A radiomics-deep learning nomogram integrating intratumoral and peritumoral DCE-MRI features for pCR prediction in breast cancer.
Authors
Affiliations (1)
Affiliations (1)
- Breast Disease Diagnosis and Treatment Center, The Affiliated Hospital of Qingdao University, Qingdao, China.
Abstract
Pathological complete response (pCR) after neoadjuvant systemic therapy (NST) is an important surrogate endpoint for breast cancer prognosis. Reliable pretreatment prediction of pCR remains difficult, as conventional imaging underutilizes the peritumoral microenvironment, including immune infiltration, matrix remodeling, and vascular dynamics. We constructed a multimodal nomogram combining intratumoral and peritumoral dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) radiomic features, 3D deep learning (DL) features, and clinical biomarkers to predict pCR. In this single-center retrospective study, 534 patients were randomly split into training (n = 373) and internal test (n = 161) sets. Four classifiers were compared at 3-, 6-, and 9-mm peritumoral extensions. A 3D ResNet-50 model with Squeeze-and-Excitation attention was used to extract DL features from the training-set-selected TP6 region. Clinical predictors were selected by multivariate logistic regression. The nomogram was evaluated using area under the curve (AUC), calibration, and decision curve analysis (DCA). The 6-mm peritumoral margin consistently yielded the numerically highest AUC across all classifiers in the training set. However, in the held-out test set, the AUC difference between Model-TP6 and the intratumoral-only model was not statistically significant (DeLong P = 0.358) and should therefore be interpreted as a numerical trend rather than a statistically confirmed advantage. The integrated nomogram achieved the numerically highest test-set AUC (0.883, 95% CI: 0.827-0.940). Although most pairwise DeLong comparisons did not reach statistical significance, it showed a favorable combination of discrimination, calibration, and decision-curve net benefit within this internal cohort. This nomogram synergizes radiomics with DL representations, offering an exploratory, preliminary pretreatment tool for stratifying NST candidates. However, in this retrospective single-center study, the model was developed and evaluated using an internal split without independent external validation, and its performance may therefore be optimistic and not generalizable to other institutions or imaging settings. Prospective multicenter external validation is essential before any clinical application.