Multiparametric MRI-based radiomics nomogram for noninvasive stratification of HER2 expression status in breast cancer

Introduction

Breast cancer (BC) is the most prevalent and potentially fatal malignancy among females (1). Among its various histological subtypes, invasive ductal carcinoma (IDC) predominates, accounting for approximately 80% of all BC cases. Within this population, it is estimated that 20–30% of patients present with a positive status for human epidermal growth factor receptor 2 (HER2) (2). Many studies have confirmed that the HER2 gene, encoding a receptor tyrosine kinase, plays a critical role in cell growth and differentiation in BC (3). HER2 positivity is linked to increased disease aggressiveness and poor prognosis (4). HER2 status has independent prognostic value and is a key indicator of the effectiveness of targeted therapies. Clinical trials targeting HER2 in BC patients have significantly reduced recurrence and metastasis in HER2-positive BC patients, marking a major milestone in BC treatment (5).

Traditionally, HER2 status is classified into two categories, HER2-positive and HER2-negative, with only positive patients eligible for HER2-targeted therapies (6). Positive results are determined by immunohistochemical (IHC) staining with 3+ or 2+ with positive fluorescence in situ hybridization (FISH), whereas IHC 0, 1+, or 2+ without FISH gene amplification is defined as HER2-negative. Over the past two years, the emergence of novel antibody-drug conjugates (ADCs) has opened new therapeutic opportunities for BC with low levels of HER2 expression that had previously been classified as “HER2-negative” (6,7). HER2-negative expression is now subclassified into HER2-low (IHC 1+/2+ with FISH negative) and HER2-zero (IHC score of 0) (8). A phase Ib clinical trial demonstrated that the novel HER2-targeted ADCs, such as trastuzumab deruxtecan (T-Dxd) exhibited significant antitumor activity in patients with HER2-low-expressing BC, yielding an objective response rate of 37.0% and a median duration of response of 10.4 months (7). These findings indicate that this agent provides a clinically meaningful treatment option even in patients with HER2-low expression, potentially revolutionizing the therapeutic landscape for this population. Consequently, this therapeutic advancement necessitates a more rigorous classification of the HER2 expression assessment system in BC to optimize patient selection and treatment outcomes. The recent reclassification of BC into three categories—HER2-zero, HER2-low, and HER2-positive—has been increasingly adopted in clinical settings.

Currently, pathological tissue biopsy is the gold standard for BC diagnosis, but it is limited by tissue heterogeneity, yielding HER2 status agreement between core needle biopsies and subsequent resection biopsy rates of 81–96% (9). Biopsy is invasive and the corresponding HER2 results take 1–2 weeks, which is time-consuming and adds to the psychological burden on patients (10). Meanwhile, patients with HER2 (2+) status often require additional FISH testing, increasing their financial burden. Consequently, there is an urgent need for a rapid and cost-effective diagnostic method to accurately assess HER2 status in BC patients.

Breast magnetic resonance imaging (MRI) is an indispensable diagnostic tool in breast imaging, which is renowned for its exceptional sensitivity in detecting BC and monitoring responses to neoadjuvant chemotherapy (11). T2-weighted imaging (T2WI) is pivotal in identifying internal features of breast abnormalities, such as bleeding, edema, and cysts, by highlighting the correlation between signal intensity and lesion composition (12). Dynamic contrast-enhanced MRI (DCE-MRI) further enhances the assessment of BC by providing detailed morphological insights and examining tumor perfusion dynamics through enhancement patterns, which are vital for understanding the key tumor characteristics of malignancy (13). In recent years, with the rapid development of artificial intelligence (AI), MRI-based radiomic algorithms have been used to extract rich radiomic features from MR images. These features are then analyzed and selected via machine and deep learning techniques. The resulting models facilitate noninvasive assessments of tumor molecular characteristics, enabling evaluations of tumor heterogeneity over time and space (14).

Recently, various MRI-based radiomic models have been developed, including those that utilize both single-parametric and multiparametric MRI, to predict the novel tri-subgroup classification of BC’s HER2 status (15,16). Zheng et al. developed three predictive radiomics models aimed at differentiating HER2-positive and others, HER2-low and others, and HER2-zero and others (15). Similarly, Liu et al. also constructed a comprehensive model that combines conventional MRI (cMRI) and radiomics features to predict the HER2 expression of invasive BC (16). However, these radiomic models have yet to show excellent performance, with areas under the curve (AUCs) ranging from 0.6 to 0.8, and all lack consideration of clinical features in the prediction model. Given that HER2-low BC exhibits distinct biological and clinicopathological features, therapeutic responses, and clinical outcomes when compared to HER2-zero and HER2-positive BC (17), it is imperative to comprehensively consider clinical characteristics to better capture the intricate heterogeneity within this tumor subset. In this context, accurate evaluation of HER2 status is critical for the appropriate use of HER2-targeted therapies, which are not only specific for differentiating patients with HER2-zero BC from those with nonzero BC (including HER2-low and HER2-positive patients) to exclude HER2-zero BC for targeted therapy, but also break the traditional HER2 dichotomies to further distinguish HER2-low BC from HER2-negative BC to find beneficiaries of new anti-HER2 drug conjugates. Unlike previous radiomic models, our study used multicenter data and incorporated clinical characteristics to improve prediction. The radiomics nomogram we constructed optimized resource allocation and reduced diagnostic delays. It provided a visualization framework which was easy to adopt in clinic for guiding HER2-targeted therapy selection.

This study aimed to develop a multiparametric MRI-based radiomic nomogram that incorporates radiomic features alongside clinical factors to assess the HER2 status of BC, especially the HER2-low status. We hypothesized that tumors with varying HER2 expression levels exhibit distinct radiomic features. These radiomic features, when integrated with clinical risk factors, could serve as robust predictive indicators. The validity of the resulting radiomic nomogram model was confirmed through a multicenter dataset, emphasizing its robustness and potential clinical applicability. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1707/rc).

Methods

Study participants

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University (No. 2024.144), and the requirement for patient informed consent was waived due to the retrospective nature of the study. The study retrospectively included a cohort of 364 BC patients sourced from a pair of distinct medical centers (Center 1, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University; Center 2, Ganzhou Cancer Hospital) with different MRI scanners between September 2022 and March 2024. The inclusion criteria were as follows: (I) patients who underwent BC surgery and whose IDC was confirmed through pathological assessment of the surgically removed tissue samples; and (II) patients who underwent breast MRI scans within two weeks prior to surgery. The exclusion criteria were as follows: (I) patients who underwent breast MRI after breast lesion biopsy or neoadjuvant therapy; (II) patients whose MR images had artifacts of low quality; and (III) patients with incomplete pathology data. Ultimately, a total of 314 patients were enrolled in the study, comprising 300 individuals with unilateral BC and 14 with bilateral BC. We devised two tasks: Task 1 aimed to differentiate BCs with HER2-zero expression from those with nonzero expression (encompassing HER2-low and HER2-positive), whereas Task 2 sought to separate BCs with HER2-low expression from those with HER2-negative expression. For Task 1, a total of 236 BC patients from Center 1 were enrolled, comprising 193 patients with HER2-low/positive expression and 43 with HER2-zero expression. These patients were then divided into training (n=165) and internal validation (n=71) cohorts via stratified sampling at a ratio of 7:3. For Task 2, a total of 160 HER2-negative BC patients were included, consisting of 117 patients with HER2-low expression and 43 with HER2-zero expression. These patients were then divided into training (n=112) and internal validation (n=48) cohorts via stratified sampling at a ratio of 7:3. A total of 78 patients from Center 2 were allocated to an external validation cohort, which included 26 patients with HER2-positive expression, 36 with HER2-low expression, and 16 with HER2-zero expression. Figure 1 shows the patient flow diagram of the study.

Figure 1 Flowchart of patient inclusion and exclusion criteria. Task 1: the distinction of HER2-low and -positive versus HER2-zero breast cancers. Task 2: the distinction of HER2-low from HER2-negative breast cancers. AUC, area under the curve; HER2, human epidermal growth factor receptor 2; IDC, invasive ductal carcinoma; MRI, magnetic resonance imaging.

