Introduction

Breast cancer is the leading type of cancer and a primary contributor to cancer-associated mortality among the female population [1]. Human epidermal growth factor receptor 2 (HER2) is a prototypic oncogene of breast cancer. It plays a crucial role in prognostic evaluation and is an established therapeutic target with high response rates [2]. Traditionally, based on immunohistochemistry (IHC) and in situ hybridization (ISH), breast cancer is dichotomously classified as HER2-overexpressing status (IHC score 3 + or IHC 2 + with ISH amplification) and HER2-negative status (IHC score 0 or 1 + , or IHC 2 + without ISH amplification) to select patients for HER2-targeted therapy [3]. However, in a recently proposed classification system, HER2-negative cases in traditional classifying criteria are sub-classified as HER2-low-expressing status (IHC 1 + or IHC 2 + without ISH amplification) and HER2-zero-expressing status (IHC score 0) [3, 4]. HER2-low-expressing, accounting for more than half of all traditionally HER2-negative breast cancers, have different biological characteristics and clinical outcomes compared with HER2-zero-expressing breast cancers or HER2-overexpressing breast cancers [2, 3, 5, 6]. Patients with HER2-zero-expressing breast cancers might be excluded from anti-HER2 therapy, and patients with HER2-overexpressing could benefit from traditional HER2-targeted therapy such as trastuzumab [3]. With the emergence of innovative HER2-targeted modalities like antibody–drug conjugates, even individuals exhibiting lower levels of HER2 expression might potentially derive therapeutic advantages from this medicinal intervention [4, 5, 7, 8]. Thus, identifying HER2-overexpressing, HER2-low-expressing, and HER2-zero-expressing statuses in patients with breast cancer could be highly desirable for determining potential candidates for anti-HER2 therapy [9].

Preoperative HER2 status assessment is usually based on pathologic analysis of biopsy specimens. However, breast cancer has considerable heterogeneity at multiple levels, affecting clinical behavior and therapeutic response [6]. The limited biopsy tissue may not comprehensively represent the entire tumor’s genetic, epigenetic, and phenotypic alterations [10, 11]. In addition, genetic profiling based on pretreatment biopsy is performed at a single time point, but tumor biology potentially changes over time and during or after treatment [7]. Furthermore, the percutaneous breast tumor biopsy is an invasive procedure. Therefore, a simple, practical, and noninvasive method is needed to assess the HER2-expressing status before treatment in breast cancer patients.

MRI is widely used for breast tumor detection, categorization, and monitoring neoadjuvant treatment effects [12]. Radiomics analysis of MR images can translate medical images into high-dimensional data, reflecting the tumoral molecular properties and tumoral heterogeneity in the subtle and spatial view [13]. Several MRI-based radiomics models, including single-parametric and multiparametric MRI, have been proposed to predict traditionally classified HER2-overexpressing and HER-negative status [13, 14]. Recently, two studies have applied MRI-based radiomics models to distinguish HER2-zero, -low, and -overexpressing statues of breast cancer [15, 16]. Ramtohul et al developed an MRI radiomics signature to discriminate the HER2-low and -positive vs. HER2-zero cancers [15]. The HER2-positive vs. HER2-low cancers could be further distinguished by a model incorporating the pathologic characteristic (histologic type) based on the pretreatment biopsy and MRI findings (presence of nonmass enhancement, multiple or single lesions) [15]. Bian et al found that the radiomics models based on the intratumoral and peritumoral radiomic features from multiparametric MRI could distinguish HER2-positive vs. HER2-negative breast cancers and HER2-low vs. HER2-negative cancers [16]. These results indicate that the three classified HER2 statuses of breast cancer may have distinct microstructure characteristics, which can be captured and discriminated by MRI-based radiomics features.

In this study, based on breast cancer data from two institutions, we developed and validated three multiparametric MRI-based radiomics predictive models for noninvasively discriminating the HER2-overexpressing vs. others, HER2-low-expressing vs. others, and HER2-zero-expressing vs. others status. Our study aims to predict HER2 status preoperatively through MRI-based radiomics, aiding in selecting candidates for traditional/new HER2-targeted therapies and identifying those who can be exempted from anti-HER2 treatment.

