Clinical evaluation of a nomogram model incorporating multimodal ultrasound features for breast cancer diagnosis

Introduction

According to global cancer statistics from 2024, breast cancer ranks as the second most common malignancy worldwide and remains the leading cause of cancer-related mortality among women (1). The condition is marked by a rising incidence, aggressive biological behavior, and a trend toward earlier age at diagnosis. Breast cancer exhibits a broad spectrum of sonographic presentations, with considerable overlap between features typically associated with benign and malignant lesions. High-frequency ultrasonography is commonly employed as the initial screening and diagnostic modality for breast disease, as it allows for detailed assessment of lesion morphology, internal structure, and parenchymal alterations in adjacent tissues. Color Doppler flow imaging (CDFI) is utilized to visualize macrovascular flow patterns within lesions, while ultrasound elastography (UE) assesses the relative stiffness of breast tissue. Contrast-enhanced ultrasound (CEUS) offers insight into microvascular architecture, vascular pathways, and hemodynamic characteristics within lesions (2).

Each ultrasound technique contributes unique diagnostic insights, although each also has inherent strengths and limitations. When combined, these multimodal imaging techniques yield complementary data that enhance the differential diagnosis of breast cancer. However, the wide array of sonographic parameters and the potential for conflicting interpretations across modalities emphasize the need for a simplified, standardized diagnostic approach. Logistic regression analysis offers an effective method for adjusting for confounding variables and identifying imaging features with the strongest associations and highest diagnostic utility (3). Nomograms represent a graphical means of expressing logistic regression outputs, quantifying the predictive value of individual variables. Against the backdrop of artificial intelligence showing great potential in the computer-aided diagnosis of breast cancer, this study aims to provide clinicians with a practical, intuitive tool for estimating the likelihood of malignancy in breast lesions (4-6).

In this context, the present investigation applied binary logistic regression analysis to imaging features obtained from B-mode ultrasound, CDFI, UE, and CEUS to identify independent predictors of breast cancer and develop a corresponding nomogram model. The overarching objective was to support accurate early diagnosis, thereby contributing to improved prognosis and quality of life for patients with breast cancer. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-471/rc).

Methods

Study population

A retrospective analysis was conducted on 141 female patients treated at Chaoyang Central Hospital, China Medical University and the First Hospital of Shanxi Medical University between September 2016 and February 2023. A total of 157 breast lesions, confirmed histopathologically via core needle biopsy or surgical excision, were included in the study. The lesions were categorized into a training cohort (n=116), evaluated using the GE LOGIQ E9 ultrasound system (Milwaukee, Wisconsin, USA), and a validation cohort (n=41), assessed with Canon Aplio 500 and Aplio i800 ultrasound systems (Ohtawara, Tochigi, Japan). Patient age ranged from 21 to 76 years, with a mean age of 45.8±11.4 years. All cases completed B-mode ultrasound, CDFI, UE, and CEUS examinations with sufficient image quality for analysis; thus, no missing data were present. Inclusion criteria were as follows: (I) lesions classified as Breast Imaging Reporting and Data System (BI-RADS) categories 3 to 5 on conventional ultrasound; and (II) individuals presenting for initial evaluation without any prior therapeutic intervention. Exclusion criteria were (I) inadequate imaging quality and (II) contraindications to CEUS. The study protocol received approval from the Ethics Committee of First Hospital of Shanxi Medical University (No. KYLL-2024-361), and written informed consent was obtained from all patients. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The other hospital was informed of and agreed to the study.

Equipment and imaging procedures

Ultrasound examinations were performed using a GE LOGIQ E9 system equipped with 9L and ML6-15 transducers, a Canon Aplio 500 with a 14L5 linear array transducer, and an Aplio i800 with an i18LX5 linear array transducer. The ultrasound contrast agent used was SonoVue (Bracco, Italy).

