Development and validation of a logistic regression model based on multiparametric MRI for differentiating triple-negative from non-triple-negative breast cancer

Abstract

Background

Triple-negative breast cancer (TNBC) is an aggressive subtype of breast cancer with a poor prognosis. The preoperative differentiation of TNBC from non-TNBC remains challenging using conventional imaging. This study aimed to develop and validate a logistic regression model based on multiparametric magnetic resonance imaging (MRI) features for distinguishing TNBC from non-TNBC.

Methods

A total of 286 female patients (mean age 42.9±5.1 years; range, 31–69 years) with pathologically confirmed breast cancer who underwent dynamic contrast-enhanced MRI (DCE-MRI) between January 2019 and September 2024 were retrospectively included. Based on immunohistochemistry, 92 patients were classified as TNBC and 194 as non-TNBC. Multiparametric MRI features, including tumor size, margin, nipple involvement, ipsilateral axillary lymph node metastasis, and Breast Imaging Reporting and Data System (BI-RADS) category, were evaluated by two radiologists. Univariate analysis was performed using rank sum test or chi-square test. Multivariable logistic regression with backward stepwise selection was used to identify independent predictors and construct the predictive model. Model performance was assessed using receiver operating characteristic (ROC) curve analysis, with area under the ROC curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) calculated. Internal validation was performed using 10-fold cross-validation.

Results

Univariate analysis showed that larger tumor size (long diameter: 20.1±14.3 vs. 17.5±4.7 mm, P<0.001; short diameter: 19.4±12.8 vs. 14.0±9.1 mm, P<0.001), unclear tumor margin (89.1% vs. 66.5%, P<0.001), nipple involvement (32.6% vs. 1.5%, P<0.001), solitary ipsilateral axillary lymph node metastasis (53.3% vs. 14.4%, P<0.001), and higher BI-RADS category (mean, 6.75 vs. 6.087, P<0.001) were significantly associated with TNBC. Multivariable logistic regression identified four independent predictors: unclear tumor margin [odds ratio (OR) =43.649; 95% confidence interval (CI): 5.191–367.001; P=0.001], ipsilateral axillary lymph node metastasis (OR =23.901, 95% CI: 6.743–84.714, P<0.001), nipple involvement (OR =9.911; 95% CI: 2.235–44.444; P=0.007), and BI-RADS category 7 (OR =3.288; 95% CI: 1.538–7.027; P=0.002). The logistic regression model achieved an AUC of 0.907 (95% CI: 0.868–0.946). At the optimal probability threshold of 0.62 (Youden index), the model yielded a sensitivity of 0.960, specificity of 0.750, accuracy of 0.887, PPV of 0.877, and NPV of 0.911. The 10-fold cross-validated AUC was 0.891 (95% CI: 0.851–0.931), supporting internal consistency.

Conclusions

The logistic regression model incorporating multiparametric MRI features—particularly unclear tumor margin, ipsilateral axillary lymph node metastasis, nipple involvement, and BI-RADS category—demonstrates good predictive performance for distinguishing TNBC from non-TNBC. This model may serve as a useful preoperative tool to assist in clinical decision-making for patients with breast cancer.

Introduction

Breast cancer incidence in China has been steadily increasing, ranking second only to lung cancer among female malignancies, with a trend toward younger age at onset, posing a serious threat to women’s health (1). Triple-negative breast cancer (TNBC) is a subtype of breast cancer, which refers to the type of breast cancer that concurrently lacks estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) expression (2,3). TNBC accounts for 10–27% of all breast cancer incidences, and compared with other subtypes, it usually has stronger invasiveness, higher recurrence rate, and poor prognosis (4-6). Recent studies have increasingly focused on correlating multimodal imaging features with breast cancer molecular subtypes (7), and functional magnetic resonance imaging (MRI) techniques have shown promise in differentiating these subtypes non-invasively (8). Therefore, accurately identifying TNBC and non-TNBC is of great significance for clinical diagnosis and treatment. Currently, MRI is the primary imaging modality for breast cancer diagnosis. This study aimed to differentiate TNBC from non-TNBC by establishing a logistic regression model for multiparametric MRI, with the hope of providing a reference for the diagnosis of TNBC. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1924/rc).

