Mammography based delta radiomics with machine learning predicts pathologic complete response after neoadjuvant chemotherapy in breast cancer
6 affiliations
- 1Department of MR-1, Affiliated Hospital of North China University of Science and Technology
- 2Department of General Surgery Ward 4 (Breast), Liuzhou Work’s Hospital
- 3Department of Archives, North China University of Science and Technology
- 4Department of Breast Surgery-3, Tangshan People’s Hospital
- 5Department of Breast Surgery-4, Tangshan People’s Hospital
- 6Department of Breast Center, Tangshan People’s Hospital (Hebei Key Laboratory of Molecular Oncology, Affiliated Tangshan People’s Hospital of North China University of Science and Technology)
Abstract
Background This study aimed to evaluate the ability of mammography-based radiomics combined with machine learning to predict pathologic complete response (pCR) in breast cancer patients receiving neoadjuvant chemotherapy (NACT), addressing the limited research on mammography radiomics. Methods A retrospective cohort of 408 female breast cancer patients who underwent neoadjuvant chemotherapy (NACT) followed by surgery from January 2022 to December 2023 was analyzed. Breast-only pCR was defined as the absence of residual invasive carcinoma in the breast tumor bed and did not incorporate axillary lymph node status. Delta radiomics features, representing quantitative changes from pre- and post-NACT mammography images, were extracted. After evaluating 18 machine learning algorithms using cross-validation, CatBoost, ExtraTrees, and Logistic Regression were selected, optimized, and validated independently. A stacking ensemble was also constructed. Model performance was assessed using AUC, accuracy, and precision, with DeLong’s test for statistical comparisons. Results Among the 408 patients, 87 (21.32%) achieved pCR. On the independent test set, ExtraTrees, CatBoost, and Logistic Regression models achieved AUCs of 0.854 (95% CI 0.773–0.936), 0.836 (95% CI 0.745–0.926), and 0.775 (95% CI 0.683–0.867), respectively. The stacking ensemble model achieved the numerically highest AUC of 0.884 (95% CI 0.807–0.961). HER2 status, hormone receptor status, and some specific radiomic features were consistently identified as important contributors to pCR assessment. DeLong’s test indicated that the stacking model had a significantly higher AUC than CatBoost (p = 0.03) and Logistic Regression (p = 0.01), whereas its difference from ExtraTrees was not statistically significant (p = 0.47). Conclusions Mammography-based delta radiomics combined with machine learning showed promising performance for predicting pCR after NACT in breast cancer patients; however, external validation is needed to confirm clinical applicability.
Related
- Magnetic resonance imaging radiomics for predicting pathologic complete response to neoadjuvant chemotherapy in breast cancer
- Fusion of Deep Learning Radiomics and Clinical Features for Individualized Prediction of Neoadjuvant Chemotherapy Response in Osteosarcoma
- Incremental Value and Cross-Center Validation of Pretreatment Tri-Modal Ultrasound Deep Learning for Predicting Breast Pathologic Complete Response to Neoadjuvant Chemotherapy: A Multicenter Study
- Pretreatment Body Composition and Its Association with Pathologic Complete Response and Survival in Breast Cancer Patients Undergoing Neoadjuvant Chemotherapy: A Retrospective Cohort Study
- Development and validation of a nomogram predicting pathologic complete response to neoadjuvant chemotherapy in luminal type breast cancer patients using MammaPrint® and clinicopathological factor
BibTeX
@article{hu2026mammography,
title = {{Mammography based delta radiomics with machine learning predicts pathologic complete response after neoadjuvant chemotherapy in breast cancer}},
author = {Xiaohang Hu and Yanfang Liu and Yuan Gao and Yuhui Ge and Yang Wang and Yongping Du and Nan Zhang and Dandan Cao and Haiping Zhang and Qi Zhou},
journal = {Discover Oncology},
year = {2026},
month = jul,
doi = {10.1007/s12672-026-05645-2},
url = {https://doi.org/10.1007/s12672-026-05645-2},
note = {CC BY},
}
RIS
TY - JOUR
TI - Mammography based delta radiomics with machine learning predicts pathologic complete response after neoadjuvant chemotherapy in breast cancer
AU - Xiaohang Hu
AU - Yanfang Liu
AU - Yuan Gao
AU - Yuhui Ge
AU - Yang Wang
AU - Yongping Du
AU - Nan Zhang
AU - Dandan Cao
AU - Haiping Zhang
AU - Qi Zhou
PY - 2026
DA - 2026/07/01
JO - Discover Oncology
DO - 10.1007/s12672-026-05645-2
UR - https://doi.org/10.1007/s12672-026-05645-2
AB - Background This study aimed to evaluate the ability of mammography-based radiomics combined with machine learning to predict pathologic complete response (pCR) in breast cancer patients receiving neoadjuvant chemotherapy (NACT), addressing the limited research on mammography radiomics. Methods A retrospective cohort of 408 female breast cancer patients who underwent neoadjuvant chemotherapy (NACT) followed by surgery from January 2022 to December 2023 was analyzed. Breast-only pCR was defined as the absence of residual invasive carcinoma in the breast tumor bed and did not incorporate axillary lymph node status. Delta radiomics features, representing quantitative changes from pre- and post-NACT mammography images, were extracted. After evaluating 18 machine learning algorithms using cross-validation, CatBoost, ExtraTrees, and Logistic Regression were selected, optimized, and validated independently. A stacking ensemble was also constructed. Model performance was assessed using AUC, accuracy, and precision, with DeLong’s test for statistical comparisons. Results Among the 408 patients, 87 (21.32%) achieved pCR. On the independent test set, ExtraTrees, CatBoost, and Logistic Regression models achieved AUCs of 0.854 (95% CI 0.773–0.936), 0.836 (95% CI 0.745–0.926), and 0.775 (95% CI 0.683–0.867), respectively. The stacking ensemble model achieved the numerically highest AUC of 0.884 (95% CI 0.807–0.961). HER2 status, hormone receptor status, and some specific radiomic features were consistently identified as important contributors to pCR assessment. DeLong’s test indicated that the stacking model had a significantly higher AUC than CatBoost (p = 0.03) and Logistic Regression (p = 0.01), whereas its difference from ExtraTrees was not statistically significant (p = 0.47). Conclusions Mammography-based delta radiomics combined with machine learning showed promising performance for predicting pCR after NACT in breast cancer patients; however, external validation is needed to confirm clinical applicability.
SN - 2730-6011
ER -