Integrated Intratumoral and Peritumoral Ultrasound Radiomics Models for Breast Nodule Diagnosis Using Machine Learning.

Authors

Abstract

The differentiation of benign and malignant breast nodules, particularly those categorized as Breast Imaging Reporting and Data System (BI-RADS) 3-4, remains a clinical challenge due to the subjectivity and operator dependence of conventional ultrasound assessment. This study aimed to evaluate the diagnostic value of intratumoral and peritumoral ultrasound radiomics features in distinguishing benign from malignant BI-RADS 3-4 breast nodules and to construct interpretable machine learning models. Ultrasound images and parameters (BI-RADS classification) of breast nodules were retrospectively collected from female patients at two institutions between January 2021 and June 2024: 571 patients (880 nodules) from Institution 1 (Shanghai Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine) and 333 patients (351 nodules) from Institution 2 (Shanghai Second People's Hospital). Included patients had BI-RADS 3-4 nodules confirmed by pathology or, for BI-RADS 3 nodules, stable findings on follow-up for at least 2 years. Nodules from Institution 1 were randomly divided into a training set (n = 615) and an internal validation set (n = 265) at a 7:3 ratio, while patients from Institution 2 constituted an external test set (n = 351). Radiomics features of intratumoral and peritumoral regions (3 and 5 pixels wide) were extracted using PyRadiomics. Features were screened using the independent <i>t</i>-test, Spearman correlation, and least absolute shrinkage and selection operator (LASSO) regression. Machine learning models-including random forest (RF), multilayer perceptron (MLP), extra trees (ET), support vector machine (SVM), logistic regression (LR), k-nearest neighbor (KNN), eXtreme Gradient Boosting (XGBoost), Adaptive Boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT), and Light Gradient Boosting Machine (LightGBM)-were trained and compared using area under the curve (AUC) and other metrics. The best-performing model was then used to compare intratumoral versus peritumoral regions. SHapley Additive exPlanations (SHAP) was applied to interpret feature importance. Across the training, internal validation, and external test sets, SVM demonstrated the most stable and balanced performance. The 3 pixels transitional zone region of interest support vector machine (EI3_SVM) model achieved a higher AUC than the tumor core region of interest (T) model in the external test set (0.875 vs. 0.787, <i>p <</i> 0.01), with accuracy 78.1%, sensitivity 42.2%, specificity 97.0%, and Brier score 0.166, outperforming other peritumoral models. Most models incorporating peritumoral features demonstrated AUCs that were higher than or comparable to those of the T model. SHAP analysis indicated that the high specificity was mainly driven by texture and shape features. Integrating intratumoral and peritumoral radiomics features significantly improves the diagnostic accuracy and objectivity for differentiating benign and malignant breast nodules. This approach may aid clinical decision-making, reduce unnecessary biopsies, and support early precision diagnosis of breast cancer.

Tags

Topics