All patients were arranged in the supine posture with arms positioned away from the body to ensure optimal exposure of the breast and axillary regions. Patient age was recorded prior to imaging. B-mode ultrasound was initiated at the upper outer quadrant and followed a clockwise radial scanning protocol centered on the nipple. Lesions were assessed in multiple planes, and lesion size, anatomical location, shape, orientation, margin characteristics, internal echogenicity, posterior acoustic features, and the presence of calcifications were all documented. Bilateral axillary regions were also examined for lymph node abnormalities. Subsequently, CDFI and UE were conducted to assess intralesional vascularity and tissue stiffness, respectively. For CEUS, the imaging plane demonstrating the highest vascular density was selected to include both the lesion and adjacent normal tissue. During CEUS, patients maintained a stable position while breathing quietly. A bolus of 4.8 mL of contrast agent was administered via the cubital vein, followed by a 5 mL flush of 0.9% saline. Dynamic enhancement patterns were recorded continuously for 3 minutes with the transducer held in a fixed position.

Image analysis and variable assignment

All images were independently interpreted in a blinded manner by two experienced sonologists with substantial expertise in both conventional ultrasound and CEUS. B-mode ultrasound features were categorized in accordance with the BI-RADS 5th edition lexicon (7). Vascularity was graded using the Adler semi-quantitative method, with CDFI grades 0 and I indicative of benignity, and grades II and III indicative of malignancy (8). Tissue stiffness was assessed using the UE scoring system, wherein scores of 1 to 3 were associated with benign lesions and scores of 4 to 5 with malignancy (9). Discrepancies in interpretation were resolved through consensus between the reviewers.

Patient age and sonographic variables from each imaging modality were coded according to the definitions presented in Table 1.

Statistical analysis

Statistical analysis was performed using SPSS version 20.0. The nomogram was developed in R (version 4.4.2; R Core Team, 2024) using RStudio (version 2024.12.0+467, Posit PBC). Key R packages included: rms (v6.8.2) for logistic regression modeling, nomogram, and calibration curves; pROC (v1.18.5) for receiver operating characteristic (ROC) analysis and area under the curve (AUC) calculation; rmda (v1.6) for decision curve analysis (DCA); Hmisc (v5.2.0) for data summarization; car (v3.1.3) for variance diagnostics; and MASS (v7.3.61) for statistical modeling. Continuous variables were expressed as mean ± standard deviation (SD), while categorical variables were compared using the Chi-squared (χ²) test. Within the training cohort, multivariable logistic regression analysis with forward selection was applied. Pathological diagnosis served as the dependent variable, and variables identified as statistically significant in univariate analysis were entered as covariates to determine independent predictors of malignancy. A diagnostic nomogram incorporating multimodal ultrasound features was developed using R software version 4.4.2. ROC curves were generated for both the training and validation cohorts to evaluate model performance, and the corresponding AUCs were calculated. Model discrimination was further assessed using the concordance index (C-index), while calibration curves were used to evaluate the agreement between predicted and observed probabilities. DCA was conducted to assess the clinical applicability of the model. A two-sided P value <0.05 was considered to indicate statistical significance.

Results

Pathological findings

Among the 157 breast lesions, 72 were confirmed as malignant. These included 51 invasive carcinomas of no special type (NST), 9 ductal carcinomas in situ (DCIS), 8 invasive carcinomas of other types, 2 mucinous carcinomas, 1 invasive tubular carcinoma, and 1 solid papillary carcinoma with invasion. The remaining 85 lesions were benign and consisted of 37 cases of fibrocystic changes or adenosis, 19 fibroadenomas, 11 fibrocystic changes with fibroadenoma formation, 6 inflammatory lesions, 3 fibrocystic changes with an inflammatory reaction, 3 intraductal papillomas, 3 cases of mammary duct ectasia, 1 plasma cell mastitis, 1 case of sclerosing adenosis, and 1 fibrous nodule with chronic inflammation.

