Personalizing adjuvant radiotherapy decisions in breast cancer: an externally test MRI radiomics approach for recurrence risk stratification Authors: Man Long, Yu Xie, Jinqiu Ruan, Rui Liao, Xia Zheng, Songliang Gao, Keyu Mao, Guili Dong, Guangjun Yang, Zhuolin Li, Zhenhui Li Published: 2026 ABSTRACT AbstractPurposeTo develop a radiomics signature based on preoperative MRI to predict recurrence-free survival (RFS) and which patients will benefit from adjuvant radiotherapy in breast cancer patients.Experimental design771 eligible patients from two independent centers were collected. The training set (n = 504) obtained was from the Yunnan Cancer Hospital, and the external test set (n = 267) was obtained from the Duke Breast Cancer MRI Study. Pearson correlation analysis and clustering were used to construct the radiomics signature Rad-score. Based on the best cut-off value, the patients were divided into a high-risk group and a low-risk group, and the prognostic value was evaluated. A combined model nomogram was developed based on the radiomics features incorporating independent clinicopathological risk factors. The combined model’s performance was assessed regarding its discrimination, calibration, and clinical usefulness. Kaplan-Meier analysis was used to identify patients who benefited from adjuvant radiotherapy.ResultsThe high-risk group was significantly associated with worse RFS in the training and test sets (p = 0.001 and 0.048). The combination model can accurately predict BC recurrence risk (AUC, 0.752; 95% confidence interval [CI], 0.698–0.807). Additionally, adjuvant radiotherapy was not associated with the outcome of high-risk patients (training and test sets: p = 0.665 and 0.903, respectively). In the low-risk group, adjuvant radiotherapy was significantly associated with RFS (training and test sets: p = 0.004 and P<0.001).ConclusionsThe Rad-score can effectively predict the prognosis of patients with breast cancer, and identify high-risk patients who do not benefit from adjuvant radiotherapy.Clinical trial numberNot applicable.Supplementary informationThe online version contains supplementary material available at 10.1186/s13014-026-02868-z.Keywords: Breast cancer, Radiomics, Magnetic resonance imaging, Breast-conserving surgery, Radical mastectomy Introduction Breast cancer is the most common cancer and the leading cause of death among women worldwide [1]. Owing to varying tumor stages, molecular subtypes, and patient drug resistance, the selection of treatment strategies differs. Some patients with breast cancer can choose direct surgery followed by adjuvant radiotherapy to minimize the risk of local recurrence and improve the survival rate of patients [2–4]. However, not all patients benefit from adjuvant radiotherapy. These breast cancer patients who do not benefit may suffer from various local adverse effects caused by radiotherapy, such as skin damage, radiation pneumonitis, and cardiac injury. They also have to pay high treatment costs and face the risk of developing a second cancer [5–8]. Therefore, accurately identifying patients with low or no benefit is crucial. Currently, the prediction of recurrence in breast cancer mainly relies on patient and primary tumor characteristics, such as tumor size, lymph node status, tumor grade, and molecular subtype. These clinical-pathological factors have been used to build clinical prognostic models [9, 10]. Despite the value of clinicopathological factors in assessing the risk of breast cancer recurrence, inter-individual differences, tumor biological behavior, and the emergence of new biomarkers and detection techniques have introduced limitations to their predictive accuracy. In addition, these methods are invasive and carry the risk of comorbid complications, and postoperative specimens are lagging and cannot guide clinical treatment decisions before surgery. Therefore, the development of a non-invasive and economical method to assess the risk of breast cancer recurrence after surgery has become imperative. In recent years, the combination of artificial intelligence and modern digital medical imaging has led to new heights in tumor diagnosis, efficacy, and prognosis prediction [11]. Conventional imaging can only reveal visually observable features such as the size, shape, signal, and enhancement patterns of tumor lesions, but it fails to provide insights into tumor heterogeneity or changes