The value of multiparametric MRI-based combined intratumoral and peritumoral radiomics in differentiating luminal and non-luminal molecular subtypes of breast cancer: a multicenter study

Introduction

Breast cancer is the most common type of malignant tumor among the global female population, and is the leading cause of cancer-related deaths in women worldwide (1). The molecular subtypes of breast cancer are critical in guiding clinical treatment strategies. According to the expression status of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor-2 (HER2), breast cancers are classified into four distinct molecular subtypes: luminal A, luminal B, HER2-enriched, and triple-negative breast cancer. Clinically, luminal A and luminal B subtypes are grouped as the luminal category, whereas HER2-enriched and triple-negative breast cancer are categorized as non-luminal subtypes (2-4). Luminal breast cancer represents the most prevalent molecular subtype, constituting approximately 70% of all breast cancer cases. This subtype is predominantly managed with endocrine therapy as the standard treatment, offering patients a favorable prognosis and reduced systemic toxicity compared to other subtypes (5,6). Non-luminal breast cancer is refractory to endocrine therapy and typically exhibits a higher histological grade and more aggressive clinical behavior, leading to poorer prognosis outcomes compared to luminal subtypes. Management requires personalized treatment strategies tailored to molecular profiles, such as HER2-targeted agents, platinum-based chemotherapy, or radiotherapy (7-9).

The current standard of care for breast cancer molecular subtyping predominantly relies on preoperative immunohistochemical analysis of core needle biopsy specimens. However, this method necessitates invasive tissue extraction, which is associated with risks such as pain, bleeding, and infection (10,11). Furthermore, core needle biopsy specimens are limited by spatial sampling bias, as they often capture only a small fraction of the tumor, failing to represent the full spectrum of intratumoral heterogeneity inherent to breast cancer. This spatial sampling constraint may compromise the diagnostic reliability of preoperative subtyping based on immunohistochemical analysis, thereby increasing the risk of misclassification of molecular subtypes compared to definitive histopathological evaluation of surgically resected specimens (12). Thus, developing a preoperative noninvasive approach for precise discrimination of luminal and non-luminal breast cancer subtypes is critically needed.

The rapid advancement of imaging technology has positioned magnetic resonance imaging (MRI) as a pivotal diagnostic and therapeutic tool in breast MRI distinguishes itself through its unparalleled soft-tissue contrast resolution, multi-parametric imaging capabilities [including T1-weighted, T2-weighted, diffusion-weighted, and dynamic contrast-enhanced (DCE) sequences], and non-ionizing radiation profile. These attributes underpin its critical role in lesion characterization, precise tumor staging, and longitudinal therapeutic response monitoring in breast cancer management (13-16). Concurrently, radiomics has emerged as a transformative computational paradigm, involving the systematic extraction of high-dimensional quantitative features from medical imaging data and their subsequent integration with advanced machine learning workflows. This framework enables the decoding of subvisual imaging biomarkers that correlate with tumor biology, thereby enhancing prognostic stratification and supporting personalized treatment strategies.

Recently, it has been found that there may be some information related to the biological characteristics of the tumor, such as angiogenesis, peritumoral infiltration of the vasculature, and mesenchymal reaction in the peritumoral region of breast cancer, which provides a great deal of valuable information for the preoperative prediction of vasculature invasion and benign-malignant differentiation of breast cancer (17-21). While optimal peritumoral region sizing remains undetermined, studies demonstrate radiomics model performance declines with expansion from 5 to 10 mm (22,23). Zhou et al. (24) also found that the proximal peritumor region was more accurate compared to larger regions. While intratumoral models show limited predictive efficacy for HER2 and Ki-67 status and lymph node metastasis in breast cancer, combined intratumoral-peritumoral radiomics demonstrate superior performance compared to single-region models (15,25,26).

Prior studies primarily focused on single-parameter analysis, with limited exploration of the complementary effects and performance optimization potential across T2-weighted imaging with fat suppression (T2WI), diffusion-weighted imaging (DWI) and DCE MRI. Several studies have shown good predictive performance of multiparametric MRI radiomics models for predicting Ki-67 and HER2 expression status. Song et al. (27) found that the multiparametric MRI model predicted luminal type breast cancer Ki-67 expression level and histologic grading with good performance, and among the established prediction models, the Naive Bayes classifier showed the highest performance. Xu et al. (28) established a clinical-radiomics nomogram model using multiparametric MRI combined with histologic grading, Ki-67, and clinical features, which had good performance in predicting HER2 expression.

Although radiomics has been used to distinguish luminal from non-luminal breast cancer subtypes, most studies rely on single-parameter or intratumoral analyses, leaving multiparametric intratumoral-peritumoral combinations largely unexamined (29-32). Therefore, this study aims to evaluate the value of preoperative non-invasive identification of luminal or non-luminal breast cancer based on DCE, T2WI, and DWI multiparametric MRI intratumoral combined with peritumoral radiomics models. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-83/rc).

Methods

Study population

The study cohort comprised female breast cancer patients who underwent breast MRI examinations across two medical centers, with data collected from December 2015 to December 2023 at the Affiliated Hospital of Qinghai University (Center 1) and from October 2019 to February 2024 at the Second Hospital of Lanzhou University (Center 2). This study also used a publicly available dataset from The Cancer Imaging Archive (33), the Investigation of Serial Studies to Predict Your Therapeutic Response with Imaging and Molecular Analysis trial conducted from May 2002 to March 2006 (34) (Center 3).

