Application of multimodal ultrasound in risk assessment of breast masses: construction and validation of the BI-RADS prediction model

Abstract

Background

Despite the fact that the diagnosis and treatment levels of breast cancer in first-tier cities have reached international standards and artificial intelligence (AI)-assisted diagnosis has been introduced in some regions, economically underdeveloped areas and primary hospitals still face significant bottlenecks in diagnosis and treatment. Standardized risk stratification tools that are tailored to the Chinese population are not yet in common use. Ultrasound, the primary equipment for community breast cancer screening in China, was adopted to develop a simplified quantitative Chinese ultrasound BI-RADS (Breast Imaging-Reporting and Data System) risk stratification model based on the epidemiological features of Chinese breast diseases for precise screening.

Methods

Data of 1,716 patients who underwent ultrasound examination at Ruijin Hospital from August 2019 to September 2020 were collected. The most suspicious lesion identified in each patient was selected for analysis. Taking the ultrasound features of breast masses as independent variables and the pathological benign/malignant status of breast masses as dependent variables, multivariate logistic regression analysis was performed. The regression coefficients were converted to integers as the assigned scores. The total score of each mass was obtained by summing the scores of all indicators, and the malignant rate of masses corresponding to each total score was calculated. The construction of the clinical prediction model for BI-RADS classification involved the implementation of stratification, which was based on the malignant rate of each total score.

Results

The following indicators were finally included in the multimodal ultrasound BI-RADS model: ill-defined margin (1 point), spiculation (2 points), microcalcification (2 points), irregular shape (1 point), abundant blood flow (1 point), posterior acoustic shadowing (1 point), and hard strain elastography or maximum shear wave elastography (SWE) ≥87.66 kPa (1 point). The total score ranged from 0 to 8 points, with higher scores, indicating higher malignant rate of the mass. The stratification of multimodal ultrasound BI-RADS was conducted according to the malignant probability. The new model demonstrated high levels of sensitivity and specificity, and was able to effectively differentiate between benign and malignant breast tumors. The area under the receiver operating characteristic curve (AUC) was 0.881.

Conclusions

The multimodal ultrasound BI-RADS model provides a novel quantitative method for the risk assessment of breast tumors and has significant clinical application value.

Introduction

Breast cancer poses a severe threat to women’s health. In 2022, the number of new cases of breast cancer in China amounted to 357,200, representing 15.5% of the global new breast cancer cases and ranking as the second most prevalent cancer among Chinese women, surpassed only by lung cancer. Additionally, 75,000 deaths were attributed to breast cancer in China, representing 11.2% of global breast cancer-related deaths (1). Ultrasound equipment is a readily available and widely utilized imaging apparatus that is well-suited for preliminary screening and diagnostic purposes in the context of breast cancer. Unlike previous studies (2-4), which mainly used the combination of multimodal ultrasound and laboratory indicators to distinguish between benign and malignant masses and predict lymph node metastasis and were highly susceptible to subjective human interpretation, the present study focuses on the practical demands of ultrasound diagnosis in primary healthcare settings in China. Firstly, it focuses on the distinctive pathological traits of breast cancer in Chinese women (5,6), including the high proportion of ductal carcinoma and the predominant location of lesions in dense breast parenchyma. Multiple multimodal ultrasound indicators, including conventional grayscale ultrasound (lesion boundary and shape), Doppler blood flow imaging (blood flow signal grading), and elastography (tissue stiffness values), were systematically screened, integrated and quantitatively analyzed. Ultimately, a simplified, quantitative Chinese-version ultrasound BI-RADS (Breast Imaging-Reporting and Data System) risk stratification model was established to intuitively predict the probability of breast cancer. Distinct from the internationally adopted version, this model is better tailored to the physiological features of Chinese women’s breasts and substantially reduces inter-observer inconsistency caused by subjective radiological evaluation.

