Establishment of prediction model for breast lesion using automated breast ultrasound system

Introduction

Breast cancer is one of the most common malignant tumors in females and has become a primary disease affecting female health worldwide (1). In 2021 the latest edition of Global Cancer Burden Report exhibited that 2.26 million newly diagnosed breast cancer cases were found, and breast cancer was one of the leading causes of cancer mortality (2). Based on the fact that early identification and early treatment could effectively improve patients’ treatment efficacy and prognosis, therefore, breast cancer screening is of ultimate importance and requires further research (3,4).

Currently, mammography and ultrasound are the most commonly used methods in breast cancer screening (5,6). Mammography has been confirmed to reduce mortality by 30–50% in 40 to 69-year-old patients with breast cancer. Whole breast ultrasound combined with mammography could significantly increase the detection rate of breast lesions in dense glands compared with single mammography (7). Handheld ultrasound (HHUS), as the most widely used ultrasound technique possesses several advantages including efficiency and high cost performance (8). However, HHUS can only provide 2D mammary gland images with limited resolution at a selected section, and no stable 3D information can be provided by this technique, which makes it not conducive for long-term follow-up for the same lesion of breast cancer patients (9).

With the development of advanced ultrasound technologies, novel techniques such as automated breast ultrasound systems (ABUS) are gaining popularity, and their presence is likely to make up for the shortcomings of traditional ultrasounds (10). ABUS can automatically acquire the volume data of the whole breast, and obtain standardized cross-sectional, sagittal, and coronal imaging information of the breast tissue after three-dimensional reconstruction, to facilitate the post-processing of ABUS data on a dedicated workstation and diagnosis of breast lesions. And many automatic ABUS imaging segmentations have been proposed in recent years. Cao et al. proposed Auto-DenseUNet (11) and Zhou et al. developed a multitask segmentation and classification platform (12). Still, these new methods should be validated in clinical scenarios.

The purpose of this research is to explore a prediction model to distinguish between malignant and benign breast lesions by an ABUS, for which a more objective and accurate corresponding relationship between the imaging features and diagnostic impression of breast lesions could be established. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1069/rc).

Methods

Study design and patients

This prospective study enrolled patients with single breast lesions identified by HHUS in Peking University Cancer Hospital between June 2010 and December 2012. The inclusion criteria were: (I) female age between 26 and 76 years; (II) patients received pathology examination of the breast lesion. The exclusion criteria were: (I) incomplete clinical data; and (II) patients refused to participate. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of Peking University Cancer Hospital (No. 2016KT14), and written informed consent was obtained from all patients.

Procedures

Patients first completed HHUS and ABUS examinations, after confirmation of breast lesions, patients further received pathological examination within the following two weeks to confirm whether the lesions were benign or malignant. Only consecutive female cases with solitary lesions were selected, totaling 2,090 cases, with a total of 2,090 lesions. The methods for obtaining pathological tissue included ultrasound-guided core needle biopsy (CNB) using HHUS or open excisional biopsy under direct visualization. All patients’ names, medical record numbers, pathology numbers, ultrasound numbers, ages, ultrasound examination dates, pathology diagnosis dates, ultrasound diagnoses, and pathology diagnosis results were verified.

HHUS and ABUS examination

Patients were maintained in a supine position, with both arms abducted and hands touching the top of their heads during HHUS examination. All HHUS examinations were conducted by three experienced physicians with expertise in breast ultrasound diagnosis using traditional 2D Doppler ultrasound.