Patients from Center 1 were randomly allocated at a 7:3 ratio to the training set and internal test set. The external test sets comprised patients from Center 2 (external test set 1) and Center 3 (external test set 2). Exclusion criteria for patients were as follows: (I) invasive breast cancer of specific types confirmed by surgical or puncture pathology; (II) tumor lesions metastatic to the breast; (III) history of puncture biopsy, surgical resection, radiotherapy, or chemotherapy prior to MRI; (IV) incomplete clinical data or insufficient pathological and immunohistochemical results; and (V) missing sequences or suboptimal image quality.

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by institutional ethics committees of the Affiliated Hospital of Qinghai University (Center 1; No. P-SL-2024-068) and the Second Hospital of Lanzhou University (Center 2; No. 2024A-1267). Data from Center 3, obtained from a public database, were exempt from ethics review. Written informed consent was waived for this retrospective analysis.

Clinical and pathological features

Clinical and pathologic data such as age, maximum tumor diameter, tumor location, lymph node metastasis determined by MRI, ER status, PR status, HER2 status, Ki-67 expression level, and menopausal status were collected from patients with breast cancer in Center 1 and Center 2. Tumor dimensions were quantified by measuring the maximal cross-sectional diameter on MRI axial images, while axillary lymph node evaluation was performed following established methodological standards as documented in references (35,36).

Immunohistochemical standards

ER, PR, and HER2 expression were detected by immunohistochemical, and the results were determined by one pathologist with more than 10 years of experience. ER and PR positive was defined as the number of tumor cells with nuclear staining ≥1%; ER and PR negative was defined as the number of tumor cells with nuclear staining <1% in the presence of a positive control (37). Immunohistochemical results classified Ki-67 into low expression (Ki-67 <20%) and high expression (Ki-67 ≥20%). HER2 expression was defined as positive when the expression status was (+++) and negative when the expression status was (−) or (+). Further fluorescence in situ hybridization (FISH) testing was required to determine the cases of (++), and those with amplification were defined as the positive group, while those without amplification were defined as the negative group (38). Breast cancer was categorized into luminal type (ER positive and/or PR positive) and non-luminal type (ER negative and PR negative) based on ER and PR expression status.

MRI image acquisition

Breast cancer patients from Center 1 and Center 2 underwent conventional MRI and DCE-MRI using a Philips Achieva 3.0 T scanner (Philips Healthcare, Amsterdam, Netherlands) with a breast-specific coil. All subjects were scanned in the prone position to allow natural draping of the breasts within the coil’s concave aperture. Key scanning sequences and parameters included: (I) T2WI fat-suppressed sequence [repetition time (TR) 5,000 ms, echo time (TE) 60 ms, slice thickness 5.5 mm, interslice gap 1 mm]; (II) DCE-MRI sequence (TR 4.6 ms, TE 2.3 ms, slice thickness 2 mm, interslice gap 1 mm); (III) DWI sequence (TR 3,000 ms, TE 59 ms, slice thickness 4.5 mm, interslice gap 1 mm, b-values 0 and 800 s/mm²). Prior to DCE scanning, a pre-contrast mask image was acquired. Gadopentetate dimeglumine contrast agent was administered via high-pressure syringe at a dose of 0.2 mmol/kg body weight with an injection rate of 2.5 mL/s. Five consecutive post-contrast phases were obtained, each with a scanning duration of 60 seconds.

Center 3 employed a 1.5 T MRI system (GE Healthcare, Signa Excite, USA) equipped with a dedicated 4- or 8-channel breast radiofrequency coil for imaging. The acquisition protocol included a localization scan, T2-weighted sequence, and contrast-enhanced T1-weighted sequences. All imaging was performed unilaterally on the symptomatic breast in the sagittal plane. The contrast-enhanced series comprised high-resolution (in-plane spatial resolution ≤1 mm) three-dimensional, fat-suppressed T1-weighted gradient-echo sequences with the following parameters: TR ≤20 ms, TE =4.5 ms, flip angle ≤45º, field of view 16–18 cm, matrix size ≥256×192, 64 slices, and slice thickness ≤2.5 mm. The T1-weighted sequence required 4.5–5 minutes for completion. A pre-contrast scan was obtained prior to contrast administration, followed by at least two post-contrast acquisitions.

Image segmentation and radiomics feature extraction

For this study, MRI images were initially exported in DICOM format from the PACS workstation and subsequently imported into 3D Slicer software (version 5.0.3, https://www.slicer.org/). This software was utilized to delineate the region of interest (ROI) and extract relevant features. To minimize feature variability, all images underwent a series of preprocessing steps. These included gray scale discretization, intensity normalization, and resampling to a voxel size of 1 mm × 1 mm × 1 mm. The volume of interest (VOI) was obtained by layer-by-layer segmentation of breast tumors on enhanced phase III or IV, T2WI, and DWI images from DCE sequences by a diagnostic radiologist with more than 2 years of experience without knowledge of tumor molecular typing. The segmentation should not go beyond the edge of the lesion, and then the segmented VOIs were proofread by a diagnostic radiologist with more than 5 years of experience. If there was a significant disagreement about the segmented area, another radiologist with more than 10 years of experience in diagnostic MRI should be consulted to discuss the segmented VOIs until a consensus was reached. Using 3D Slicer software, the tumor VOIs of all patients were automatically expanded outward by 3 mm to generate the intratumor combined peritumor 3 mm VOI (VOI_Peri3mm). Only the lesion with the largest diameter was segmented when there were multiple breast tumors in the same patient.