in the microenvironment. In contrast, radiomics enables high-throughput extraction of diverse imaging data from annotated tumor images, offering a non-invasive means of assessment [12]. A study combining conventional imaging features of preoperative MRI with pathological findings showed that some women with breast cancer do not benefit from postoperative adjuvant radiotherapy [13]. However, this study used only conventional MRI imaging features and did not incorporate MRI-based radiomics features, which may introduce subjectivity. Another study used MRI-based radiomics features to predict recurrence in early breast cancer but failed to stratify breast cancer patients or identify subgroups that could benefit from postoperative adjuvant radiotherapy [14]. Therefore, this study aims to develop an MRI-based radiomics signature for evaluating the risk of postoperative recurrence in patients with early-stage breast cancer and to perform risk stratification analysis based on the radiomics signature. In addition, we envision using this signature to screen subgroups that would not benefit from postoperative adjuvant radiotherapy. Materials and methods Study sample This retrospective multicenter study was approved by the local ethics committee (Project Ethics number, KYLX2023-163), and written informed consent was waived for patients. This study cohort included female breast cancer patients who underwent direct surgery without neoadjuvant therapy at Yunnan Cancer Hospital from January 2013 to December 2018. Additionally, a publicly available dataset from The Cancer Imaging Archive, sourced from the Duke Breast Cancer MRI Study (January 2000 to March 2014), was utilized for analysis. The Yunnan Cancer Hospital dataset served as the training set, while the Duke University dataset acted as the test set. All patients underwent surgical treatment, including breast-conserving surgery and mastectomy. All patients were eligible as follows: patients with breast cancer confirmed by biopsy or surgery; female patients who did not receive neoadjuvant therapy after diagnosis and underwent surgery directly; The exclusion criteria were as follows: patients without indications for radiotherapy after radical mastectomy; patients with missing preoperative MRI or poor image quality; patients with missing follow-up information or follow-up time less than 90 days; Information on adjuvant radiotherapy was missing (Fig. 1). Patient screening flowchart. The Duke breast cancer MRI data set was publicly available and downloaded from the cancer imaging archive MRI procedure and segmentation To obtain a good contrast between the tumor and the surrounding breast parenchyma, the first phase contrast-enhanced T1-weighted imaging (T1+C) images were used in this study. The specific data acquisition parameters for magnetic resonance imaging are shown in Appendix S1 and S1 Table. Figure 2 shows the workflow of this study. Two in-training radiologists (L.M. and Z.X., each with 2 years of experience in breast MRI) utilized the 3D Slicer software (version 4.11.0) to perform semi-automatic tumor volume delineation. During the contouring process, coil artifacts, normal breast glands, and fat were manually excluded. If a patient had multifocal lesions, only the largest mass was delineated for subsequent image analysis and segmentation. To assess interobserver reproducibility, tumor volume segmentation was performed independently by another radiologist in training (G.S.L., with 2 years of experience in breast MRI) on 30 lesions randomly selected from the training set. The radiomics flow chart of the study. GLCM: gray-level cooccurrence matrix; GLDM: gray-level dependence matrix; GLRLM: gray-level run-length matrix; GLSZM: gray-level size zone matrix; NGTDM: neighboring gray-tone difference matrix Radiomics feature extraction A total of 1316 Radiomics features were extracted from each image using the derived Radiomics plugin in the 3D Slicer software. Before feature extraction, the image was resampled to a voxel size of 1 mm × 1 mm × 1 mm to normalize the voxel spacing. The Bin Width was set to 25. Finally, LoG kernel sizes were set to “1, 1.5, 2, 2.5”. Feature selection and radiomics signature construction Feature selection was conducted based on the training set. Initially, the inter-observer repeatability of radiomics feature extraction was evaluated using the intraclass correlation coefficient (ICC), with only features achieving an ICC greater than 0.8 being retained for further analysis. Subsequently, to improve the comparability and stability of these features, we applied normalization to the remaining radiomics features. Finally, we conducted dimensionality reduction and feature selection using Pearson correlation analysis and the Cluster method to eliminate high collinearity (>0.99) among the radiomics features. Variable selection was performed using a 10-fold cross-validation approach. Ultimately, six radiomics features were selected and utilized to construct and validate the model using the Cox proportional hazards regression model. A radiomics score (Rad-score) was calculated for each patient using a linear combination of the coefficient weights of the selected features. Radiomics modeling was performed using the FeAture Explorer (FAE, version 0.5.14) open-source software packages [15]. Prognostic value of Rad-score The potential association between the Rad-score and recurrence-free survival (RFS) was evaluated in both the training and test sets. In the training set, patients were divided into low-risk and high-risk groups based on the optimal cut-off value for the Rad-score determined using the X-tile software [16]. The same Rad-score cut-off values were applied to the test set. Kaplan-Meier survival analysis was used to compare RFS between the different risk groups in both cohorts. The chi-square test was employed to compare the baseline characteristics between the two groups. Univariate and multivariate Cox proportional hazards models were utilized in this study. Factors with p-values less than 0.05 in the univariate analyses were included in the multivariate analyses. Development and validation of radiomics nomogram We built a nomogram in the training set. This nomogram combines radiomics features and independent risk factors based on multivariate Cox analysis with stepwise selection. The accuracy of the nomogram in predicting patient survival was evaluated using the receiver operating characteristic curve (ROC). Additionally, the performance of the nomogram was assessed using calibration curves and decision curve analysis (DCA). Clinicopathological variables and follow-up Both cohorts systematically gathered comprehensive clinicopathological data about the subjects, including age, menopausal status, pathological type, estrogen receptor (ER) status, progesterone receptor (PR) status, human epidermal growth factor receptor 2 (HER2) status, molecular classification, clinical T stage, clinical N stage, adjuvant chemotherapy regimens, endocrine therapy protocols, and targeted therapy interventions. The training set’s data were extracted from the electronic medical records of Yunnan Cancer Hospital, whereas the test set’s clinicopathological information was sourced from The Cancer Imaging Archive. The endpoint of this study was RFS, defined as the time interval from the date of surgery to the occurrence of local-regional recurrence, distant metastasis, patient death, or the last follow-up visit. Tumor recurrence was identified by typical imaging signs of recurrence, confirmed pathologically during follow-up, or by tumor-related death. Patients treated at our center underwent follow-up visits every 3 months for the first 2 years, every 6 months for years 3 to 5, and annually thereafter. Each follow-up included a physical examination, laboratory tests, and ultrasonography of the liver, breast region, and lymphatic drainage area. Chest CT, mammography, or MRI examinations were also performed annually. Each patient was followed for a minimum of 5 years, with the last follow-up conducted on February 1, 2024. Statistical analysis Categorical and continuous variables were expressed as frequencies with percentages and medians with IQRs, respectively. The Fisher’s exact test or chi-square test, and the Mann-Whitney U test were used to compare baseline characteristics between the two groups. Univariate and multivariate Cox regression analyses were used to evaluate the relationship between each patient’s characteristics and RFS. ROC curve, calibration curve, and Hosmer-Lemeshow test were used to evaluate the calibration and discrimination of the prediction model. DCA was used to evaluate the clinical application value of the prediction model. The Kaplan-Meier method and the log-rank test were used to assess RFS. Statistical analyses were carried out