Methods

General information

A retrospective analysis was conducted on the imaging and clinical data of 286 patients with breast cancer who underwent dynamic contrast-enhanced MRI (DCE-MRI) scans and had pathological confirmation of breast cancer after biopsy or surgical resection at Yantaishan Hospital (Yantai, China) from January 2019 to September 2024. All patients were female, aged 31–69 years, with an average age of 42.9±5.1 years. The inclusion criteria required pathological confirmation of breast cancer after surgical resection or ultrasound-guided biopsy with available molecular subtyping results, both biopsy and surgery performed within 1 week after DCE-MRI examination, and acceptable image quality. The exclusion criteria included absence of receptor detection results precluding molecular subtyping and poor image quality that could affect analysis and interpretation.

A total of 286 patients with breast cancer met the above criteria and were included in the study. According to the results of immunohistochemistry, 92 patients with early-stage breast cancer who expressed negative for ER, PR, and HER2 were classified as the TNBC group, whereas the remaining 194 breast cancer patients were classified as the non-TNBC group.

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Medical Ethics Committee of Yantaishan Hospital (No. YTSH-202411-19) and written informed consent was provided by all participants.

Examination methods

For MRI, a GE Discovery MR750 3.0T (GE Healthcare, Chicago, IL, USA) was used for breast flat-field scanning combined with dynamic enhancement scanning. The patients were placed in prone position, with their breasts naturally hanging in the breast coil. T1-weighted imaging (T1WI), fat-suppressed T2-weighted imaging (T2WI), single-b value (1,000 s/mm²) diffusion-weighted imaging (DWI), and DCE-MRI were performed. All sequences were axially oriented, and a coronal fat-suppressed T2WI scan was performed last. The DCE-MRI scan was performed using the Vibrant technology in axial orientation, with a total of 7 phases (1 phase of pre-enhancement attenuation map and 5 phases of dynamic enhancement scans and 1 phase of delayed high-resolution scans), and the scanning parameters were as follows: repetition time (TR) =3.9 ms, echo time (TE) =2.1 ms, flip angle (FA) =12°, field of view (FOV) =36 cm × 36 cm, matrix =256×256, number of excitations (NEX) =1, slice thickness =7 mm, no gap scanning, and the scan time for each phase was 10 seconds. Contrast agent gadolinium-diethylenetriamine pentaacetic acid (Gd-DTPA) was used for the enhancement scan, with a dose of 0.2 mL/kg body weight, which was injected via the antecubital vein at a flow rate of 2.0 mL/s. After the injection, 20 mL of 0.9% physiological saline was immediately injected to flush the line.

Image analysis

Two radiologists with more than 5 years of experience in mammographic diagnosis of breast diseases independently read the MRI images on the Picture Archiving and Communications System (PACS) workstation without knowing the pathological results. When the observers’ opinions were inconsistent (e.g., on the long or short signal of T1WI and T2WI, whether the diffusion-limited was consistent), a third senior radiologist in breast imaging diagnosis was included to consult and make a decision.

On the MRI images, the location, shape, size, margin, T1WI and T2WI signals, uniformity of enhancement, time-signal intensity curve (TIC), diffusion-limited status, and Breast Imaging Reporting and Data System (BI-RADS) classification of the lesion were recorded. At the same time, the presence of nipple involvement and ipsilateral axillary lymph node metastasis were evaluated.

Inter-rater reliability between the two radiologists was assessed using Cohen’s kappa coefficient for categorical variables (tumor boundary, nipple involvement, lymph node metastasis, BI-RADS category) and the intraclass correlation coefficient (ICC) for continuous variables (tumor diameter). A kappa value >0.80 was considered excellent agreement.

Statistical analysis

All statistical analyses were performed using SPSS 26.0 (IBM Corp., Armonk, NY, USA). Continuous variables were compared between the TNBC and non-TNBC groups using the rank sum test, and categorical variables were analyzed using the Chi-squared test or Fisher’s exact test, as appropriate. Variables with a P value <0.05 in univariate analysis were considered candidate predictors for multivariable analysis.