Univariate analysis of multimodal ultrasound features

In the training cohort, univariate analysis identified 16 variables that demonstrated statistical significance (P<0.05). These included patient age; B-mode ultrasound features (lesion morphology, orientation, margin characteristics, posterior acoustic features, presence of microcalcifications, and lymph node status); CDFI grades; UE scores; and CEUS parameters (enhancement pattern, enhancement intensity, contrast distribution, post-contrast lesion boundaries, morphology, size changes, and presence of penetrating vessels or radial enhancement) (Table 2).

Multivariate analysis of multimodal ultrasound features

Forward stepwise logistic regression analysis was performed on the 16 significant variables identified in the training cohort. Six independent predictors of malignancy were identified in descending order of odds ratios (ORs): lesion size expansion on CEUS, non-circumscribed margins, penetrating vessels or radial enhancement patterns, heterogeneous contrast agent distribution, CDFI grade II/III, and UE score of 4 to 5 (Table 3).

Development and validation of the nomogram

A diagnostic nomogram was constructed using the six independent predictors derived from multivariate logistic regression analysis. Each variable was assigned a line segment proportional in length to its relative weight in the model, and total scores were used to estimate the probability of malignancy (Figure 1).

Figure 1 Nomogram for predicting the probability of malignancy in breast lesions categorized as BI-RADS 3–5. To use the nomogram, locate the patient’s value for each predictor variable on its corresponding axis and draw a vertical line to the top “Points” scale to determine the number of points assigned. Sum all points and locate this total on the “Total Points” axis. Draw a vertical line down to the “Predicted Probability of Malignancy” axis to obtain the estimated risk. BI-RADS, Breast Imaging Reporting and Data System; CDFI, color Doppler flow imaging; UE, ultrasound elastography.

The nomogram demonstrated excellent discriminative ability in the training cohort, with an area under the ROC curve of 0.966 [95% confidence interval (CI): 0.939–0.992]. In the validation cohort, the AUC was 0.819 (95% CI: 0.687–0.952), indicating strong discriminative performance, albeit with a reduction in predictive accuracy (Figure 2A,2B).

Figure 2 ROC curve for the diagnostic performance of the nomogram. (A) ROC curve for the diagnostic performance of the nomogram in the training cohort. (B) ROC curve for the diagnostic performance of the nomogram in the validation cohort. AUC, area under the curve; CI, confidence interval; ROC, receiver operating characteristic.

Calibration curves indicated a high degree of agreement between predicted and observed probabilities in both the training and validation cohorts, with standard errors of 0.041 and 0.035, respectively, confirming the predictive accuracy of the nomogram (Figure 3A,3B).

Figure 3 Calibration curve for the nomogram. Calibration curves assessing the agreement between the nomogram’s predicted probabilities of malignancy and the actual observed outcomes. The diagonal reference line (dashed) represents perfect prediction. The solid line represents the performance of the nomogram model. (A) Calibration curve for the training cohort. (B) Calibration curve for the validation cohort.

DCA demonstrated favorable clinical utility of the model across a wide threshold probability range (2–99%), indicating that clinical decisions informed by the nomogram would likely result in improved outcomes (Figure 4A,4B).

Figure 4 Decision curve analysis of the nomogram. The x-axis represents the probability threshold for classifying a lesion as high-risk. The y-axis shows the standardized net benefit. The red line represents the nomogram model; the gray dashed line represents treating all patients as malignant (“All”); and the black solid line represents treating no one as malignant (“None”). (A) Decision curve analysis of the nomogram in the training cohort. (B) Decision curve analysis of the nomogram in the validation cohort.

In the training cohort, there were 53 malignant cases, and six variables were included in the final multivariable model, yielding an events-per-variable (EPV) ratio of approximately 8.83, slightly below the conventional threshold of 10 but still within an acceptable range. To mitigate overfitting risk, we employed forward stepwise selection and validated the model in an independent cohort acquired using different ultrasound systems, supporting its generalizability.