by L.M. using SPSS (version 25.0) and R (version 4.2.4) software. Two-sided p values of less than 0.05 were considered to indicate statistical significance. Results Patient characteristics A total of 771 patients were included in the study, with 504 patients in the training set (median age, 50 years; IQR, 44–60 years) and 267 patients in the test set (median age, 56 years; IQR, 48–63 years). Among them, 241 patients (31%) underwent breast-conserving surgery, while 530 patients (69%) underwent mastectomy. Regarding survival outcomes, with a mean follow-up of 69 months (range, 3–130 months) for the entire cohort, there were 95 recurrences in the training set (19 local-regional recurrences and 76 distant metastatic recurrences) and 27 recurrences in the test set (8 local-regional recurrences and 19 distant metastatic recurrences). The patient characteristics in the training and test sets are summarized in Table 1. Statistical differences were observed between the training and test sets for all clinicopathological characteristics except for PR status and clinical T-stage (p < 0.05). Table 1Clinical characteristics of patients with breast cancer in the training and test setsVariablesTotal(n = 771)Training Set(n = 504)Test Set(n = 267)PValueAge (y)a51(45–60)50(44–57)56(48–63)<0.001bMenstrual status<0.001 Premenopausal377 (49%)279 (55%)98 (37%) Postmenopausal390 (50%)225 (45%)165 (62%) Not available4 (1%)0 (0%)4 (1%)Histological grade<0.001 IDC594 (77%)360 (71%)234 (88%) Other types169 (22%)144 (29%)25 (9%) Not available8 (1%)0 (0%)8 (3%)ER status<0.001 Negative178 (23%)141 (28%)37 (14%) Positive586 (76%)356 (71%)230 (86%) Not available7 (1%)7 (1%)0 (0.00%)PR status0.08 Negative212 (28%)146 (29%)66 (25%) Positive553 (71%)352 (70%)201 (75%) Not available6 (1%)6 (1%)0 (0%)HER2 status<0.001 Negative447 (58%)217 (43%)230 (86%) Positive173 (22%)136 (27%)37 (14%) Not available151 (20%)151 (30%)0 (0%)Molecular subtype<0.001 Luminal576 (75%)374 (74%)202 (76%) HER2+68 (9%)31 (6%)37 (14%) TN46 (6%)18 (4%)28 (10%) Not available81 (10%)81 (16%)0 (0%)cT-stage0.35 cT1 or T2714 (93%)470 (93%)244 (91%) cT3 or T457 (7%)34 (7%)23 (9%)cN-stage<0.001 cN0185 (24%)38 (7%)147 (55%) cN1-2352 (46%)267 (53%)85 (32%) cN3234 (30%)199 (40%)35 (13%)Adjuvant Chemotherapy<0.001 Yes629 (81%)483 (96%)146 (55%) No138 (18%)21 (4%)117 (44%) Not available4 (1%)0 (0%)4 (1%)Targeted Therapy<0.001 Yes82 (11%)50 (10%)32 (12%) No684 (88%)454 (90%)230 (86%) Not available5 (1%)0 (0%)5 (2%)Endocrine Therapy<0.001 Yes305 (40%)105 (21%)200 (75%) No458 (59%)399 (79%)59 (22%) Not available8 (1%)0 (0%)8 (3%)Note.—Categorical data are presented as numbers of women, with percentages in parentheses. P values were calculated using the χ2 test or Fisher exact test for categorical variables. IDC = invasive ductal carcinoma, ER = estrogen receptor, PR = progesterone receptor, HER2 = human epidermal growth factor receptor 2, TN = triple-negative, cT-stage = clinical T stage, cT-stage = clinical N stageaContinuous data is presented as median (IQR)bKruskal-Wallis test Clinical characteristics of patients with breast cancer in the training and test sets Note.—Categorical data are presented as numbers of women, with percentages in parentheses. P values were calculated using the χ2 test or Fisher exact test for categorical variables. IDC = invasive ductal carcinoma, ER = estrogen receptor, PR = progesterone receptor, HER2 = human epidermal growth factor receptor 2, TN = triple-negative, cT-stage = clinical T stage, cT-stage = clinical N stage aContinuous data is presented as median (IQR) bKruskal-Wallis test Building the Rad-score and predictive model Feature selection and Rad-score building These six features were then used to develop a Rad-score. The specific formula for calculating the Rad-score is shown in Appendix 1. The contribution of the selected features and their regression coefficients to this Rad-score construction is presented as a histogram in Fig. 3A. The correlation heatmap in Fig. 3B shows the relative independence of the selected features. The selected radiomics features. A: Histogram shows the role of individual features that contribute to the radiomics signature; histogram shows the role of individual features that contribute to the radiomics signature; B: The correlation heatmap Analysis of the prognostic value of rad-score Based on the optimal cut-off value of 0.32 for the Rad-score determined by