Features were extracted from the generated VOIs using PyRadiomics, an open-source Python package. The extracted features included: (I) morphological characteristics; (II) first-order features; (III) gray-level co-occurrence matrix (GLCM) features; (IV) gray-level run-length matrix (GLRLM) features; (V) neighboring gray-tone difference matrix (NGTDM) features; (VI) gray-level dependence matrix (GLDM) features; and (VII) gray-level size-zone matrix (GLSZM) features.

Feature selection and model development

Z-score standardization was performed before radiomics feature screening to solve the problem of different nature of data and improve the comparability of data, accelerate the speed of gradient descent to find the optimal solution, and improve the accuracy of the model. The inter-class correlation coefficient (ICC) was utilized to assess the features derived from the two segmentations (pre- and post-processing). Features with an ICC value exceeding 0.8 were retained, as these were deemed to exhibit superior consistency and stability. In this study, each sequence was screened individually. First, one-way analysis was performed using the Mann-Whitney U test, and features with a P value <0.1 were retained. Secondly, features with a correlation >0.9 were removed using correlation analysis. The correlation analysis used in this study was the Spearman rank correlation coefficient. Finally, the best radiomics features were selected using the simulated annealing algorithm. After all features were screened, the random forest (RF) machine learning algorithm was used to construct radiomics models (including 8 models) to identify luminal or non-luminal breast cancer: single-parameter intratumor radiomics models, single-parameter intratumor + peritumor 3 mm (intratumor_Peri3mm) radiomics models, multiparameter intratumor fusion radiomics models, multiparameter intratumor + peritumor 3 mm (multiparameter intratumor_Peri3mm) fusion radiomics models, and the above models were used in test sets to validate the efficacy of the models.

Statistical analysis

R 4.2.3 software and SPSS 27.0 statistical software were used for statistical analysis. Measurement information that conformed to normal distribution was expressed as mean ± standard deviation, and groups were compared with each other using a two-sample independent t-test. Measurement information that did not conform to normal distribution was expressed as median (P25, P75), and groups were compared with each other using the Mann-Whitney U test. Counting information was expressed by frequency statistics, and the χ² test was used to compare the differences between groups. Consistency analysis was performed using ICC. The validity of the proposed model was rigorously assessed through a comprehensive set of performance metrics, incorporating the area under the receiver operating characteristic curve (AUC), classification accuracy, sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV). Statistical analysis of inter-model AUC differentials was performed using the DeLong test to evaluate significance levels. The goodness of fit of the model was assessed using calibration curves, and in addition, decision curve analysis (DCA) was performed to visualize the net benefit of clinical decision-making. The flowchart of this study is shown in Figure 1.

Figure 1 Flowchart of the processing steps for identifying luminal or non-luminal breast cancer using radiomics. In the ROI segmentation images, the orange color represents the intratumoral region, while the purple color denotes the peritumoral region. GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLRLM, gray-level run-length matrix; GLSZM, gray-level size-zone matrix; NGTDM, neighboring gray-tone difference matrix; ROI, region of interest.

Results

Participant selection

The study initially included 537 female breast cancer patients who underwent MRI screening across the three participating centers. Patients were excluded based on the following criteria: (I) specific types of invasive breast cancer confirmed by surgical or puncture pathology (n=3); (II) secondary tumor lesions in the breast (n=4); (III) history of puncture biopsy, surgical resection, radiotherapy, or chemotherapy prior to MRI (n=9); (IV) incomplete clinical data or pathological/immunohistochemical results (n=9); and (V) incomplete imaging sequences or poor image quality (n=207). Ultimately, the final cohort included 199 patients from Center 1, 67 from Center 2, and 39 from Center 3 (Figure 2).

Figure 2 Flowchart shows patient exclusion for each data set. MRI, magnetic resonance imaging.

Analysis of general clinical characteristics

A total of 305 patients with breast cancer were enrolled in this study Centers 1, 2 and 3. Patients were established in Center 1 (luminal type 114, non-luminal type 85), Center 2 (luminal type 52, non-luminal type 15), and Center 3 (luminal type 21, non-luminal type 18). Comparison of the clinical and pathological characteristics between the three centers showed that the difference between tumor location (P=0.02) and HER2 expression status (P<0.001) between the luminal and non-luminal groups in Center 1 was statistically significant. Differences between the remaining characteristics were not statistically significant (P>0.05) (Table 1).

