Integrating Deep Feature Extraction and MRI Radiomics for Survival Prediction in Breast Cancer After Neoadjuvant Chemotherapy.

[RATIONALE AND OBJECTIVES] Breast cancer (BC) remains a leading contributor to the global cancer burden among women, with neoadjuvant chemotherapy (NAC) established as the standard of care for early-stage disease. However, substantial interpatient variability in treatment outcomes persists, primarily driven by inherent tumor biological heterogeneity. This underscores an urgent need for more precise prognostic tools to optimize clinical decision-making.

[MATERIALS AND METHODS] This multicenter study included 216 BC patients who completed NAC, with no overlap in datasets with previous research. We extracted four-dimensional data: clinical characteristics, pathomics features, deep learning-derived pathological features (via ResNet50), and multiparametric MRI (mpMRI) radiomics. A multimodal Cox model integrating deep feature representations and radiomic variables was constructed to combine these data. Notably, this approach differs from prior studies, which have predominantly focused on single-modality inputs (eg, radiomics or pathomics alone) or short-term endpoints such as pathological complete response (pCR).

[RESULTS] The proposed model, leveraging deep feature representations derived from CNNs and radiomic fusion, achieved superior prognostic accuracy in predicting 5-year and 7-year overall survival (OS) compared to both single-modality models and findings from previous research. For 5-year OS, it achieved an area under the receiver operating characteristic curve (AUC) of 0.890 in the training set and 0.820 in the validation set; for 7-year OS, the AUC values were 0.910 (training) and 0.870 (validation), with statistically significant superiority over unidimensional models. Calibration curves and decision curve analyses further confirmed its robust clinical utility.

[CONCLUSION] The multimodal integration of imaging, pathology, and clinical data, particularly the inclusion of CNN-derived deep features, provides complementary information that improves survival prediction in NAC-treated BC patients. This represents a meaningful advancement over existing models that rely on single-modality data or focus on short-term outcomes.

[RESEARCH REGISTRATION UNIQUE IDENTIFYING NUMBER] The study is registered at https://www.chictr.org.cn and has acquired only Identifier: ChiCTR2500098023.