Before entering variables into the multivariable logistic regression model, multicollinearity among candidate predictors was assessed using the variance inflation factor (VIF); a VIF greater than 5 was considered indicative of concerning collinearity. Multivariable logistic regression with backward stepwise selection was then performed to identify independent predictors of TNBC and to construct the predictive model. The area under the receiver operating characteristic (ROC) curve (AUC) was used to evaluate the discriminative ability of the model, and its 95% confidence interval (CI) was calculated using the DeLong method. The optimal probability cutoff value was determined by maximizing the Youden index (sensitivity + specificity − 1). At this threshold, we calculated the accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of the model.

For the BI-RADS classification, we adopted a numerical coding scheme to convert the categorical assessment into an analyzable ordinal variable: 1, BI-RADS 1 (negative); 2, BI-RADS 2 (benign); 3, BI-RADS 3 (probably benign); 4, BI-RADS 4a (low suspicion); 5, BI-RADS 4b (intermediate suspicion); 6, BI-RADS 4c (high suspicion); 7, BI-RADS 5 (highly suggestive of malignancy). This linear coding preserved the ordinal nature of increasing suspicion across categories. In univariate analysis, BI-RADS was treated as a continuous variable. For the multivariable logistic regression model, BI-RADS was dichotomized as <7 (BI-RADS 4a, 4b, or 4c) vs. 7 (BI-RADS 5) because the vast majority of malignant lesions in our cohort were classified as either BI-RADS 4 subcategories or BI-RADS 5, and this dichotomization provided the most clinically relevant stratification while ensuring adequate cell counts for stable model estimation.

To assess internal validity and model stability, we performed 10-fold cross-validation using a nested validation approach. In each fold, the dataset was randomly partitioned into 10 equal subsets. For each iteration, nine subsets were used as the training set and the remaining subset constituted the test set. Within each training set, the complete modeling pipeline was repeated independently: univariate screening, multicollinearity assessment, and backward stepwise variable selection were performed to select predictors and build the model, which was then evaluated on the held-out test set. This process was repeated 10 times to ensure that each subset served as the test set exactly once. The average AUC with 95% CI across all 10 folds was computed. This nested approach ensures that no information from the test set influenced variable selection, thereby providing an unbiased estimate of model performance. A P value <0.05 was considered statistically significant for all analyses.

Results

Comparison of MRI findings between TNBC and non-TNBC

Univariate analysis showed that TNBC and non-TNBC had statistically significant differences in the long axis, short axis, tumor margin, MRI BI-RADS classification, nipple involvement, and the presence of solitary ipsilateral axillary lymph node metastasis (P<0.05). For each predictor, TNBC was more likely to occur when the tumor was larger, the tumor margin was unclear, the nipple was involved, accompanied by solitary ipsilateral axillary lymph node metastasis, and the MRI BI-RADS classification was higher (Table 1, Figure 1).

Table 1. Comparison of MRI findings between TNBC and non-TNBC.

Parameters	TNBC group (n=92)	Non-TNBC group (n=194)	χ2/Z	P value
Age (years)	41.8±3.2	43.4±5.9	−1.247	0.289
Long diameter (mm)	20.1±14.3	17.5±4.7	−4.728	<0.001
Short diameter (mm)	19.4±12.8	14.0±9.1	−3.561	<0.001
BI-RADS	6.75	6.087	−5.161	<0.001
Side			2.286	0.093
Left	50	129
Right	42	65
Location			3.064	0.547
Upper external	46	88
Lower external	12	28
Upper outer	14	28
Lower outer	1	14
Central	19	36
T1 signal			0.034	0.854
Hypointensity	23	46
Isointensity	69	148
T2 signal			0.309	0.578
Isointensity	78	169
Hyperintensity	14	25
Tumor margin			15.309	<0.001
Clear	10	65
Blurred	82	129
Tumor morphology			0.568	0.451
Regular	13	19
Irregular	79	175
Diffusion restricted			1.403	0.236
No	9	8
Yes	83	186
Enhanced			0.022	0.882
No	10	11
Yes	82	183
TIC			2.348	0.309
I	2	10
II	59	112
III	31	72
Nipple involvement			55.632	<0.001
No	62	191
Yes	30	3
Ipsilateral axillary lymph node metastasis			40.673	<0.001
No	43	166
Yes	49	28