Clinicopathologic characteristics of patients

The clinical characteristics, including age, menstrual history, reproductive history, and a thorough evaluation of family history, emphasizing BC occurrence among first-degree relatives, were retrieved from the electronic medical record system. Pathological details of BC, such as lesion location, HER2 status, hormone receptor status [including estrogen receptor (ER) and progesterone receptor (PR)], histological grade, axillary lymph node (ALN) status, and the Ki-67 index, were obtained from the hospital’s pathology system. Histopathological findings were reviewed independently by two experienced pathologists, one with 3 years and the other with 12 years of specialized experience. All patients in this study underwent surgical procedures, and all pathological assessments were based on surgical tissue samples. Histological grades were classified as low (Grade I) or high (Grades II or III). The ER and PR status and the Ki-67 index were evaluated according to established guidelines from previous publications (18,19). Specifically, ER and PR statuses were categorized as negative (≤1%) or positive (>1%), and Ki-67 was classified as indicative of low proliferation (<14%) or high proliferation (≥14%).

HER2 status classification

According to the standardized testing guidelines outlined by the 2018 American Society of Clinical Oncology and the College of American Pathologists, all IHC scores indicative of HER2 expression were meticulously evaluated by specialized BC pathologists during the core needle biopsy procedure (20,21). For tumors exhibiting borderline IHC scores of 2+, further analysis was conducted via FISH to provide a more definitive assessment. Patients were categorized into three groups on the basis of FISH score: (I) HER2-zero (IHC score of 0); (II) HER2-low (IHC score of 1+ or 2+ with negative FISH results); and (III) HER2-positive (IHC score of 3+ or 2+ with positive FISH results).

MRI acquisition and image analysis

MRI examinations were conducted using a 3.0 Tesla scanner (SIGNA Architect; GE Healthcare, Chicago, IL, USA) equipped with an 8-channel breast coil at Center 1 and a 3.0 Tesla MR scanner [Omega; United Imaging (uMR), Shanghai, China] equipped with an 8-channel breast coil at Center 2. The patient was positioned in a prone posture, with arms extended alongside the head and breasts naturally suspended within the coils, feet first. Imaging protocols at both institutions included a comprehensive suite of sequences: axial T1-weighted imaging (T1WI), axial T2WI with fat suppression (FS), and DCE imaging. For DCE, we employed the differential sub-sampling with cartesian ordering (DISCO) technique, which utilizes sampling pattern in combination with view sharing, two-point Dixon fat-water separation, and parallel imaging (22). This method provides improved temporal resolution while maintaining the high spatial resolution used in the clinical setting. Many studies have demonstrated that this technique reduces the mutual constraints between temporal and spatial resolution, allows the capture of kinetic information of the lesion in the very early post-contrast phase, obtains more precise lesion enhancement features, and facilitates the assessment of BC heterogeneity (23).

Before DCE-MRI, an initial precontrast T1WI scan was conducted. This was followed by the injection of intravenous gadolinium contrast agent (gadopentetate dimeglumine, Magnevist, 0.1 mmol/kg at 2.5 mL/s at Center 1; Gd-DTPA-BMA, Omniscan, 0.1 mmol/kg at 2.0 mL/s at Center 2), along with a 20 mL saline flush. At Center 1, the DCE component consisted of 3 precontrast phases followed by 57 postcontrast phases. Each phase lasted for 4 seconds, resulting in a total of 60 phases. In contrast, at Center 2, the DCE sequence included 1 precontrast phase and 6 postcontrast phases. Each phase had a duration of 75 seconds, resulting in a total of 8 phases. The specifics of the MRI sequences are delineated in Table S1. To address MRI intensity inhomogeneity artifacts affecting model accuracy, N4 bias field correction (24) was applied to correct low-frequency spatial intensity variations, enhancing uniformity while preserving anatomical details crucial for segmentation and quantitative analysis.

Breast lesion segmentation

The segmentation of the entire tumor’s volume of interest (VOI) was independently performed by two seasoned radiologists via the three-dimensional (3D) slicer PyRadiomics module (https://pyradiomics.readthedocs.io/en/latest/index.html) on the peak-enhanced phase of both DCE and T2WI images. Reader 1, with 5 years of specialized experience in breast MRI from Center 1, and Reader 2, with 6 years of expertise in the same field from Center 2, conducted these segmentations independently and without access to clinicopathological information. When the assessments differed between the readers, a senior radiologist provided the final determination. Breast lesions were delineated independently on both DCE and T2WI sequences. For DCE, the VOIs were manually outlined on the slices depicting tumor contours during the peak enhancement phase (at the 20th phase, 80 s postcontrast at Center 1, and at the 2nd phase, 90 s postcontrast at Center 2). These specific phases were chosen to ensure accurate differentiation of the tumor from neighboring tissues (25). The tumor volume was determined by manually tracing the tumor contour layer by layer, excluding necrosis, calcification, edema, blood vessels, and normal tissue. T2WI was used as a guide for delineation in patients with ambiguous boundaries. In cases with multiple lesions, only the largest lesions were analyzed. Figure 2 illustrates the experimental workflow of this study.

Figure 2 Workflow of radiomics analysis. Model A: clinical characteristics only; Model B: DCE radiomics features only; Model C: T2WI radiomics features only; Model D: combined DCE and T2WI radiomics features; Model E: combined clinical characteristics and radiomics features (DCE + T2WI). DCA, decision curve analysis; DCE, dynamic contrast-enhanced imaging; ICC, intra-class correlation coefficient; LASSO, least absolute shrinkage and selection operator; MRI, magnetic resonance imaging; ROC, receiver operating characteristic; ROI, region of interest; T2WI, T2-weighted imaging.

Radiomics feature extraction and selection

Using the 3D slicer PyRadiomics module, we conducted feature extraction from the regions of interest (ROIs) in each patient’s omics imaging data. This process culminated in the extraction of 851 radiomic features per imaging sequence. These features encompassed a diverse range, including 18 first-order features, 14 shape features, 75 texture features, and 744 wavelet features. Among the texture features, we analyzed the gray level co-occurrence matrix (GLCM), gray level area matrix (GLSZM), gray level run-length matrix (GLRLM), neighborhood gray level difference matrix (NGTDM), and gray level dependence matrix (GLDM). The original features were transformed through the Wavelet filter to produce the wavelet features. Reader 1 and Reader 2 independently conducted VOI segmentation on a randomly selected subset of 40 patients from the training set. Both readers repeated the segmentation process after 2 weeks to validate consistency. The intraclass correlation coefficient (ICC) was calculated for the radiomic features to assess inter- and intra-rater reproducibility, with an ICC >0.75 signifying adequate reliability, and were eligible for inclusion in the MRI radiomics signature (RS) development.

First, the samples were partitioned into training and validation sets at a 7:3 ratio, with low-reproducibility features (ICC ≤0.75) being excluded to ensure data quality. Data standardization was subsequently performed, which involved resampling each voxel to a uniform size of 1×1×1 mm³ and normalizing the gray values via the Darwin Scientific Research Platform (Beijing Yizhun Intelligent Technology Co., Ltd., Beijing, China). The retained radiomic features were subsequently standardized, and the maximum relevance minimum redundancy (mRMR) method was used to scale the values to [0, 1] to eliminate the sizes of different features to initially eliminate redundant features. Since feature selection plays an important role in training classifiers, reducing computational complexity and improving classification accuracy, the optimal feature filter (sample variance F value; https://scikit-learn.org/stable/modules/generated/sklearn.feature_selection.SelectKBest.html) was used to evaluate the linear correlation between each feature and the category label, and the most relevant features were selected. Next, the least absolute shrinkage and selection operator (LASSO) and logistic regression were applied, with 5-fold cross-validation used to adjust the penalty parameters and select the optimal lambda value, further refining the feature set. The radiomic features from individual sequences (DCE and T2WI) were separately modeled. By integrating the salient features from each sequence and applying an additional dimensionality reduction process of feature selection, we identified the key features essential for joint modeling. This approach aimed to optimize both diagnostic accuracy and stability. These meticulously chosen features were subsequently linearly combined, with their respective regression coefficients serving as weights, to create a robust RS. For each patient, a radiomics score (Rad-score) was computed via the intercepts and coefficients derived from logistic regression. Further details about the features and Rad-score calculation formulas are provided in Appendix 1. A logistic regression model incorporating these features was subsequently developed. Ultimately, the training cohort served as the foundation for constructing the radiomic model, whereas the validation cohort was employed to assess its predictive ability.

