Intratumoral heterogeneity derived from synthetic MRI for characterization of HER2 subtypes in breast cancer.

Authors

Abstract

The human epidermal growth factor receptor 2 (HER2) plays a pivotal role in determining both the prognosis and therapeutic strategies for breast cancer patients. This study aimed to develop a habitat model based on synthetic magnetic resonance imaging (MRI) to quantitatively assess intratumoral heterogeneity (ITH) and to evaluate its value in predicting HER2 expression status. A retrospective analysis was conducted from September 2023 to September 2025, including 233 patients with pathologically confirmed invasive breast cancer. The investigation was organized into three key tasks: distinguishing HER2-positive from HER2-negative tumors; differentiating HER2-low from HER2-zero/positive cases; and separating HER2-zero from HER2-low/positive cases. Tumor regions were identified on synthetic MRI images and underwent clustering analysis. The resulting clusters, together with global pixel distribution patterns, were used to calculate ITH scores-specifically, ITHT1, ITHT2, and ITHPD. Patients were randomly allocated to either a training or validation set in a 7:3 ratio. Clinical independent predictors were identified through multivariable logistic regression and subsequently integrated with ITH signatures to construct comprehensive predictive models. Nine distinct machine learning algorithms were evaluated to determine the most effective approach. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) as well as decision curve analysis (DCA). Additional evaluation metrics included the Kolmogorov-Smirnov (KS) statistic and calibration curves to appraise clinical utility. The integrated model, developed using the random forest algorithm, demonstrated superior predictive capability across all three tasks. In Task 1, the model achieved AUC values of 0.855 and 0.835 for the training and validation sets, respectively. For Task 2, the model yielded AUCs of 0.905 and 0.801, while in Task 3, AUCs of 0.825 and 0.814 were observed in the training and validation sets, respectively. Notably, the model's clinical utility was substantiated by favorable Brier scores and validation through KS statistics and DCA. The integration of clinical predictors with synthetic MRI-derived ITH signatures resulted in robust, noninvasive models for preoperatively predicting HER2 expression status in breast cancer patients.

Tags

Topics