# Multicenter study provides radiomic and biological insights into neoadjuvant chemotherapy response and prognosis in luminal breast cancer

## TL;DR

A new radiomics model accurately predicts chemotherapy response and long-term prognosis in luminal breast cancer patients using MRI data and biological insights.

## Contribution

A subregion-aware, multitemporal radiomics model is introduced for predicting neoadjuvant chemotherapy response and prognosis in luminal breast cancer.

## Key findings

- The radiomics model outperformed traditional MRI and clinical models in predicting pathologic complete response (pCR) with an AUC of 0.83.
- The model also showed superior prognostic discrimination with a C-index of 0.84 for recurrence prediction in non-pCR patients.
- Radiogenomic analysis linked high-perfusion radiomic features to drug metabolism and estrogen signaling pathways, while moderate-perfusion features aligned with hypoxia and immune evasion.

## Abstract

Luminal breast cancer shows limited sensitivity to neoadjuvant chemotherapy (NAC) and substantial risk of late recurrence among non-pCR patients. Accurate tools to predict both NAC response and long-term prognosis are urgently needed.

We retrospectively analyzed 850 patients from three cohorts (FUSCC, YNCC,tures were used for pCR prediction with XGBoost, whereas pre-, post-, and delta (Δ) features informed prognostic modeling with Cox-XGBoost. Performance was evaluated by AUC and C-index with internal and external validation. Independent value of the radiomics-derived RadScore was tested by I-SPY2). A subregion-aware, multitemporal radiomics model was built from DCE- and DWI-MRI, complemented by conventional MRI descriptors (e.g., breast edema, shrinkage pattern) and clinicopathologic variables. Pre-NAC feamultivariable regression. Radiogenomic analyses explored biological underpinnings.

For response prediction, 253 patients from FUSCC (pCR 38/253) and 222 from YNCC (pCR 34/222) were included. Seven radiomics features were retained, mainly from high-perfusion subregions (4/7). The combined model achieved the best performance across cohorts, surpassing radiomics, traditional MRI, and clinical models, with AUCs of 0.83 [0.78-0.88], 0.78 [0.73-0.83], 0.63 [0.57-0.69] and 0.61 [0.55-0.67] in the validation cohort. RadScore derived from the radiomics remained an independent predictor of pCR after adjusting for clinical and MRI variables (OR = 2.06 [1.28–2.15]; P = 0.001). For prognosis, 318 non-pCR patients with ≥ 5 years follow-up were analyzed (FUSCC n = 160, 44 events; YNCC n = 158, 48 events). Nine radiomics features were retained, dominated by delta (Δ) features (5/9) and moderate-perfusion subregions (4/9). The combined model showed the highest prognostic discrimination outperforming radiomics, traditional MRI and clinical models in the validation cohort (0.84 [0.73-0.90] vs 0.81 [0.61, 0.84], 0.60 [0.73-0.90], 0.58 [0.47, 0.69]). Prognostic RadScore remained independentlnedy associated with recurrence (HR 4.78, 95% CI 2.54–5.89; P = 0.001), along with post-NAC Ki-67, diffuse edema, and non-concentric shrinkage. Radiogenomic validation confirmed that high-perfusion features driving response were enriched in drug-metabolism, PI3K-Akt, and estrogen signaling pathways, whereas moderate-perfusion delta (Δ) features driving prognosis aligned with hypoxia and immune-evasion programs associated with recurrence.

Subregion-aware, multitemporal radiomics accurately and interpretabily predicts NAC response and long-term prognosis in luminal breast cancer, supporting individualized treatment selection and risk stratification.

The online version contains supplementary material available at 10.1186/s40644-026-00994-1.