Model development and validation

For each task, a distinct RS was formulated by aggregating the chosen features derived separately from DCE (RS-DCE) and T2WI (RS-T2WI) images. An integrated radiomics signature (RS-Com) was also created by consolidating all the selected features from both the DCE and T2WI sequences. RS development occurred using the training cohort, followed by its validation on both the internal and external validation cohorts.

Univariate and multivariate logistic regression were used to analyze the 11 clinical characteristics (age, body type, body position, PR status, ER status, Ki-67 index, ALN, histological grade, menstrual history, family history, reproductive history) and MRI RS of patients to determine the factors associated with HER2+ expression status. A clinical model and radiomics nomogram were established.

Consequently, five predictive models were devised to ascertain the HER2 expression in BC: a clinical model (Model A), RS-DCE (Model B), RS-T2WI (Model C), RS-Com (Model D), and a radiomics nomogram model (Model E). The model’s performance was assessed in terms of discrimination, calibration, and clinical utility via receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA), respectively. The DeLong test was used to compare the differences in AUC between each model pairing. Additionally, a series of quantitative metrics, including accuracy (ACC), sensitivity (SEN), and specificity (SPE), were calculated to further evaluate the model’s performance.

Statistical analysis

The R programming language (version 4.1.0, https://www.programmingr.com/) was utilized for conducting the statistical analyses and manipulating the data. For the comparison of continuous variables such as age and size, the independent samples t-test or the Mann-Whitney U test was employed, with results presented as the mean with its associated standard deviation (SD). Categorical variables including PR, ER, Ki-67, ALN, histological grade, menstrual history, family history, and reproductive history were compared between groups via the chi-square test or Fisher’s exact test, with outcomes depicted in terms of absolute counts (n) and their respective percentages (%). Univariate and multivariate logistic regression were used to assess the correlations between HER2 expression and various clinical factors. The statistical significance was determined via two-tailed tests, with P values of <0.05 indicating statistical significance.

Results

Clinicopathological characteristics

The clinicopathological characteristics of the training, internal, and external validation groups corresponding to both tasks are detailed in Tables 1,2, respectively. Across all cohorts for Task 1, HER2-low and HER2-positive tumors presented significantly greater frequencies of high Ki-67 expression and ER positivity. Additionally, in the training and internal validation cohorts, HER2-low and HER2-positive tumors were more likely to be PR-positive and ALN-positive than HER2-zero tumors were. The remaining clinicopathological features did not significantly differ (Table 1). For Task 2, in addition to the menstrual history, which varied in the external validation group but not in the training and internal validation groups, the HER2-low tumors were more likely to be ER-positive and ALN-positive and presented a greater incidence of high Ki-67 expression than the HER2-zero tumors in both the training and external validation groups. Right-sided lesions were more common in HER2-low tumors in the internal validation cohort, and PR positivity was greater in the training cohort. No other clinicopathological differences were noted between HER2-low and HER2-zero tumors across the three cohorts (Table 2).

Radiomics features analysis and RS construction

The ICC analysis demonstrated that the features chosen by both readers exhibited a high level of agreement, with an ICC >0.8. For Task 1, two distinct radiomic models, RS-DCE and RS-T2WI, were established utilizing five and six features, respectively. The 15 features from DCE and T2WI were combined to construct the RS-Com. For Task 2, two distinct radiomic models, RS-DCE and RS-T2WI, were established utilizing five and seven features, respectively. The 15 features from DCE and T2WI were combined to construct the RS-Com. Figures S1,S2 illustrate the LASSO feature selection from DCE and T2WI in both tasks. Tables S2,S3 list the detailed radiomics features of RS-Com in Tasks 1 and 2. The feature importance ranking in both tasks is shown in Figure 3. For each task, the RS-Com was formulated by aggregating the chosen features, each assigned a weight corresponding to their non-zero coefficients, as detailed in Appendix 1.

Figure 3 The importance ranking of selected features in the RS-DCE, RS-T2WI, and RS-Com model for Task 1 (A-C) and Task 2 (D-F). Task 1: differentiate HER2-zero breast cancers from HER2-nonzero breast cancers; Task 2: differentiate HER2-low breast cancers from HER2-negative breast cancers. DCE, dynamic contrast-enhanced; RS-Com, the combined radiomics signature constructed with all selected features from DCE and T2WI; RS, radiomics signature; T2WI, T2-weighted image.

For Task 1, when the three radiomic models were compared, no significant disparities were observed. Nevertheless, the RS-DCE demonstrated a marginally superior discriminatory ability over the RS-T2WI in terms of the AUC within the training cohort (0.743 vs. 0.741, P=0.978) and the internal validation cohort (0.732 vs. 0.663, P=0.541). Conversely, in the external validation cohort, the RS-DCE exhibited modestly reduced performance (0.629 vs. 0.652, P=0.839). Notably, the RS-Com model demonstrated optimal performance across all cohorts, achieving AUCs of 0.861, 0.784, and 0.794 for training, internal validation, and external validation, respectively, outperforming both the RS-DCE (P=0.067, 0.668, 0.031) and the RS-T2WI (P=0.077, 0.278, 0.235) across all three cohorts, as detailed in Table 3 and Table S4.

For Task 2, despite the absence of statistically significant disparities, the RS-DCE demonstrated a slightly enhanced capacity for discrimination over the RS-T2WI, as evidenced by the AUC in the training (0.774 vs. 0.771, P=0.971) and the external validation (0.738 vs. 0.637, P=0.399) cohorts. The RS-DCE exhibited a marginally reduced AUC in the internal validation cohort (0.668 vs. 0.677, P=0.948). In terms of overall performance, the RS-Com model surpassed both RS-DCE and RS-T2WI across all three cohorts—training, internal validation, and external validation—achieving AUCs of 0.822, 0.725, and 0.773, respectively, and outperformed both the RS-DCE (P=0.472, 0.628, 0.675) and the RS-T2WI (P=0.326, 0.712, 0.228) across all three cohorts. These findings are further detailed in Table 3 and Table S5.

Construction of a clinical predictive model

For Task 1, univariate and multivariate logistic regression analyses revealed that PR [odds ratio (OR) =0.308; 95% confidence interval (CI): 0.118–0.803; P=0.016] and ER (OR =0.050; 95% CI: 0.004–0.564; P=0.015) were independent clinical predictors, as shown in Table S6. These two variables were subsequently utilized to establish a clinical prediction model (Figure 4, Model A). The AUCs of the clinical model across the training, internal validation, and external validation cohorts were 0.688 (95% CI: 0.592–0.785), 0.692 (95% CI: 0.544–0.841), and 0.638 (95% CI: 0.504–0.772), respectively. Additionally, the model’s ACC, SEN, and SPE in the training cohort were 74.5%, 60.0%, and 77.8%, respectively; in the internal validation cohort, these values were 74.6%, 61.5%, and 77.6%; and in the external validation cohort, they were 75.6%, 43.8%, and 83.9%, as summarized in Table 3.

Figure 4 ROC curves for five models from Task 1 (A-C) and Task 2 (D-F) in the training, internal validation, and external validation cohorts. Task 1: differentiate HER2-zero breast cancers from HER2-nonzero breast cancers; Task 2: differentiate HER2-low breast cancers from HER2-negative breast cancers. Model A: clinical characteristics only; Model B: DCE radiomics features only; Model C: T2WI radiomics features only; Model D: combined DCE and T2WI radiomics features; Model E: combined clinical characteristics and radiomics features (DCE + T2WI). AUC, area under the curve; DCE, dynamic contrast-enhanced; ROC, receiver operating characteristic; T2WI, T2-weighted image.

