Breast cancer remains one of the most prevalent malignancies across the globe and contributes substantially to cancer-related deaths among women (1). Neoadjuvant chemotherapy (NAC) is frequently employed in cases of locally advanced breast cancer to reduce tumor size and enhance the feasibility of breast-conserving surgery (BCS) (2, 3). However, responses to NAC vary greatly among patients because of tumor heterogeneity. Tumor shrinkage patterns after NAC are prognostically relevant and increasingly used to guide individualized treatment (4).

Differences in regression patterns after NAC strongly influence surgical decision-making (5). Tumor regression is commonly categorized as either concentric regression (CR) or non-concentric regression (NCR). A typical feature of CR is a consistent, inward pattern of shrinkage, often leaving behind a solitary residual lesion or achieving full pathological resolution. Such regression allows clearer tumor boundary identification, improving surgical outcomes. Research involving MDCT has shown BCS success rates reaching up to 94% (6). In contrast, NCR is often associated with irregular tumor shrinkage, fragmented residual foci, or a mesh-like appearance, complicating the evaluation of residual disease extent (6). The RCB system standardizes evaluation of tumor burden after NAC; an RCB-III score reflects heavy residual disease and greater recurrence risk (7, 8). Emerging evidence suggests that the immune context of the tumor microenvironment can affect the trajectory of tumor regression. Notably, an increased presence of tumor-infiltrating lymphocytes (TILs) has been linked to enhanced responsiveness to neoadjuvant chemotherapy (9). Thus, precise evaluation of regression patterns using imaging modalities is critical for guiding personalized treatment strategies.

MRI has become integral to breast cancer evaluation, as it captures high-resolution insights into tumor vascular structures and functional characteristics (10). Compared to traditional imaging methods, MRI offers enhanced sensitivity in detecting non-mass enhancement, delineating tumor edges, and monitoring morphological changes during therapy, making it valuable for pre-surgical evaluation (11). Nevertheless, MRI’s accuracy in identifying residual lesions post-NAC is sometimes compromised by misinterpretations—both false positives and negatives—which may hinder optimal surgical planning (12). This limitation partly stems from MRI’s reduced sensitivity to post-treatment histological changes like necrosis and fibrosis (13). In cases of NCR, irregular tumor cell dispersion and complex stromal architecture may mask enhancement signals, thereby raising the likelihood of diagnostic errors (14). Moreover, NCR-type tumors often exhibit poorly defined boundaries and trigger immune reactions that minimally impact perfusion, limiting the effectiveness of quantitative MRI metrics (15). Integrating radiomics features with clinical data presents a potential approach to bridge diagnostic limitations and refine MRI-based classification of regression types.

Radiomics refers to the process of extracting large-scale quantitative data from routine medical images, enabling non-invasive insights into tumor biological characteristics such as spatial heterogeneity and therapeutic response (16). Evidence from multiple centers indicates that when radiomic features are integrated with clinical indicators, they can support accurate prediction of recurrence-free survival (RFS) and overall survival (OS) in individuals with breast cancer (17). This research proposes a predictive model that integrates radiomic attributes derived from MRI with clinicopathological factors to classify tumor regression patterns following NAC, aiming to identify dependable imaging indicators for informing personalized treatment strategies.

In this work, we developed a nomogram that incorporates radiomic parameters from pre-NAC T1-weighted MRI alongside selected clinicopathological variables to classify tumor regression patterns—namely, CR and NCR responses—in breast cancer. The model yielded high classification efficacy, with AUC values of 0.99 and 0.75 in the training and external test sets, respectively, demonstrating its reliability and adaptability to different patient populations. The relatively large cohort size enhanced the statistical robustness and strengthened the clinical generalizability of the findings. Feature extraction was carried out using a rigorously controlled procedure, which included expert-guided segmentation of baseline MRI scans, followed by multi-step filtering with ANOVA, Pearson correlation, and LASSO-based selection. This comprehensive pipeline ensured methodological consistency and reproducibility. By fusing quantitative imaging traits with pathological profiles, the resulting nomogram supports anticipatory surgical planning and individualized therapy selection.

Recognizing the spatial variability in tumor response after NAC, we examined the predictive capacity of clinicopathological indicators in breast cancer. Univariate statistical testing (p < 0.1) revealed estrogen receptor (ER) positivity (p = 0.002) and cN stage (p = 0.093) as significant correlates, aligning with the biological principles of the RCB system (18). Notably, ER-positive cases were more frequently associated with non-concentric regression (63.91 ± 34.60%) compared to ER-negative tumors (43.34 ± 39.75%, p = 0.002), suggesting a potential connection between hormonal activity and chemotherapy response (19). To assess model efficacy, several supervised classification algorithms were applied, including Random Forest and XGBoost. In the training dataset, Random Forest achieved the highest accuracy (AUC = 0.993, 95% CI: 0.986–0.999), outperforming logistic regression (AUC = 0.70). However, in the independent test cohort, its predictive power diminished (AUC = 0.57), reflecting the limitations of models based exclusively on clinical variables. This evident gap between training and validation performance indicates that the Random Forest model may have partially overfitted the training data. Such behavior is frequently observed when the number of predictors outweighs the sample size, causing the model to learn cohort-specific variations rather than generalizable patterns. Introducing stricter feature filtering, cross-validation, and careful parameter tuning could improve the model’s robustness and consistency when applied to independent datasets. These results are consistent with Bitencourt et al. (20), who reported improved diagnostic accuracy in HER2-positive patients using integrated radiomic-clinical frameworks (AUC = 0.89), compared to clinical-only approaches (AUC = 0.61). Moreover, the significance of cN staging echoes conclusions from a multicenter investigation by Yu et al., which demonstrated the association between nodal involvement and NAC response (21). While age (p = 0.086) and progesterone receptor (PR) status (p = 0.06) were not statistically significant, older patients exhibited a higher incidence of CR (52.52 ± 9.38 vs. 49.55 ± 10.12 years), potentially reflecting immunological differences with age that warrant further research.

