spec), accuracy (ACC), positive predictive values (PPV）and negative predictive values (NPV) of each model.

In terms of clinical and imaging manifestations, the patients with borderline/malignant PTs were older and their maximum tumor diameters were longer than those in patients with benign PTs in the training and validation sets (P < 0.05). There were significant differences in the MG features between benign and borderline/malignant PTs in the training and validation sets. Borderline/malignant PTs had more indistinct margins and more heterogeneous density than benign PTs. The differences in US features were also significant. Borderline/malignant PTs were more likely to show cystic changes and noncircumscribed margins than benign PTs. The radiomics model in the training cohort demonstrated the highest diagnostic performance with an AUC of 1.0, outperforming the nomogram model (AUC 0.939). Both the radiomics and nomogram models showed superior diagnostic performance compared to that of the US (AUC: 0.93), MLO (AUC: 0.913), CC (AUC: 0.888), and clinical (AUC: 0.832) models. The nomogram model in the validation cohort had the highest diagnostic performance with an AUC of 0.791, outperforming both the MLO (AUC: 0.777) and radiomics (AUC: 0.766) models. The radiomics, nomogram, and MLO models all showed superior diagnostic performance compared to that of the US (AUC: 0.681), CC (AUC: 0.699), and clinical (AUC: 0.699) models.

The radiomics and nomogram models demonstrated potential capability in differentiating benign and borderline/malignant PTs, which may have contributed to guiding treatment strategies. For preoperative assessment and surgical planning of PTs, a combined approach utilizing both core needle biopsy and the nomogram model may provide optimal decision support.