Abstract
Purpose
To evaluate the added value of a synthetic radiomics model (S-Radiomics) that integrates topologic and Hessian-based features with traditional radiomic features for predicting a pathologic complete response (pCR) in human epidermal growth factor receptor 2 (HER2)–positive breast cancer.
Materials and Methods
Patients with HER2-positive breast cancer from three centers were retrospectively included between January 2018 and January 2024. Patients from centers 1 and 2 (n = 201) were randomly assigned to training (n = 150) and internal testing (n = 51) sets, whereas patients from center 3 (n = 82) constituted external testing set I. An additional genomic set (n = 64, I-SPY 2) and external testing set II (n = 57, I-SPY 1) were used for independent validation. Multivariable logistic regression with all clinical variables was used to identify independent predictors (P < .05) and develop the clinical model, adjusting for potential confounders. Traditional radiomics model (T-Radiomics) and S-Radiomics were constructed. Performance was evaluated using area under the receiver operating characteristic curve (AUC), integrated discrimination improvement, net reclassification index, and survival analysis.
Results
A total of 404 female patients with HER2-positive breast cancer (mean age ± SD, 50 years ± 9) were included. The S-Radiomics outperformed the T-Radiomics on both the internal (AUC, 0.86 vs 0.78; P = .37) and external test (AUC, 0.84 vs 0.73; P = .11) sets, with improvements in the integrated discrimination improvement (P = .24 and .01) and net reclassification index (P = .21 and .01). Shapley additive explanations analysis revealed topologic features as the most important predictors. For center 1, survival analysis revealed longer disease-free survival in predicted responders (P = .02).
Conclusion
The S-Radiomics outperformed the T-Radiomics in predicting pathologic complete responses in patients with HER2-positive breast cancer.
Keywords: Radiogenomics, MRI, Breast, Human Epidermal Growth Factor Receptor 2, Breast Cancer, Neoadjuvant Chemotherapy, Pathologic Complete Response, Radiomics
Supplemental material is available for this article.
© RSNA, 2026
See also commentary by Hu in this issue.
Keywords: Radiogenomics, MRI, Breast, Human Epidermal Growth Factor Receptor 2, Breast Cancer, Neoadjuvant Chemotherapy, Pathologic Complete Response, Radiomics
Summary
Synthetic MRI radiomics accurately predicted pathologic complete response to neoadjuvant therapy in human epidermal growth factor receptor 2–positive breast cancer, outperformed traditional radiomics models, and was associated with immune-related pathways and recurrence-free survival.
Key Points
■ In 283 patients with human epidermal growth factor receptor 2–positive breast cancer, the synthetic radiomics model achieved area under the receiver operating characteristic curve values of 0.84 for the training set, 0.86 for the internal set, and 0.84 for the external set, outperforming a traditional model with values of 0.76 for the training set, 0.78 for the internal set, and 0.73 for the external set.
■ The mean Gaussian curvature was the most influential predictor; higher curvature favored a pathologic complete response.
■ RNA processing–related pathway upregulation, greater immune cell infiltration, and longer recurrence-free survival (log-rank P = .02) were observed in predicted low-risk groups.
Introduction
Human epidermal growth factor receptor 2 (HER2)–positive breast cancer, which accounts for approximately 15%–20% of all patients with breast cancer, is characterized by high invasiveness and poor prognosis (1). Standard treatment involves combining neoadjuvant chemotherapy (NAC) with anti-HER2 targeted therapies, yielding high pathologic complete response (pCR) rates and improved survival outcomes (2). Although the pCR serves as a reliable surrogate end point for prognosis, it can be assessed only after six to eight cycles of NAC. However, some patients are resistant to treatment, and NAC is associated with considerable adverse effects (3). Thus, early identification of patients unlikely to achieve a pCR, ideally before the start of treatment, is clinically important.