This model is designed to address the shortage of advanced equipment in primary hospitals and further reduce inconsistencies in assessment results caused by interpretive variations among different physicians. The system is capable of intuitively displaying the risk probability of breast cancer and corresponding stratification grades. Its core value lies in eliminating the reliance on high-end imaging equipment—instead, it can be applied using ultrasound devices already popularized in primary healthcare settings, thereby facilitating “precision preliminary screening” at the grassroots level. Furthermore, the utilization of standardized quantitative logic results in a substantial reduction of interpretive differences between physicians with varying levels of experience and avoids assessment inconsistencies caused by subjective judgment. Ultimately, this model provides a practical and implementable tool for improving the primary breast cancer screening system in China. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2831/rc).

Methods

This retrospective analysis was granted approval by the Ethics Committee of Ruijin Hospital, Affiliated with Shanghai Jiao Tong University School of Medicine (No. 1.0/2021-11-3), with a waiver for the requirement of written informed consent from participants due to the retrospective nature. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Patients

Patients who underwent preoperative ultrasound examination in our hospital between August 2019 and September 2020 were enrolled. A total of 1,716 patients were included, with one most suspicious lesion selected per patient for the study. All participants received a comprehensive multimodal ultrasound examination prior to surgery. The mean age of the 1,716 patients was 51.2±14.3 years (range, 16–88 years), and the distribution of benign and malignant masses was 658 cases and 1,058 cases, respectively. Inclusion criteria: Clear imaging images; Complete multimodal ultrasound examination data; Complete clinicopathological data. Exclusion criteria: Poor-quality or missing imaging images; Incomplete multimodal ultrasound records (i.e., missing data for any ultrasound modality); Incomplete clinical data; Lesions already categorized as BI-RADS category 6 (pathologically confirmed malignancy via biopsy).

Breast ultrasound examination

The ultrasound devices used in this study included Esaote MyLab90, Philips iU22, and Mindray Resona7, all equipped with linear array transducers with a frequency range of 4.0–14.0 MHz. Breast ultrasound examinations were performed using a standardized institutional protocol. Conventional 2D grayscale ultrasound, color Doppler, strain elastography (SE), and shear wave elastography (SWE) were used for lesion assessment. Complete scanning of bilateral breasts, axillary tail, accessory breast tissue, and axillary lymph nodes was performed. All images were recorded, stored on a hard drive, and archived in the Digital Imaging and Communications in Medicine (DICOM) system. According to the 5th edition BI-RADS guidelines (American College of Radiology, 2013) while blinded to pathological results. On this basis, strain elastography was dichotomized into soft and hard, blood supply was classified as absent/minimal or abundant, and quantitative values of SWE were additionally enrolled. In cases of discrepancies in their conclusions, a third radiologist (with more than 10 years of clinical experience) was consulted for discussion until a consensus was reached.

Statistical analysis

Statistical analysis was conducted using SPSS Statistics (version 26.0, USA). In the univariate analysis, continuous data were described as mean ± standard deviation (x̄±s), and comparisons between groups were performed using the independent-samples t-test or Mann-Whitney U test (for non-normal distributed data). The optimal cutoff value of the SWE (max) for distinguishing benign from malignant breast masses was determined using receiver operating characteristic (ROC) curve analysis. The threshold was selected by maximizing the Youden index (Youden index = sensitivity + specificity – 1). Categorical data were presented as constituent ratios or rates, and group comparisons were analyzed using the Chi-squared (χ²) test or Fisher’s exact test (when the expected frequency was <5). Variables with a P value <0.05 were considered statistically significant and included in subsequent analyses. Logistic regression analysis was employed for the screening of multimodal ultrasound indicators: ultrasound features of breast masses served as independent variables, while the pathological nature (benign vs. malignant) of breast masses was defined as the dependent variable. The screened indicators were then subjected to binary logistic regression analysis, and the regression coefficient of each indicator was rounded to the nearest integer to assign a corresponding score.

