Development and Validation of a Combined 2D-US and CDUS Nomogram for Differentiating Benign and Malignant Breast Masses: A Retrospective Study

Introduction

Breast masses are the most common clinical manifestation of breast diseases, and accurate differentiation between benign and malignant lesions is critical for guiding subsequent diagnostic and therapeutic strategies.¹ With the steadily increasing incidence of breast cancer worldwide, achieving high diagnostic accuracy while minimizing unnecessary invasive procedures has become a major focus in breast imaging.^2,3 Ultrasonography, characterized by its noninvasiveness, convenience, and high repeatability, is widely applied in the screening, diagnosis, and follow-up of breast diseases, playing a key role in the initial evaluation of breast masses.^4,5

Two-dimensional ultrasonography (2D-US) provides essential imaging information for distinguishing benign from malignant masses through the assessment of morphology, boundary, margin, internal echogenicity, and posterior acoustic features. Framework such as the Breast Imaging Reporting and Data System (BI-RADS) classification system have been developed to standardize evaluation.⁶ However, despite its widespread use, BI-RADS is limited by interobserver variability, particularly in intermediate categories, and may lead to unnecessary biopsies or inconsistent risk estimation in clinical practice. Color Doppler ultrasonography (CDUS) offers complementary information by depicting lesion vascularity, which may reflect biological behavior and aggressiveness.^3,7 Nevertheless, previous studies have demonstrated that relying on a single ultrasonographic feature or modality has inherent limitations in differentiating benign from malignant masses, and cannot fully facilitate individualized risk assessment.^8,9

Recent advances have attempted to improve diagnostic accuracy by integrating multiple ultrasonographic features and developing predictive models, including machine learning, radiomics, and nomograms.¹⁰ However, there remains a lack of standardized approaches regarding variable selection, statistical methodology, and model construction. Many studies are limited to univariate analyses or simple combinations of imaging features without systematically evaluating the independent contribution of each characteristic.¹¹ Furthermore, predictive models based on combined imaging features often lack adequate calibration and external validation, limiting their clinical applicability in risk stratification.

Given this context, there is a clear gap in systematically integrating 2D-US and CDUS features using a rigorous statistical framework, to identify imaging factors independently associated with malignancy, and to develop a predictive model with robust diagnostic performance and reliability. We hypothesize that combining conventional 2D-US and CDUS features into a multivariable predictive nomogram will improve differentiation between benign and malignant breast masses and provide a practical tool for individualized malignancy risk assessment. Therefore, this study retrospectively analyzed the 2D-US and CDUS features of benign and malignant breast masses, identified independent imaging predictors, and developed a combined imaging-based predictive nomogram to evaluate malignancy risk.

Materials and Methods

General Information

This was a single-center retrospective observational study. Clinical and imaging data were retrospectively collected from patients with breast masses who underwent 2D-US and CDUS at our hospital between May 2023 and June 2025. All patients obtained definitive pathological diagnoses within 4 weeks after imaging examination. All cases were confirmed by surgical pathology or core needle biopsy pathology as the reference standard. A total of 258 female patients were included, comprising 177 cases of benign breast masses and 81 cases of malignant breast masses.

Inclusion criteria: (1) a definite mass lesion detected on breast ultrasonography; (2) definitive pathological diagnosis obtained within 4 weeks after ultrasonographic examination; and (3) complete imaging data with adequate image quality for diagnostic evaluation and analysis.

Exclusion criteria: (1) prior history of breast surgery, radiotherapy, or chemotherapy; (2) pregnancy or lactation; (3) multiple breast masses with imaging findings that could not be individually matched with pathological results; and (4) presence of significant inflammation, traumatic lesions, or other conditions potentially affecting imaging interpretation. A flowchart of patient inclusion and exclusion is provided in Figure 1.

Figure 1 Flowchart of patient selection and study enrollment.