For Task 2, univariate and multivariate logistic regression analyses revealed that PR (OR=0.310; 95% CI: 0.113–0.853; P=0.023) and ER (OR=0.329; 95% CI: 0.111–0.970; P=0.044) were independent clinical predictors, as shown in Table S7. These two variables were subsequently utilized to establish a clinical prediction model (Figure 4, Model A). The AUCs of the clinical model across the training, internal validation, and external validation cohorts were 0.699 (95% CI: 0.594–0.804), 0.693 (95% CI: 0.543–0.844), and 0.656 (95% CI: 0.501–0.812), respectively. Additionally, the model’s ACC, SEN, and SPE in the training cohort were 68.8%, 66.7%, and 69.5%, respectively; in the internal validation cohort, these values were 45.8%, 92.3%, and 28.6%; and in the external validation cohort, they were 65.4%, 56.3%, and 69.4%, as summarized in Table 3.

Establishment and validation of radiomics nomogram

For both Task 1 and Task 2, radiomics nomograms integrating the MRI RS with two key clinical features were developed (Figure 5A and Figure 5B, respectively). Both radiomics nomograms demonstrated significantly superior predictive performance compared to models using clinical features alone or RS alone across all cohorts. For Task 1, the radiomics nomogram achieved AUCs of 0.916 (95% CI: 0.866–0.965), 0.940 (95% CI: 0.857–0.983), and 0.820 (95% CI: 0.701–0.938) in training, internal validation, and external validation cohorts, respectively, with corresponding ACC (78.2%, 91.5%, 78.2%), SEN (93.3%, 92.3%, 81.3%), and SPE (74.8%, 91.4%, 77.4%) (Table 3). Statistical analysis confirmed significant improvements over the clinical model (training cohort Model E vs. A: P<0.001; internal cohort Model E vs. A: P=0.013; external cohort Model E vs. A: P=0.003) and RS-only models. Similarly, for Task 2, the radiomics nomogram yielded AUCs of 0.863 (95% CI: 0.791–0.935), 0.892 (95% CI: 0.796–0.988), and 0.833 (95% CI: 0.719–0.948) across training, internal validation, and external validation cohorts, respectively, with ACC (71.4%, 83.3%, 80.8%), SEN (96.7%, 76.9%, 81.3%), and SPE (62.2%, 85.7%, 80.6%), significantly outperforming both clinical (training cohort Model E vs. A: P=0.002; internal cohort Model E vs. A: P=0.025; external cohort Model E vs. A: P=0.040) and RS-only models, as detailed in Tables S4,S5.

Figure 5 Radiomics nomogram generated from the training set data in Tasks 1–2 (A,B). (A) Radiomics nomogram shows the incorporation of MRI radiomics signature and two pathologic features (ER and PR status) for the distinction of HER2-low and -positive versus HER2-zero cancers, as developed in training cohorts for Task 1. (B) Radiomics nomogram shows the incorporation of MRI radiomics signature and two pathologic features (ER and PR status) for differentiation of HER2-zero and HER2-negative cancers, as developed in training cohorts for Task 2. Task 1: differentiate HER2-zero breast cancers from HER2-nonzero breast cancers; Task 2: differentiate HER2-low breast cancers from HER2-negative breast cancers. ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; MRI, magnetic resonance imaging; PR, progesterone receptor.

For both tasks, DCA demonstrated the high clinical utility of the radiomics nomograms. In Task 1, DCA (Figure 6A-6C) confirmed the radiomics nomogram’s superior net benefit across the training, internal validation, and external validation cohorts. Similarly, for Task 2 (HER2-low status assessment in BC), DCA (Figure 6D-6F) highlighted its clinical reliability in all cohorts. Simultaneously, calibration curves validated the accuracy of the radiomics nomograms. In Task 1, calibration curves (Figure 7A-7C) showed strong concordance between predicted and observed probabilities across all cohorts. Likewise, for Task 2, calibration curves (Figure 7D-7F) demonstrated excellent agreement between nomogram-predicted probabilities and actual observations in all three validation sets.

Figure 6 DCA illustrating the net clinical benefit of five models across three cohorts (training, internal validation, and external validation) for Task 1 (A-C) and Task 2 (D-F). Task 1: differentiate HER2-zero breast cancers from HER2-nonzero breast cancers; Task 2: differentiate HER2-low breast cancers from HER2-negative breast cancers. All: the net benefit when all patients undergo pathological diagnosis; None: the net benefit if no patients receive pathological diagnosis; Model A: clinical characteristics only; Model B: DCE radiomics features only; Model C: T2WI radiomics features only; Model D: combined DCE and T2WI radiomics features; Model E: combined clinical characteristics and radiomics features (DCE + T2WI). DCA, decision curve analysis; DCE, dynamic contrast-enhanced; T2WI, T2-weighted image.

Figure 7 Calibration performance of model E (multiparametric radiomics combined with clinical characteristics) in Tasks 1 and 2 across training, internal, and external validation cohorts. (A-C) Task 1. (D-F) Task 2. Task 1: differentiate HER2-zero breast cancers from HER2-nonzero breast cancers; Task 2: differentiate HER2-low breast cancers from HER2-negative breast cancers.

Discussion

In this study, we conducted a comprehensive analysis of clinical and radiomic features to predict different HER2 expression statuses in BC patients. We successfully developed and verified a multiparametric MRI-based radiomics nomogram, which showed remarkable performance across training, internal validation, and external validation cohorts in differentiating HER2-low and HER2-positive BCs from HER2-zero expression BCs, as well as distinguishing HER2-low BCs from HER2-negative BCs. This approach may be useful to accurately assess HER2 expression status and thus select candidates for novel or traditional HER2-targeted therapies when IHC staining results are equivocal or FISH analysis is limited.

The efficacy of HER2-targeted therapies such as trastuzumab has been validated in HER2-positive BC, but they have been shown to be ineffective in HER2-negative BC (26). The newly proposed classification system highlights that the traditional “HER2-negative” category includes both low and zero expression of HER2 (27). With the development of trastuzumab dokamazine (SYD-985), a new treatment option has emerged for patients with HER2-low expression BC (7). Notably, this innovative strategy has revolutionized treatment for HER2-negative BC patients (28). Recent studies have shown that HER2-low BC has unique clinical and biological characteristics, including varying rates of pathological complete response (pCR), treatment responsiveness, and clinical outcomes compared with HER2-zero expression, necessitating its distinction and precise HER2 assessment for optimal treatment and prognosis prediction (17,29,30).

BC is a heterogeneous disease with diverse morphological and biological features (31). Understanding its molecular subtype is crucial for guiding treatment and predicting outcomes (32). However, the limited quantity of biopsy-derived tumor tissue might not fully capture tumor heterogeneity. Moreover, changes in receptor status and subtype can occur as the disease progresses or in response to therapy, further complicating the clinical picture (33). Among the various imaging modalities for BC diagnosis, mammography is a two-dimensional (2D) method but has a notably high false-positive rate (34). Ultrasound offers 3D assessment with broad applicability and quick examination times; however, its positive predictive value for BC screening is limited, and its accuracy can be influenced by the operator’s proficiency and expertise (35). In contrast, MRI excels in assessing lesion morphology, functionality, and the tissue environment, offering high sensitivity (36). Multiparametric MRI enables thorough analysis of tumor morphology, enhancement patterns, and diffusion restriction, offering a comprehensive understanding of tumor characteristics. In recent years, the rapid progress of AI has unlocked significant potential for radiomics, deep learning, and medical imaging (37). Radiomics, specifically, enables the noninvasive extraction of high-dimensional, visually imperceptible data from medical images, identifying tumor heterogeneity characteristics within ROIs, and facilitating objective tumor assessment (38). MRI-based radiomics has become a widely used tool in BC research, showing promising performance in differentiating between benign and malignant tumors, identifying molecular subtypes, detecting lymph node metastasis, evaluating prognostic factors, and assessing treatment efficacy after neoadjuvant chemotherapy (39-43). HER2 expression in BC has garnered significant attention from oncologists and pathologists, propelling it to the forefront of BC research. Several studies have applied MRI-based radiomics analysis to predict HER2 expression levels (44,45). For ongoing clinical trials involving HER2-low BC, an accurate assessment of HER2-low status is essential for generating reliable results.