Extraction and selection of radiomics features

From each patient’s intratumoral and intratumoral_Peri3mm regions, 2,252 radiomic features were systematically extracted across three distinct imaging modalities. The feature selection pipeline employed a tripartite analytical framework: univariate analysis to identify statistically significant features, correlation analysis to eliminate redundancy, and a simulated annealing algorithm to optimize combinatorial efficiency. Following this rigorous procedure, modality-specific feature retention was achieved: DWI retained 1 feature and DWI_Peri3mm retained 2 features; T2WI retained 1 feature and T2WI_Peri3mm retained 3 features; DCE retained 2 features and DCE_Peri3mm retained 3 features. Subsequent integration of these single-parameter-derived features through cross-modal multivariate screening resulted in an optimized feature ensemble comprising 5 features from the combined intratumoral (DWI + T2WI + DCE) dataset and 8 features from the peritumoral-enhanced (DWI_Peri3mm + T2WI_Peri3mm + DCE_Peri3mm) composite. This hierarchical selection architecture (Tables 2,3, Figures 3,4) demonstrates the enhanced dimensionality and biological relevance achieved through multiparametric peritumoral radiomic profiling.

Figure 3 The violin plot demonstrates the distribution of radiomic features retained after multiparameter combined intratumoral and peritumoral (3 mm) screening. Significant inter-subtype differences (P<0.05) were observed in these features across the following sequences: T2WI (A,D,G), T2WI_Peri3 (B), DWI_Peri3 (C,H), and DCE-MRI (E,F). The vertical axis displays standardized feature values. DCE-MRI, dynamic contrast enhancement magnetic resonance imaging; DWI, diffusion-weighted imaging; Luminal, luminal breast cancer; Non, non-luminal breast cancer; Peri3, intratumor combined peritumor 3 mm; T2WI, T2-weighted imaging with fat suppression.

Figure 4 Correlation heatmaps of final retained features after multiparameter-based intratumor combined peritumor screening in the training set (A), internal test set (B), external test set 1 (C), and external test set 2 (D).

Construction of radiomics models and evaluation of efficacy

Using the RF algorithm, eight distinct radiomic prediction models were developed from the aforementioned feature sets, encompassing two categories: four models focusing on non-expanded tumor regions (DCE RF, DWI RF, T2WI RF, and their combined DWI + T2WI + DCE RF) and four models incorporating 3 mm peritumoral expansion zones (DWI_Peri3 RF, DCE_Peri3 RF, T2WI_Peri3 RF, and their integrated DWI_Peri3 + T2WI_Peri3 + DCE_Peri3 RF). Comparative analysis revealed that among single-parameter configurations, the T2WI_Peri3 RF model demonstrated superior discriminative performance for luminal or non-luminal classification in the training cohort, achieving an AUC of 0.774 [95% confidence interval (CI): 0.698–0.849]. DeLong’s test confirmed statistically significant differential performance between T2WI_Peri3 RF and DCE RF (P=0.004), though no significant differences were observed relative to other single-parameter models (P>0.05). The integrated multiparametric model combining intratumoral and peritumoral features (DWI_Peri3 + T2WI_Peri3 + DCE_Peri3 RF) demonstrated superior predictive performance, achieving an AUC value of 0.819 (95% CI: 0.748–0.889) in the training cohort, along with balanced diagnostic metrics of 75.0% accuracy, 66.7% sensitivity, and 81.3% specificity. The composite model exhibited statistically significant performance enhancement relative to its intratumoral-only counterpart (DWI + T2WI + DCE RF) (P<0.001), demonstrating predictive values of 72.7% for positive cases and 76.5% for negative classifications. Notably, while the multiparametric intratumoral-peritumoral integrated model showed numerically higher AUC values than the optimal single-parameter model (T2WI_Peri3 RF), this difference did not reach statistical significance (P>0.05), suggesting complementary rather than mutually exclusive diagnostic value between single-parametric and multi-parametric approaches.

Our quantitative analysis revealed that multiparametric intratumoral-peritumoral integrated models consistently demonstrated numerically elevated AUC values compared to their single-parameter intratumoral counterparts in the training cohort, though no statistically significant inter-model differentials were observed (P>0.05) (Table 4, Figure 5). DCA further substantiated the clinical superiority of the DWI_Peri3 + T2WI_Peri3 + DCE_Peri3 RF model, which provided greater net benefit across clinically relevant threshold probability ranges compared to alternative configurations (Figure 6). Model calibration assessment revealed excellent agreement between predicted probabilities and observed outcomes in both the training set and internal validation cohort. However, performance degradation was evident in the external validation cohort, indicating potential domain shift challenges in model generalizability (Figure 7). The discordance between internal and external validation metrics underscores the necessity for multicenter validation to ensure clinical deployability.

Figure 5 The receiver operating characteristic curves for the training set (A), the internal test set (B), the external test set 1 (C), and the external test set 2 (D). AUC, area under the receiver operating characteristic curve; DCE, dynamic contrast enhancement; DWI, diffusion-weighted imaging; Peri3, intratumor combined peritumor 3 mm; T2WI, T2-weighted imaging with fat suppression.

Figure 6 The DCA of the radiomics models in the training set (A), the internal test set (B), the external test set 1 (C), and the external test set 2 (D). DCA, decision curve analysis; DCE, dynamic contrast enhancement; DWI, diffusion-weighted imaging; Peri3, intratumor combined peritumor 3 mm; T2WI, T2-weighted imaging with fat suppression.

Figure 7 The calibration curves of the radiomics models in the training set (A), the internal test set (B), the external test set 1 (C), and the external test set 2 (D). DCE, dynamic contrast enhancement; DWI, diffusion-weighted imaging; Peri3, intratumor combined peritumor 3 mm; T2WI, T2-weighted imaging with fat suppression.