Ultrasonographic Examination Methods

All patients underwent breast ultrasonography using high-frequency linear array transducers. The equipment included the Philips EPIQ 7 color Doppler ultrasound diagnostic system (Philips Healthcare, the Netherlands) and the GE LOGIQ E9 color Doppler ultrasound diagnostic system (GE Healthcare, United States). The probe frequencies ranged from 7 to 12 MHz. During examination, patients were positioned in the supine or oblique supine position, with full exposure of both breasts and the axillary regions. Bilateral breasts were systematically scanned in multiple planes. Under 2D-US mode, we recorded the location, number, and maximum diameter of each breast mass. The following features were evaluated systematically: morphology, boundary, margin characteristics, internal echogenicity, aspect ratio (anteroposterior diameter to transverse diameter), posterior acoustic changes, and calcification status. All masses were categorized according to the American College of Radiology Breast Imaging Reporting and Data System (BI-RADS) ultrasonography classification standard.¹² For statistical analysis, categories were grouped as 3–4A and 4B–5 according to the study design.

After completion of the 2D-US assessment, the examination was switched to CDUS mode. Under settings optimized for detecting low-velocity blood flow (low wall filter and low pulse repetition frequency), intralesional and perilesional vascular signals were evaluated.

Observational Indicators

The 2D-US variables included: (1) Morphology: regular or irregular; (2) Boundary: circumscribed or indistinct; (3) Margin: smooth or lobulated/spiculated; (4) Internal echogenicity: homogeneous or heterogeneous; (5) Aspect ratio: ≤1 or >1 (vertical diameter to transverse diameter); (6) Posterior acoustic changes: enhancement, no significant change, or attenuation; (7) Calcification: absence of calcification or presence of microcalcification; (8) CDUS variables included the presence or absence of vascular signals, vascular signal grading, and vascular distribution pattern.

Vascular grading was performed according to the Adler blood flow grading system, classifying vascularity into grades 0–III. For statistical analysis, grades were grouped as 0–I and II–III. The vascular distribution pattern was categorized as peripheral or central/mixed based on the intratumoral distribution of blood flow signals. All ultrasonographic images were independently interpreted by two physicians with more than 5 years of experience in breast ultrasonographic diagnosis, both blinded to the pathological results. In cases of disagreement, a consensus was reached through joint review and discussion, and the consensus result was used for final analysis. Interobserver agreement for ultrasonographic feature assessment was evaluated using the intraclass correlation coefficient (ICC). The overall ICC was 0.784.

Statistical Analysis

Statistical analyses were performed using SPSS version 26.0 (IBM Corp., United States) and R software (version 4.2.2; R Foundation for Statistical Computing, Austria). After testing for normality, continuous variables that conformed to a normal distribution were expressed as mean ± standard deviation , and comparisons between groups were performed using the independent samples t-test. Variables that did not conform to a normal distribution were expressed as median [M (P₂₅, P₇₅)], and comparisons between groups were conducted using the Mann–Whitney U-test. Categorical variables were presented as counts and percentages [n (%)], and intergroup comparisons were performed using the χ²-test. The benign or malignant outcome of breast masses was defined as the dependent variable. Variables with P<0.05 in the univariate analysis were entered into the multivariate logistic regression model to identify independent predictors of malignancy. Odds ratios (OR) and their 95% confidence intervals (CI) were calculated. Based on the multivariate logistic regression model, a nomogram model for predicting the risk of malignancy in breast masses was constructed using the “rms” package in R software. Receiver operating characteristic (ROC) curve analysis was then performed to evaluate the diagnostic performance of individual imaging indicators and the combined model. The area under the curve (AUC), sensitivity, specificity, and optimal cutoff value were calculated. Model calibration was assessed using the Hosmer–Lemeshow goodness-of-fit test, and calibration curves were generated. Internal validation was conducted using bootstrap resampling to evaluate the agreement between predicted and observed probabilities. A two-sided P<0.05 was considered statistically significant.

Results

Comparison of Baseline Characteristics

The results showed that patients in the malignant group were significantly older than those in the benign group (t=5.15, P<0.001), and the proportion of postmenopausal patients was significantly higher in the malignant group (χ²=6.11, P=0.013). In addition, the distribution of BI-RADS categories differed significantly between the two groups (χ²=3.98, P=0.046). No statistically significant differences were observed between the two groups in terms of body mass index (BMI), menopausal status, family history of breast cancer, lesion laterality, or lesion number (all P>0.05). Different pathological types in the benign and malignant groups were described using descriptive statistics only and were not included in the intergroup comparative analysis (Table 1).