For Task 1, Ramtohul et al. (44) reported that the multiparametric MRI radiomics model (DCE T1-weighted + T2-weighted) yielded high efficacy (AUC =0.80) with respect to the external test set. In contrast with prior research, our comprehensive model, which leverages DCE and T2WI, has matched the efficacy demonstrated in previous studies. A significant addition to our work is that our radiomics nomogram model also takes into account clinically relevant risk factors and achieves excellent performance, which affirmatively demonstrates our model’s strength in discerning HER2-zero from nonzero (including HER2-low and HER2-positive) expression. This precision is pivotal for pinpointing individuals who are ineligible for HER2-targeted treatments, facilitating more tailored and effective clinical interventions.

For Task 2, we gave full consideration to the key subgroup of HER2-low patients. Building upon the traditional binary classification of HER2 expression established by previous researchers, our study delved deeper into the distinctions between HER2-low and HER2-zero patients. We acknowledge that these subgroups have unique prognostic outcomes and emphasize the significance of developing more effective treatment strategies for them (17). Previous radiomics studies have primarily focused on differentiating between HER2-negative and HER2-positive expression, achieving moderate to high diagnostic accuracy via DCE, diffusion-weighted imaging (DWI), or T2WI (45,46). Xu et al. (45) used these methods to distinguish 70 HER2-positive patients from 144 HER2-negative patients. Similarly, Zhou et al. (46) developed multiparametric MRI-based radiomic signatures, achieving AUCs of 0.860 and 0.810 in the training and validation cohorts, respectively. However, only a subset of current imaging studies have addressed low HER2 expression, even though this category accounts for most BCs (47) and serves as a therapeutic target for novel ADCs. Zheng et al. utilized T2WI, DCE, DWI, and ADC, achieving an AUC of 0.782 for discriminating HER2-low expression from other levels in the validation cohort, which was lower than the present findings (15). Their model relied primarily on ADC features, whereas the present model relied primarily on early-phase DCE features. Bian et al. (48) investigated the potential of intratumoral and peritumoral features extracted from multiparametric MRI to distinguish HER2-positive from HER2-negative BCs, achieving an AUC of 0.76 in the external validation but a lower AUC of 0.71 for identifying the more subtle HER2-low and HER2-zero subgroups. Moreover, Guo et al. (49) employed MRI-based deep learning radiomics to differentiate between HER2-negative and HER2-overexpressing BCs, achieving AUCs of 0.868 and 0.763 for the training and validation cohorts, respectively. When distinguishing between HER2-low and HER2-zero, the AUCs were 0.855 and 0.750, respectively. The diagnostic performance, particularly in identifying HER2-low status, remains modest, and room for improvement exists. Consequently, there is a pressing need to enhance the diagnostic capabilities of these models to ensure more accurate identification of HER2-low status in BC patients. Recent intensive research has revealed the effectiveness of a radiomics nomogram model that integrates clinical factors and the Rad-score derived from radiomic features, significantly enhancing its predictive ability (50,51). Our findings revealed that several clinicopathologic variables, specifically the ER/PR status, were correlated with the HER2 expression status, which was consistent with previous research findings (17). Consequently, we successfully developed and validated a radiomics nomogram that exhibited robust performance in identifying HER2-low status. In particular, the radiomics nomogram achieved AUCs of 0.863, 0.892, and 0.833 in the training cohort, internal validation cohort, and external validation cohort, respectively. These results highlight the potential of our nomogram as a dependable tool for differentiating between HER2-low and HER2-zero BCs, thereby identifying patients who may benefit from novel anti-HER2 ADCs.

Previous studies have revealed the predictive power of DCE-MRI for BC molecular subtype, prognosis, lesion morphology, and hemodynamics (52). Our study investigated the value of MRI radiomics in predicting HER2 expression from DCE, T2WI, and combined sequences, along with relevant clinical indicators. The radiomics nomogram model robustly predicted HER2-low and HER2-positive status, suggesting shared molecular characteristics on the surface of these cells. These findings indicate that HER2-low and HER2-positive BCs may share pathological and clinical traits (7). Additionally, T2WI-based radiomic models (53) can detect HER2-positive expression, suggesting that HER2-low BC cells may also exhibit invasive characteristics.

In this study, the majority of the extracted features were derived from wavelet transform, underscoring its crucial role in predicting HER2 status. This finding aligns with previous research that has established a correlation between high-order features obtained through wavelet transformation and HER2 expression levels (15). The mechanistic basis of this correlation can be attributed to the inherent multi-scale nature of the wavelet transform, which enables it to capture frequency information across spatial scales. This capability allows for the revelation of subtle tumor heterogeneity, including textural and morphological patterns that remain imperceptible through conventional imaging analysis (54,55). By amplifying the quantitative features of tumor heterogeneity—particularly in morphological features such as irregular margins and infiltrative growth patterns characteristic of HER2-overexpressing tumors—wavelet transform features provided a robust basis for predicting HER2 status in BC patients. Our results show a significant association between wavelet transform features extracted from BC MRI images and HER2 different expression, validating their utility as a non-invasive biomarker for HER2 stratification. The specific texture features utilized to construct the RS were primarily the GLCM. Previous MRI-based radiomics studies have also highlighted the significance of wavelet transform and GLCM in feature extraction (56). GLCM features, which are widely used in radiomics and machine learning, reveal subtle changes in tumor histology. They characterize gray-level intensities in 3D images, enabling quantitative predictions/classifications and subjective assessments (57). Prior research has highlighted the utility of GLCM parameters derived from MRI imaging in characterizing the intricate heterogeneity and structural complexity of the intratumorally microenvironment (58). Building on this foundation, our results demonstrated that GLCM features derived from T2WI effectively captured the microenvironmental heterogeneity of BC, such as uneven cell density and necrotic regions, and could serve as robust imaging biomarkers, facilitating more accurate patient stratification in HER2 different expression, consistent with findings reported in prior studies (59).

The RS extracted from medical images, including GLCM and wavelet transform features, have been shown to be repeatable, non-invasive, and capable of providing valuable insights into the underlying biology of tumors. Our findings advocate the integration of wavelet transform and GLCM features in the construction of RS. This approach synergistically captures macrostructural heterogeneity and microtextural irregularities to comprehensively characterize HER2-associated radiological phenotypes. It not only enhances the predictive power of RS but also ensures biological interpretability consistent with established mechanisms of HER2 signaling and tumor heterogeneity. The non-invasive and reproducible nature of these radiological features provides a quantitative window into tumor biology, complementing traditional clinical and pathological assessment. In our extracted feature importance ranking analysis, we found that wavelet-HLH and wavelet-LLL are of high significance. This may be related to its highlighting the fine-grained textural changes associated with microvessel density and irregular blood flow patterns between different expressions of HER2 (60), as well as the associated differences between regions of cell proliferation and necrosis (60,61).

Our study suggests that subtle differences in tumor microcirculation structure may be revealed under different levels of HER2 expression. Notably, varying HER2 expression alters vascular endothelial growth factor (VEGF) levels, triggering microvessel changes that impact tumor growth, invasion, and metastasis, which aligns with the findings of Chen et al. (62). In clinical practice, BC often manifests with cystic necrosis and aggressive characteristics such as rapid proliferation and inadequate central blood supply. Hormone-dependent ER and PR are vital prognostic markers (63). Our study showed that ER and PR were significantly expressed in both HER2-low and HER2-positive BC patients, and their levels could be used as reliable indicators of disease severity, which is consistent with previous studies showing that continuous stress responses in BC patients can lead to elevated ER and PR levels (64,65). Although the ROC analysis indicated that these clinicopathological variables possessed only modest discriminative power, we integrated radiomic features with clinicopathological variables to develop a radiomics nomogram. The radiomics nomogram model excelled in predicting HER2 expression, indicating its potential to comprehensively depict the biological characteristics of tumors with HER2-low expression and aid in the trinary categorization of HER2 expression levels, which aligns with the findings of Peng et al. (51). Meanwhile, it could help clinicians to select patients for novel or traditional HER2-targeted therapy, especially in the presence of ambiguous IHC staining results or limited access to FISH analysis.