Data are presented as mean ± SD, mean, or number. For statistical comparisons, χ² values are reported for categorical variables and Z values for continuous variables. BI-RADS classification categories: 1, BI-RADS 1 (negative); 2, BI-RADS 2 (benign); 3, BI-RADS 3 (probably benign); 4, BI-RADS 4a (low suspicion); 5, BI-RADS 4b (intermediate suspicion); 6, BI-RADS 4c (high suspicion); 7, BI-RADS 5 (highly suggestive of malignancy). BI-RADS, Breast Imaging Reporting and Data System; MRI, magnetic resonance imaging; SD, standard deviation; TIC, time-signal intensity curve; TNBC, triple-negative breast cancer.

Figure 1.

Patient, female, 64 years old, invasive ductal carcinoma of the right breast (triple-negative). (A,B) T1WI and fat-suppressed T2WI show a heterogeneously hypointense lesion in the right breast with clear, slightly lobulated margins. The overlying skin is not involved. (C) DWI shows a high-signal lesion. (D) ADC map shows a corresponding hypointense lesion. (E) Enhanced T1WI shows a markedly enhanced lesion with heterogeneous enhancement and clear margins. (F) IHC: ER (−), PR (−), HER2-RX (0), Ki-67 (+, positive rate about 70%), CK5/6 (−), E-cad (+), p120 membrane (+), SMA (−), p63 (−), E-cadherin (+), VEGF (−), and AR (+) (HE, ×100). ADC, apparent diffusion coefficient; AR, adrenergic receptor; CK, cytokeratin; DWI, diffusion-weighted imaging; ER, estrogen receptor; HE, hematoxylin and eosin; HER2, human epidermal growth factor receptor 2; IHC, immunohistochemistry; PR, progesterone receptor; SMA, smooth muscle actin; T1WI, T1-weighted imaging; T2WI, T2-weighted imaging; VEGF, vascular endothelial growth factor.

A total of 286 breast cancer cases were included in this study, of which 92 were triple-negative and 194 were non-triple-negative. The results of univariate analysis of clinical indicators and MRI dynamic enhancement examination results between the TNBC group and the non-TNBC group are presented in Table 1.

Diagnostic model establishment and effect assessment

Inter-rater reliability analysis demonstrated excellent agreement between the two radiologists for all MRI features: tumor boundary (κ=0.89), nipple involvement (κ=0.94), ipsilateral axillary lymph node metastasis (κ=0.86), BI-RADS category (κ=0.81), and tumor size measurements (ICC =0.92).

Based on the results of the rank sum test, six factors were selected, including tumor long and short diameters, tumor margin clarity, nipple involvement, ipsilateral axillary lymph node metastasis, and BI-RADS classification, to be included in the logistic regression analysis. The results are shown in Table 2 and the ROC curve (Figure 2) is plotted. The AUC for the logistic regression model was 0.907 (95% CI: 0.868–0.946; DeLong method). The optimal probability threshold, determined by the Youden index, was 0.62, yielding a sensitivity of 0.960 and specificity of 0.750. The ROC curve with the 95% confidence band and the optimal threshold point is shown in Figure 2. The results of this study indicate that among the predictive variables for TNBC, the risk index of the two variables of ipsilateral axillary lymph node metastasis and tumor margin clarity is the highest, followed by BI-RADS classification and nipple involvement. The combined risk factors were input into the model as subvariables, the logistic regression probability threshold was divided, and the confusion matrix was plotted. Finally, the optimal regression model was selected. The accuracy, sensitivity, specificity, PPV, and NPV of predicting TNBC were 0.887, 0.960, 0.750, 0.877, and 0.911, respectively, and the AUC was 0.907. Multicollinearity diagnostics showed that all predictors included in the final model had VIF values <2 (BI-RADS: 1.23; tumor boundary: 1.41; nipple involvement: 1.17; ipsilateral axillary lymph node metastasis: 1.35), indicating no significant multicollinearity. The 10-fold cross-validated AUC was 0.891 (95% CI: 0.851–0.931), supporting the internal consistency and robustness of the model.