X-tile analysis in the training set, patients were categorized into low-risk (<0.32) and high-risk (≥0.32) groups. In the training set, high-risk patients exhibited significantly worse RFS compared to low-risk patients (hazard ratio [HR], 2.98 [95% CI: 1.71–5.20]; p = 0.001; Fig. 4A). Consistent prognostic outcomes were observed in the test set (HR, 4.02 [95% CI: 1.01–15.99]; p = 0.048; Fig. 4B). These findings demonstrate the high predictive accuracy of the Rad-score. Survival analysis of patients stratified according to Rad-score. A: Analysis of relapse-free survival curves between high- and low-risk groups in the training cohort; B: Analysis of recurrence-free survival curves between high- and low-risk groups in the external test cohort Univariate Cox analysis identified that menopausal status, ER status, clinical N stage, targeted therapy, and Rad-score were associated with RFS in the training set. Multivariate analysis further revealed that menopausal status, clinical N stage, targeted therapy, and Rad-score were independent prognostic factors for RFS (p < 0.05) (Table 2). In the multivariate analysis of the test set, the radiation score was also a prognostic factor for RFS. (HR,2.62 [95% CI]: 1.14–6.02; p < 0.024) (Table 3). Table 2Univariate and multifactorial Cox regression analysis of breast cancer patients in the training setVariablesUnivariable AnalysisMultivariable AnalysisHazard RadioPValueHazard RadioPValueAge (y)1.02 (1.00–1.03)0.077Menstrual status PremenopausalReferenceReference Postmenopausal1.54 (1.05–2.24)0.0261.65 (1.12–2.43)0.011Histological grade IDCReference Other types0.81 (0.52–1.25)0.337ER status NegativeReferenceReference Positive0.65 (0.44–0.97)0.0340.78 (0.51–1.18)0.233 Not available0.91 (0.22–3.79)0.9000.97 (0.23–4.07)0.964PR status NegativeReference Positive1.27 (0.82–1.97)0.281 Not available1.68 (0.40–7.09)0.480HER2 status NegativeReference Positive1.35 (0.86–2.12)0.193 Not available1.11 (0.71–1.76)0.647Molecular subtype LuminalReference HER2+1.08 (0.50–2.34)0.851 TN1.07 (0.39–2.94)0.890 Not available1.24 (0.77–2.01)0.378cT-stage cT1 or T2Reference cT3 or T41.45 (0.7–2.77)0.268cN-stage cN0ReferenceReference cN1-20.71 (0.31–1.60)0.4090.78 (0.34, 1.81)0.569 cN32.22 (1.02–4.84)0.0452.25 (1.01, 5.00)0.048Adjuvant Chemotherapy NoReference Yes0.90 (0.37–2.20)0.813Targeted Therapy NoReferenceReference Yes2.27 (1.38–3.74)0.0012.28 (1.36–3.82)0.002Endocrine Therapy NoReference Yes0.64 (0.37–1.08)0.095Rad-score4.12 (2.44–6.97)<0.0014.07 (2.34–7.05)<0.001Note.—Data in parentheses are 95% CIs. Only variables identified as significant (p < 0.05) in the univariable analysis were included in the multivariable analysis. IDC = invasive ductal carcinoma, ER = estrogen receptor, PR = progesterone receptor, HER2 = human epidermal growth factor receptor 2, TN = triple-negative, cT-stage = clinical T stage, cN-stage = clinical N stage, Rad-score = radiomics score Univariate and multifactorial Cox regression analysis of breast cancer patients in the training set Note.—Data in parentheses are 95% CIs. Only variables identified as significant (p < 0.05) in the univariable analysis were included in the multivariable analysis. IDC = invasive ductal carcinoma, ER = estrogen receptor, PR = progesterone receptor, HER2 = human epidermal growth factor receptor 2, TN = triple-negative, cT-stage = clinical T stage, cN-stage = clinical N stage, Rad-score = radiomics score Table 3Multivariate Cox regression analysis of breast cancer patients in the test setVariablesMultivariable Analysisβ CoefficientHazard RatioP valueMenstrual status PremenopausalReference Postmenopausal−0.200.82 (0.37–1.81)0.681 Not available−16.330.00 (0.00-Inf b)0.998ER status NegativeReference Positive−0.440.64 (0.25–1.62)0.349cN-stage cN0Reference cN1-2−0.060.94 (0.39–2.25)0.886 cN30.141.15 (0.37–3.55)0.809Targeted Therapy NoReference Yes−0.330.80 (0.23–2.76)0.728 Not available1.062.90 (0.37–22.58)0.310Rad-score0.962.62 (1.14–6.02)0.024aNote.—Data in parentheses are 95% CIs. ER = estrogen receptor, cN-stage = clinical N stage, Rad-score = radiomics scoreaIndicates statistical significance; p < 0.05bInf: Positive infinity Multivariate Cox regression analysis of breast cancer patients in the test set Note.