The ABUS was performed with a high-end ultrasound scanner (somo-v ABUS, U-systems Inc.) equipped with a 10 MHz probe, a flexible arm with the transducer at the end, a touch screen, and a 3D workstation. During ABUS examination, patients were also maintained in a supine position. After a single sweep with the wide-aperture linear array transducer, ABUS will automatically acquire a 17 cm × 15 cm × 5 cm volume dataset of breasts. After ABUS data acquisition, one experienced surgical oncologist (5 years of breast ultrasound experience) reviewed multi-planar compounded breast lesion images in all three planes (coronal, sagittal and axial) on the dedicated VIEWe software. Meanwhile, the advanced 3D Workstation uses a unique descriptor from the 3D-US and BI-RADS-US lexicons (13) (including lesion form, size, shape, orientation, margin, boundary, echo pattern, and posterior acoustic features) to describe the lesion characteristics. The lesion margin was divided into two secondary variables, margin (A) and margin (B). Specifically, positive margin (A) was defined as circumscribed, while negative margin (A) included Indistinct + angular + microlobulated + spiculated shape; positive margin (B) was defined as indistinct, while negative margin (B) included Circumscribed + angular + microlobulated + spiculated. There was only one interpreter in the entire study. All parameters of the collected images are shown in Figure 1.

Figure 1 The input interface of EpiData 3.1.

Pathological biopsy & sonogram evaluation

Patients with abnormal breast lesions (1 cm ≤ dual diameter ≤5 cm) detected by HHUS and ABUS were further examined by CNB, and open biopsy (OB) was performed when the pathological result could not be confirmed by CNB.

Data collection and definitions

Concerning the BI-RADS classification in 2013 and combined with the characteristic term dictionary of the ultrasound part of the fifth edition of the system, EpiData 3.1 software was used to make an ABUS film reading and recording tool. The input interface is shown in Figure 1. At the top of the interface is the basic information of the patient and the date of examination. The ABUS description information entry section is displayed in the middle of the page. Items 1 and 2 are the basic characteristics of the breast, including the description of the breast background and whether the breast duct is dilated or not. Entry 3 is the core part of the input information, that is, the description of lesion characteristics, which is mainly divided into location, echo type, lesion manifestations, size, shape, direction, boundary, edge and rear echo.

Statistical analysis

Statistical analyses were performed using SPSS software (version 19.0, IBM, Inc., USA). Categorical variables were selected according to clinical practical significance, described as n (percentage) and examined by the chi-square test. We defined benign and malignant primary tumors as the dependent variable (1 for malignant, 0 for benign), adjusted the significant variables in univariate logistic regression analysis, and then established the prediction equation. Finally, we used the receiver operating characteristic (ROC) and calculated the area under the curve (AUC) for model evaluation. The cut-off value was decided according to the maximum principle of the Youden index and then evaluated the authenticity, reliability and application value of the prediction model by calculating the sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and accuracy. A two-sided P value <0.05 was regarded as statistical significance in the present study.

Results

Patients’ characteristics

A total of 13,520 patients with breast lesions who received ABUS were initially included, after the exclusion of pathology unclear and lost follow-up patients, 2,090 females showed single lesions between 1 and 5 cm. Among them, 630 consecutive cases at the top of the sequence number (30.1%) were selected for this preliminary research (the remaining 1,460 cases are intended for follow-up validation), and all patients were randomly split in training set and validation set at a 1:1 ratio (both n=315). The patient’s inclusion flowchart is provided in Figure 2. Because most patients were not hospitalized for information registration, baseline information was not available for all patients.

Figure 2 Flowchart of case selection. ABUS, automated breast ultrasound system.

Pathological diagnosis

The histopathological diagnosis of 630 cases is shown in Table 1. No false negatives occurred after at least 5 years of follow-up. All malignant cases (n=387, 61.4%) received standard treatment after histopathological diagnosis.

Correlation between lexicons and pathological diagnosis

The correlation between each lexicon and pathological diagnosis of 630 cases is shown in Table 2. The median lesion size was 1.8 cm. The results of univariate analysis showed lexicons about lesion form (P<0.001), shape (P<0.001), orientation (P<0.001), margin (P<0.001), boundary (P<0.001), and posterior acoustic feature (P=0.001) had statistical significance with the pathological results (P<0.05). Compared with benign lesions, the malignant lesions tend to have a more disordered form, irregular shape, more vertical orientation, and less circumscribed margin, while the boundary shows a more ecogenic halo and the posterior acoustic features tend to have more shadow. According to the clinical significance, we finally chose three lexicons describing lesion form, margin, and boundary for multivariate regression analysis.

