Ultrasound-based multiregional radiomics nomogram for predicting recurrence in HER2-positive breast cancer

Accurate prediction of recurrence risk is crucial for personalizing therapy in human epidermal growth factor receptor 2-positive (HER2-positive) breast cancer. We aimed to develop an interpretable machine learning model integrating multimodal data to address this need.

This retrospective study enrolled 148 patients with human epidermal growth factor receptor 2 (HER2)-positive breast cancer between 2017 and 2021. On preoperative ultrasound images, the intratumoral region (region of interest B, ROIB) was manually delineated by experienced radiologists. Based on the manually segmented ROIB, the 5 mm inward inner peritumoral region (ROIC) and 5 mm outward outer peritumoral region (ROIA) were automatically generated via the built-in adaptive tool of 3D Slicer, with non-breast tissues excluded manually. Radiomic features were extracted from each ROI using PyRadiomics. A pre-fusion strategy (i.e. feature fusion) was used to construct combined feature sets (ROIA+B and ROIA+B+C [ROI All]), and the radiomic model based on ROI All was defined as Rad All. After feature selection via Student's t-test/Mann-Whitney U test, Pearson correlation analysis, maximum relevance minimum redundancy (mRMR) algorithm, and least absolute shrinkage and selection operator (LASSO) regression, multiple machine learning models were compared, and the support vector machine (SVM) was selected as the optimal algorithm for radiomic model construction. Independent clinical risk factors were screened by Cox regression analysis, and a post-fusion strategy integrated these factors with the radiomic signature to build a combined nomogram model. Model performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), calibration curves, decision curve analysis (DCA), DeLong test, net reclassification improvement (NRI), and integrated discrimination improvement (IDI). Survival analysis was performed with the Kaplan-Meier method, and SHapley Additive exPlanations (SHAP) analysis was applied to enhance model interpretability.

Among the tested radiomic models, the Rad All model that integrated intratumoral, inner peritumoral, and outer peritumoral features showed the best overall performance in the internal validation set, with an AUC of 0.820 (95% CI: 0.689-0.951). The combined nomogram incorporating the Rad All signature and independent clinical risk factors (irregular mass shape, abnormal posterior echo features, lymph node metastasis, and progesterone receptor[PR] negativity) achieved an AUC of 0.905 (95% CI: 0.817-0.992), which was higher than the radiomic‑only and clinical‑only models. Statistical assessments including DeLong test, NRI, and IDI supported the incremental predictive value of the combined model (all p < 0.05). The model exhibited acceptable calibration and favorable clinical net benefit in decision curve analysis. Survival analysis using the Rad All model enabled effective risk stratification in both training and internal validation sets. SHAP analysis identified lbp‑2D_Agrlrm_ARunLengthNonUniformityNormalized as the most important radiomic feature, with high values related to increased recurrence risk.

In conclusion, integrating intratumoral, inner peritumoral, and outer peritumoral radiomic features (Rad All model) may provide more reliable predictive performance for postoperative recurrence in HER2-positive breast cancer compared with single-region radiomic models. The combined nomogram incorporating Rad All and clinical risk factors further improves predictive efficacy and may serve as a supplementary tool for clinical decision-making. SHAP analysis enhances model interpretability by identifying key predictive features. However, given the lack of independent external validation, the generalizability of our models needs to be further verified in future prospective multi-center studies.