MRI has been increasingly used as a noninvasive approach for the early prediction of treatment response in patients with breast cancer, particularly through imaging biomarkers derived from tumor morphology and enhancement patterns (4,5). HER2-positive tumors frequently appear as irregular masses with washout kinetics on MR images, reflecting pronounced morphologic and vascular heterogeneity (6,7). These imaging patterns reflect underlying biologic variability, which influences drug delivery and treatment sensitivity (8). Radiomics offers a quantitative framework for characterizing tumor phenotype by extracting high-throughput features from MR images (9). Although traditional radiomic shape descriptors enable baseline morphologic assessments, a prior study has reported limited ability to characterize breast tumors, which exhibit complex morphologies, and its suboptimal performance in predicting the response to NAC (10). These challenges highlight the need for more informative imaging descriptors that better capture the structural and vascular heterogeneity of HER2-positive tumors.
To address these limitations, topologic and Hessian-based features may serve as complementary descriptors that enrich the characterization of tumor morphologies and internal architectures. Topologic descriptors quantify global shape complexity and structural continuity, providing a more integrated view of tumor architectures (11,12). In contrast, Hessian-based features emphasize tubular structures and fine-scale spatial gradients, which may correspond to vascular morphology and enhancement-related heterogeneity (13). Integrating these features, which capture distinct yet biologically relevant attributes, with traditional radiomic features may provide incremental value in predicting pCRs in patients with HER2-positive breast cancer.
The aim of this study was to evaluate the added value of a synthetic radiomics strategy that integrates traditional, topologic, and Hessian-based features for predicting a pCR in patients with HER2-positive breast cancer. Additionally, the biologic interpretability and clinical prognostic value of the developed model were explored.
Materials and Methods
Study Sample
This retrospective, multicenter study was approved by the institutional review board at each participating center, and the requirement for informed consent was waived. Data from 332 consecutive patients with biopsy-confirmed HER2-positive invasive breast cancer who underwent pretreatment breast MRI were initially collected from three centers: 136 from Guangzhou First People’s Hospital (center 1; January 2018–January 2024), 102 from Henan Cancer Hospital (center 2; March 2019–December 2022), and 94 from Liuzhou People’s Hospital (center 3; January 2020–March 2024). Samples were excluded if MR images showed obvious artifacts, no definitive surgery was performed, or NAC was incomplete (fewer than six or eight cycles before surgery). Finally, 283 samples were included in the final analysis: 118 from center 1, 83 from center 2, and 82 from center 3. Samples from centers 1 and 2 were randomly assigned to a training set (n = 150) and an internal test set (n = 51) at a 3:1 ratio. Samples from center 3 (n = 82) were included in external test set I. All samples received HER2-targeted therapy (trastuzumab with or without pertuzumab) combined with chemotherapy (either taxane-based or anthracycline- and taxane-based), administered in six or eight 21-day cycles before surgery.
To evaluate the biologic and prognostic relevance of the imaging biomarkers, two independent HER2-positive breast cancer datasets were obtained from The Cancer Imaging Archive. The gene test set included 88 patients from the ACRIN 6698/I-SPY 2 trial (NCT01564368) (https://www.cancerimagingarchive.net/collection/acrin-6698/), whose paired dynamic contrast-enhanced (DCE) MRI and clinical data were derived from the GPL20078 and GPL30493 platforms. External test set II included DCE MRI and clinical data from 67 samples from the ACRIN 6657/I-SPY 1 trial (NCT00033397) (https://www.cancerimagingarchive.net/collection/ispy1/). Samples with substantial image artifacts or missing clinical data were excluded. A total of 121 patients were included after exclusion: 64 in the gene test set and 57 in external test set II. The present study includes 164 samples that overlap with our previous work (14). An additional 240 dataset entries were accrued for this study, comprising 119 newly included samples from multicenter and 121 samples from public databases. Unlike our prior research, which focused on longitudinal morphologic and functional changes using multiparametric MRI, the current work focuses on baseline DCE MRI radiomics. Furthermore, this study integrates transcriptomic data and survival analysis to explore the biologic underpinnings and clinical utility.
MRI Acquisition
MRI scans at centers 1 and 2 were performed using either a 1.5-T scanner (uMR 560; United Imaging) or a 3.0-T scanner (Skyra; Siemens Healthcare). The images in the external test set (center 3) were obtained using a 3.0-T scanner (Philips Healthcare). DCE MRI was conducted using an axial three-dimensional fat-saturated T1-weighted sequence, following intravenous injection of a gadolinium-based contrast agent (Magnevist; Bayer Healthcare) at 0.2 mL/kg body weight and a rate of 1.5 mL/sec. All patients underwent MRI within 1 week before NAC initiation and before biopsy.