—Data in parentheses are 95% CIs. ER = estrogen receptor, cN-stage = clinical N stage, Rad-score = radiomics score aIndicates statistical significance; p < 0.05 bInf: Positive infinity Development and validation of the combined model nomogram We constructed a combined model nomogram based on the multivariate analysis results (Fig. 5A). The nomogram demonstrated good discriminative ability between high- and low-risk patients, with AUC values of 0.752 (95% CI: 0.698–0.807) in the training set and 0.611 (95% CI: 0.495–0.727) in the test set (Fig. 5B). The calibration curves for the nomogram model showed good agreement between observed and predicted outcomes in both cohorts (Hosmer-Lemeshow test p > 0.05) (Fig. 5C). Additionally, DCA (Fig. 5D) revealed that the nomogram model provided greater net benefit than the treat-all and treat-none strategies across the most reasonable threshold probabilities, indicating the potential clinical utility of the combined model in decision-making. A: A nomogram based on univariate analysis to estimate the RFS of breast cancer patients. B: Receiver operating characteristic curves in the training and test cohorts. C: Calibration curve of the nomogram in the training cohort. D: Decision curve analysis of the nomogram in the training cohort. cN-stage = clinical N stage To identify the benefit-to-risk ratio of radiotherapy In the training set, patients who received adjuvant radiotherapy after surgery had a better prognosis (HR, 1.90 [95% CI: 1.22–2.96]; p = 0.004; Fig. 6A) for low-risk patients. However, in the high-risk group, adjuvant radiotherapy did not improve RFS (HR, 0.85[95% CI: 0.401–1.793]; p = 0.665; Fig. 6B). Similar results were observed in the test set, where low-risk patients had a significantly better prognosis with adjuvant radiotherapy (HR, 19.90[95% CI: 5.96–66.42]; p < 0.001; Figures 6C), while high-risk patients did not benefit significantly (HR, 1.16[95% CI: 0.11–12.15]; p = 0.903; Figures 6D). Thus, the effectiveness of adjuvant radiotherapy varies across different recurrence risk strata. Survival analysis of adjuvant radiotherapy in the different risk groups of the two cohorts. Survival analysis of the (A) low-risk group and (B) high-risk group with adjuvant radiotherapy in the training cohort. Survival analysis of the (C) low-risk and (D) high-risk groups with adjuvant radiotherapy in the test cohort The distribution of distant metastasis in the total sample showed that the prevalence of the high-risk group was higher than that of the low-risk group in the training set (88% vs 12%). The results were similar in the test set (22% vs 6%). Among those receiving adjuvant radiotherapy, the training set showed higher distant metastasis in the high-risk group than in the low-risk group (21% vs. 14%), while the test set had similar rates in both groups (17%) (Table S2). Discussion This study constructed six radiomics features associated with RFS of breast cancer patients. It developed and validated a combined model nomogram that integrates clinicopathological and radiomics features for individualized prediction of recurrence risk in breast cancer. Importantly, the Rad-score derived from these six radiomics features demonstrates reliable predictive value for recurrence risk stratification and identification of patients unlikely to benefit from adjuvant radiotherapy following surgery, thereby providing effective support for personalized treatment strategies in breast cancer patients. Radiomics is a non-invasive method that can rapidly extract quantitative features from medical images in a high-throughput manner and use highly diverse statistical analysis methods to reflect tumor heterogeneity. At present, radiomics has been widely used in tumor identification, efficacy evaluation of neoadjuvant chemotherapy, and prognosis in breast cancer [17, 18]. Most breast cancer-related radiomics studies are based on MRI examination. Texture analysis has been used in breast MRI to detect microcalcifications [19], to differentiate between benign and malignant lesions [20, 21], and to differentiate breast cancer subtypes [22]. In addition, this method has been used to predict efficacy response in patients undergoing neoadjuvant chemotherapy [23, 24]. The prognosis of patients with breast cancer varies. Although the prognosis of patients with early breast cancer is usually optimistic, the specific prognosis depends on many factors. The estimation of RFS is of great significance for the individualized treatment of breast cancer patients. Some studies have used radiomics to study the prognosis of breast