Breast carcinoma (BC) is among the tumors with the highest morbidity and mortality in women, which accounts for one in four cancer cases and for one in six cancer deaths in the great majority of countries (1). BC is a heterogeneous disease with different clinical characteristics, clinical behaviors, and treatment response profiles (2). There are four major subtypes of breast carcinoma based on estrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor receptor 2 (HER2) status. These intrinsic molecular subtypes are defined as: luminal A and B, HER2-enriched, and basal-like subtypes. Luminal A and B are luminal type, while HER2-enriched and basal-like are non-luminal type (3).

Luminal type shows better prognosis and response to hormone receptor-targeted therapies (4), while non-luminal type is more aggressive, has poorer outcomes than luminal type, and reveals a higher rate of locoregional recurrence and lower survival rate after distant metastasis (5–7). Since the prognosis of BC differs according to the molecular subtypes, guidelines recommend for immunohistochemistry (IHC) assessment of the molecular subtypes of BC during initial diagnosis (8, 9).

Currently, the preoperative assessment of the molecular subtypes of BC commonly relies on IHC results after core needle biopsy. Whereas, biopsy is invasive and vulnerable to sampling error that might negatively affect the treatment decision and increase health care costs of BC diagnosis when repeated biopsy is required (10, 11). Moreover, the limited biopsy tissue makes it difficult to fully evaluate the heterogeneity within the tumor (12). Therefore, a noninvasive approach that could precisely predict the molecular subtypes of BC, would be high valuable in the early management of patients with BC.

Radiomics is an innovative tool that usually extracts a large number of quantitative features from medical images by using mathematical algorithms, and the quantitative extracted features can represent the shape, intensity, and texture of tumors. It assumes that the quantitative features may be the reflection of mechanisms occurring at a genetic and molecular level, and relevant to tumor behavior or patient’s prognosis (13, 14). Radiomics features have been employed to noninvasively evaluate intratumoral heterogeneity. In recent years, studies have found that radiomics analysis can be utilized to determine sentinel lymph node metastasis (15, 16), distinguish between the benign and malignant breast tumor (17, 18), identify the molecular subtype (19), and evaluate the response to neoadjuvant chemotherapy (20).

Several studies have demonstrated that radiomics features based on computed tomography (CT) or magnetic resonance imaging (MRI) have the ability to discriminate the luminal from non-luminal BC. Wang and colleagues (21) established a radiomics model based on chest CT to distinguish the luminal from non-luminal type and achieved area under the curve (AUC) values of 0.842 in the training set and 0.757 in the test set, demonstrating that chest CT radiomics may provide a new concept for the identification of breast cancer molecular subtypes. The results of a former study by Leithner et al. (22) indicated the potential of radiomics signatures from multiparametric MRI to allow the separation of the hormone receptor-positive and hormone receptor-negative BC. However, no study based on ultrasound radiomics has been published on distinguishing between the luminal and non-luminal BC. Furthermore, compared to CT and MRI, ultrasound regarded as a nonradioactive, available, and low-cost tool is commonly adopted for breast tumor screening and diagnosis.

Therefore, based on the above background, we studied whether ultrasound radiomics features could be adopted as a predictive biomarker for discriminating the luminal from non-luminal type in patients with BC, and the purpose of this study was to develop and validate an ultrasound-based radiomics nomogram by integrating the clinical risk factors and radiomics score (Rad-score) to preoperatively predict the luminal BC.

The study was approved by our Institutional Ethics Committee and performed on the basis of the Helsinki Declaration, and patient informed consent requirement was waived due to the retrospective nature of this study.

In the present radiomics study, feature extraction was carried out on the basis of ultrasound images, and three prediction models were established using clinical feature, Rad-score and combined clinical feature and Rad-score, respectively, for predicting the luminal BC. The results demonstrated that the Rad-score and nomogram had appreciable predictive performance and could be applied as useful methods for preoperatively differentiating between the luminal and non-luminal type in patients with BC. The consistency between the nomogram-predicted probability of luminal BC and actual results was assessed by the calibration curve. On the calibration curve in our study, both p-values were > 0.05 in the training and test sets, indicating that the stability of the nomogram was well. Decision curve analysis revealed that the net benefit of using the nomogram (blue curve) and Rad-score (orange curve) for predicting luminal BC was more considerable than that of the treat-all or treat none approach, indicating well clinical application of the novel models.