Methods

Ethics

This retrospective investigation obtained approval from the Ethics Committees of the respective institutions, with informed consent waived due to the study’s retrospective design.

Study population

From January 2016 to July 2022, 651 breast cancer patients from Shantou Center Hospital (hereafter, Institution I) were collected; between January 2019 and September 2020, 275 breast cancer patients from Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University (hereafter, Institution II), were collected. The patient enrollment pathway is presented in Fig. 1. The inclusion criteria comprised the following: (1) female individuals diagnosed with primary breast cancer through pathological examination; (2) patients who underwent breast multiparametric MRI; (3) HER2 status was determined by the IHC and/or ISH of postsurgical specimens. The exclusion criteria were as follows: (1) pathologic results of HER2 expressing status were incomplete or not available; (2) incomplete MRI data or poor quality of MRI images such as motion artifacts; (3) patients underwent breast MRI after breast lesion biopsy, surgery, or neoadjuvant therapy; (4) patients underwent neoadjuvant therapy before breast surgery and pathologic examination. The tumor size was not an inclusion or exclusion criterion in this study.

Fig. 1
figure 1

The patient enrollment pathway in two institutions. n1, number of patients in institution I; n2, number of patients in institution II

Full size image

Clinicopathologic characteristics of patients

The clinical characteristics, including the patient’s age and family history of breast cancer, were obtained from the electronic medical record system. The pathologic characteristics of breast cancer, including histology (invasive ductal carcinoma or others), HER2 status, hormone receptor (estrogen receptor [ER] and progesterone receptor [PR]) status, and Ki-67 index were collected from the hospital’s pathology system. ER or PR negativity was defined as less than 1% of tumor cells showing positive nuclear staining, while ER or PR positivity was defined as 1% or more of tumor cells exhibiting positive nuclear staining. The threshold for Ki-67 was set at less than 14% for low Ki-67 level and 14% or more for high Ki-67 level.

HER2 status classification

HER2 status was determined through IHC or ISH, following the clinical practice guidelines for HER2 testing in breast cancer issued by the American Society of Clinical Oncology/College of American Pathologists [4, 17]. Further ISH testing was done for IHC 2 + results. Patients were grouped into HER2-overexpressing (IHC 3 + or IHC 2 + with ISH amplification), HER2-low-expressing (IHC 1 + or IHC 2 + without ISH amplification), and HER2-zero expressing (IHC 0) [3, 14].

MRI acquisition

MRI was conducted using a 3.0-T MRI scanner (Siemens Magnetom Verio) or 1.5-T MRI scanner (Siemens Magnetom Area) with an 8-channel breast coil in Institution I, and 1.5 T MR scanner (Siemens Magnetom Avanto) with an 8-channel breast coil in Institution II. Imaging sequences included axial T1-weighted imaging (T1WI), axial turbo-spin-echo T2-weighted imaging (T2WI) with fat suppression (FS), axial single-shot diffusion-weighted imaging (DWI), and axial dynamic contrast-enhanced (DCE) imaging. The apparent diffusion coefficient (ADC) map was calculated in a pixel-by-pixel manner from DWI via the manufacturer’s software (syngo.via; Siemens Healthcare), based on the formula ADC =  − In [S(b2) − S(b1)] ÷ (b2 − b1), where S(b1) and S(b2) were the signal intensity corresponding to b values of 50 and 800 s/mm2 at Institution I, and 0 and 800 s/mm2 at Institution II, respectively. Before DCE, an initial T1WI-FS pre-contrast scan was performed, followed by intravenous gadolinium contrast injection (Gadopentetate Dimeglumine, Magenvist, 0.2 mmol/kg at 2 mL/s in Institution I; Gd-DTPA-BMA, Omniscan, 0.1 mmol/kg at 3.5 mL/s in Institution II), with 20 mL saline flush. DCE consisted of one pre-contrast and five post-contrast phases with each phase for 60 s at Institution I, and two pre-contrast and 38 post-contrast phases with 8-s temporal resolution at Institution II. MRI sequence details are given in Table 1.

Table 1 Multiparametric breast MRI sequences in two institutions
Full size table

Two radiologists (reader 1: Y.W.Z., 13 years of breast lesion expertise from Institution I; reader 2: Y.Q., 7 years of breast lesion experience from Institution II) measured the long tumor diameter on the peak-enhanced DCE phase (2nd phase in Institution I, 12th in Institution II) [18] on ITK-SNAP software (version 3.8.0, http://www.itksnap.org). The average value of each breast tumor was recorded.