cancer. We found that the RFS of different risk groups was significantly different in both the training and test sets when grouped according to the cut-off value of the Rad-score, indicating that the score is capable of prognostic stratification of breast cancer patients. Kim et al. [25] reported that there was a certain relationship between MRI texture features and poor RFS in 203 primary breast cancer patients evaluated by them. The results of this study are similar to ours. Similarly, Park et al. [26] reported that the rad-score based on radiomics features could predict disease-free survival in patients with invasive breast cancer. Although we used different approaches to select the radiomics features and Rad-score cut-off values, consistent with our results, a higher Rad-score indicates a worse prognosis. Interestingly, in our study, a total of six radiomics features were selected, and the well-known radiomics feature entropy was not included. However, many previous studies have emphasized the importance of entropy [27, 28]. We suspect that the absence of entropy in the Pearson correlation analysis and Cluster methods may be due to multicollinearity. A selected shape feature, which indirectly reflects the geometry of the tumor. A more irregular tumor shape may be associated with a worse prognosis. The two selected first-order features mainly reflect the basic properties of the image and explain the distribution of voxel intensities within the image region. In addition, we selected three higher-order texture features, all of which reflect the textural heterogeneity within the tumor. The size zone distribution in the GLSZM measures the uniformity of textural patterns; the variance in the GLRLM uses variations in gray-level values to assess the variability of textural information; and the inverse variance in the GLCM is associated with the smoothness of textural information within the tumor. These higher-order texture features have different numerical definitions, so they measure different aspects of tumor structural heterogeneity. Moreover, the six radiomic features we selected were all calculated from contrast-enhanced T1-weighted images. This suggests that contrast-enhanced T1-weighted images better depict tumor heterogeneity, which is consistent with the previous speculation by Chamming et al. [29], but inconsistent with the study by Parikh et al. [27], who believed that T2-weighted images better reveal tumor heterogeneity. This study developed and validated a nomogram based on a combination of radiomic features, offering a more precise approach for predicting breast cancer recurrence. The nomogram comprises two components: six radiomic features and clinical pathological outcomes. In our research, the nomogram model demonstrated robust predictive performance. It can facilitate individualized preoperative prediction of poor RFS outcomes and serve as an effective tool to support personalized clinical treatment for patients. In addition to the above, our findings also showed that postoperative adjuvant radiotherapy did not significantly improve RFS in high-risk breast cancer patients. Conversely, it significantly improved the RFS of low-risk patients. This conclusion was also validated in the test set. In other words, adjuvant radiotherapy after surgery does not benefit this group of high-risk patients and may instead add to their burden. This may be because more than 80% of patients in the high-risk group will develop distant metastases after surgery, while the low-risk group will not develop distant metastases. Similar results were obtained in the external test dataset. At present, adjuvant radiotherapy is effective only in patients with locoregional recurrence and not in those with distant metastases [30, 31]. Distant metastasis still requires chemotherapy to be controlled. Additionally, it may also be because the number of patients in the high-risk group in this study was relatively small, which did not allow its statistical significance to be reflected. Future trials need to assess and quantify the impact of different preoperative treatment regimens on distant metastasis and survival rates. Whether radiotherapy plays a role in reducing distant metastasis remains a priority area for future research. Our study has Several limitations. First, this study is retrospective, so there may be bias in sample selection, and there is a lack of a multicenter