Model development and validation

The three lexicons were defined as categorical variables (Table S1) before further analysis. The nodule and mass were combined into one variable.

Data from the 315 training samples were used for the establishment of the prediction model by multivariate logistic regression analysis, and the final equation was Y=1.604 × boundary + 1.045 × lesion form − 5.436 × margin (A) − 2.166 × margin (B). In the training cohort, the AUC was 0.882 [95% confidence interval (CI): 0.844–0.919], indicating a good clinical application value (Figure 3). According to the maximum principle of Youden-index, when using Y=1.0 as a diagnostic boundary value, Y≥1.0 was diagnosed as malignant, Y<1.0 was diagnosed as benign, the sensitivity, specificity, PPV, NPV, and accuracy were 0.771, 0.842, 0.891, 0.687, and 0.798, respectively. A nomogram was also plotted according to the multivariate logistic regression analysis results (Figure 4).

Figure 3 ROC curve of the training sample model. CI, confidence interval; ROC receiver operating characteristic.

Figure 4 Nomogram developed for malignant breast lesion prediction.

The model validation was performed using the validation cohort. The AUC was 0.866 (95% CI: 0.824–0.928) (Figure 5). According to the maximum principle of Youden-index, when using Y=1.0 as a diagnostic boundary value, cases with Y≥1.0 were diagnosed as malignant, and cases with Y<1.0 were diagnosed as benign, the sensitivity, specificity, PPV, NPV, and accuracy were 0.773, 0.805, 0.856, 0.702, and 0.783.

Figure 5 ROC curve of the validation sample model. CI, confidence interval; ROC receiver operating characteristic.

Discussion

This study consecutively included female patients with pathology-confirmed breast lesions 1–5 cm in diameter. The lesion form, shape, orientation, margin, boundary, and posterior acoustic feature were significantly different between malignant and benign lesions. After the selection of clinically important features, four features were used for the development of multivariate logistic regression, and the AUC reached 0.882 in the training cohort, while the validation cohort showed AUC of 0.866.

In this study, after univariate analysis, we found several factors associated with different pathological types. We further combined some classification variables into binary classification variables due to the limited sample size. For example, the ‘nodule’ and ‘mass’ were combined into one variable, and the ‘lesion margin’ was divided into two secondary variables, margin (A) and margin (B). The feasibility of these features has also been verified in past research (14-16).

Breast malignant lesion prediction is widely researched in previous studies, Militello et al. reported 3D dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is useful for malignant lesion prediction and reached an AUC of 0.725, sensitivity of 0.709, and specificity of 0.741 (17); Zhong et al. developed a classification model based on breast ultrasound and a novel algorithm achieving an AUC of 0.906 and 0.790, sensitivity of 0.852 and 0.723, and specificity of 0.886 and 0.804 in two datasets, indicating the possible overfitting of their model (18). In our present study, after feature selection, boundary, lesion form, and two margin information were included in the multivariate logistic regression model, the AUC reached 0.882 in the training cohort and 0.866 in the validation cohort, indicating a great generalization ability of our proposed model, the underlying reason could be partially explained by the feature used in our study, unlike most published work, we directly extract the morphological description for disease classification, which made our model has better explainability. From the equation, we found that margin (A) and margin (B) provided higher weights for the lesion classification. While boundary and lesion forms showed less significance. This might indicate that a clear distinction is necessary between these lexicons in clinical work (19), especially when an indistinct, angular, microlobulated, and spiculated lesion is found, which would directly lead to a >50% predicted risk of malignant lesion. Although shape and orientation were also different between benign and malignant lesions, they seemed to have little predictive value and were not included in the equation because of their small contributions. Still, clinical practical application of the prediction model might reduce the bias of human judgment and maximize the decrease in the misdiagnosis rate in the process of screening. Meanwhile, although the AUC was high, we noticed that the NPV only achieved 0.687 in the present study, while Militello et al. showed their 3D-DCE-based prediction model achieved an NPV of 0.75 (17), the radiomics model developed by Guo et al. even achieved an NPV of 0.93, although the AUC was only 0.811 (20), these findings indicating that patients with negative prediction should also receive further examination and follow-up, and the algorithm should be further optimized to improve the NPV. Moreover, compared to previous models, the model used in the present study was more clinically friendly, the margin, boundary, and lesion form information was easy to acquire during clinical practice, with no further information required, which could facilitate the accurate prediction of malignant lesions in basic medical institutions. Also, unlike complex “black-box” radiomics or deep learning models, our approach prioritizes high clinical interpretability and seamless integration into routine workflow (21,22).