Breast lesion segmentation

The whole tumor’s volume-of-interest (VOI) was segmented on the peak-enhanced phase of DCE, T2WI-FS, DWI, and ADC images via ITK-SNAP software (version 3.8.0, http://www.itksnap.org) independently by two experienced radiologists in breast MRI (reader 1: Y.W.Z. from Institution I; and reader 2: Y.Q. from Institution II), independently. Readers 1 and 2 independently performed VOI segmentation on a randomly chosen subset of 50 patients from the training dataset, repeated after 2 weeks. The intra-class correlation coefficients (ICCs) of radiomics features were calculated to assess the intra-rater reproducibility (first vs. second segmentations by each reader) and inter-rater reproducibility (first segmentation by reader 1 vs. first segmentation by reader 2, second segmentation by reader 1 vs. second segmentation by reader 2) of image segmentation. ICC > 0.75 indicated satisfactory agreement.

The breast lesion segmentation was performed separately on DCE, T2WI-FS, and DWI images. For DCE, the VOIs were manually delineated on tumor-contouring slices in the peak-enhanced phase of DCE (the 2nd phase, 60 s post contrast enhancement in Institution I and the 12th phase, 96 s post contrast enhancement in Institution II), on which the tumor can be better differentiated from the neighboring tissues [18]. For DWI, the VOIs were manually drawn on DWI images (b values = 800 s/mm2). Then, VOIs of DWI were directly applied to ADC maps and manually corrected in the case of mismatched areas. During the segmentation, the VOIs were delineated to cover all the tumor areas while excluding edema, blood vessels, and normal fibroglandular tissue.

Radiomics feature extraction and selection

The radiomics analysis workflow is shown in Fig. 2. Radiomics feature extraction was conducted utilizing the Pyradiomics package (version 3.8.3, http://pyradiomics.readthedocs.io). Before radiomics feature extraction, all segmented VOIs underwent z-score normalization to standardize the distribution of image voxels [19]. A total of 6432 radiomics features (1608 radiomics features each from DWI, ADC, T2WI-FS, and DCE) were obtained for each patient [20]. The extracted radiomics features included morphological features and texture features. Details of radiomics features are shown in Supplementary Table S1.

Fig. 2
figure 2

Workflow of radiomics analysis. Abbreviations: DCE, dynamic contrast-enhanced imaging; T2WI-FS, T2-weighted imaging fat suppression; DWI, diffusion weighted imaging; ADC, apparent diffusion coefficient; VOI, volume of interest; LASSO, shrinkage and selection shrinkage and selection operator; ROC, receiver operating characteristic

Full size image

Radiomics model development and validation

Three MRI-based radiomics models were developed to differentiate (1) HER2-overexpressing vs. others (HER2-low-expressing + HER2-zero-expressing); (2) HER2-low-expressing vs. others (HER2-overexpressing + HER2-zero-expressing); and (3) HER2-zero-expressing vs. others (HER2-overexpressing + HER2-low-expressing). The least absolute shrinkage and selection operator (LASSO) regression method was used to select the optimal HER2 status–related radiomics features from the training dataset. The radiomics signature was represented as a radiomics score, which was computed using a formula that integrated chosen radiomics features, weighted by their corresponding coefficients obtained from LASSO regression. The predictive models were built by the selected radiomics signature in the training dataset and then tested in the independent and external validation datasets, respectively. The predictive performance of each model was quantified by the receiver operating characteristic (ROC) curve with the calculation of the area under the ROC curve (AUC).

Statistical analysis

All statistical analyses were conducted using SPSS (version 25.0; IBM Corp) and R software (version 4.1.2, https://www.r-project.org). Continuous variables were presented as means ± standard deviation, while categorical variables were conveyed as absolute numbers (n) and their corresponding proportions (%). Differences in categorical variables among HER2-expressing statuses were evaluated with Fisher’s exact or Pearson chi-square tests (i.e., family history of breast cancer, histology of breast cancer, hormone receptor status, and Ki-67 level). Differences in continuous variables among HER2-expressing statuses were evaluated with ANOVA (i.e., age and MRI-determined long diameter of breast cancer). The LASSO regression analysis was used to select the optimal HER2 status-related radiomics features and build the predictive models for discriminating different HER2 statuses. The ROC curve, along with sensitivity, specificity, and accuracy, was employed to evaluate the discriminative performance of each predictive model. p < 0.05 was considered statistically significant.

