A Multimodal Fusion Model of Radiomics and Deep Learning Integrating the Tumor Microenvironment Accurately Predicts Pathological Complete Response in Breast Cancer.
Authors
Affiliations (3)
Affiliations (3)
- Institute of Computational Science and Technology, Guangzhou University, Guangzhou, China (D.H., P.X., W.L., Z.Y.).
- Department of Radiology, Guangdong Provincial People's Hospital, Guangzhou, China (J.P., Z.L.).
- Institute of Computational Science and Technology, Guangzhou University, Guangzhou, China (D.H., P.X., W.L., Z.Y.). Electronic address: [email protected].
Abstract
Pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) is a critical prognostic marker in breast cancer, yet its prediction remains challenging due to tumor heterogeneity and limitations of conventional imaging. While radiomics and deep learning (DL) have shown promise, prior studies often neglect the peritumoral microenvironment, a key determinant of therapeutic response. We developed a multimodal model integrating intratumoral radiomics, peritumoral features (9-mm expansion), and DL-derived patterns from pre-NAC MRI. The model was trained and internally validated on a high-quality, multicenter cohort from the I-SPY2 trial (n = 929) and externally validated on an independent cohort (n = 95). We extracted 3190 radiomic and 2048 DL features, selecting optimal subsets via Lasso regression and bidirectional selection. Nine machine learning algorithms were evaluated, with logistic regression (LR) emerging as the top performer. The final integrated model (Intra-Peri-DL) demonstrated favorable performance, achieving an area under the curve (AUC) of 0.888 (95% CI: 0.841-0.933) in internal validation and 0.890 (95% CI: 0.804-0.958) in external validation. This performance was statistically superior to single-modality models (intratumoral radiomics, peritumoral radiomics, or DL features alone; all P<0.05), although the improvement over the combined DL+Intra model did not reach statistical significance. The model achieved high sensitivity (>0.91) in both cohorts and suggested potential clinical utility in decision curve analysis. By synergizing radiomics and DL to capture tumor-microenvironment interplay, our model enhances pCR prediction accuracy, offering a potential clinically actionable tool for personalized NAC decision-making. This framework bridges imaging phenotypes with biological insights, paving the way for precision oncology in breast cancer.