A number of studies have demonstrated that radiomics is regarded as an useful and noninvasive method for predicting molecular subtypes in patients with BC, most of which are on the basis of mammography and MRI imaging. Son and colleagues (27) built radiomics signatures based on synthetic mammography reconstructed from digital breast tomosynthesis to predict molecular subtypes of breast cancer. In the validation cohort, the radiomics signature yielded an AUC of 0.838, 0.556, and 0.645 for the triple-negative, HER2 and luminal subtypes, respectively. In our study, the AUC value of Rad-score predicting luminal BC was higher than theirs (0.786 vs. 0.645). This might have some relationship with low spatial resolution of mammography and limited display effect for breast tumor. A prior study by Li et al. (28) including a total of 351 patients and developing radiomics models for predicting the HER2 status preoperatively, found that the intratumoral and peritumoral radiomics scores achieved AUCs of 0.683 and 0.690 in the validation cohort, respectively. Furthermore, the combined radiomics score improved the predictive performance and yielded an AUC of 0.713. Huang and colleagues (29) reported that the multi-parametric MRI-based radiomics models could predict molecular subtypes of breast cancer. The highest performances were obtained for discriminating basal-like vs. non-basal-like (AUC, 0.965), HER2-enriched vs. non-HER2-enriched (AUC, 0.840), and hormone receptor-positive/non-HER2-enriched vs. others (AUC, 0.860) using multilayer perceptron. Compared to mammography and MRI imaging, ultrasound regarded as a nonradiative, convenient, and low-cost technology is universally used for breast tumor screening and diagnosis (30, 31). As far as we know, few studies have evaluated the feasibility of utilizing an ultrasound radiomics method in breast carcinoma to predict the luminal BC, and currently most of these focus on predicting triple-negative breast carcinoma, HER2 status, Ki-67 index, etc.

Leithner et al.’ study (22) including 91 breast cancers adopted a multi-layer perceptron feed-forward artificial neural network (MLP-ANN) to differentiate the hormone receptor-positive from hormone receptor-negative BC based on multiparametric MRI images, yielding an overall median AUC of 0.69, with median accuracies of 64.7% in the training dataset and 60.0% in the validation dataset. As compared to the above study, more samples were included in our study and applied the tenfold cross validation, the nomogram model achieved significantly higher median accuracies of 83.36% in the training set and 80.05% in the test set, with a median AUC value of 0.765 in the test set. This might lead by that the ultimate effect of the model is closely relevant to the generalization of the neural network and sample size. If the sample dataset is weakly representative, there are a number of conflictive or redundant samples, and then it is difficult for the neural network to obtain the expected result.

For the prediction of the luminal type, the AUC values of the Rad-score and nomogram were 0.786 and 0.767 in the test set. In a recent non-contrast-enhanced chest CT radiomics study (21), Wang and colleagues established forty-two models to predict the luminal type of breast cancer by the combination of six feature screening methods and seven machine learning classifiers. The final optimal model for external validation on the independent test set obtained an AUC value of 0.757. With regard to feature dimensionality reduction algorithms, they found that the overall performance of the LASSO regression was better than other dimensionality reduction methods in the field of AUC and accuracy. In our study, we also adopted LASSO algorithm for feature dimensionality reduction. In their study, the prediction model was established by seven different machine learning classifiers, among which support vector machine achieved the highest AUC value in both internal and external validations. On the contrast, in our study, models were built by using logistic regression only, and thus, further studies adopting other machine learning classifiers should be taken account in future. However, in our study, the Rad-score and nomogram all showed appreciable performance in the training and test sets. We believe that they could be utilized as a reliable technique in discriminating the luminal from non-luminal BC and may promote to assist clinicians for preoperative decision-making.

