A shapley additive exPlanations (SHAP)-driven quantitative ultrasound radiomics model for non-invasive HER2 status prediction in invasive breast cancer based on intratumoral and peritumoral habitat features

This study evaluates a predictive model using quantitative ultrasound features from intratumoral and peritumoral habitats to non-invasively determine HER2 status in invasive breast cancer patients. This retrospective study included 669 patients. The intra-tumoral ROI (Intra-ROI) was manually delineated, and the peri-tumoral ROI (Peri-ROI) was created by expanding 5 mm outward from the tumor boundary. Intra- and peri-tumoral subregions were established via K-Means clustering. Models were developed based on features from the intra-tumoral microenvironment, peri-tumoral microenvironment, entire intra-tumoral area, and entire peri-tumoral area. A habitat integration model was constructed by integrating two subregions, and a comprehensive model was built by combining clinical factors. The study aimed to distinguish HER2-positive and -negative cases and classify IHC 0 and 1 + subgroups among HER2-negative patients. Model effectiveness was evaluated using AUC, accuracy, sensitivity, specificity, DeLong test, Hosmer-Lemeshow test, decision curve analysis (DCA), Precision-Recall Curve and SHAP value analysis. For Task 1 (distinguishing HER2-positive from negative), the habitat + clinical combined model achieved superior performance with a test AUC of 0.757, ACC of 0.692, sensitivity of 0.692, specificity of 0.691, F1-score of 0.466, and MCC of 0.337. The intra- and peritumor habitat ensemble model also demonstrated strong performance with AUC of 0.669, ACC of 0.697, sensitivity of 0.436, specificity of 0.759, F1-score of 0.358, and MCC of 0.251. Both models significantly outperformed single-region models (Intratumor Habitat Model: AUC 0.642, F1-score 0.372; Peritumor Habitat Model: AUC 0.593, F1-score 0.335; Whole Intratumor Model: AUC 0.609, F1-score 0.344; Whole Peritumor Model: AUC 0.577, F1-score 0.326). For Task 2 (distinguishing IHC 0 from 1 + in HER2-negative patients), both the habitat + clinical combined model (test AUC 0.786, ACC 0.745, sensitivity 0.793, specificity 0.640, F1-score 0.811, MCC 0.449) and the intra- and peritumor habitat ensemble model (test AUC 0.778, ACC 0.752, sensitivity 0.838, specificity 0.560, F1-score 0.823, MCC 0.458) demonstrated robust predictive power, significantly outperforming single-region models (Intratumor Habitat Model: AUC 0.693, F1-score 0.751; Peritumor Habitat Model: AUC 0.713, F1-score 0.763; Whole Intratumor Model: AUC 0.671, F1-score 0.697; Whole Peritumor Model: AUC 0.630, F1-score 0.687). SHAP analysis highlighted progesterone receptor (PR), estrogen receptor (ER), and ensemble model probabilities as key predictive factors. Integrating intratumoral and peritumoral ultrasound features with clinical data demonstrates promising potential for non-invasive HER2 status prediction, with the ability to capture intratumoral heterogeneity. However, these findings should be interpreted with caution given the limitations of the current study, including its retrospective single-center design. Further validation through multi-center prospective studies is warranted before any clinical application. This approach may contribute to improving personalized breast cancer treatment strategies pending robust clinical validation.