Compared to traditional HHUS, ABUS can provide 3D images of the whole breast, which makes ABUS more suitable for accurate clinical research than HHUS. In our study, the HHUS was used as a preliminary examination tool, and all patients included should receive an ABUS examination. In addition, previous studies have proven that the retraction phenomenon in the coronal plane (Figure 6) could be a predictive factor for malignant lesions (23,24), however, due to anatomical structure, the coronal plane was unique to ABUS and unavailable for HHUS, but in China, the vast majority of breast ultrasonic examinations are still performed solely with HHUS. Moreover, the ABUS is a more accurate, precise, and rapid imaging tool which facilitates detailed description of shape, orientation, margin, boundary information, in a previous study, Sebastian et al. showed that the HHUS costs approximately 20–30 min more than the ABUS (25), while a complete ABUS process including case reading and data recording performed by researcher requires only 10–12 min. However, to better compare the performance of ABUS, this study dropped the coronal plane for further analysis, but the data collection included all the 3D information. More importantly, the lexicon-based prediction model showed unique value in some rural areas with limited medical facilities. Under such circumstances, the raw data were difficult to preserve in a low-resource setting; however, the examination reported with lexicon features could help clinicians improve the prediction of breast lesions and support the health management of potential high-risk patients (26).

Figure 6 The retraction phenomenon in the coronal plane of ABUS. ABUS, automated breast ultrasound system.

There are still several limitations in this study: firstly, this study was a single center study with patients enrolled between 2010 and 2012, further evaluation of the results is necessitated in another large-scale multi-center; secondly, some baseline characteristics were not complete in the present during to outpatients reason, therefore the model was developed only with ABUS features; thirdly, no inter- or intra-observer correlation analysis was performed, and future study will incorporate this part for improved reliability; Fourth, only solitary lesions were included in the developmental process, and patients with multifocal or bilateral lesions can be included in future study; finally, although the model proposed in this study achieved a high AUC, the NPV was relatively low, if a patient presented with a negative prediction result but highly clinical susceptibility, further confirmation examination should be performed.

Conclusions

In summary, the characteristics of lesion form, boundary, and margin appear to be the primary features distinguishing between benign and malignant lesions on ABUS. While the resultant malignant lesion prediction model demonstrated strong diagnostic performance across both the training and validation cohorts, further external validation remains essential.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1069/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1069/dss

Funding: This work was supported by the Beijing Municipal Science & Technology Commission (No. D161100000816006) and the Capital’s Funds for Health Improvement and Research (No. 2018–2-2152).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1069/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 in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of Peking University Cancer Hospital (No. 2016KT14), and written informed consents were obtained from all patients.