Results

Clinicopathologic characteristics of patients

In total, 581 patients in Institution I and 190 patients in Institution II met the inclusion criteria of our study. Patients in Institution I were randomly divided into a training dataset (n = 407) and an independent validation dataset (n = 174) with a ratio of 7:3. One hundred and ninety patients from Institution II were allocated to the external validation dataset. The clinical characteristics of patients in the training, independent validation, and external validation datasets are shown in Table 2. Among the whole dataset, 256 (33.2%), 360 (46.7%), and 155 (20.1%) patients had HER2-overexpressing, HER2-low-expressing, and HER2-zero-expressing breast cancer, respectively. In the training dataset, 128 (31.4%), 201 (49.4%), and 78 (19.2%) patients had HER2-overexpressing, HER2-low-expressing, and HER2-zero-expressing breast cancer, respectively. In the independent validation dataset, 64 (36.8%), 72 (41.4%), and 38 (21.8%) patients had HER2-overexpressing, HER2-low-expressing, and HER2-zero-expressing breast cancer, respectively. In the external validation dataset, 64 (33.7%), 87 (45.8%), and 39 (20.5%) patients had HER2-overexpressing, HER2-low-expressing, and HER2-zero-expressing breast cancer, respectively. Hormone receptor status and Ki-67 level were significantly different among the three HER2-expressing groups in training (both p < 0.001), independent validation (p < 0.001 and p = 0.007, respectively), and external validation (p = 0.001 and p = 0.008, respectively) datasets. The MRI-determined long diameter of breast cancer significantly differed among the three HER2-expressing groups in the training dataset (p < 0.001). There were no statistically significant differences in age (p = 0.664, 0.916, 0.084, respectively), family history of breast cancer (p = 0.463, 0.411, 0.165, respectively), and histology of breast cancer (p = 0.389, 0.715, 0.510, respectively) among three datasets.

Table 2 Clinicopathologic characteristics of patients in training, independent validation, and external validation dataset
Full size table

Development of MRI-based radiomics models

The development and validation of the MRI-based radiomics predictive models are shown in Figs. 1 and 2. The MRI-based radiomics features associated with HER2-overexpressing, HER2-low-expressing, and HER2-zero-expressing statuses were respectively selected in the training dataset (Fig. 3). Finally, 11 radiomics features were selected for differentiating HER2-overexpressing breast cancers from others; one radiomics feature was selected for differentiating HER2-low-expressing breast cancers from others; 17 radiomics features were selected for differentiating HER2-zero-expressing breast cancers from others. The terms and corresponding coefficients of the radiomics feature in the LASSO regression model are listed in Table 3. In the training dataset, optimal MRI radiomics score cutoffs were − 0.775 for HER2-overexpressing vs. others, 0.021 for HER2-low-expressing vs. others, and − 1.397 for HER2-zero-expressing vs. others. MRI radiomics models in the training dataset achieved AUCs of 0.809 for HER2-overexpressing breast cancers vs. others, 0.779 for HER2-low-expressing vs. others, and 0.889 for HER2-zero-expressing vs. others, listed in Table 4.

Fig. 3
figure 3

Radiomics feature selection using the least absolute shrinkage and selection operator (LASSO) logistic regression. A, D, G The selection of tuning parameter (λ) in the LASSO model (A HER2-overexpressing vs. others; D HER2-low-expressing vs. others; G HER2-zero-expressing vs. others) using tenfold cross-validation via minimum criteria. The area under the curve (AUC) was plotted versus log (λ). A dotted vertical line was drawn at the optimal values using the minimum criteria and the one standard error of the minimum criteria. Log (λ) was set as − 2.707408, − 1.508029, and − 2.828045, respectively, according to tenfold cross-validation in this study. B, E, H The dotted vertical line was plotted at the selected λ value, resulting in 11, 1, and 17 non-zero-coefficient features, respectively. C, F, I The correlation coefficients of the 11, 1, and 17 selected radiomics features in discriminating HER2-overexpressing status vs. others, HER2-low-expressing status vs. others, and HER2-zero-expressing status vs. others, respectively