For the publicly available cohorts, DCE MRI was performed using 1.5-T or 3.0-T clinical scanners equipped with dedicated breast coils, according to each trial’s standardized imaging protocol (detailed acquisition parameters provided in Table S1) (15,16).
Histologic Analysis and Clinical Data Collection
Histopathologic assessments were performed with biopsy before NAC. These evaluations were conducted by dedicated breast pathologists at each participating center who were masked to the radiomics data. Immunohistochemistry was used to evaluate the expression of the estrogen receptor, progesterone receptor (PR), and HER2 and the Ki-67 index. The estrogen receptor and PR were considered positive if more than 1% of invasive tumor cells exhibited nuclear staining. HER2-positive status was defined as immunohistochemistry 3+ or immunohistochemistry 2+ with HER2 amplification by fluorescence in situ hybridization. The pCR was determined on the basis of surgical pathology results within 1 month after completing NAC and was defined as the absence of residual invasive cancer in both the breast and axillary lymph nodes, regardless of the presence of residual ductal carcinoma in situ (17). Clinical data, including patient age, menopausal status, tumor location, clinical T stage, and clinical N stage, were extracted from electronic medical records.
Tumor Segmentation and Feature Extraction
A junior radiologist (S.C.) independently delineated the regions of interest for each tumor on the DCE images using ITK-SNAP (version 4.0.2; University of Pennsylvania). For DCE imaging, early-phase contrast-enhanced images were selected because this phase typically exhibits the most pronounced tumor enhancement. The tumors were manually delineated on all the sections exhibiting tumor enhancement, and the entire tumor boundary was outlined. If multiple lesions were present, only the largest tumor was segmented. A senior radiologist (Y.G.) subsequently randomly selected 30 images for resegmentation, and intraclass correlation coefficient analysis was performed.
All DCE MR images were first harmonized using ComBat to reduce scanner variability and then isotropically resampled to 1 × 1 × 1 mm3, followed by intensity normalization for each patient. Extraction of features, comprising traditional, topologic, and Hessian-based features was performed. Three types of traditional radiomic features were included: first-order statistics, shape-based features, and texture-based features. A topologic radiomic-based feature extraction method was developed, with differential geometry and topology used to characterize the curvature and morphologic features of the tumor surfaces. This approach involves standardizing and resampling three-dimensional tumor masks, followed by the calculation of topologic features such as the Gaussian curvature, mean curvature, principal curvature, curvedness, shape index, and total curvature. In addition, a Hessian radiomic-based feature extraction method was developed, with the Hessian matrix used to capture tumor heterogeneity. This method involves calculating the number of negative eigenvalues (Hessian index) at each voxel within the DCE MR images; then, concave, saddle, and convex points on the tumor surface are identified. By computing various Hessian-derived features such as the blobness, vesselness, and structural contributions across multiple sigma scales (1.0, 2.0, and 3.0), detailed spatial information related to the tumor morphology was captured (Appendix S1). Radiomic features were extracted from DCE MR images using the PyRadiomics (version 3.0.1) package following the standardization and definitions provided by the image biomarker standardization initiative.
Feature Selection and Model Construction
A three-step procedure was used to select robust radiomic features. First, the intraclass correlation coefficient was calculated to assess the stability of the radiomic features, and features with an intraclass correlation coefficient less than 0.75 were excluded. Next, Pearson correlation analysis was employed to remove redundant features, using a threshold of ρ greater than 0.90. Finally, the genetic algorithm was used to identify the optimal feature subset (Appendix S2). All code used for modeling and data analysis is deposited in a publicly accessible repository (https://github.com/JBnoJB/S-radiomics).
Three logistic regression models for predicting the pCR were constructed: clinical model, traditional radiomics model (T-Radiomics), and synthetic radiomics model (S-Radiomics). All clinical variables were initially included in the multivariable logistic regression model. To control for potential bias, patient age, menstrual status, and tumor location were adjusted for as confounding variables. Other variables were evaluated as candidates, and significant variables (P < .05) were identified as independent predictors to construct the final clinical model. The T-Radiomics was developed using traditional radiomic features, whereas the S-Radiomics was developed using a combination of traditional, topologic, and Hessian-based radiomic features (Tables S2 and S3).