Discussion

Breast cancer demonstrates considerable heterogeneity in molecular subtypes, genetic alterations, biological behavior, therapeutic responses, and prognostic outcomes. This heterogeneity reflects the complex progression of the disease and presents significant challenges in clinical decision-making. Early detection, accurate diagnosis, and timely intervention remain essential strategies for improving clinical outcomes and reducing mortality.

Analysis of sonographic features revealed notable distinctions between malignant and benign breast lesions. Malignant lesions were more likely to exhibit irregular morphology, non-parallel orientation, and non-circumscribed margins—characterized as blurred, angular, spiculated, or microlobulated—along with posterior acoustic shadowing, microcalcifications, and abnormal axillary lymph nodes. UE findings frequently indicated high tissue stiffness, with scores of 4 to 5. Conversely, benign lesions were generally associated with regular morphology, oval shape, parallel orientation, circumscribed margins, unchanged or enhanced posterior acoustic features, and low UE scores [1–3]. These sonographic characteristics are consistent with the underlying histopathological features.

Invasive breast carcinomas often display infiltrative growth patterns, extending along connective tissue and adipose planes and leading to disruption and remodeling of surrounding tissues. These processes are reflected sonographically as irregular morphology, non-circumscribed margins, and echogenic halos. Cellular proliferation and stromal invasion are typically accompanied by desmoplastic reactions involving increased collagen deposition and altered extracellular matrix organization, resulting in greater tissue stiffness (10). This increased stiffness is manifested sonographically by acoustic shadowing and elevated UE scores. In contrast, benign breast lesions tend to exhibit expansile growth, as observed in fibrocystic changes and fibroadenomas, which were prevalent in the present cohort. Adenosis is histologically characterized by an increased number of terminal ductules and acini with minimal collagen content (11). Fibroadenomas, composed of epithelial and stromal elements, are generally encapsulated within a thin fibrous layer (12). These structural features contribute to the regular morphology, well-demarcated margins, and lower tissue stiffness typically observed in benign lesions, along with unchanged or enhanced posterior acoustic properties.

Despite these characteristic patterns, diagnostic challenges remain due to the broad histopathological spectrum of benign breast conditions. Fibroadenomas may undergo secondary changes such as apocrine metaplasia, cystic degeneration, calcification, and hemorrhagic necrosis, which can result in complex and atypical sonographic appearances. Additionally, benign lesions with extensive fibrotic components may exert compressive effects on surrounding ductal structures, leading to increased tissue stiffness that may mimic the sonographic features of malignant lesions (13,14).

Vascular evaluation of breast lesions revealed distinct patterns associated with malignancy and benignity. Malignant lesions were frequently associated with Adler grades II/III on CDFI and exhibited hyperenhancement on CEUS, characterized by centripetal filling, ill-defined post-contrast margins, irregular morphology with lesion enlargement, and the presence of feeding vessels or radial enhancement. In contrast, benign lesions typically demonstrated Adler grades 0/I on CDFI, with CEUS findings including hypoenhancement or isoenhancement, homogeneous contrast distribution, well-defined lesion margins, regular morphology without significant dimensional changes, and infrequent detection of feeding vessels or radial enhancement.

These imaging patterns are consistent with the underlying pathophysiological mechanisms. Breast carcinomas are characterized by sustained hypermetabolic activity and tumor-associated angiogenesis, predominantly mediated by upregulated expression of vascular endothelial growth factor (VEGF). This process results in the formation of abnormal neovasculature with thin walls, incomplete basement membranes, and arteriovenous shunts (15). On CDFI, this vascular architecture appears like disorganized or network-type signals, while CEUS reveals corresponding hyperenhancement. Additionally, breast cancer vasculature often exhibits branching at acute or perpendicular angles from parent arteries, with centripetal directional flow, observable on imaging as feeding vessels on CDFI and centripetal enhancement on CEUS (16). Vascular density is generally higher at the periphery of malignant tumors than in their central regions (17). Although B-mode ultrasound is limited in detecting microscopic tumor infiltration at lesion margins, CEUS facilitates visualization of ill-defined boundaries, irregular morphology, lesion size expansion, and radial enhancement—all indicative of peripheral neovascular invasion.

