Development and External Validation of an Interpretable Nomogram Integrating Deep Learning and Clinicopathological Signatures on DCE-MRI for Predicting Pathological Complete Response to Neoadjuvant Therapy in Breast Cancer

Abstract Objective: To systematically compare the predictive performance of two-dimensional (2D) and 2.5-dimensional (2.5D) deep learning input strategies based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) for predicting pathological complete response (pCR) after neoadjuvant therapy (NAT) in breast cancer under a fixed backbone network, and to evaluate the cross-center generalizability of a multimodal model integrating clinicopathological information with deep learning signatures. Methods: A total of 524 patients with breast cancer who received NAT at two centers were retrospectively enrolled. The 345 patients from center one were divided into a training set (n = 241) and an internal validation set (n = 104) at a 7:3 ratio, and 179 patients from center two served as an independent external test set. Pre-treatment DCE-MRI was used to construct 2D (largest cross-section) and 2.5D (central slice ±2, five slices as multi-channel input) deep learning signatures. Nine convolutional neural network architectures were systematically compared, and the optimal backbone (DenseNet161) was selected. Univariable and multivariable logistic regression analyses of 11 clinicopathological and MRI semantic features were performed to identify independent predictors of pCR, and a clinical signature was constructed using extreme gradient boosting (XGBoost). The independent predictors were then fused with the 2D and 2.5D deep learning signatures via logistic regression to build a combined signature, which was visualized as a nomogram. The receiver operating characteristic (ROC) curve, DeLong test, calibration curve, Hosmer-Lemeshow test, and decision curve analysis were used to evaluate the discrimination, calibration, and clinical net benefit of each signature. Gradient-weighted class activation mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) were applied to interpret the deep learning and clinical signatures at the pixel and feature levels, respectively. Results: Multivariable analysis identified axillary lymph node metastasis status (OR = 0.250, 95% CI: 0.127–0.492, P < 0.001), human epidermal growth factor receptor 2 (HER-2) status (OR = 3.516, 95% CI: 1.590–7.777, P = 0.002), and Ki-67 expression (OR = 4.035, 95% CI: 1.019–15.979, P = 0.047) as independent predictors of pCR. The 2.5D signature achieved higher areas under the curve (AUCs) than the 2D signature in the training, internal validation, and external test sets (0.758 vs 0.701, 0.551 vs 0.534, and 0.597 vs 0.526, respectively), although the differences were not statistically significant (all P > 0.05). The combined signature yielded the highest discrimination in the training set (AUC = 0.865, 95% CI: 0.819–0.910), significantly outperformed either single deep learning signature across all three cohorts (all P < 0.01), and achieved discrimination comparable to the clinical signature in the external test set (0.726 vs 0.711, P = 0.665). Decision curve analysis showed that both the combined and clinical signatures provided clinical net benefit over a wide range of threshold probabilities. Stratified analysis indicated that HER-2 positivity was the dominant driver of pCR, whereas the incremental predictive value of Ki-67 was mainly observed in the HER-2-negative population (pCR rate, 18.3% vs 7.3%). Conclusion: With a fixed backbone network, the 2.5D multi-channel input strategy yields consistent but limited performance gains over the 2D input strategy. The combined signature integrating three independent clinical predictors with deep learning signatures demonstrates robust cross-center discrimination and a wide range of clinical net benefit, and may serve as an auxiliary tool for individualized treatment decision-making before NAT in breast cancer.