Full size image
Table 3 Selected radiomics features for predictive model development
Full size table
Table 4 Performances of predictive models for HER2 status discrimination
Full size table

Validation of MRI-based radiomics predictive models

The predictive performance in differentiating different HER2-expressing statuses is shown in Table 4 and Fig. 4. MRI radiomics models achieved an AUC of 0.737 and 0.725 for differentiating HER2-overexpressing breast cancers from others, 0.778 and 0.782 for differentiating HER2-low-expressing breast cancers from others, and 0.867 and 0.813 for differentiating HER2-zero-expressing breast cancers from others in the independent validation and external validation datasets, respectively.

Fig. 4
figure 4

Receiver operating characteristic (ROC) curves of the MRI-based radiomics predictive models for HER2 status discrimination in training (A), independent validation (B), and external validation (C) datasets. AUC, area under the curve; CI, confidence interval

Full size image

Discussion

Our study findings demonstrated that the multiparametric MRI-based radiomics predictive models had good performance in identifying HER2-overexpressing status, HER2-low-expressing status, and HER2-zero-expressing status in breast cancers with an AUC ranging from 0.725 to 0.813 in the external validation cohort. These multiparametric MRI radiomics models hold the potential to classify the HER2 status noninvasively and may assist in clinical decision-making for breast cancer.

The effectiveness of HER2-targeted therapies like trastuzumab is widely recognized in HER2-overexpressing breast cancer patients, whereas it lacks efficacy in those with HER2-negative breast cancers. However, the traditional HER2-negative breast cancer includes the population of HER2-low-expressing plus HER2-zero-expressing breast cancer, as indicated in the newly proposed classification system [21]. Previously, several single-parametric or multiparametric MRI radiomics models have been developed to differentiate HER2-overexpressing from HER2-negative breast cancers. The single-parametric MRI radiomics models showed an AUC ranging from 0.65 to 0.69 [22,23,24]. For example, Li et al reported an MRI radiomics model based on DCE to differentiate HER2-overexpressing from HER2-negative disease, with an AUC of 0.65 [22]. Comparatively, multiparametric MRI radiomics models showed higher performance with an AUC ranging from 0.69 to 0.887 [12,13,14]. For example, Xu et al developed a multiparametric (T2WI, DCE, and ADC) MRI radiomics model incorporating a machine learning algorithm for identifying HER2-overexpressing and HER2-negative breast cancers and showed the highest AUC of 0.887 [13]. Our study developed a multiparametric MRI-based radiomics model to discriminate HER2 overexpressing breast cancers from traditionally negative HER2 status (HER2-low-expressing plus HER2-zero-expressing). The MRI radiomics model in our study incorporated radiomics features from DCE, DWI, and ADC and showed good performance with an AUC of 0.809 in the external validation dataset. The performance of our multiparametric MRI radiomics model was comparable with the previously reported multiparametric MRI radiomics models [12,13,14] while superior to the previously described single-parametric MRI radiomics models [23, 24]. The better performance of multiparametric MRI radiomics models may be attributed to the fact that more complementary information can be derived from multiparametric sequences, such as morphologic findings, micro-vessel permeability, and water diffusion properties [14, 25].

Notably, the new HER2-targeted therapy approach and its clinical benefits in HER2-low-expressing breast cancer have dramatically revolutionized the clinical treatment landscape of traditional HER2-negative breast cancers [8]. A new group of patients, those with HER2-low-expressing breast cancers, is now recognized as having potential benefits from novel HER2-targeted therapies. The clinicopathologic characteristics and targeted therapies of this new group of breast cancer have been elucidated [2,3,4, 26, 27]. A combined assessment of patients from four clinical trials revealed notable distinctions between HER2-low-expressing breast cancers and HER2-zero-expressing breast cancers in biologic features, clinicopathologic findings, therapeutic response, and clinical outcome [2]. A recent phase III clinical trial showed that trastuzumab deruxtecan (a kind of antibody–drug conjugate) significantly improved the survival time in individuals with advanced HER2-low-expressing breast cancer compared with the chemotherapy, highlighting the clinical significance of the HER2-low-expressing, and the need to redefine this subgroup from the traditional HER2-negative breast cancers [3].