The total score for each breast mass was determined by calculating the sum of the scores for all the ultrasound indicators that were included, resulting in a score range of 0 to 8. The malignancy rates were calculated on a per-total-score basis, with elevated scores denoting an increased probability of malignancy. All lesions were stratified into risk grades according to the newly established multimodal ultrasound-based BI-RADS classification. Internal validation of the logistic regression model was performed via ROC analysis. Model performance was quantified using AUC, sensitivity, specificity.

Results

Clinical information

The results of the histopathological examination revealed the presence of 658 benign lesions and 1,058 malignant lesions in total. These malignant lesions included 629 cases of invasive ductal carcinoma (IDC), 25 cases of invasive lobular carcinoma (ILC), 206 cases of ductal carcinoma in situ (DCIS), and other types of malignant lesions such as papillary carcinoma, metaplastic carcinoma, mucinous carcinoma, squamous cell carcinoma, and lymphoma. A total of 1,716 patients were enrolled, with an age of 51.2±14.3 years (range, 16–88 years). A higher age was associated with a higher malignant rate of breast masses. Additionally, the malignant rate of masses was significantly higher in postmenopausal patients (Table 1).

Table 1. Clinical characteristics of patients.

Clinical characteristic	Group	Benign (n)	Malignant (n)	Total (n)	χ2	P
Age	<40 years	216	134	350	170.499	<0.001
	40–60 years	345	497	842
	>60 years	97	427	524
Menstrual status	Premenopausal	448	405	853	144.164	<0.001
	Postmenopausal	210	653	863

Stratification of multimodal ultrasound BI-RADS based on malignant probability

For the continuous variable of maximum shear-wave elastography (SWEmax), the optimal cutoff value of 86.77 kPa for differentiating benign and malignant masses was determined using the maximum Youden’s index. Given that strain elastography (soft/hard) and maximum shear-wave elastography values essentially represent two alternative metrics reflecting the same clinical dimension, only the more commonly used strain elastography was included in the statistical analysis. All included parameters are listed in Table 2. The inclusion criterion for variables was a P value <0.05.

Table 2. Indicators included in multimodal ultrasound BI-RADS and their quantitative scoring.

Indicator	Benign	Malignant	χ2	P1	B	P2	Value
Orientation
Parallel	646 (39.2)	1,004 (60.8)	11.804	0.001
No parallel	12 (18.2)	54 (81.8)			−0.412	0.286	–
Echo pattern
No posterior	575 (48.7)	605 (51.3)	172.295	<0.001
Attenuation	83 (15.5)	453 (84.5)			0.786	<0.001	1
Calcification
No/coarse calcification	587 (56.8)	447 (43.2)	373.580	<0.001
Microcalcification	71 (10.4)	611 (89.6)			1.660	<0.001	2
Vascularity
Absent/minimal	279 (73.4)	101 (26.6)	254.003	<0.001
Rich blood flow	379 (28.4)	957 (71.6)			0.929	<0.001	1
Shape
Regular	277 (80.8)	66 (19.2)	326.184	<0.001
Irregular	381 (27.7)	992 (72.3)			1.055	<0.001	1
Margin
Circumscribe	125 (89.9)	14 (10.1)	337.673	<0.001
Indistinct	502 (43.3)	657 (56.7)			0.921	0.006	1
Spiculated	31 (7.4)	387 (92.6)			2.441	<0.001	2
Elastography
Soft	496 (55.9)	391 (44.1)	239.852	<0.001
Hard	162 (19.5)	667 (80.5)			1.143	<0.001	1

Data are presented as n (%). P1: univariate analysis using the Chi-squared test; P2: multivariate analysis using logistic regression. “–”: indicates no value assignment. BI-RADS, Breast Imaging-Reporting and Data System.

Using pathology as the gold standard, the total score and corresponding malignant probability of each mass were compiled, as shown in Table 3. The total scores were then classified and summarized based on the malignant probability stratified by BI-RADS.