In contrast, benign lesions are associated with limited angiogenic activity and a more organized vascular architecture with sparse distribution. These vascular features are characterized by orderly branching and lower vascular density (18). Consequently, CDFI often reveals minimal or absent vascular signals, and CEUS findings typically include regular morphology, well-demarcated margins, and lesion dimensions consistent with B-mode ultrasound. Feeding vessels are rarely visualized in benign lesions.

Through logistic regression analysis incorporating age and multimodal sonographic features, six independent predictors of malignancy were identified. Ranked by descending OR, these predictors included lesion size expansion on CEUS, non-circumscribed lesion margins, presence of feeding vessels or radial enhancement, heterogeneous contrast enhancement, Adler grade II/III on CDFI, and UE scores of 4 to 5. A diagnostic nomogram was constructed based on these variables and demonstrated high discriminatory capacity in the training cohort (AUC =0.966), with calibration curves indicating strong concordance between predicted probabilities and observed outcomes. The model provided risk stratification across a wide range of predicted probabilities (2–99%). Clinicians can use the nomogram by assigning points based on a patient’s specific ultrasound features, summing the total score, and mapping it to the estimated probability of malignancy. This provides a quantitative reference to inform biopsy decisions for BI-RADS 3–5 lesions, serving as an adjunct—not a substitute—for comprehensive clinical evaluation. In the validation cohort, the model yielded a lower AUC of 0.819, reflecting good discriminative performance with preserved calibration. This reduction in diagnostic performance may be attributable to variations in ultrasound equipment used across cohorts, highlighting the need for additional external validation to evaluate generalizability across platforms.

In recent years, machine learning and deep learning methods have demonstrated great potential in automatically extracting complex features from multimodal ultrasound images and constructing high-accuracy predictive models. However, these approaches are often regarded as “black boxes”, with decision processes that are difficult to interpret. In contrast, the nomogram model developed in this study offers high interpretability and clinical transparency, with clearly defined weights for each predictor, facilitating clinician understanding, trust, and practical application in clinical decision-making. Future studies will explore machine learning and deep learning approaches to optimize multimodal ultrasound feature extraction and potentially combine them with nomogram frameworks to further improve diagnostic performance.

Limitations of this study

(I) To ensure adequate statistical power and sufficient sample size in each category, some sonographic features were grouped based on clinical experience. This may have obscured individual parameters with high diagnostic value, causing them to demonstrate statistical insignificance in the analysis. (II) Discrepancies in image interpretation were resolved by consensus without formally assessing interobserver agreement, which may have introduced variations in the findings. (III) This study is retrospective and subject to selection bias. Moreover, the training and validation cohorts were imaged using different ultrasound systems (GE vs. Canon), which may affect the reproducibility of CEUS features. (IV) The relatively small sample size highlights the need for future studies with larger cohorts and ongoing refinement of feature improve to enhance model performance.

Conclusions

The nomogram presents a practical and interpretable tool for differentiating between benign and malignant lesions classified as BI-RADS categories 3 to 5. Its application may improve efficiency in communication between technical and clinical teams and provide reliable imaging-based support for clinical decision-making.

Acknowledgments

We would like to acknowledge the hard and dedicated work of all the staff who implemented the intervention and evaluation components of the study.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-471/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-471/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-471/prf

Funding: This study was supported by the Research Project of the Health Commission of Shanxi Province (No. 2024066).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-471/coif). The authors have no conflicts of interest to declare.