To build a robust model for predicting tumor regression subtypes following NAC, we employed a range of supervised learning methods, including LR, SVM, RF, and XGBoost. These algorithms were selected due to their established performance in radiomics-based cancer prediction tasks (18, 20, 21). To mitigate overfitting, a two-stage variable reduction strategy was adopted: initial univariate analysis (p < 0.05) and Pearson correlation filtering (r > 0.9) were used to eliminate redundant features, followed by LASSO regression to retain the most predictive variables. Out of 1,196 extracted radiomic features, 84 independent parameters remained, from which the top 8 were incorporated into the final classifier. Among the models tested, RF delivered the most consistent predictive ability, with AUCs of 0.816 and 0.75 in the training and validation sets, respectively. In contrast, although XGBoost performed well in the training group (AUC = 0.95), it demonstrated reduced generalization capacity in the validation data (AUC = 0.65), suggesting overfitting—a recognized challenge in radiomics applications involving high-dimensional inputs (20, 21). Likewise, the noticeable reduction in XGBoost performance from training to validation cohorts reinforces the need for stronger regularization and systematic hyperparameter optimization. Implementing nested cross-validation, adjusting learning rates, or constraining tree depth may further mitigate model variance and enhance predictive reliability across unseen data. These results underscore the need for larger datasets and algorithm refinement to enhance external validity and clinical applicability.

In this study, radiomic features were extracted from various domains, such as first-order statistical measures, texture descriptors like the gray-level co-occurrence matrix (GLCM), and morphological characteristics, including sphericity and surface area. Among these, “original-shape-Sphericity” and “wavelet-LLH-ngtdm-Busyness” exhibited the highest ability to distinguish between tumor regression patterns, emphasizing the relevance of tumor shape and internal heterogeneity in defining CR and NCR (13, 22, 23). This finding aligns with previous studies, such as those by Li et al., who highlighted the importance of morphological features in evaluating treatment response (14). In addition, Braman et al. demonstrated the value of texture-based metrics in reflecting tumor microenvironment complexity (23). One of the strengths of our model is the integration of eight radiomic features—selected through LASSO regularization—with four key clinical variables (age, ER, PR, and cN stage). This composite model achieved an AUC of 0.75 in the validation cohort, with sensitivity and specificity values of 0.64 and 0.88, respectively. Importantly, the specificity of the combined model was significantly higher than the clinical-only model (0.88 vs. 0.50), which underscores its potential for guiding clinical decisions. When predicting centripetal regression, the model achieved a positive predictive value of 94%, demonstrating its practical value in preoperative planning, especially for decisions regarding breast conservation. These results are consistent with findings from Yu et al. and Bitencourt et al., both of whom reported improved predictive accuracy when combining radiomic and clinical features (20, 21). Additionally, the correlation between ER-negative status or advanced cN stage with non-concentric regression is in line with previous studies that identified these factors as associated with poorer chemotherapy responses (19, 24). The nomogram based on this integrated model offers a straightforward and clinically relevant tool for assessing individual patient risk, aiding in personalized treatment planning.

To further ensure clinical translatability, the predictive framework should be validated on larger, independent, and multi-institutional cohorts. Expanding the dataset and standardizing MRI acquisition parameters will help reduce bias and confirm the model’s stability under varying imaging conditions, thereby improving its generalizability for real-world applications.

Although the combined model demonstrated favorable predictive capability, enhancing statistical clarity and validating its performance with external cohorts remain important. Incorporating calibration assessment and interval estimation could improve interpretability. The nomogram performed steadily compared with single-model approaches, yet confirmation through multicenter data is required to ensure broader applicability.

This study has several limitations. Due to its retrospective design at a single institution, there is a potential risk of selection bias, which may affect the external applicability of the results. The relatively small size of the validation cohort (n = 61) further limits the statistical power and generalizability of the findings. To validate and extend these observations, future research should incorporate a multicenter, prospective design, such as the I-SPY2 framework, which will allow for more robust conclusions across diverse patient populations (25). Additionally, incorporating complementary imaging modalities like dynamic contrast-enhanced MRI (DCE-MRI), apparent diffusion coefficient (ADC) mapping, and T2-weighted imaging (T2WI) could improve the model’s predictive power by offering a more comprehensive view of tumor biology.

In conclusion, the MRI-based clinical–radiomic fusion model developed in this study successfully stratified tumor regression patterns in breast cancer patients undergoing NAC. By integrating both imaging features and clinical data, the model provides a practical decision-support tool for personalized treatment planning, particularly for breast-conserving surgeries. Future efforts should focus on validating the model in larger, more diverse cohorts and exploring the integration of genomic or molecular biomarkers to further enhance its clinical relevance and translational potential.

The integration of imaging biomarkers and clinical profiles enables reliable prediction of tumor response post-NAC, supporting more informed and tailored treatment strategies.