HER2 expression is typically assessed via IHC and FISH, which may capture only a snapshot of potentially heterogeneous tumors (66). Also, due to the FISH assay, which is an expensive and time-consuming test to perform, the radiomics nomogram is particularly important in areas where FISH infrastructure is lacking or when histological examination is delayed due to resource-limited conditions, reducing unnecessary procedures and costs as well as complementing the limited histopathological evaluation. The incorporation of a radiomics nomogram into the diagnostic process can help to evaluate the heterogeneity of an entire tumor, identify suitable biopsy targets, and monitor temporal and spatial changes in tumor biology as the disease progresses. The radiomics nomogram can provide valuable spatiotemporal monitoring, and unlike single time-point biopsies, it can longitudinally track HER2 dynamics during neoadjuvant therapy, which is critical for timely adjustments to treatment plans. Our study revealed that a multiparametric MRI radiomics model outperforms clinical and single-sequence models in AUC, ACC, SEN, and SPE, capturing richer lesion information and more accurately predicting outcomes. Remarkably, the radiomics nomogram model achieved the highest level of predictive accuracy, emphasizing the importance of integrating clinical parameters to enhance its predictive ability, which aligns with our research objective. The calibration curve and the Hosmer-Lemeshow test confirmed the robust fit of our model, demonstrating its validity. Additionally, the DCA curve demonstrated greater net clinical benefits of the radiomics nomogram model than both the radiomics model alone and the clinical model across most threshold ranges. In contrast to existing studies, our radiomics nomogram combines the spatial richness of RS with the specificity of histopathology features, allowing for a comprehensive description of tumor phenotypes and heterogeneity, capturing spatial information that is critical for understanding cancer progression. It can assist in the ternary classification of HER2 expression levels and help select patients for novel or conventional HER2-targeted therapies. It can also simplify diagnostic workflow and improve patient prognosis by reducing reliance on invasive and time-consuming IHC and FISH tests. Its adoption by radiologists and oncologists is further supported by the interpretability of the model output through clear visualization, making it an important adjunct to traditional pathology assessment.

However, this study still has certain limitations. First, our study focused on IDCs of no special type, which account for approximately 80% of all BCs, minimizing confounding factors from diverse subtypes. Second, in our DCE-MRI analysis, we concentrated on the most significant enhancing phase without considering precontrast phases or other enhanced sequences. Finally, the retrospective design and modest sample size introduce a degree of bias to our conclusions. Furthermore, to accurately differentiate the three expression levels of HER2, we will need to build a triple classification model in the future or implement an approach similar to that of previous studies, utilizing multiple binary classification models and a wider range of BC types to strengthen the credibility and applicability of our findings.

Conclusions

Our multiparametric MRI-based radiomics nomogram provides a rapid and accurate tool for classifying HER2 status in IDC patients. By distinguishing between HER2 -zero patients (unsuitable for targeted therapy) and HER2- low patients (potential candidates for novel ADCs), this tool helps clinicians to develop personalized treatment strategies, reduces reliance on invasive histopathological examinations, minimizes diagnostic delays, and accelerates treatment decisions. Its integration into the clinical workflow can optimize resource allocation, increase treatment precision, and improve the prognosis of IDC patients.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-24-1707/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-24-1707/dss

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1707/coif). J.D. reports that he is an employee of GE Healthcare. Y.D. reports that he is an employee of Yizhun Medical AI Co. Ltd. The other 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. The study was approved by the Ethics Committee of The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University (No. 2024.144), and the requirement for patient informed consent was waived due to the retrospective nature of the study.

