BACKGROUND

Radiomics research based on whole tumors is limited by the unclear biological significance of radiomics features, which therefore lack clinical interpretability.

PURPOSE

We aimed to determine whether features extracted from subregions defined by habitat imaging, reflecting tumor heterogeneity, could identify breast cancer patients who will benefit from neoadjuvant chemotherapy (NAC), to optimize treatment.

METHODS

143 women with stage II-III breast cancer were divided into a training set (100 patients, 36 with pathologic complete response [pCR]) and a test set (43 patients, 16 with pCR). Patients underwent 3-T magnetic resonance imaging (MRI) before NAC. With the pathological results as the gold standard, we used the training set to build models for predicting pCR based on whole-tumor radiomics (Model), intravoxel incoherent motion (IVIM)-based habitat imaging (Model), conventional MRI features (Model), and immunohistochemical findings (Model). We also built the combined models Model and Model. In the test set, we compared the performance of the combined models with that of the invasive Model by using the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA). Receiver operating characteristic (ROC) curve analysis was performed to evaluate the predictive value of the model. The DeLong test was used to compare diagnostic efficiency across different parameters.

RESULTS

In the prediction of pCR, Model, Model, Model, Model, Model, Model and Model achieved AUCs of 0.895, 0.757, 0.705, 0.807, 0.800, 0.856, and 0.891 respectively, in the training set and 0.549, 0.708, 0.700, 0.788, 0.745, 0.909, and 0.891 respectively, in the test set. The DeLong test revealed no significant difference between Model versus Model (p = 0.695) and Model versus Model (p = 0.382) but showed a significant difference between Model and Model (p = 0.043).

CONCLUSION

The habitat model we established from first-order features combined with conventional MRI features and IHC findings accurately predicted pCR before NAC. This model can facilitate decision-making during individualized treatment for breast cancer.