Table 2. Multivariable logistic regression analysis for predicting TNBC.

Parameters	β	OR (95% CI)	P value
BI-RADS classification (ref: <7)			0.002
7	1.190	3.288 (1.538–7.027)
Tumor boundary (ref: clear)			0.001
Blurred	3.776	43.649 (5.191–367.001)
Nipple involvement (ref: no)			0.007
Yes	2.294	9.911 (2.235–44.444)
Ipsilateral axillary lymph node metastasis (ref: no)			<0.001
Yes	3.174	23.901 (6.743–84.714)
Long diameter (mm) (ref: >24)			0.160
≤24	−0.079	0.924 (0.829–1.031)
Short diameter (mm) (ref: >17)			0.475
≤17	0.059	1.061 (0.901–1.250)

Variables with P<0.05 in univariate analysis were entered into multivariable logistic regression. Multicollinearity was assessed using VIF (all VIF <2). BI-RADS classification categories: 1, BI-RADS 1 (negative); 2, BI-RADS 2 (benign); 3, BI-RADS 3 (probably benign); 4, BI-RADS 4a (low suspicion); 5, BI-RADS 4b (intermediate suspicion); 6, BI-RADS 4c (high suspicion); 7, BI-RADS 5 (highly suggestive of malignancy). β, regression coefficient; BI-RADS, Breast Imaging Reporting and Data System; CI, confidence interval; OR, odds ratio; TNBC, triple-negative breast cancer; VIF, variance inflation factor.

Figure 2.

ROC curve of the multivariable logistic regression model for differentiating TNBC from non-TNBC. The AUC is 0.907 (95% CI: 0.868–0.946). The optimal threshold (probability =0.62) based on the Youden index, corresponding to a sensitivity of 0.960 and specificity of 0.750. AUC, area under the ROC curve; CI, confidence interval; ROC, receiver operating characteristic; TNBC, triple-negative breast cancer.

To facilitate clinical application, we developed a nomogram based on the final logistic regression model (Figure 3). The nomogram incorporates the four independent predictors—tumor boundary, nipple involvement, ipsilateral axillary lymph node metastasis, and BI-RADS classification—allowing for individualized prediction of TNBC probability. Each predictor is assigned a point value on the top scale; the total points accumulated across all predictors correspond to a predicted probability of TNBC on the bottom scale.

Figure 3.

Nomogram for predicting TNBC based on multiparametric MRI features. To use the nomogram, locate the patient’s status for each predictor on the corresponding axis, draw a vertical line upward to the “Points” bar to assign points for each predictor, sum the points for all predictors, locate the total points on the “Total points” axis, and draw a vertical line downward to the “Probability of TNBC” axis to obtain the predicted probability. BI-RADS classification categories: 1, BI-RADS 1 (negative); 2, BI-RADS 2 (benign); 3, BI-RADS 3 (probably benign); 4, BI-RADS 4a (low suspicion); 5, BI-RADS 4b (intermediate suspicion); 6, BI-RADS 4c (high suspicion); 7, BI-RADS 5 (highly suggestive of malignancy). BI-RADS, Breast Imaging Reporting and Data System; LN_Meta, lymph node metastasis; MRI, magnetic resonance imaging; TNBC, triple-negative breast cancer.

Discussion

Currently, logistic regression models are widely used in the medical academic field, particularly for predicting various types of tumor diseases, such as logistic model prediction of papillary thyroid carcinoma and prostate cancer (9,10). After conducting a logistic regression analysis, this study selected feature variables such as unclear tumor margin, ipsilateral axillary lymph node metastasis, BI-RADS classification, and nipple involvement. The odds ratio (OR) value of unclear tumor margin was 43.649, indicating that unclear tumor margin has the highest value in differentiating TNBC from non-TNBC; ipsilateral axillary lymph node metastasis and nipple involvement (OR value >5); the OR value of long axis was only 0.924, which has a lower correlation in differentiating TNBC from non-TNBC. Using the logistic regression model, the AUC was 0.907, and the prediction accuracy of TNBC was 0.887, sensitivity was 0.960, specificity was 0.750, PPV was 0.877, and NPV was 0.911. This indicates that the model has good predictive value for differentiating TNBC from non-TNBC. The multivariate analysis in this group suggests that unclear tumor margin is an independent risk factor for differentiating TNBC from non-TNBC. TNBC often grows in an invasive manner (11,12), and the tumor grows in all directions, resulting in irregular tumor shape and unclear tumor margin. Cancer cells intermingle with normal glandular tissue.

