A multimodal deep learning radiomics nomogram integrating clinicopathological features for predicting pathological complete response after neoadjuvant therapy in triple-negative breast cancer: a multicenter retrospective study

Abstract

Background Accurate noninvasive estimation of pathological complete response (pCR) after neoadjuvant therapy (NAT) in triple-negative breast cancer (TNBC) may support multidisciplinary discussions, but multimodal MRI-ultrasound evidence remains limited. We developed and externally validated a clinicopathological deep learning radiomics nomogram and explored its potential auxiliary value in a strictly hypothetical post-hoc surgical-planning scenario. Methods With institutional review-board approval and waived consent, this retrospective multicenter diagnostic accuracy study included 436 biopsy-proven TNBC patients who underwent NAT and surgery: Center 1 training cohort (n = 268) and internal validation cohort (n = 86), plus historical cases from cooperating Centers 2–3 for external validation (n = 82). pCR was defined as ypT0/is ypN0. Pretreatment dynamic contrast-enhanced MRI (DCE-MRI), diffusion-weighted imaging (DWI)/apparent diffusion coefficient (ADC) maps, T2-weighted imaging (T2WI), and B-mode ultrasound were analyzed. Intratumoral and 5-mm peritumoral ring radiomics features were extracted and selected by LASSO. DenseNet-121 extracted modality-specific deep features, fused using multi-head self-attention. Age, tumor size, Ki-67, tumor-infiltrating lymphocytes (TILs), cT/N stage, and histological grade were integrated with imaging scores in a logistic-regression nomogram. Parameters and thresholds were locked after Center 1 development and applied once to external data. No model output was used clinically. Feature standardization and ComBat harmonization parameters were strictly locked to the training cohort before application to validation data. Results pCR occurred in 182/436 patients (41.7%). The nomogram achieved areas under the receiver operating characteristic curve (AUCs) of 0.917 (95% CI, 0.881-0.947) in training, 0.883 (0.814-0.937) in internal validation, and 0.856 (0.773-0.921) in external validation, exceeding clinical (external AUC, 0.721), radiomics (0.786), and deep learning (0.819) models. External sensitivity, specificity, and accuracy at the locked threshold were 0.765, 0.813, and 0.793, respectively. Calibration was acceptable (slope, 0.93; intercept, −0.06; Brier score, 0.16). Decision curve analysis showed net benefit for thresholds of 0.20-0.70. Multi-threshold analysis revealed that a high-specificity threshold (0.65) yielded a positive predictive value of 0.85, potentially identifying a subset of patients with very high confidence of achieving pCR. Compared with the clinical model, external reclassification improved (continuous net reclassification improvement [NRI], 0.34; integrated discrimination improvement [IDI], 0.11; P = 0.003 and P = 0.008, respectively). Conclusion Among the few multicenter TNBC studies combining MRI, ultrasound, intratumoral/peritumoral radiomics, and attention-based deep learning, this nomogram showed good retrospective accuracy for pCR prediction after NAT. The model is presented as an aid for pCR prediction and preoperative stratification rather than as direct support for surgical decision-making. Prospective validation and clinical-impact studies are required before implementation.