Recently, two studies indicated the potential utility of MRI-based radiomics models for predicting HER2 status, including the HER2-low-expressing category [15, 16]. Ramtohul et al reported that the MRI-based radiomics signature achieved an AUC ranging from 0.78 to 0.79 for discriminating HER2-low plus HER2-over vs. HER2-zero breast cancers [15]. The established MRI-based radiomics signature involved 6 radiomic features from T2WI and one radiomic feature from T1 contrast-enhanced images [15]. Further, they built a combined model incorporating the pathologic feature (i.e., histology type) and MRI findings (i.e., nonmass enhancement and multiple lesions) for the prediction of HER2-low vs. HER2-positive cancers, with a reported AUC of 0.78 and 0.77 in the training and external dataset, respectively [15]. Bian et al reported that two MRI radiomics signatures were able to distinguish HER2-overexpressing from HER2-negative (AUCs range, 0.760 to 0.793) and HER2-low from HER2-negative breast cancers (AUCs range, 0.711 to 0.820) [16]. One radiomics signature consisting of 13 radiomic features (6 features from T1-weighted contrast-enhanced images and 7 features from ADC images) was found to be able to distinguish HER2-overexpressing vs. HER2-negative cancers. Another radiomics signature consisting of 9 radiomics features (6 features from T1-weighted contrast-enhanced images and 3 features from ADC images) was developed to differentiate the HER2-low vs. HER2-negative cancers [16]. In our study, we extracted high-throughput radiomics features from four imaging sequences to construct MRI radiomics models to differentiate the newly proposed HER2 expression statutes of breast cancer. A total of 6432 radiomics features (1608 radiomics features each from DWI, ADC, T2WI-FS, and DCE) were extracted from each patient. Then, we developed and validated MRI-based radiomics models to differentiate each HER2-expressing status (i.e., HER2-overexpressing status, HER2-low-expressing status, and HER2-zero-expressing) from other statuses. Our study showed that the predictive model with 11 radiomics features, including 3 features from DWI, 4 from ADC, and 4 from DCE images, had AUCs ranging from 0.725 to 0.809 in differentiating HER2 overexpressing breast cancers from HER2-low-expressing plus HER2-zero-expressing breast cancers. In addition, a predictive model with one radiomics feature from DWI had AUCs ranging from 0.778 to 0.782 in differentiating HER2-low-expressing breast cancers from HER2-overexpressing plus HER2-zero-expressing breast cancers. Finally, a predictive model with 17 radiomics features, including 2 features from DCE, 8 from DWI, and 7 from ADC images, was developed and validated to differentiate HER2-zero-expressing from HER2-overexpressing plus HER2-low-expressing breast cancers with AUCs ranging from 0.813 to 0.889. However, no radiomics feature from T2WI was selected to develop the radiomics model, which may indicate that the radiomics features derived from DWI, ADC, and DCE have stronger capability for discriminating different HER2-expressing statuses. These results showed that features derived from multiparametric (i.e., DWI, ADC, and DCE) MRI radiomics hold the potential as a noninvasive tool for identifying the HER2-overexpressing, HER2-low-expressing, and HER2-zero-expressing breast cancers, which might help to determine potential candidates for traditional anti-HER2 therapy, new anti-HER2 antibody–drug conjugates therapy, or patients who may be excluded from anti-HER2 treatment.

There were some limitations in our study. First, since the study was conducted retrospectively, potential bias in patient selection may exist. Second, to compute the ROC curve for a three-classification event, we changed the three-classification to two-classification to distinguish different HER2-expressing statuses. More advanced statistical methods might solve this drawback. Third, the tumor VOI segmentation was performed manually, and the traditional handcrafted radiomics feature extraction was time-consuming and complicated. These might be overcome with a more automatic segmentation approach, such as deep learning frameworks.

In summary, our study demonstrated that multiparametric MRI-based radiomics models hold potential as a noninvasive tool to discriminate the HER2-expressing status in breast cancer patients. Pretreatment identification of HER2-low-expressing from other kinds of HER2-expressing breast cancers will facilitate the refinement of anti-HER2 therapy in breast cancer patients.