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. This study was conducted with approval from the Ethics Committee of First Hospital of Shanxi Medical University (No. KYLL-2024-361). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Written informed consent was obtained from all participants.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Medical Policy and Administration Agency, National Health Commission of the People’s Republic of China. Breast cancer treatment guidelines(2022 edition). Zhonghua Zhong Liu Za Zhi 2023;45:803-33. [Crossref] [PubMed]
3. Shibahara T, Wada C, Yamashita Y, et al. Deep learning generates custom-made logistic regression models for explaining how breast cancer subtypes are classified. PLoS One 2023;18:e0286072. [Crossref] [PubMed]
4. Moldovanu S, Munteanu D, Biswas KC, et al. Breast Lesion Detection Using Weakly Dependent Customized Features and Machine Learning Models with Explainable Artificial Intelligence. J Imaging 2025;11:135. [Crossref] [PubMed]
5. Tăbăcaru G, Moldovanu S, Munteanu D, et al. New hybrid features extracted from US images for breast cancer classification. Sci Rep 2025;15:25690. [Crossref] [PubMed]
6. Liu Y, Zhu H, Yuan J, et al. A nomogram for predicting breast cancer based on hematologic and ultrasound parameters. Am J Transl Res 2023;15:5602-12.
7. American College of Radiology. Breast Imaging Reporting and Data System(BI-RADS), Atlas. 5th ed. Reston, VA: American College of Radiology; 2013.
8. Adler DD, Carson PL, Rubin JM, et al. Doppler ultrasound color flow imaging in the study of breast cancer: preliminary findings. Ultrasound Med Biol 1990;16:553-9. [Crossref] [PubMed]
9. Zhu Y, Jiang L, Chen YN, et al. The diagnostic value of real-time ultrasound elastography and contrast-enhanced ultrasound in BI-RADS 4A breast lesions. Gland Surg 2025;14:488-97. [Crossref] [PubMed]
10. Liu WL, Wang JJ, Zhan YY, et al. A Preliminary Study on the Clinical Value of nomogram based on Ultrasound Features and Reversal Score in Differentiating Breast Adenosis from small Breast Cancer. Chinese Journal of Ultrasound in Medicine 2024;(7):40, 737-41.
11. Chu RX, Peng M. The diagnostic value of BI-RADS classification combined with ultrasound elastography technology for breast cancer in young women. Chinese Journal of Ultrasound in Medicine 2023;39:765-8.
12. Han X. The value of high frequency color Doppler ultrasound in differential diagnosis of phyllodes tumor and fibroadenoma of the breast. Chinese Journal of Practical Medicine 2018;45:47-9.
13. Wang XR, Wang JJ, Zhan YY, et al. Construction and application of nomogram model for differentiating breast adenosis from small breast cancer based on multimodal ultrasound. Chinese Journal of Ultrasound in Medicine 2024;40:993-7.
14. Liang T, Cong S, Yi Z, et al. Ultrasound-Based Nomogram for Distinguishing Malignant Tumors from Nodular Sclerosing Adenoses in Solid Breast Lesions. J Ultrasound Med 2021;40:2189-200. [Crossref] [PubMed]
15. de Heer EC, Jalving M, Harris AL. HIFs, angiogenesis, and metabolism: elusive enemies in breast cancer. J Clin Invest 2020;130:5074-87. [Crossref] [PubMed]
16. Wang C, Che Y. A ultrasonic nomogram of quantitative parameters for diagnosing breast cancer. Sci Rep 2023;13:12340. [Crossref] [PubMed]
17. Chong L, Xia LH, Li HY. The characteristics of tumor-associated macrophages in different regions of contrast-enhanced ultrasound of breast cancer and their relationship with clinicopathology. Chinese Journal of Medical Imaging 2025;33:158-64.
18. Zheng X, Li F, Xuan ZD, et al. Combination of shear wave elastography and BI-RADS in identification of solid breast masses. BMC Med Imaging 2021;21:183. [Crossref] [PubMed]