Discussion

The luminal and non-luminal molecular subtyping of breast cancer is an important basis for the selection of clinical treatment modalities, as well as an important prognostic factor for the tumor, which has a non-negligible clinical application value. In addition, inaccurate preoperative molecular typing may lead to incorrect clinical decisions and reduced treatment outcomes. In this multicenter retrospective investigation, we constructed and evaluated a radiomics model integrating multiparametric intratumoral and peritumoral features. This model enables non-invasive and precise differentiation between luminal and non-luminal breast cancers, offering crucial guidance for clinical decision-making processes.

Prior investigations (39,40) have predominantly employed radiomic analyses of mammographic and ultrasonographic imaging for breast cancer molecular subtyping, yet both modalities exhibit substantial intrinsic limitations. Conventional two-dimensional acquisitions in mammography and ultrasound inherently compromise the volumetric characterization of tumor heterogeneity. Mammographic protocols further suffer from critical deficiencies: (I) lack of hemodynamic characterization capabilities, confined to morphological assessment of lesions; (II) unavoidable ionizing radiation exposure with non-trivial biological risks. In contrast, breast MRI, with its superior soft-tissue contrast resolution, zero ionizing radiation emission, and multiparametric imaging capacity (DWI, T2WI, DCE), offers unprecedented opportunities for multidimensional tumor profiling. This advanced modality enables comprehensive interrogation of tumor pathobiology across angiogenesis, cellularity, and stromal remodeling axes, thereby establishing itself as the emerging gold standard for precision breast oncology applications spanning early detection, molecular classification, and therapeutic response prediction.

This study innovatively established a multiparametric intratumoral-peritumoral radiomic framework, with the DWI_Peri3 + T2WI_Peri3 + DCE_Peri3 RF model exhibiting strong discriminative capability across validation cohorts. Quantitative assessment showed AUC values of 0.819 (95% CI: 0.748–0.889) in the training set, 0.795 (95% CI: 0.676–0.915) in internal validation, 0.771 (95% CI: 0.640–0.902) in external cohort 1, and 0.590 (95% CI: 0.407–0.773) in external cohort 2. Future investigations will expand clinical data integration to establish hybrid radiomic-clinical nomograms for enhanced classification accuracy. Notably, the model maintained generalizability in external cohort 1, while reduced efficacy in external cohort 2 potentially stemmed from limited cohort size (n=39) meeting exclusion criteria at Center 3. Calibration analysis revealed suboptimal agreement in external cohorts and fundamental limitations of discriminative RF models in probability calibration. Multivariate analysis confirmed that conventional clinical parameters, age, tumor diameter, and location, lacked independent predictive capacity, aligning with Son et al.’s findings (41) that radiomic signatures supersede clinicopathological features in molecular subtyping.

In our study, the DWI_Peri3 + T2WI_Peri3 + DCE_Peri3 RF model was developed based on 8 correlation features, including 3 first-order features, 2 GLSZM features, 1 GLCM feature, and 2 GLDM features, which suggests that the different radiomics features have complementary value in capturing tumor biological information. Higher-order texture features seem to be an important feature in distinguishing luminal from non-luminal breast cancer, which is similar to the study of Huang et al. (31), in which the prediction model was composed of one first-order feature, 5 GLSZM features, 1 GLRLM feature, and 1 GLDM feature. The tumor peripheral environment is closely related to the process of tumor neoangiogenesis, which can lead to an increase in neovascular permeability and promote the secretion of peritumoral cytokines, which play an important role in tumor aggressiveness (42). Intratumor combined with peritumor information provides a more complete response to the biology of the tumor. Therefore, extracting and fusing tumor and peri-tumor radiomics features may improve the discriminatory performance of the model. Peri-tumor regions, along with intra-tumor regions, were also found to serve as complementary information for predicting molecular subtypes of tumors in a study by Niu (43), which is consistent with our study.

As a novel imaging modality for breast cancer evaluation, 2-deoxy-2-[¹⁸F]fluoro-D-glucose positron emission tomography/computed tomography (18F-FDG PET/CT) demonstrates significant clinical value in prognostic assessment. Urso et al. (44) confirmed that PET/CT-detected disease recurrence was significantly associated with reduced overall survival, with this prognostic value remaining independent of traditional tumor marker CA15.3 levels. In treatment response prediction, Hou et al. (45) innovatively integrated PET/CT-derived intratumoral and peritumoral radiomics features to construct a support vector machine model, which exhibited superior predictive performance (training set AUC =0.95; test set AUC =0.83). Notably, subtype-specific models tailored for different molecular subtypes (luminal, HER2-positive, and triple-negative) all demonstrated consistent discriminatory performance (AUC values of 0.90, 0.86, and 0.92, respectively), findings that corroborate the results of our MRI-based peritumoral feature analysis. These results suggest that although luminal primary breast cancers exhibit low FDG avidity, a multi-regional feature extraction strategy can effectively compensate for the limitations of relying solely on FDG uptake intensity. Based on current findings, we propose that future studies should explore the integration of MRI and PET radiomics features to establish a multimodal assessment framework, thereby advancing the development of precision diagnosis and treatment in breast cancer.