Table 1 Comparison of Baseline Characteristics [, n (%)]

Comparison of 2D-US Features Between Benign and Malignant Masses

Significant differences were observed between the two groups regarding morphology, boundary, margin, and calcification (χ²=45.32, P=0.001; χ²=59.83, P<0.001; χ²=60.81, P<0.001, respectively). Compared with the benign group, malignant masses more frequently presented with irregular morphology, indistinct boundaries, lobulated or spiculated margins, and microcalcifications. No statistically significant differences were found between the two groups in terms of morphology, internal echogenicity, aspect ratio, or posterior acoustic changes (all P>0.05) (Table 2). Representative grayscale ultrasound images of benign and malignant breast lesions are shown in Figure 2.

Table 2 Comparison of 2D-US Features Between Benign and Malignant Masses [n (%)]

Figure 2 Representative grayscale ultrasound images of benign and malignant breast lesions. (A) A 35-year-old woman with a benign fibroadenoma confirmed by postoperative pathology. Grayscale ultrasound shows an oval-shaped mass with a regular shape, smooth margins, well-defined borders, and posterior acoustic enhancement. (B) A 69-year-old woman with invasive carcinoma confirmed by postoperative pathology. Grayscale ultrasound demonstrates an irregular mass with lobulated and angulated margins, partially indistinct borders, multiple punctate hyperechoic foci within the lesion, and localized posterior acoustic attenuation.

Comparison of CDUS Features Between Benign and Malignant Masses

No significant difference was observed between the two groups in vascular distribution pattern (χ²=1.59, P=0.21). Although the proportion of masses with detectable vascular signals was higher in the malignant group than in the benign group, the difference did not reach statistical significance (χ²=2.30, P=0.13). In contrast, a significant difference was observed in Adler grade between the two groups (χ²=11.41, P=0.001), with a markedly higher proportion of grade II–III vascular signals in the malignant group (Table 3). Representative CDUS features of benign and malignant breast lesions are presented in Figure 3.

Table 3 Comparison of CDUS Features Between Benign and Malignant Masses [n (%)]

Figure 3 Representative CDUS features of benign and malignant breast lesions. (A) A benign fibroadenoma showing minimal internal vascularity, corresponding to Adler grade I. (B) An invasive carcinoma demonstrating abundant intralesional blood flow with central vascular distribution, corresponding to Adler grade III.

Multivariate Logistic Regression Analysis and Nomogram Construction Based on 2D-US Combined with CDUS

Variables that were statistically significant in the univariate analysis were entered into a multivariate logistic regression model. The results demonstrated that age, maximum tumor diameter, BI-RADS category, tumor shape, boundary, margin, and Adler grading were independent predictors of malignant outcomes. Specifically, age (OR = 1.135, 95% CI: 1.049–1.228, P=0.002) and maximum tumor diameter (OR = 1.357, 95% CI: 1.127–1.634, P=0.001) were positively associated with malignancy. Regarding imaging features, a higher BI-RADS category (OR = 2.897, 95% CI: 1.222–6.864, P=0.016), irregular tumor shape (OR = 5.208, 95% CI: 2.243–12.093, P<0.001), indistinct boundaries (OR = 11.400, 95% CI: 4.553–28.546, P<0.001), and lobulated or spiculated margins (OR = 6.885, 95% CI: 2.976–15.928, P<0.001) were all significant predictors of malignancy. A higher Adler grade was identified as an independent risk factor (OR = 2.473, 95% CI: 1.096–5.576, P=0.029). In contrast, calcification was not significant (OR = 0.360, 95% CI: 0.121–1.074, P=0.067) (Table 4).