Table 3. Actual malignant rate for each total score.

Total score	Benign	Malignant	Number	Malignant rate (%)
0	51	1	52	1.9
1	108	12	120	10.0
2	164	39	203	19.2
3	184	96	280	34.3
4	81	174	255	68.2
5	34	178	212	84.0
6	27	263	290	90.7
7	5	180	185	97.3
8	4	115	119	96.6

• A total score of 0: malignant rate of <2%;

• A total score of 1: malignant rate of 2–10%;

• A total score of 2–3: malignant rate of 10–50%;

• A total score of 4–6: malignant rate of 50–95%;

• A total score of 7–8: malignant rate of >95%.

Finally included indicators: indistinct margin: 1 point; spiculated margin: 2 points; Microcalcification: 2 points; Irregular shape: 1 point; rich blood flow: 1 point; posterior echo attenuation: 1 point; hard texture on strain elastography or maximum SWE value ≥87.66 kPa: 1 point. The total score obtained for each mass and the corresponding BI-RADS stratification were clearly displayed in Figure 1.

Figure 1.

Quantitative version of BI-RADS. BI‑RADS, Breast Imaging‑Reporting and Data System; SWE, shear wave elastography.

A 78-year-old woman presented with a breast mass at the 12 o’clock position of the right breast. The ultrasonic findings of the mass are shown in Figure 2, as follows: irregular shape (1 point); spiculated margin due to glandular retraction (2 points); posterior echo attenuation (1 point); microcalcifications visible within the mass (2 points); a thick perforating vessel observed inside the mass (1 point); Hard texture on strain elastography/maximum SWE value of 227.6 kPa (1 point). With a comprehensive score of 8, the mass was classified as BI-RADS Category 5. The final pathological result confirmed it as IDC of the breast.

Figure 2.

IDC classified as BI-RADS Category 5 with a comprehensive score of 8. BI‑RADS, Breast Imaging‑Reporting and Data System; IDC, invasive ductal carcinoma.

Validation results and model stability

The total score of ultrasound indicators was used as the test variable, and the benign/malignant nature of breast masses as the state variable. The diagnostic performance of multimodal ultrasound BI-RADS was evaluated using the receiver operating characteristic (ROC) curve. The area under the receiver operating characteristic curve (AUC) was 0.881 [95% confidence interval (CI): 0.865, 0.896], with a sensitivity of 86.0% and a specificity of 77.1% (Figure 3).

Figure 3.

AUC. AUC, area under the receiver operating characteristic curve; US-BIRADS, Ultrasound Breast Imaging Reporting and Data System.

Discussion

The full name of the BI-RADS system is the Breast Imaging Reporting and Data System, which was established by the American College of Radiology (ACR) in 1992. Its original purpose was to standardize mammographic imaging (breast X-ray) reports, making them more standardized and interpretable (7). With the continuous advancement of medical technology, the BI-RADS system has been expanded to cover multiple aspects, including ultrasound, magnetic resonance imaging (MRI), follow-up, and prognostic monitoring, and has become a crucial tool for the diagnosis of breast diseases (8,9). Consistent and appropriate use of BI-RADS vocabulary terms and assessment categories can effectively communicate findings, estimate the risk of malignant tumors, and provide management recommendations to patients and referring clinicians.

Although artificial intelligence (AI) and computer-aided detection (CAD) systems are increasingly investigated and have achieved diagnostic accuracy for breast masses comparable to that of dedicated breast radiologists (10-12), standardized use of BI-RADS terminology and classification can effectively convey imaging findings, stratify malignancy risk, and guide clinical management for patients and referring physicians. Despite the promising performance of such AI-based tools (13), they are not the focus of the present study. Conversely, our research is predicated on well-defined ultrasound parameters with a view to developing a novel quantitative method: a multimodal ultrasound BI-RADS model, aimed at serving primary healthcare institutions. Given the relative scarcity of high-end equipment in primary hospitals, ultrasound is the preferred screening modality for breast diseases due to its ease of operation, non-radiative nature, and cost-effectiveness. Since our data were obtained from a well-known tertiary grade A general hospital in China—relying on the hospital’s profound expertise in relevant disease fields and the high-level diagnostic and therapeutic capabilities of its expert team—the constructed diagnostic model is strongly supported in terms of the reliability of data sources and the scientific validity of diagnostic logic.