In the study conducted by Wang et al. (46), predictive models for molecular subtypes of breast cancer were developed using seven distinct machine learning algorithms. Among these, support vector machines demonstrated superior performance by achieving the highest AUC values in both internal validation and external validation cohorts. Similarly, Sheng et al. (47) investigated multiple machine learning approaches for constructing predictive models of luminal staging in breast cancer, with the eXtreme Gradient Boosting algorithm emerging as the most effective methodology. Comparatively, our current investigation employed a univariate modeling strategy utilizing solely RF for model development. This methodological limitation underscores the necessity for future research to explore and incorporate alternative machine learning methodologies to enhance model robustness and predictive accuracy.

Leithner et al. (48) explored the utility of multiparametric MRI-derived radiomic histology integrated with breast cancer receptor status and molecular subtypes. Their methodology was constrained by exclusively sectioning tumors along diameter planes, 2D approach that may inadequately account for intratumoral heterogeneity. In contrast, our study employed comprehensive 3D tumor segmentation to radiomic features, thereby yielding more biologically informative data and enabling the development of a robust predictive model. This methodological advancement therefore aligns with the findings of Xu et al. (49), whose comparative analysis demonstrated the superior reliability of 3D radiomic feature extraction over conventional 2D or maximal cross-sectional delineation techniques. Feng et al. (29) extracted radiomic features from both the entire tumor and three intratumoral subregions to construct predictive models for distinguishing luminal and non-luminal breast cancers. Their results revealed that logistic regression models based on single subregional features outperformed those derived from whole-tumor analyses. This contrasts with our current framework, which relies solely on whole-tumor radiomics without subregional analysis. Future investigations will extend our methodology by incorporating tumor subregion segmentation to develop subregion-specific predictive models, thereby rigorously evaluating their comparative prognostic performance.

There are some limitations of this study. This is a retrospective study, and the sample data were collected retrospectively, which may introduce some bias. In this study, manual segmentation of ROIs may have introduced segmentation inaccuracies, potentially resulting in loss of image information. Therefore, future research should adopt more precise lesion segmentation methods, such as semi-automatic segmentation techniques. Although this study is a multicenter study, the sample size is small, and more samples should be included in the future to validate the model stability and clinical applicability. In addition, the data source in the external test set 2 is from 2002–2006, which resulted in a large difference from the training set images. Therefore, we need the latest imaging information from Western populations to validate the generalizability of the model. The main reason for selecting DCE-MRI stage III or IV images in this study is that breast cancer intensifies significantly in stage III or IV. However, other periods of DCE-MRI may provide different and important information, which can be further verified in the future.

Conclusions

The multiparametric intratumoral combined peritumoral fusion model has high clinical value in distinguishing luminal or non-luminal breast cancer and can provide the corresponding imaging basis for the clinic, thereby providing a more reasonable treatment plan. In future studies, we hope to collect datasets with high image quality and large sample sizes to further validate our model.

Acknowledgments

The authors sincerely thank the patients who participated in this trial. We also sincerely thank the individuals and investigators of the I-SPY1 trial who made significant contributions to this study.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-83/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-83/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-83/prf

Funding: This work was supported by the National Natural Science Foundation of China (No. 82260346), Qinghai Province “Kunlun Talents High-end Innovation and Entrepreneurial Talents” Top Talent Cultivation Project (No. 13,2021), Qinghai Provincial Department of Science and Technology of China (No. 2023-ZJ-918M), West Light Foundation of the Chinese Academy of Sciences (No. 2023-1-14), and the National Clinical Key Specialty Construction Program (Document No. 90, Office of Health Commission of Qinghai Province, 2024).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-83/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by institutional ethics committees of the Affiliated Hospital of Qinghai University (Center 1; No. P-SL-2024-068) and the Second Hospital of Lanzhou University (Center 2; No. 2024A-1267). Data from Center 3 were obtained from a public database and were exempt from ethics review. Written informed consent was waived for this retrospective analysis.