Regarding age distribution, our finding of similar mean ages between TNBC and non-TNBC groups (41.8 vs. 43.4 years, P=0.289) differs from some previous studies reporting that TNBC tends to occur in younger women (13). Several factors may account for this discrepancy. First, our single-center Chinese cohort may reflect regional demographic characteristics distinct from Western populations. Second, both groups in our study consisted predominantly of premenopausal women, which may have minimized intergroup age differences. Third, our non-TNBC group included all other molecular subtypes (luminal A, luminal B, HER2-enriched), whereas some studies have specifically compared TNBC to luminal A cancers, which typically occur in older women. Finally, our sample size may have been insufficient to detect small but clinically meaningful age differences. These considerations highlight the importance of multicenter studies with diverse populations to fully characterize the demographic features of breast cancer molecular subtypes.

Breast cancer with lymph node metastasis is not only an important factor affecting prognosis, but also a major factor in making clinical treatment decisions. It represents one of the research hotspots in recent artificial intelligence and genomics studies. Some studies have shown that by using artificial intelligence deep learning non-invasive imaging biomarkers, the degree of axillary lymph node metastasis in early breast cancer patients can be predicted (14,15). Other studies have shown that lymph node metastasis in breast cancer is related to its molecular subtype (16). The multivariate analysis in this study revealed ipsilateral axillary lymph node metastasis as an independent risk factor for differentiating TNBC from non-TNBC, with an OR value of 23.901, and has a predictive advantage. Ma et al. (17) reported that TNBC in younger patients is associated with higher rates of lymph node metastasis and distinct biological features compared to elderly patients. The aggressive biological behavior of TNBC, including its association with higher histological grade and lower hormone receptor expression, contributes to its increased propensity for lymph node involvement (18). The univariate analysis results of this study also show that axillary lymph node metastasis has certain predictive value (P<0.05).

As societal expectations evolve, increasing emphasis on both oncologic safety and cosmetic outcomes has made nipple-sparing surgical approaches more desirable among breast cancer patients; therefore, accurate preoperative assessment of the involvement of the nipple-areola complex (NAC) is necessary. The incidence of nipple involvement in the TNBC group (30/92, 32.6%) was substantially higher than that in the non-TNBC group (3/194, 1.5%), and the univariate analysis showed statistical significance, which is consistent with previous literature reports. The nipple involvement in this study mainly manifested as inward depression of the nipple, which is caused by the shortening of the mammary duct, suspension ligament or connective tissue due to the invasion of the tumor, and the multivariate analysis shows that it has certain reference value in the differential diagnosis of the two. At the same time, the univariate analysis also showed that the tumor diameter and the BI-RADS classification have statistical significance, and they show no statistical significance between age groups, and there is no advantage in prediction, which may be related to the small sample size of the case group in this study, and further prospective studies with a large sample size are needed.

The observed association between unclear tumor margins and TNBC can be explained by the distinct growth pattern of this molecular subtype. Pathologically, TNBC often demonstrates pushing borders with surrounding desmoplastic reaction and lymphocytic infiltration, rather than the well-circumscribed encapsulated growth seen in some luminal-type cancers (12). This infiltrative growth pattern disrupts the normal glandular architecture, creating imaging interfaces that appear spiculated or indistinct on MRI. Similarly, the higher incidence of nipple involvement in TNBC (32.6% vs. 1.5% in non-TNBC) reflects the tumor’s tendency for ductal spread and invasion of the NAC, consistent with its aggressive biological behavior and higher histological grade. Recent advances in radiomics and deep learning have shown promise in predicting TNBC molecular subtypes from medical images. For example, several studies have demonstrated that textural features extracted from DCE-MRI or DWI can achieve AUCs ranging from 0.85 to 0.94 for TNBC identification (19). Compared to these black box approaches, our logistic regression model offers distinct advantages in clinical interpretability. The explicit identification of four key imaging features—tumor margin, nipple involvement, lymph node status, and BI-RADS category—provides clinicians with transparent, understandable criteria that can be readily applied at the reading station without requiring specialized software or computational expertise. Furthermore, the strong performance of these conventional features (AUC =0.907) suggests that the predictive signal captured by complex radiomics signatures may, in part, reflect these same fundamental morphologic characteristics. However, radiomics approaches may detect subtle textural heterogeneity not apparent to the human eye, potentially offering complementary value. Future work could explore hybrid models combining interpretable clinical features with selected radiomics signatures to further improve diagnostic accuracy.

