Breast cancer is a heterogeneous and complex disease. The main molecular subtypes of breast cancer are luminal A (LA), luminal B (LB), human epidermal growth factor receptor2-positive (HER2), and triple-negative (TN) (1, 2).This heterogeneity leads to vast differences in disease progression, treatment response, and prognosis. Therefore, preoperatively identifying different molecular subtypes of breast cancer can be significant in clinical practice. Nomograms have been used extensively for visualizing predictive models in cancer. They present user-friendly graphic presentation of the estimated probabilities of the molecular subtypes. Moreover, nomograms may guide clinical diagnosis and treatment and can help facilitate precision medicine (3).

Hand-held ultrasonography (HHUS) has some limitations because of the lack of standardization (4). Meanwhile, three-dimensional ultrasonography (3D-US) has been essential as a preoperative tool because of its reproducibility and reduced operator dependence. Furthermore, its unique coronal plane can provide additional diagnostic information and potentially improve the characterization of breast lesions. Moreover, the availability of automated breast US in clinical practice is increasing (5, 6). Determining whether preoperative 3D-US can distinguish tumour subtypes has important clinical significance. Previous studies have shown that the characteristics of 3D-US correlated with molecular classification, for example, calcification is associated with the HER2-positive subtype, while the retraction phenomenon is more related to the LA subtype (7–9).

In recent years, artificial intelligence for 3D-US has been applied in the differential diagnosis of benign and malignant breast masses, however there are only a studies on 3D–US in molecular typing (10).

The study aimed to investigate the relevance of the 3D–US imaging characteristics of breast cancer associated with specific molecular subtypes in order to establish nomograms that distinguish the molecular subtypes of breast cancer. This study explored the possibility of predicting molecular typing models with large samples and provided insights into the future artificial intelligence prediction of molecular typing.

US is the most widely used auxiliary examination technique for the preoperative evaluation of breast cancer. However, certain errors may occur during the preoperative puncture, which can have adverse effects. Therefore, identifying the molecular classification of breast cancer is vital to guide the selection of an individualized clinical plan. Moreover, it is difficult to utilize the traditional classification for these treatment plans. Therefore, establishing a model that incorporates multiple factors to identify the molecular subtype can be used to predict the risk of mortality and facilitate the implementation of individualized treatment (18, 19). Since the ultrasonography features of different molecular types overlap, it is more appropriate to use a multivariate prediction model. The model in this study uses more parameters, thereby avoiding the limitations of single-parameter evaluation. The nomograms created here integrated relevant factors that affect molecular typing in a simple graphical manner, allowing clinicians and patients to understand the relationship between various factors and molecular typing. Given a set of known quantitative conditions, they can easily calculate the probability value of the corresponding outcome (20). Therefore, the realization of individualized classification prediction is in line with the pursuit of individualized treatment.

LA and LB cancers account for approximately 70% of breast cancer cases (21). Similar findings were observed in this cohort, where these subtypes accounted for 72.8% of cases (LA: 26.6% and LB:44.3%). Therefore, the distribution of molecular subtypes in this single-center study is consistent with those in other studies with larger samples.

ABVS is a 3D–US imaging system that overcomes the main shortcomings of HHUS, such as lack of standardization, non-repeatability, small field of view, and excessive time investment. Additionally, for women with dense breast tissue, ABVS can recognize calcification and the retraction phenomenon more due to its the unique coronal reconstruction. Furthermore, ABVS is more accurate than HHUS in assessing the extent of the disease, average lesion size (22), and largest diameter (23). Hence, it can better assess the ture scope of the disease. Recent studies have also shown that the imaging features acquired by ABVS are related to the molecular subtypes of breast cancer. In this study, there was a significant difference among the four molecular subtypes in the maximum diameter of the tumor. Patients with the LA subtype had tumors<2cm, whereas in patients with other subtypes, the largest diameter was 2–5cm. The HER2 molecular subtype had more calcification than the other subtypes. The retraction phenomenon was more likely to be seen in the LA subtype than in any other subtypes. The TN subtype did not display the retraction phenomenon orpost-acoustic enhancement. In the LA subtype, the body has more time to respond to cancer cells and form fibrosis,which leads to a contraction pattern (24, 25). Since ABVS has these advantages, it is moreusefulin evaluating breast cancer preoperatively.

A nomogram is a graphical calculation model that uses known predictive factors to calculate the numerical probability of a clinical event. Such prediction models are valuable. A multivariate logistic regression model we constructed via step wise analysis, Nomograms were subsequently developed based on the fitted multivariate logistic regression model.Tumor size, calcification, post-acoustic enhancement, and the retraction phenomenon were used to construct the nomogram model for distinguishing the molecular subtype of breast cancer. The scores of the influencing factors could show the individualized prediction results. The discriminative power of the nomogram was quantified using the AUC, exhibiting the accuracy of the test. Previous studies have shown that models with AUCs of 0.5–0.7 have low predictive value, models with AUCs of 0.7–0.85 have better predictive value, and models with AUCs of 0.85–0.95 have the best predictive value (26). The model in this study has AUCs of 0.68–0.78,indicating that the prediction model had a good degree of discrimination. The nomograms also contained information for clinical use; therefore, they might serve as tools to calculate the probabilities of the various molecular subtypes.

Recent studies have developed models to predict the molecular classification of breast cancer through radiomics or machine learning approaches using image segmentation (27, 28). These generally employ two-dimensional US modelling for a single molecular classification or deep learning magnetic resonance image modeling. Although these have been shown to be predictive, the feature extraction is complicated and has certain limitations. Compared with HHUS, ABVS allows the use of computer-aided design (29) and artificial intelligence technology to improve the diagnostic performance of deep machine learning due to its repeatability and image storage method. 3D-USmay potentially be useful in the field of artificial intelligence, however, this needs further verification. This study predicted the four molecular classifications using only 3D-US features. Nomograms we also used to visualize the predictive multinomial model of the molecular classifications.

The study has several limitations. First, this was a retrospective study performed at a single institution. A multicenter prospective study with a large sample size needs to be performed to validate this study’s results. Second, the sample sizes of the four molecular subtypes were unbalanced. As a result, the ROC curves of subtypes with fewer samples are unsatisfactory. Further studies should consider having balanced sample sizes.

In conclusion, sonographic features such as calcification and posterior acoustic features were significantly associated with the breast cancer molecular subtypes. In addition, the presence of the retraction phenomenon was the most important predictor for the LA subtype. Nomogramsfor the prediction of the molecular subtypes were established based on the results of the multifactor analysis, The calibration and ROC curves showed that the model had good performance. Further multicenter studies will be useful for updating and validating these nomograms to improve the predictions of molecular subtypes.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Hangzhou First people’s Hospital ethics committee. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

LB contributions to design of the work. YT and LY analysis of data for the work, XX and LL, drafting the work. LZ revising it critically for important intellectual content. All authors agree to be 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 supported by The Construction Fund of Medical Key Disciplines of Hangzhou.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.