Integrating delta radiomics and changes in multimodal ultrasound features for predicting response to neoadjuvant
Jingchao Chen1, Kangjian Wang1, Ming He1
1Department of Ultrasound, Zhangzhou Affiliated Hospital of Fujian Medical University, Zhangzhou, China.
Background:
To evaluate whether a multimodal, noninvasive approach can enable early prediction of response to neoadjuvant chemotherapy (NAC) in patients with breast cancer. This study aimed to develop and validate a model combining delta radiomics features (RFs), immunohistochemical (IHC) markers, tumor shrinkage patterns, and multimodal ultrasound (US) changes for early and noninvasive prediction of NAC response in breast cancer, and to preliminarily assess the association between shrinkage patterns and IHC characteristics.
Methods:
A total of 101 patients with breast cancer treated with NAC were included. US examinations performed before treatment and at mid-treatment were used to assess multimodal imaging changes, tumor shrinkage patterns, and radiomics alterations. Delta-RFs were derived from the two time points, and a delta radiomics score (delta Rad-score) was built using reproducible features selected by intraclass correlation coefficient (ICC) and least absolute shrinkage and selection operator (LASSO). Clinicopathological variables and US changes were further combined to develop three models, including a delta-radiomics model, an US-IHC model, and an integrated model. Model discrimination was quantified using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. SHapley Additive exPlanations (SHAP) analysis was further performed to explain model predictions.
Results:
Progesterone receptor (PR) status, shrinkage pattern, change in maximum tumor diameter, and change in enhancement area were independently associated with major histopathological response (MHR), and three delta-RFs were retained to construct the delta Rad-score. The combined model achieved the highest discrimination in both the training (AUC, 0.949) and validation (AUC, 0.911) cohorts, outperforming the delta-radiomics model (AUC, 0.751 and 0.670) and showing performance comparable to the US-IHC model (AUC, 0.941 and 0.893). It also yielded the lowest Brier scores (0.092 and 0.143), together with favorable calibration and net clinical benefit on decision curve analysis.
Conclusions:
Integrating delta radiomics with changes in multimodal US features enables accurate, noninvasive, mid-treatment prediction of NAC response in breast cancer, potentially supporting earlier identification of poor responders and timely therapeutic adjustment.