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. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
2. Casey MC, Prakash A, Holian E, McGuire A, Kalinina O, Shalaby A, Curran C, Webber M, Callagy G, Bourke E, Kerin MJ, Brown JA. Quantifying Argonaute 2 (Ago2) expression to stratify breast cancer. BMC Cancer 2019;19:712. [Crossref] [PubMed]
3. Banerji U, van Herpen CML, Saura C, Thistlethwaite F, Lord S, Moreno V, Macpherson IR, Boni V, Rolfo C, de Vries EGE, Rottey S, Geenen J, Eskens F, Gil-Martin M, Mommers EC, Koper NP, Aftimos P. Trastuzumab duocarmazine in locally advanced and metastatic solid tumours and HER2-expressing breast cancer: a phase 1 dose-escalation and dose-expansion study. Lancet Oncol 2019;20:1124-35. [Crossref] [PubMed]
4. Bhattacharjee A, Rajendra J, Dikshit R, Dutt S. HER2 borderline is a negative prognostic factor for primary malignant breast cancer. Breast Cancer Res Treat 2020;181:225-31. [Crossref] [PubMed]
5. Wildiers H, Tryfonidis K, Dal Lago L, Vuylsteke P, Curigliano G, Waters S, Brouwers B, Altintas S, Touati N, Cardoso F, Brain E. Pertuzumab and trastuzumab with or without metronomic chemotherapy for older patients with HER2-positive metastatic breast cancer (EORTC 75111-10114): an open-label, randomised, phase 2 trial from the Elderly Task Force/Breast Cancer Group. Lancet Oncol 2018;19:323-36. [Crossref] [PubMed]
6. Modi S, Jacot W, Yamashita T, Sohn J, Vidal M, Tokunaga E, et al. Trastuzumab Deruxtecan in Previously Treated HER2-Low Advanced Breast Cancer. N Engl J Med 2022;387:9-20. [Crossref] [PubMed]
7. Modi S, Park H, Murthy RK, Iwata H, Tamura K, Tsurutani J, Moreno-Aspitia A, Doi T, Sagara Y, Redfern C, Krop IE, Lee C, Fujisaki Y, Sugihara M, Zhang L, Shahidi J, Takahashi S. Antitumor Activity and Safety of Trastuzumab Deruxtecan in Patients With HER2-Low-Expressing Advanced Breast Cancer: Results From a Phase Ib Study. J Clin Oncol 2020;38:1887-96. [Crossref] [PubMed]
8. Eiger D, Agostinetto E, Saúde-Conde R, de Azambuja E. The Exciting New Field of HER2-Low Breast Cancer Treatment. Cancers (Basel) 2021.
9. D'Alfonso T, Liu YF, Monni S, Rosen PP, Shin SJ. Accurately assessing her-2/neu status in needle core biopsies of breast cancer patients in the era of neoadjuvant therapy: emerging questions and considerations addressed. Am J Surg Pathol 2010;34:575-81. [Crossref] [PubMed]
10. Chen R, Qi Y, Huang Y, Liu W, Yang R, Zhao X, Wu Y, Li Q, Wang Z, Sun X, Wei B, Chen J. Diagnostic value of core needle biopsy for determining HER2 status in breast cancer, especially in the HER2-low population. Breast Cancer Res Treat 2023;197:189-200. [Crossref] [PubMed]
11. Fang C, Zhang J, Li J, Shang H, Li K, Jiao T, Yin D, Li F, Cui Y, Zeng Q. Clinical-radiomics nomogram for identifying HER2 status in patients with breast cancer: A multicenter study. Front Oncol 2022;12:922185. [Crossref] [PubMed]
12. Harada TL, Uematsu T, Nakashima K, Kawabata T, Nishimura S, Takahashi K, Tadokoro Y, Hayashi T, Tsuchiya K, Watanabe J, Sugino T. Evaluation of Breast Edema Findings at T2-weighted Breast MRI Is Useful for Diagnosing Occult Inflammatory Breast Cancer and Can Predict Prognosis after Neoadjuvant Chemotherapy. Radiology 2021;299:53-62. [Crossref] [PubMed]
13. Li H, Sun H, Liu S, Zhang W, Arukalam FM, Ma H, Qian W. Assessing the performance of benign and malignant breast lesion classification with bilateral TIC differentiation and other effective features in DCE-MRI. J Magn Reson Imaging 2019;50:465-73. [Crossref] [PubMed]
14. Leithner D, Mayerhoefer ME, Martinez DF, Jochelson MS, Morris EA, Thakur SB, Pinker K. Non-Invasive Assessment of Breast Cancer Molecular Subtypes with Multiparametric Magnetic Resonance Imaging Radiomics. J Clin Med 2020;9:1853. [Crossref] [PubMed]
15. Zheng S, Yang Z, Du G, Zhang Y, Jiang C, Xu T, Li B, Wang D, Qiu Y, Lin D, Zhang X, Shen J. Discrimination between HER2-overexpressing, -low-expressing, and -zero-expressing statuses in breast cancer using multiparametric MRI-based radiomics. Eur Radiol 2024;34:6132-44. [Crossref] [PubMed]
16. Liu W, Yang Y, Wang X, Li C, Liu C, Li X, Wen J, Lin X, Qin J. A Comprehensive Model Outperformed the Single Radiomics Model in Noninvasively Predicting the HER2 Status in Patients with Breast Cancer. Acad Radiol 2025;32:24-36. [Crossref] [PubMed]
17. Denkert C, Seither F, Schneeweiss A, Link T, Blohmer JU, Just M, et al. Clinical and molecular characteristics of HER2-low-positive breast cancer: pooled analysis of individual patient data from four prospective, neoadjuvant clinical trials. Lancet Oncol 2021;22:1151-61. [Crossref] [PubMed]
18. Hammond ME, Hayes DF, Dowsett M, Allred DC, Hagerty KL, Badve S, et al. American Society of Clinical Oncology/College Of American Pathologists guideline recommendations for immunohistochemical testing of estrogen and progesterone receptors in breast cancer. J Clin Oncol 2010;28:2784-95. [Crossref] [PubMed]
19. Dowsett M, Nielsen TO, A'Hern R, Bartlett J, Coombes RC, Cuzick J, et al. Assessment of Ki67 in breast cancer: recommendations from the International Ki67 in Breast Cancer working group. J Natl Cancer Inst 2011;103:1656-64. [Crossref] [PubMed]
20. Wolff AC, Hammond MEH, Allison KH, Harvey BE, Mangu PB, Bartlett JMS, Bilous M, Ellis IO, Fitzgibbons P, Hanna W, Jenkins RB, Press MF, Spears PA, Vance GH, Viale G, McShane LM, Dowsett M. Human Epidermal Growth Factor Receptor 2 Testing in Breast Cancer: American Society of Clinical Oncology/College of American Pathologists Clinical Practice Guideline Focused Update. J Clin Oncol 2018;36:2105-22. [Crossref] [PubMed]
21. Tarantino P, Hamilton E, Tolaney SM, Cortes J, Morganti S, Ferraro E, Marra A, Viale G, Trapani D, Cardoso F, Penault-Llorca F, Viale G, Andrè F, Curigliano G. HER2-Low Breast Cancer: Pathological and Clinical Landscape. J Clin Oncol 2020;38:1951-62. [Crossref] [PubMed]
22. Morrison CK, Henze Bancroft LC, DeMartini WB, Holmes JH, Wang K, Bosca RJ, Korosec FR, Strigel RM. Novel High Spatiotemporal Resolution Versus Standard-of-Care Dynamic Contrast-Enhanced Breast MRI: Comparison of Image Quality. Invest Radiol 2017;52:198-205. [Crossref] [PubMed]
23. Kim JJ, Kim JY, Hwangbo L, Suh HB, Son Y, Nickel MD, Grimm R. Ultrafast Dynamic Contrast-Enhanced MRI Using Compressed Sensing: Associations of Early Kinetic Parameters With Prognostic Factors of Breast Cancer. AJR Am J Roentgenol 2021;217:56-63. [Crossref] [PubMed]
24. Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, Gee JC. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging 2010;29:1310-20. [Crossref] [PubMed]
25. Peng S, Chen L, Tao J, Liu J, Zhu W, Liu H, Yang F. Radiomics Analysis of Multi-Phase DCE-MRI in Predicting Tumor Response to Neoadjuvant Therapy in Breast Cancer. Diagnostics (Basel) 2021.
26. Mamounas EP, Untch M, Mano MS, Huang CS, Geyer CE Jr, von Minckwitz G, et al. Adjuvant T-DM1 versus trastuzumab in patients with residual invasive disease after neoadjuvant therapy for HER2-positive breast cancer: subgroup analyses from KATHERINE. Ann Oncol 2021;32:1005-14. [Crossref] [PubMed]
27. Tarantino P, Viale G, Press MF, Hu X, Penault-Llorca F, Bardia A, et al. ESMO expert consensus statements (ECS) on the definition, diagnosis, and management of HER2-low breast cancer. Ann Oncol 2023;34:645-59. [Crossref] [PubMed]
28. Prat A, Bardia A, Curigliano G, Hammond MEH, Loibl S, Tolaney SM, Viale G. An Overview of Clinical Development of Agents for Metastatic or Advanced Breast Cancer Without ERBB2 Amplification (HER2-Low). JAMA Oncol 2022; Epub ahead of print. [Crossref]
29. Zhang G, Ren C, Li C, Wang Y, Chen B, Wen L, Jia M, Li K, Mok H, Cao L, Chen X, Lin J, Wei G, Li Y, Zhang Y, Balch CM, Liao N. Distinct clinical and somatic mutational features of breast tumors with high-, low-, or non-expressing human epidermal growth factor receptor 2 status. BMC Med 2022;20:142. [Crossref] [PubMed]
30. Zhang H, Peng Y. Current Biological, Pathological and Clinical Landscape of HER2-Low Breast Cancer. Cancers (Basel) 2022.
31. Abramson VG, Lehmann BD, Ballinger TJ, Pietenpol JA. Subtyping of triple-negative breast cancer: implications for therapy. Cancer 2015;121:8-16. [Crossref] [PubMed]
32. Sanges F, Floris M, Cossu-Rocca P, Muroni MR, Pira G, Urru SAM, et al. Histologic subtyping affecting outcome of triple negative breast cancer: a large Sardinian population-based analysis. BMC Cancer 2020;20:491. [Crossref] [PubMed]