The present study incorporated strain elastography or SWE, as outlined in the 5th edition of the Breast Ultrasound BI-RADS, and categorized the blood supply of masses into two types: absent/minimal blood flow and abundant blood flow. A multi-angle and multi-parameter analysis of lesions was conducted, thereby significantly improving the accuracy of diagnosis and reducing the likelihood of misdiagnosis and missed diagnosis. Indicators were screened using multivariate logistic regression, and the selected indicators included mass shape, margin, calcification, blood supply, posterior acoustic features, and shear wave or strain elastography, as demonstrated in Table 2. Benign masses characteristically manifest expansile patterns, with approximately 94% of benign nodules exhibiting a round or oval morphology and clear margins (14). In contrast, malignant masses demonstrate invasive growth, and the traction they exert on surrounding tissues can manifest as ill-defined margins or spiculation. Calcifications in benign lesions principally occur in proliferative nodules, fibroadenomas, papillomas, duct ectasia, and vascular calcifications (15). Calcifications in breast cancer are predominantly microcalcifications, presenting as clustered or scattered tiny calcifications. The vast majority of such cases are duct-derived malignant tumors, including IDC and DCIS. In the cases collected in this study, IDC and DCIS accounted for 78.9% of all malignant tumors (835/1,058), which is consistent with the results of previous studies (5). Masses accompanied by calcifications are a common cause of posterior acoustic attenuation; therefore, on ultrasound images, acoustic shadowing is generally obvious behind calcifications. In comparison with macrocalcifications, microcalcifications are frequently characterized by a less pronounced posterior acoustic shadowing. However, malignant tumors themselves have a strong ability to absorb ultrasound waves, which also results in relatively obvious posterior acoustic shadowing. It is evident that the most prevalent pathological subtypes of breast tumors in China are frequently associated with microcalcifications. Consequently, posterior acoustic attenuation of masses is more indicative of malignancy than posterior acoustic enhancement or no change. Previous literature reports have shown that the sensitivity of posterior acoustic attenuation in diagnosing malignant masses ranges from 48% to 67%, and the specificity ranges from 74% to 98% (16).

A prospective multicentre study has demonstrated that the positive predictive value (PPV) of strain elastography can reach 80% (17). SWE is recognized as a significant addition to B-mode ultrasound (18); meanwhile, incorporating SWE into the feature analysis of the BI-RADS guidelines improves the specificity of breast ultrasound quality assessment without reducing its sensitivity (19,20). In our study, a cutoff value of SWE (Emax) ≥87.66 kPa was assigned 1 point, which is similar to previous studies that reported a cutoff value of SWE (Emax) at 82.6 kPa for differentiating benign and malignant breast masses (21). This study objectively reproduced real-world clinical breast ultrasound practice, where clinicians may apply strain elastography alone or combined with shear-wave elastography. Since the two elastographic techniques have highly overlapping clinical implications for assessing lesion stiffness, only the more commonly used strain elastography was included in statistical analyses to avoid multicollinearity, with one point assigned for positive elastographic results. Similarly, one point was given if the maximum shear-wave elastography value was ≥86.77 kPa. In order to avoid the occurrence of both duplicate counting and overestimation, it was imperative to record a single positive score for each modality that was positive. Super-resolution ultrasound imaging contributes to the differential diagnosis of breast masses by calculating the microvascular flow velocity and microvascular density within the masses (22). Similarly, Jiang et al. (23) derived the vascularization index (VI) of masses through studying the three-dimensional ultrasound features of breast cancer; their results all indicated that a high degree of blood supply provides sufficient nutrients for the rapid growth of masses, which is consistent with the growth characteristics of malignant masses.