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, et al. 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. Su Y, Zheng Y, Zheng W, et al. Distinct distribution and prognostic significance of molecular subtypes of breast cancer in Chinese women: a population-based cohort study. BMC Cancer 2011;11:292. [Crossref] [PubMed]
3. Engstrøm MJ, Opdahl S, Hagen AI, et al. Molecular subtypes, histopathological grade and survival in a historic cohort of breast cancer patients. Breast Cancer Res Treat 2013;140:463-73. [Crossref] [PubMed]
4. Sali AP, Sharma N, Verma A, et al. Identification of Luminal Subtypes of Breast Carcinoma Using Surrogate Immunohistochemical Markers and Ascertaining Their Prognostic Relevance. Clin Breast Cancer 2020;20:382-9. [Crossref] [PubMed]
5. Pellegrino B, Hlavata Z, Migali C, et al. Luminal Breast Cancer: Risk of Recurrence and Tumor-Associated Immune Suppression. Mol Diagn Ther 2021;25:409-24. [Crossref] [PubMed]
6. Goldner M, Pandolfi N, Maciel D, et al. Combined endocrine and targeted therapy in luminal breast cancer. Expert Rev Anticancer Ther 2021;21:1237-51. [Crossref] [PubMed]
7. Khan Y, Rizvi S, Raza A, et al. Tailored therapies for triple-negative breast cancer: current landscape and future perceptions. Naunyn Schmiedebergs Arch Pharmacol 2025; Epub ahead of print. [Crossref]
8. Li YW, Dai LJ, Wu XR, et al. Molecular Characterization and Classification of HER2-Positive Breast Cancer Inform Tailored Therapeutic Strategies. Cancer Res 2024;84:3669-83. [Crossref] [PubMed]
9. Bianchini G, De Angelis C, Licata L, et al. Treatment landscape of triple-negative breast cancer - expanded options, evolving needs. Nat Rev Clin Oncol 2022;19:91-113. [Crossref] [PubMed]
10. Vučković L, Klisic A, Raonić J, et al. Comparative study of immunohistochemical determination of breast cancer molecular subtypes on core biopsy and surgical specimens. Eur Rev Med Pharmacol Sci 2021;25:3990-6. [Crossref] [PubMed]
11. Pölcher M, Braun M, Tischitz M, et al. Concordance of the molecular subtype classification between core needle biopsy and surgical specimen in primary breast cancer. Arch Gynecol Obstet 2021;304:783-90. [Crossref] [PubMed]
12. Park HL, Hong J. Vacuum-assisted breast biopsy for breast cancer. Gland Surg 2014;3:120-7. [Crossref] [PubMed]
13. Witowski J, Heacock L, Reig B, et al. Improving breast cancer diagnostics with deep learning for MRI. Sci Transl Med 2022;14:eabo4802. [Crossref] [PubMed]
14. Wang X, Xie T, Luo J, et al. Radiomics predicts the prognosis of patients with locally advanced breast cancer by reflecting the heterogeneity of tumor cells and the tumor microenvironment. Breast Cancer Res 2022;24:20. [Crossref] [PubMed]
15. Pesapane F, Sorce A, Battaglia O, et al. Contrast Agents in Breast MRI: State of the Art and Future Perspectives. Biomedicines 2025;13:829. [Crossref] [PubMed]
16. Wang Q, Lin Y, Ding C, et al. Multi-modality radiomics model predicts axillary lymph node metastasis of breast cancer using MRI and mammography. Eur Radiol 2024;34:6121-31. [Crossref] [PubMed]
17. Liu W, Li L, Deng J, et al. A comprehensive approach for evaluating lymphovascular invasion in invasive breast cancer: Leveraging multimodal MRI findings, radiomics, and deep learning analysis of intra- and peritumoral regions. Comput Med Imaging Graph 2024;116:102415. [Crossref] [PubMed]
18. Yu T, Yu R, Liu M, et al. Integrating intratumoral and peritumoral radiomics with deep transfer learning for DCE-MRI breast lesion differentiation: A multicenter study comparing performance with radiologists. Eur J Radiol 2024;177:111556. [Crossref] [PubMed]
19. Yu H, Meng X, Chen H, et al. Correlation Between Mammographic Radiomics Features and the Level of Tumor-Infiltrating Lymphocytes in Patients With Triple-Negative Breast Cancer. Front Oncol 2020;10:412. [Crossref] [PubMed]
20. Braman NM, Etesami M, Prasanna P, et al. Erratum to: Intratumoral and peritumoral radiomics for the pretreatment prediction of pathological complete response to neoadjuvant chemotherapy based on breast DCE-MRI. Breast Cancer Res 2017;19:80. [Crossref] [PubMed]
21. Hwang KT, Kim YA, Kim J, et al. The influences of peritumoral lymphatic invasion and vascular invasion on the survival and recurrence according to the molecular subtypes of breast cancer. Breast Cancer Res Treat 2017;163:71-82. [Crossref] [PubMed]
22. Wang Z, Zhang H, Lin F, et al. Intra- and Peritumoral Radiomics of Contrast-Enhanced Mammography Predicts Axillary Lymph Node Metastasis in Patients With Breast Cancer: A Multicenter Study. Acad Radiol 2023;30:S133-42. [Crossref] [PubMed]
23. Mao N, Shi Y, Lian C, et al. Intratumoral and peritumoral radiomics for preoperative prediction of neoadjuvant chemotherapy effect in breast cancer based on contrast-enhanced spectral mammography. Eur Radiol 2022;32:3207-19. [Crossref] [PubMed]
24. Zhou J, Zhang Y, Chang KT, et al. Diagnosis of Benign and Malignant Breast Lesions on DCE-MRI by Using Radiomics and Deep Learning With Consideration of Peritumor Tissue. J Magn Reson Imaging 2020;51:798-809. [Crossref] [PubMed]