Clinical features including US-reported tumor size, age, tumor location, ultrasound equipment and ultrasound-reported lymph node metastasis were assessed in this study. Among them, US-reported tumor size (p < 0.001) showed a significantly statistical difference between the luminal and non-luminal BC in the training set. Furthermore, several previous studies have demonstrated that there was statistical difference in the terms of tumor size between the luminal and non-luminal BC (32, 33). The nomogram integrated with the US-reported tumor size showed a little higher predictive performance than that of Rad-score (AUC, 0.832 vs. 0.828) in the training set, but a little lower (AUC, 0.767 vs. 0.786) in the test set. However, no matter in the training set or in the test set, there was no statistical difference in the field of predictive performance between the nomogram and Rad-score in our study, indicating that the US-reported tumor size had limited effect to predictive performance of the nomogram. Furthermore, the US-reported tumor size could be acquired preoperatively, and then the model could be utilized for individualized prediction of the luminal type in patients with BC.

In the present study, the Rad-score was developed based on 11 luminal-related features, among which 1 first order feature, 5 glszm features, 1 glrlm feature and 1 gldm feature were included. A mix of first-order, texture and wavelet features seemed to be of importance for differentiation between the luminal and non-luminal BC, suggesting the complementary value of the combination of different radiomics features to capture different functional aspects of tumor biology. Six out of eleven radiomics features were texture features that showed the value of inter-tumor heterogeneity in predicting the gene expression (34, 35). At the same time, eight out of eleven radiomics features were wavelet features, which can also be utilized to extensively quantify heterogeneity of the tumor through different spatial scales at respective directional orientations (36). Furthermore, the first-order statistics feature such as Kurtosis appeared among the final radiomics features. Kurtosis describes the intensity value of the tumor, which is applied to many classification tasks (37, 38). Besides, radiomics features including ‘wavelet-HHL_gldm_DependenceNonUniformityNormalized’, ‘original_glrlm_HighGrayLevelRunEmphasis’, ‘wavelet-LHL_glszm_SizeZoneNonUniformity’ and ‘original_glszm_SmallAreaEmphasis’ were significantly associated with molecular subtypes of BC in ultrasound images, which had a higher proportion of the weight coefficient. However, the shape-based feature was not selected to constitute the radiomics model, indicating that the morphological characterization might be less relevant to the luminal type. Hence, we believe that tumor molecular level information can be obtained from tumor radiomics analysis and radiomics features extracted from ultrasound images of BC are available to predict molecular subtypes.

Our study had several limitations that should be taken into account. First, the retrospective nature of the analyses might have introduced selection bias. In addition, this was a single-center study with a limited number of 264 patients. Hence, future prospective studies are needed to further validate the predictive performance of the models by using a large, multi-center cohort. Second, the contour of tumor’s ROI was manually depicted by sonographers, but the judgment of tumor’s contour was easily influenced by personal subjective experience. However, we believe that this was partially solved by interobserver reproducibility assessment. Third, only gray-scale ultrasound images were adopted in our study, and in the future, we will add radiomics features of multimodal ultrasound to further studies. For example, contrast-enhanced (39) and elastography (40) ultrasound images, which may contain more radiomics features than gray-scale ultrasound images. Finally, our study was only conducted with two dimensional analysis of the largest plane of the tumor, which might not comprehensively capture the heterogeneous features of the tumor as compared to a model on the basis of three dimensional analysis. Future studies should focus on the establishment of a three-dimensional model for distinguishing the luminal from non-luminal BC.

In summary, both the Rad-score and nomogram can be applied as useful, noninvasive tool for preoperatively discriminating the luminal from non-luminal type in patients with BC. Our study may provide a novel method to evaluate molecular subtypes of BC. However, further studies with a prospective design and larger population are required to validate the conclusions.

The original contributions presented in the study are included in the article/ Supplementary Material . Further inquiries can be directed to the corresponding authors.

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

JW and LG collected the clinical and radiomics data. YW and YJ preprocessed patients’ ultrasound images and drew the ROI. JW and YJ analyzed the data and developed the prediction model. JW wrote the manuscript. ZW, DX, and LH designed the study. All authors contributed to the article and approved the submitted version.

Jinhua Science and Technology Bureau Scientific Research Project (2022-3-019).

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.