The SWE (Emax) values and blood supply of malignant breast masses were significantly higher than those of benign masses, which is consistent with the findings of our study. Analysis of patients’ age and menopausal status (see Table 1) revealed a positive correlation between increasing age and malignancy rate of masses. Furthermore; the malignancy rate of masses was also higher in postmenopausal patients. Following the application of the statistical analysis, neither of these two indicators demonstrated statistical significance and as a consequence, they were both excluded from further consideration. This may be attributed to the influence of confounding interactions; therefore, clinical factors were not included in the model. Nevertheless, it is well recognized that patient age cannot be ignored in the classification of mass types, a view supported by studies conducted by He and Su et al. (24,25). Points were assigned according to the weight of each indicator; the higher the score, the higher the malignancy rate of the mass. Multimodal ultrasound BI-RADS stratification was performed based on the probability of malignancy.

The innovations of our study are as follows: Firstly, we constructed a clinical prediction model for BI-RADS classification of breast masses using only multimodal ultrasound indicators. In order to ascertain the most effective ultrasound features in identifying breast masses as either benign or malignant, a multivariate logistic regression analysis was performed, with the ultrasound features of the breast masses acting as independent variables and the pathological benign/malignant status of the breast masses as the dependent variable. The regression coefficients were rounded to integers for score assignment, thereby providing a novel quantitative method for breast tumor risk assessment. The probability of breast cancer can be predicted intuitively through simple score summation, offering reliable references for further examinations.

The scoring method simplifies the application of BI-RADS, which differs from AI models that rely on large volumes of annotated data—AI models often show decreased accuracy when dealing with rare cases, special patients, or variations in equipment models. In contrast, this quantitative method is developed under the leadership of senior physicians from tertiary hospitals; the selection of parameters inherently covers common complex scenarios in clinical practice. Moreover, the model’s independence from fixed data distributions ensures its capacity to sustain reliable predictive performance in contexts involving equipment updates or substantial individual variations among patients, obviating the necessity for recurrent model retraining.

Secondly, this method features low implementation costs and a much lower promotion threshold compared to AI. Based on ultrasound equipment and routine measurement parameters familiar to physicians, it requires no additional hardware investment. It is evident that primary hospitals have the capacity to expeditiously assimilate this innovation through the facilitation of rudimentary training, thereby facilitating its propagation across a diverse array of medical institutions at varying echelons. This approach is instrumental in attaining “technological inclusiveness.”

This study has certain limitations. First, as the data were collected from tertiary hospitals where we collaborated, the malignancy rate in the dataset may be higher than that in the general population. This has the potential to result in selection bias, thereby giving rise to potential issues. Secondly, the indicators screened in this study lack discussions on clinical factors and relevant signs. In the future, new ultrasound technologies can be further incorporated to explore the feasibility of predicting the pathological subtypes of masses. Finally, additional multicenter and prospective cohort studies are required to verify the feasibility of the model and its applicability for dissemination to other regions.

Conclusions

Based on conventional multimodal ultrasound, quantitative assignment was performed for malignant-related indicators. Breast cancer risk was predicted via simple summation of the scores, which yielded favorable sensitivity and specificity. Furthermore, this approach reduces interobserver variability in subjective assessment and is suitable for popularization in primary-level hospitals.

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 retrospective analysis was granted approval by the Ethics Committee of Ruijin Hospital, Affiliated with Shanghai Jiao Tong University School of Medicine (No. 1.0/2021-11-3), with a waiver for the requirement of written informed consent from participants due to the retrospective nature.

Data Sharing Statement

Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2831/dss

qims-16-11-876-dss.pdf ^{(70.6KB, pdf)}

DOI: 10.21037/qims-2025-1-2831