25. Liu Y, Li X, Zhu L, et al. Preoperative Prediction of Axillary Lymph Node Metastasis in Breast Cancer Based on Intratumoral and Peritumoral DCE-MRI Radiomics Nomogram. Contrast Media Mol Imaging 2022;2022:6729473. [Crossref] [PubMed]
26. Li C, Song L, Yin J. Intratumoral and Peritumoral Radiomics Based on Functional Parametric Maps from Breast DCE-MRI for Prediction of HER-2 and Ki-67 Status. J Magn Reson Imaging 2021;54:703-14. [Crossref] [PubMed]
27. Song SE, Cho KR, Cho Y, et al. Machine learning with multiparametric breast MRI for prediction of Ki-67 and histologic grade in early-stage luminal breast cancer. Eur Radiol 2022;32:853-63. [Crossref] [PubMed]
28. Xu A, Chu X, Zhang S, et al. 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]
29. Feng S, Yin J. Dynamic contrast-enhanced magnetic resonance imaging radiomics analysis based on intratumoral subregions for predicting luminal and nonluminal breast cancer. Quant Imaging Med Surg 2023;13:6735-49. [Crossref] [PubMed]
30. Lafcı O, Celepli P, Seher Öztekin P, et al. DCE-MRI Radiomics Analysis in Differentiating Luminal A and Luminal B Breast Cancer Molecular Subtypes. Acad Radiol 2023;30:22-9. [Crossref] [PubMed]
31. Huang T, Fan B, Qiu Y, et al. Application of DCE-MRI radiomics signature analysis in differentiating molecular subtypes of luminal and non-luminal breast cancer. Front Med (Lausanne) 2023;10:1140514. [Crossref] [PubMed]
32. Yuen S, Monzawa S, Yanai S, et al. The association between MRI findings and breast cancer subtypes: focused on the combination patterns on diffusion-weighted and T2-weighted images. Breast Cancer 2020;27:1029-37. [Crossref] [PubMed]
33. Clark K, Vendt B, Smith K, et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J Digit Imaging 2013;26:1045-57. [Crossref] [PubMed]
34. Esserman LJ, Berry DA, Cheang MC, et al. Chemotherapy response and recurrence-free survival in neoadjuvant breast cancer depends on biomarker profiles: results from the I-SPY 1 TRIAL (CALGB 150007/150012; ACRIN 6657). Breast Cancer Res Treat 2012;132:1049-62. [Crossref] [PubMed]
35. Mortellaro VE, Marshall J, Singer L, et al. Magnetic resonance imaging for axillary staging in patients with breast cancer. J Magn Reson Imaging 2009;30:309-12. [Crossref] [PubMed]
36. Javid S, Segara D, Lotfi P, et al. Can breast MRI predict axillary lymph node metastasis in women undergoing neoadjuvant chemotherapy. Ann Surg Oncol 2010;17:1841-6. [Crossref] [PubMed]
37. Allison KH, Hammond MEH, Dowsett M, et al. Estrogen and Progesterone Receptor Testing in Breast Cancer: ASCO/CAP Guideline Update. J Clin Oncol 2020;38:1346-66. [Crossref] [PubMed]
38. Wolff AC, Hammond MEH, Allison KH, et al. HER2 Testing in Breast Cancer: American Society of Clinical Oncology/College of American Pathologists Clinical Practice Guideline Focused Update Summary. J Oncol Pract 2018;14:437-41. [Crossref] [PubMed]
39. Ma W, Zhao Y, Ji Y, et al. Breast Cancer Molecular Subtype Prediction by Mammographic Radiomic Features. Acad Radiol 2019;26:196-201. [Crossref] [PubMed]
40. Huang Y, Yao Z, Li L, et al. Deep learning radiopathomics based on preoperative US images and biopsy whole slide images can distinguish between luminal and non-luminal tumors in early-stage breast cancers. EBioMedicine 2023;94:104706. [Crossref] [PubMed]
41. Son J, Lee SE, Kim EK, et al. Prediction of breast cancer molecular subtypes using radiomics signatures of synthetic mammography from digital breast tomosynthesis. Sci Rep 2020;10:21566. [Crossref] [PubMed]
42. Santucci D, Faiella E, Cordelli E, et al. The Impact of Tumor Edema on T2-Weighted 3T-MRI Invasive Breast Cancer Histological Characterization: A Pilot Radiomics Study. Cancers (Basel) 2021;13:4635. [Crossref] [PubMed]
43. Niu S, Jiang W, Zhao N, et al. Intra- and peritumoral radiomics on assessment of breast cancer molecular subtypes based on mammography and MRI. J Cancer Res Clin Oncol 2022;148:97-106. [Crossref] [PubMed]
44. Urso L, Quartuccio N, Caracciolo M, et al. Impact on the long-term prognosis of FDG PET/CT in luminal-A and luminal-B breast cancer. Nucl Med Commun 2022;43:212-9. [Crossref] [PubMed]
45. Hou X, Chen K, Wan X, et al. Intratumoral and peritumoral radiomics for preoperative prediction of neoadjuvant chemotherapy effect in breast cancer based on (18)F-FDG PET/CT. J Cancer Res Clin Oncol 2024;150:484. [Crossref] [PubMed]
46. Wang F, Wang D, Xu Y, et al. Potential of the Non-Contrast-Enhanced Chest CT Radiomics to Distinguish Molecular Subtypes of Breast Cancer: A Retrospective Study. Front Oncol 2022;12:848726. [Crossref] [PubMed]
47. Sheng W, Xia S, Wang Y, et al. Invasive ductal breast cancer molecular subtype prediction by MRI radiomic and clinical features based on machine learning. Front Oncol 2022;12:964605. [Crossref] [PubMed]
48. Leithner D, Mayerhoefer ME, Martinez DF, et al. Non-Invasive Assessment of Breast Cancer Molecular Subtypes with Multiparametric Magnetic Resonance Imaging Radiomics. J Clin Med 2020;9:1853. [Crossref] [PubMed]
49. Xu A, Chu X, Zhang S, et al. Prediction Breast Molecular Typing of Invasive Ductal Carcinoma Based on Dynamic Contrast Enhancement Magnetic Resonance Imaging Radiomics Characteristics: A Feasibility Study. Front Oncol 2022;12:799232. [Crossref] [PubMed]