Table 4 Multivariate Logistic Regression Analysis of Factors Associated with Malignant Outcome of Breast Masses

Based on the above multivariate logistic regression model, a nomogram was constructed to predict the risk of malignancy in breast masses (Figure 4). The nomogram integrates age, maximum tumor diameter, BI-RADS category, tumor shape, boundary, margin, calcification status, and Adler grade. Different weights were assigned to each predictor to enable visualized assessment of individual malignancy risk. As illustrated in the nomogram, the contributions of individual variables differ within the model. Indistinct boundaries, lobulated or spiculated margins, and irregular shape correspond to higher point values, indicating greater weights in malignancy risk prediction. BI-RADS category and Adler grade also demonstrate good discriminatory ability. With increasing age and tumor diameter, the corresponding scores gradually rise, suggesting a progressive increase in malignancy risk. In practical application, the points assigned to each variable are summed to obtain a total score for an individual patient. This total score is then mapped onto the probability scale at the bottom of the nomogram to estimate the likelihood of malignancy in breast masses, thereby enabling individualized quantitative risk prediction.

Figure 4 Nomogram based on multivariate logistic regression for predicting malignancy risk of breast masses.

Diagnostic Performance of 2D-US Combined with CDUS in Differentiating Benign and Malignant Masses

ROC curve analysis demonstrated that the discriminatory ability of individual ultrasound features for differentiating benign and malignant breast masses varied. Age (AUC = 0.692) and maximum tumor diameter (AUC = 0.674) showed moderate diagnostic performance. Tumor shape (AUC = 0.721), boundary (AUC = 0.756), and margin (AUC = 0.760) exhibited relatively good discriminatory ability. In contrast, BI-RADS category (AUC = 0.565) and Adler grade (AUC = 0.612) showed comparatively limited diagnostic performance when used alone. The combined imaging-based predictive model constructed from the above variables demonstrated significantly superior diagnostic performance compared with any single parameter, with an AUC of 0.926 (95% CI: 0.892–0.959), indicating excellent discriminatory ability. At the optimal cut-off value of 0.354, the model achieved a sensitivity of 83.95% and a specificity of 89.27%, reflecting robust diagnostic performance (Table 5 and Figure 5).

Table 5 ROC Curve Analysis of the 2D-US Combined with CDUS Model for Differentiating Benign and Malignant Breast Masses

Figure 5 ROC curves of the predictive model for malignant breast lesions.

Calibration Performance and Internal Validation of the Combined Imaging Model

The Hosmer–Lemeshow goodness-of-fit test indicated no significant difference between predicted and observed probabilities of malignancy (χ²=4.75, P=0.784), suggesting good model fit. The calibration curve showed good agreement between predicted and observed probabilities after bootstrap resampling correction. Among the 258 cases, the mean absolute error between predicted and observed probabilities was 0.003, and the mean squared error was 0.00157. In 90% of cases, the absolute prediction error did not exceed 0.061, indicating a small overall deviation between predicted and actual malignancy probabilities (Figure 6).

Figure 6 Calibration curve of the combined imaging model.

Discussion

Based on single-center retrospective data, this study systematically analyzed the differences in imaging features of benign and malignant breast masses on 2D-US and CDUS, and developed a combined imaging predictive model (nomogram) to assess individualized malignancy risk. The results demonstrated that age, maximum tumor diameter, BI-RADS category, tumor morphology, boundary, margin, and Adler grade remained significantly associated with malignant outcomes in multivariate analysis. The nomogram constructed from these variables showed favorable discriminative ability (AUC = 0.926) and favorable calibration performance, indicating its potential utility as a clinical risk assessment tool. These findings support the value of integrating multidimensional US features for quantitative breast mass risk stratification. While advanced computational approaches such as radiomics and machine learning have been explored in breast imaging, our study focused on routinely obtainable 2D-US and CDUS features to provide a practical and interpretable predictive tool.