Regarding the TIC patterns, our univariate analysis showed no significant difference between TNBC and non-TNBC groups (P=0.309), with type II (plateau) and type III (washout) curves predominating in both groups. TIC patterns reflect hemodynamic characteristics such as tumor microvascular density, permeability, and perfusion (20). Although some studies have suggested that TNBC may more frequently demonstrate washout kinetics due to its high angiogenic activity (21), our findings indicate that TIC patterns alone may not reliably distinguish molecular subtypes, as considerable overlap exists among different breast cancer subtypes. This is consistent with previous research showing that enhancement kinetics are influenced by multiple factors including tumor grade, stromal composition, and necrotic components, which may vary independently of receptor status (22). The lack of significant association in our study underscores that TIC patterns should be interpreted in conjunction with morphological features rather than as standalone predictors of molecular subtype.

This study has several limitations that should be acknowledged. First, as a single-center retrospective study, there is potential for selection bias, and the findings may not be generalizable to other populations or institutions. Although 10-fold cross-validation demonstrated internal consistency, the lack of external validation in an independent cohort limits the robustness and generalizability of our model. Second, the relatively small sample size, particularly for the TNBC group (n=92), may have limited statistical power and precluded subgroup analyses. Third, we did not include certain potentially informative imaging features such as pectoralis muscle involvement, tumor vascularity, quantitative perfusion parameters, or texture features derived from radiomics analysis. Fourth, some pathological diagnoses were based on core needle biopsy rather than surgical specimens, which may introduce sampling error. Fifth, the numerical coding of BI-RADS categories [1–7] as an ordinal variable and subsequent dichotomization at BI-RADS 5, although necessary for model building, may not fully capture the clinical nuances of this classification system. Alternative approaches, such as treating BI-RADS subcategories as separate dummy variables or using data-driven ordinal collapsing methods, could be explored in future studies with larger sample sizes. Sixth, although we did not apply Bonferroni correction for multiple comparisons in the multivariable logistic regression, the variables were selected based on a priori univariate screening (P<0.05), and the sample size provided adequate events per variable (EPV ≈15) to support model stability. Nonetheless, future validation studies are needed to confirm the robustness of these predictors. Finally, the moderate specificity (75%) at the optimal threshold suggests that the model may be better suited for ruling in TNBC than for ruling it out. Future multicenter prospective studies with larger, more diverse populations are warranted to externally validate our findings, incorporate additional imaging and clinical variables, and further refine the model’s clinical utility.

Conclusions

Multivariable logistic regression analysis identified four key MRI features for differentiating TNBC from non-TNBC: unclear tumor margins, nipple involvement, BI-RADS classification, and ipsilateral axillary lymph node metastasis. The nomogram developed in this study provides a practical tool for clinicians to estimate individual patient risk of TNBC preoperatively, facilitating personalized treatment planning. These findings have potential clinical utility and warrant broader validation. It is important to emphasize that these MRI features serve as imaging surrogates and are not intended to replace pathological assessment, which remains the gold standard for molecular subtyping and definitive diagnosis.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Medical Ethics Committee of Yantaishan Hospital (No. YTSH-202411-19) and written informed consent was obtained from all participants.

Data Sharing Statement

Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1924/dss

qims-16-08-631-dss.pdf ^{(74.8KB, pdf)}

DOI: 10.21037/qims-2025-1924