33. van de Ven S, Smit VT, Dekker TJ, Nortier JW, Kroep JR. Discordances in ER, PR and HER2 receptors after neoadjuvant chemotherapy in breast cancer. Cancer Treat Rev 2011;37:422-30. [Crossref] [PubMed]
34. Miller AB. Digital mammography. J Natl Cancer Inst 2014;106:dju125. [Crossref] [PubMed]
35. Melnikow J, Fenton JJ, Whitlock EP, Miglioretti DL, Weyrich MS, Thompson JH, Shah K. Supplemental Screening for Breast Cancer in Women With Dense Breasts: A Systematic Review for the U.S. Preventive Services Task Force. Ann Intern Med 2016;164:268-78. [Crossref] [PubMed]
36. Zhang L, Tang M, Min Z, Lu J, Lei X, Zhang X. Accuracy of combined dynamic contrast-enhanced magnetic resonance imaging and diffusion-weighted imaging for breast cancer detection: a meta-analysis. Acta Radiol 2016;57:651-60. [Crossref] [PubMed]
37. Karim S, Qadir A, Farooq U, Shakir M, Laghari AA. Hyperspectral Imaging: A Review and Trends towards Medical Imaging. Curr Med Imaging 2022;19:417-27. [Crossref] [PubMed]
38. Kuhl CK, Truhn D. The Long Route to Standardized Radiomics: Unraveling the Knot from the End. Radiology 2020;295:339-41. [Crossref] [PubMed]
39. Baselga J, Swain SM. Novel anticancer targets: revisiting ERBB2 and discovering ERBB3. Nat Rev Cancer 2009;9:463-75. [Crossref] [PubMed]
40. Kang SS, Ko EY, Han BK, Shin JH, Hahn SY, Ko ES. Background parenchymal enhancement on breast MRI: influence of menstrual cycle and breast composition. J Magn Reson Imaging 2014;39:526-34. [Crossref] [PubMed]
41. Kröger N, Milde-Langosch K, Riethdorf S, Schmoor C, Schumacher M, Zander AR, Löning T. Prognostic and predictive effects of immunohistochemical factors in high-risk primary breast cancer patients. Clin Cancer Res 2006;12:159-68. [Crossref] [PubMed]
42. Goldhirsch A, Wood WC, Coates AS, Gelber RD, Thürlimann B, Senn HJ. Panel members. Strategies for subtypes--dealing with the diversity of breast cancer: highlights of the St. Gallen International Expert Consensus on the Primary Therapy of Early Breast Cancer 2011. Ann Oncol 2011;22:1736-47. [Crossref] [PubMed]
43. Borgan E, Sitter B, Lingjærde OC, Johnsen H, Lundgren S, Bathen TF, Sørlie T, Børresen-Dale AL, Gribbestad IS. Merging transcriptomics and metabolomics--advances in breast cancer profiling. BMC Cancer 2010;10:628. [Crossref] [PubMed]
44. Ramtohul T, Djerroudi L, Lissavalid E, Nhy C, Redon L, Ikni L, Djelouah M, Journo G, Menet E, Cabel L, Malhaire C, Tardivon A. Multiparametric MRI and Radiomics for the Prediction of HER2-Zero, -Low, and -Positive Breast Cancers. Radiology 2023;308:e222646. [Crossref] [PubMed]
45. Xu A, Chu X, Zhang S, Zheng J, Shi D, Lv S, Li F, Weng X. Development and validation of a clinicoradiomic nomogram to assess the HER2 status of patients with invasive ductal carcinoma. BMC Cancer 2022;22:872. [Crossref] [PubMed]
46. Zhou J, Tan H, Li W, Liu Z, Wu Y, Bai Y, Fu F, Jia X, Feng A, Liu H, Wang M. Radiomics Signatures Based on Multiparametric MRI for the Preoperative Prediction of the HER2 Status of Patients with Breast Cancer. Acad Radiol 2021;28:1352-60. [Crossref] [PubMed]
47. Viale G, Basik M, Niikura N, Tokunaga E, Brucker S, Penault-Llorca F, Hayashi N, Sohn J, Teixeira de Sousa R, Brufsky AM, O'Brien CS, Schmitt F, Higgins G, Varghese D, James GD, Moh A, Livingston A, de Giorgio-Miller V. Retrospective study to estimate the prevalence and describe the clinicopathological characteristics, treatments received, and outcomes of HER2-low breast cancer. ESMO Open 2023;8:101615. [Crossref] [PubMed]
48. Bian X, Du S, Yue Z, Gao S, Zhao R, Huang G, Guo L, Peng C, Zhang L. Potential Antihuman Epidermal Growth Factor Receptor 2 Target Therapy Beneficiaries: The Role of MRI-Based Radiomics in Distinguishing Human Epidermal Growth Factor Receptor 2-Low Status of Breast Cancer. J Magn Reson Imaging 2023;58:1603-14. [Crossref] [PubMed]
49. Guo Y, Xie X, Tang W, Chen S, Wang M, Fan Y, Lin C, Hu W, Yang J, Xiang J, Jiang K, Wei X, Huang B, Jiang X. Noninvasive identification of HER2-low-positive status by MRI-based deep learning radiomics predicts the disease-free survival of patients with breast cancer. Eur Radiol 2024;34:899-913. [Crossref] [PubMed]
50. Mao N, Yin P, Li Q, Wang Q, Liu M, Ma H, Dong J, Che K, Wang Z, Duan S, Zhang X, Hong N, Xie H. Radiomics nomogram of contrast-enhanced spectral mammography for prediction of axillary lymph node metastasis in breast cancer: a multicenter study. Eur Radiol 2020;30:6732-9. [Crossref] [PubMed]
51. Peng Y, Zhang X, Qiu Y, Li B, Yang Z, Huang J, Lin J, Zheng C, Hu L, Shen J. Development and Validation of MRI Radiomics Models to Differentiate HER2-Zero, -Low, and -Positive Breast Cancer. AJR Am J Roentgenol 2024;222:e2330603. [Crossref] [PubMed]
52. Guarneri V, Bras-Maristany F, Dieci MV, Griguolo G, Par L, Mar Ín-Aguilera M, Miglietta F, Bottosso M, Giorgi CA, Blasco P, Castillo O, Galv N P, Vivancos A, Villagrasa P, Parker JS, Perou CM, Conte P, Prat A. HER2DX genomic test in HER2-positive/hormone receptor-positive breast cancer treated with neoadjuvant trastuzumab and pertuzumab: A correlative analysis from the PerELISA trial. EBioMedicine 2022;85:104320. [Crossref] [PubMed]
53. Wu Z, Xu S, Zhou L, Yin W, Lin Y, Du Y, Wang Y, Jiang Y, Yin K, Zhang J, Lu J. Clinical significance of quantitative HER2 gene amplification as related to its predictive value in breast cancer patients in neoadjuvant setting. Onco Targets Ther 2018;11:801-8. [Crossref] [PubMed]
54. Jing R, Wang J, Li J, Wang X, Li B, Xue F, Shao G, Xue H. A wavelet features derived radiomics nomogram for prediction of malignant and benign early-stage lung nodules. Sci Rep 2021;11:22330. [Crossref] [PubMed]
55. Song SE, Bae MS, Chang JM, Cho N, Ryu HS, Moon WK. MR and mammographic imaging features of HER2-positive breast cancers according to hormone receptor status: a retrospective comparative study. Acta Radiol 2017;58:792-9. [Crossref] [PubMed]
56. Ghose JC, Mitra SK, Chhetri MK. Systolic time intervals in the differential diagnosis of constrictive pericarditis and cardiomyopathy. Br Heart J 1976;38:47-50. [Crossref] [PubMed]
57. Alibabaei S, Rahmani M, Tahmasbi M, Tahmasebi Birgani MJ, Razmjoo S. Evaluating the Gray Level Co-Occurrence Matrix-Based Texture Features of Magnetic Resonance Images for Glioblastoma Multiform Patients' Treatment Response Assessment. J Med Signals Sens 2023;13:261-71. [Crossref] [PubMed]
58. Yang YS, Feng F, Qiu YJ, Zheng GH, Ge YQ, Wang YT. High-resolution MRI-based radiomics analysis to predict lymph node metastasis and tumor deposits respectively in rectal cancer. Abdom Radiol (NY) 2021;46:873-84. [Crossref] [PubMed]
59. Weaver O, Leung JWT. Biomarkers and Imaging of Breast Cancer. AJR Am J Roentgenol 2018;210:271-8. [Crossref] [PubMed]
60. Lubner MG, Smith AD, Sandrasegaran K, Sahani DV, Pickhardt PJ. CT Texture Analysis: Definitions, Applications, Biologic Correlates, and Challenges. Radiographics 2017;37:1483-503. [Crossref] [PubMed]
61. Nissan N, Ochoa Albiztegui RE, Fruchtman-Brot H, Gluskin J, Arita Y, Amir T, Reiner JS, Feigin K, Mango VL, Jochelson MS, Sung JS. Extremely dense breasts: A comprehensive review of increased cancer risk and supplementary screening methods. Eur J Radiol 2025;182:111837. [Crossref] [PubMed]
62. Chen P, Zhao S, Guo W, Shao G. Dynamic contrast-enhanced magnetic resonance imaging features and apparent diffusion coefficient value of HER2-positive/HR-negative breast carcinoma. Quant Imaging Med Surg 2023;13:4816-25. [Crossref] [PubMed]
63. Teifke A, Behr O, Schmidt M, Victor A, Vomweg TW, Thelen M, Lehr HA. Dynamic MR imaging of breast lesions: correlation with microvessel distribution pattern and histologic characteristics of prognosis. Radiology 2006;239:351-60. [Crossref] [PubMed]
64. Valkonen M, Isola J, Ylinen O, Muhonen V, Saxlin A, Tolonen T, Nykter M, Ruusuvuori P. Cytokeratin-Supervised Deep Learning for Automatic Recognition of Epithelial Cells in Breast Cancers Stained for ER, PR, and Ki-67. IEEE Trans Med Imaging 2020;39:534-42. [Crossref] [PubMed]
65. Yi C, Li L, Ma J. Tumor enhancement and shrinkage pattern in dynamic contrast-enhanced magnetic resonance imaging for predicting pathologic complete response after human epidermal growth factor receptor 2 (HER2)-targeted therapy in breast cancer. Quant Imaging Med Surg 2024;14:6734-44. [Crossref] [PubMed]
66. Seol H, Lee HJ, Choi Y, Lee HE, Kim YJ, Kim JH, Kang E, Kim SW, Park SY. Intratumoral heterogeneity of HER2 gene amplification in breast cancer: its clinicopathological significance. Mod Pathol 2012;25:938-48. [Crossref] [PubMed]