Regarding morphology on 2D-US, malignant masses more frequently presented with indistinct boundaries, lobulated or spiculated margins, and microcalcifications. Multivariate analysis confirmed that irregular morphology, indistinct boundaries, and lobulated or spiculated margins were independent predictors, indicating stable predictive value beyond univariate comparisons. These findings are consistent with previously reported typical malignant imaging signs.¹³ Indistinct boundaries and spiculated margins reflect infiltrative tumor growth, whereas lobulated or spiculated contours may be associated with stromal reaction and fibrous proliferation.¹⁴ Microcalcifications were also more common in malignant masses and may relate to tumor necrosis, calcium salt deposition, or intraductal lesions.¹⁵

CDUS results showed that malignant masses had significantly higher Adler grades than benign lesions, which remained an independent predictor in multivariate analysis. Although a direct causal link with tumor angiogenesis cannot be established, this association suggests that quantitative blood flow grading provides incremental predictive value. Tumor-associated angiogenesis is a fundamental process underlying tumor growth and invasion, and malignant tumors are often accompanied by increased vascularity and enhanced blood flow signals.¹⁶

In multivariate analysis, BI-RADS category demonstrated strong predictive weight, reflecting its established role in breast imaging evaluation. As BI-RADS integrates multiple imaging features, its inclusion in the nomogram helps enhance individualized risk stratification when combined with quantitative morphological and vascular characteristics.¹⁷ Incorporating structured imaging features into the model based on BI-RADS classification may further improve predictive precision and clinical applicability.

The combined imaging model constructed in this study demonstrated good discriminative performance in ROC analysis, and both the calibration curve and the Hosmer–Lemeshow test indicated good agreement between predicted and observed probabilities. Compared with studies focusing solely on AUC, this study systematically evaluated calibration performance, which is critical for clinical applicability.¹⁸ Internal validation using bootstrap resampling confirmed the consistency of model predictions, and the number of malignant events provided an adequate events-per-variable ratio for the multivariate logistic regression analysis. From a practical perspective, such a model may support decision-making in lesions located within a diagnostic “gray zone,” guiding the need for biopsy or follow-up without increasing examination burden.

From a clinical perspective, these findings suggest that routinely obtainable information from 2D-US and CDUS, when appropriately integrated and modeled, may provide supportive evidence for individualized assessment of breast mass malignancy risk. For lesions located within a diagnostic “gray zone,” the combined model may assist in decision-making regarding invasive procedures or follow-up strategies. It should be emphasized that this model is not intended to replace pathological diagnosis but rather to serve as an adjunctive tool in clinical decision-making, improving objectivity and consistency in risk evaluation without increasing examination burden.

Several limitations should be acknowledged. First, the retrospective design and single-center setting may introduce selection bias related to patient composition and imaging interpretation patterns, and caution is warranted when generalizing these findings. Second, although independent double reading was performed to reduce subjective bias, assessment of imaging features remains partially operator-dependent. Third, the sample size in this study was relatively limited, and the estimation of some variables may therefore be subject to a certain degree of instability. Although relevant clinical factors such as menopausal status and family history were included in the analysis, other potential confounders not captured in this study may still have influenced the observed associations and model performance. Fourth, the predictive model was only internally validated using bootstrap resampling; external validation in independent cohorts is needed to confirm its generalizability and stability across different populations, imaging equipment, and examination conditions. Accordingly, the reported predictive performance should be interpreted with appropriate caution until external validation is available. Future multicenter, prospective studies with larger sample sizes and external validation are warranted to further evaluate the robustness and clinical applicability of the model. In addition, the integration of radiomics or artificial intelligence approaches may further enhance feature extraction and risk prediction performance.

In summary, by systematically integrating 2D-US and CDUS features, this study identified independent imaging predictors associated with malignant outcomes and developed a combined imaging predictive model for differentiating benign and malignant breast masses. Although further validation is required, these findings provide a practical risk assessment framework based on routine imaging information and lay the groundwork for future research in individualized breast mass evaluation.

Conclusion

This study demonstrated that 2D-US combined with CDUS features has clinical value in differentiating benign and malignant breast masses. The multivariable-based combined imaging predictive model showed favorable diagnostic performance and calibration. As an adjunctive tool, it may assist individualized malignancy risk assessment, particularly for diagnostically indeterminate lesions. Further multicenter, prospective validation is warranted to confirm its stability and generalizability.