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. Cathcart-Rake EJ, Tevaarwerk AJ, Haddad TC, D'Andre SD, Ruddy KJ. Advances in the care of breast cancer survivors. BMJ 2023;382:e071565. [Crossref] [PubMed]
2. Wilkinson L, Gathani T. Understanding breast cancer as a global health concern. Br J Radiol 2022;95:20211033. [Crossref] [PubMed]
3. Huntley C, Torr B, Sud A, Houlston RS, Hingorani AD, Jones ME, Turnbull C. The impact of risk stratification by polygenic risk and age on breast cancer screening in women aged 40-49 years: a modelling study. Lancet 2023;402:S54. [Crossref] [PubMed]
4. Mittra I, Mishra GA, Dikshit RP, Gupta S, Kulkarni VY, Shaikh HKA, Shastri SS, Hawaldar R, Gupta S, Pramesh CS, Badwe RA. Effect of screening by clinical breast examination on breast cancer incidence and mortality after 20 years: prospective, cluster randomised controlled trial in Mumbai. BMJ 2021;372: [Crossref] [PubMed]
5. Nicholson WK, Silverstein M, Wong JB, Barry MJ, Chelmow D, Coker TR, Davis EM, Jaén CR, Krousel-Wood M, Lee S, Li L, Mangione CM, Rao G, Ruiz JM, Stevermer JJ, Tsevat J, Underwood SM, Wiehe S. Screening for Breast Cancer: US Preventive Services Task Force Recommendation Statement. JAMA 2024;331:1918-30. [Crossref] [PubMed]
6. Henderson JT, Webber EM, Weyrich MS, Miller M, Melnikow J. Screening for Breast Cancer: Evidence Report and Systematic Review for the US Preventive Services Task Force. JAMA 2024;331:1931-46. [Crossref] [PubMed]
7. Berg WA, Blume JD, Cormack JB, Mendelson EB, Lehrer D, Böhm-Vélez M, Pisano ED, Jong RA, Evans WP, Morton MJ, Mahoney MC, Larsen LH, Barr RG, Farria DM, Marques HS, Boparai K. Combined screening with ultrasound and mammography vs mammography alone in women at elevated risk of breast cancer. JAMA 2008;299:2151-63. [Crossref] [PubMed]
8. Hareendranathan AR, Tripathi A, Panicker MR, Zhou Y, Knight J, Jaremko JL. Domain-aware contrastive learning for ultrasound hip image analysis. Comput Biol Med 2022;149:106004. [Crossref] [PubMed]
9. Kratkiewicz K, Pattyn A, Alijabbari N, Mehrmohammadi M. Ultrasound and Photoacoustic Imaging of Breast Cancer: Clinical Systems, Challenges, and Future Outlook. J Clin Med 2022;11:1165. [Crossref] [PubMed]
10. Brunetti N, De Giorgis S, Zawaideh J, Rossi F, Calabrese M, Tagliafico AS. Comparison between execution and reading time of 3D ABUS versus HHUS. Radiol Med 2020;125:1243-8. [Crossref] [PubMed]
11. Cao X, Chen H, Li Y, Peng Y, Zhou Y, Cheng L, Liu T, Shen D. Auto-DenseUNet: Searchable neural network architecture for mass segmentation in 3D automated breast ultrasound. Med Image Anal 2022;82:102589. [Crossref] [PubMed]
12. Zhou Y, Chen H, Li Y, Liu Q, Xu X, Wang S, Yap PT, Shen D. Multi-task learning for segmentation and classification of tumors in 3D automated breast ultrasound images. Med Image Anal 2021;70:101918. [Crossref] [PubMed]
13. Eremici I, Borlea A, Dumitru C, Stoian D. Factors Associated with False Positive Breast Cancer Results in the Real-Time Sonoelastography Evaluation of Solid Breast Lesions. Medicina (Kaunas) 2024;60:1023. [Crossref] [PubMed]