prospective study with a larger data set for clinical application validation. Second, although we used another center as the test set, data were collected early in this cohort, and there were problems, such as nonstandard treatment protocols. Given the continuous improvement in medical technology and treatment measures, it could potentially affect patient outcomes. This problem also exists in part of the data in the training set. Future research should incorporate more comprehensive and up-to-date datasets. Third, physicians with different experiences contoured images, which may impact the stability of radiomic features. The use of fully automated and accurate tumor segmentation methods is also critical. This can not only improve the efficiency of quantitative imaging analysis, but also ensure stability and consistency throughout the process. Fourth, the core limitation of the six radiomics features used to construct the Rad-score is the disconnect between mathematical statistical properties and physical/biological significance. Higher-order texture features such as RunVariance, InverseVariance, and SZNUN mathematically quantify tumor heterogeneity, but they lack clear biological correspondence. The increase or decrease of these feature values may reflect either true tumor hypoxia, necrosis, or cell density changes, or they may be artifacts caused by image noise, partial volume effects, or improper selection of gray discretization parameters. This disconnect between mathematical statistical properties and pathophysiological states makes it difficult for risk scores based on these features to directly guide clinical decision-making. In conclusion, our research findings indicate that the constructed Rad-score has the potential to serve as a biomarker for recurrence risk in breast cancer patients. More importantly, for patients with high risk identified by Rad-score, postoperative adjuvant radiotherapy has no significance for their prognosis, and other treatment options should be sought to reduce the recurrence of patients. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements This study is a joint effort by many investigators and staff members, and their con-tribution is greatly appreciated. We especially thank all patients who participated in this study. Abbreviations Area Under the ROC Curve Breast Cancer Confidence Interval Decision Curve Analysis Estrogen Receptor Human Epidermal Growth Factor Receptor 2 Hazard Ratio Intra-class Correlation Coefficient Invasive Ductal Carcinoma Magnetic Resonance Imaging Progesterone Receptor Recurrence-free survival Receiver Operating Characteristic Curve Volume of interest Author contributions Conception and design: Man Long, Zhenhui Li, Zhuolin Li, Guangjun Yang. Provision of study materials or patients: Yu Xie, Jinqiu Ruan. Data collection and assembly: Xia Zheng, Songliang Gao, Keyu Mao, Guili Dong. Data analysis and interpretation: Man Long, Jinqiu Ruan, Zhenhui Li. Manuscript writing: Man Long, Yu Xie, Jinqiu Ruan. Final approval of manuscript: All authors. Accountable for all aspects of the work: All authors. Funding This research was funded by Yunnan Fundamental Research Projects(202401AT070006), Major Science and Technology Projects in Yunnan Province (202201AY070001-148), and Yunnan Fundamental Research Projects (202401AT070009). Data availability The external test cohort from the Duke Breast Cancer MRI Research is available at the Cancer Imaging Archive. Additional data generated or analyzed during the study are available from the corresponding author by request. Declarations Ethics approval and consent to participate This study was conducted in accordance with the Declaration of Helsinki (2013 revision). This retrospective multicenter study was approved by the local ethics committee (Project Ethics number, KYLX2023-163), and written informed consent was waived for patients. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Man Long, Yu Xie, Jinqiu Ruan, and Rui Liao Joint first authors. Contributor Information Guangjun Yang, Email: ygj350493684@163.com. Zhuolin Li, Email: lizhuolin0327@163.com. Zhenhui Li, Email: lizhenhui@Kmmu.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The external test cohort from the Duke Breast Cancer MRI Research is available at the Cancer Imaging Archive. Additional data generated or analyzed during the study are available from the corresponding author by request.