14. Hong AS, Rosen EL, Soo MS, Baker JA. BI-RADS for sonography: positive and negative predictive values of sonographic features. AJR Am J Roentgenol 2005;184:1260-5. [Crossref] [PubMed]
15. Costantini M, Belli P, Lombardi R, Franceschini G, Mulè A, Bonomo L. Characterization of solid breast masses: use of the sonographic breast imaging reporting and data system lexicon. J Ultrasound Med 2006;25:649-59; quiz 661. [Crossref] [PubMed]
16. Gokalp G, Topal U, Kizilkaya E. Power Doppler sonography: anything to add to BI-RADS US in solid breast masses? Eur J Radiol 2009;70:77-85. [Crossref] [PubMed]
17. Militello C, Rundo L, Dimarco M, Orlando A, Woitek R, D’Angelo I, Russo G, Bartolotta TV. 3D DCE-MRI Radiomic Analysis for Malignant Lesion Prediction in Breast Cancer Patients. Acad Radiol 2022;29:830-40. [Crossref] [PubMed]
18. Zhong S, Tu C, Dong X, Feng Q, Chen W, Zhang Y. MsGoF: Breast lesion classification on ultrasound images by multi-scale gradational-order fusion framework. Comput Methods Programs Biomed 2023;230:107346. [Crossref] [PubMed]
19. Magni V, Cozzi A, Muscogiuri G, Benedek A, Rossini G, Fanizza M, Di Giulio G, Sardanelli F. Background parenchymal enhancement on contrast-enhanced mammography: associations with breast density and patient’s characteristics. Radiol Med 2024;129:1303-12. [Crossref] [PubMed]
20. Guo R, Yu Y, Huang Y, Lin M, Liao Y, Hu Y, Li Q, Peng C, Zhou J. A nomogram model combining ultrasound-based radiomics features and clinicopathological factors to identify germline BRCA1/2 mutation in invasive breast cancer patients. Heliyon 2024;10:e23383. [Crossref] [PubMed]
21. Zhang L, Du Q, Shen M, He X, Zhang D, Huang X. Interpretable model based on MRI radiomics to predict the expression of Ki-67 in breast cancer. Sci Rep 2025;15:13318. [Crossref] [PubMed]
22. Xu W, Zheng B, Wen C, Zeng H, Wang S, He Z, Liao X, Chen W, Li Y, Qin G. Enhancing Specificity in Predicting Axillary Lymph Node Metastasis in Breast Cancer through an Interpretable Machine Learning Model with CEM and Ultrasound Integration. Technol Cancer Res Treat 2025;24:15330338251334735. [Crossref] [PubMed]
23. Cho KR, Seo BK, Lee JY, Pisano ED, Je BK, Lee JY, Choi EJ, Chung KB, Whan Oh Y. A comparative study of 2D and 3D ultrasonography for evaluation of solid breast masses. Eur J Radiol 2005;54:365-70. [Crossref] [PubMed]
24. Lin X, Wang J, Han F, Fu J, Li A. Analysis of eighty-one cases with breast lesions using automated breast volume scanner and comparison with handheld ultrasound. Eur J Radiol 2012;81:873-8. [Crossref] [PubMed]
25. Wojcinski S, Farrokh A, Hille U, Wiskirchen J, Gyapong S, Soliman AA, Degenhardt F, Hillemanns P. The Automated Breast Volume Scanner (ABVS): initial experiences in lesion detection compared with conventional handheld B-mode ultrasound: a pilot study of 50 cases. Int J Womens Health 2011;3:337-46. [Crossref] [PubMed]
26. Zhang W, Li N, Li J, Zhao Y, Long Y, He C, Zhang C, Li B, Zhao Y, Lai S, Ding W, Gao M, Tan L, Wei X, Yang R, Jiang X. Noninvasive identification of proliferative hepatocellular carcinoma on multiphase dynamic CT: quantitative and LI-RADS lexicon-based evaluation. Eur Radiol 2025;35:3460-75. [Crossref] [PubMed]