Shapley Additive Explanations Analysis
Shapley additive explanations (SHAP) is a model-agnostic interpretability method, grounded in cooperative game theory, that quantifies the contribution of each feature to model predictions (18). SHAP summary plots were generated to rank the global importance of the radiomic features on the basis of their average absolute contribution to the model predictions. To examine how individual features influenced predictions across patients, swarm plots were used to visualize the distribution and direction of the SHAP values at the patient level.
Radiogenomic Analysis and Survival Analysis
To assess whether the S-Radiomics reflects the biologic behavior of tumors and associations with gene expression, paired messenger RNA expression data were collected from the National Center for Biotechnology Information. Patients were categorized into pCR and non-pCR groups on the basis of the binary predictions of the S-Radiomics. Gene set enrichment analysis was conducted based on the entire I-SPY 2 gene expression dataset, with response categories used as phenotype labels. Kyoto Encyclopedia of Genes and Genomes gene sets were used to identify pathways associated with treatment response.
To explore the relationship between the S-Radiomics and the tumor microenvironment, immune cell abundance scores for 10 cell types were calculated using the microenvironment cell population counter method and normalized by count. Differences in immune cell abundance between the pCR and non-pCR groups were evaluated using the rank-sum test and t tests.
Recurrence-free survival was used to evaluate the prognostic value of the model based on the center 1 cohort. Recurrence-free survival was defined as the interval between the initial surgery and the first occurrence of local or distant recurrence. Patients were excluded from the survival analysis if follow-up records were not available. The remaining patients were followed up through March 1, 2026. The Kaplan-Meier method and log-rank test were used in survival analyses, with risk stratification performed using an objective probability threshold of 50%.
Sample Size Evaluation
A minimum sample size of 34 cases (17 pCR and 17 non-pCR) was required, assuming a 90% power, a two-sided significance level of .05, an alternative hypothesis area under the receiver operating characteristic curve (AUC) of 0.800 against a null AUC of 0.500, and an allocation ratio of 1. Thus, a total sample size of 50 cases was sufficient to detect an AUC difference from the null hypothesis of 0.500 with 90% power and a two-sided significance level of .05 if the true AUC was 0.800, exceeding the minimum statistical requirement of 34 cases (19). Sample analyses were performed using MedCalc (version 22.016; https://www.medcalc.org).
Statistical Analysis
All the statistical analyses were conducted using R (version 4.4.2; R Foundation for Statistical Computing) and Python (version 3.7; Python Software Foundation). For intergroup comparisons of clinical features, Student t test or the Mann-Whitney U test was used for continuous variables, whereas the χ2 test or Fisher exact test was used for categorical variables. The receiver operating characteristic curve was plotted, and the AUC was calculated to evaluate model performance. Additionally, the sensitivity, accuracy, specificity, positive predictive value, and negative predictive value were calculated to assess the model’s diagnostic performance. Decision curve analysis was performed to assess the net clinical benefit of the model. To evaluate changes in model performance, the integrated discrimination improvement (IDI) and net reclassification index were calculated. A P value less than .05 was considered to indicate statistical significance.
Results
Patient Characteristics
A flowchart of the participant selection is illustrated in Figure 1. The workflow of this study is depicted in Figure 2. A total of 283 patients with HER2-positive breast cancer were included in this study. In the training set (mean age ± SD, 51.7 years ± 9.2, 150 female), 67 of 150 patients (44.7%) achieved a pCR after NAC. In the internal test (mean age ± SD, 50.7 years ± 9.8, 51 female) and external test (mean age ± SD, 50.7 ± 9.9 years, 82 female) sets, 23 of 51 patients (45.1%) and 39 of 82 patients (47.6%) achieved a pCR, respectively. In the public datasets, the pCR rate was 43.9% (25 of 57 patients) in the ACRIN 6657/I-SPY 1 set and 53.2% (34 of 64 patients) in the ACRIN 6698/I-SPY 2 set. The clinicopathologic characteristics of the patients across all cohorts are presented in Table 1.
Figure 1:
Flowchart shows inclusion and exclusion criteria. BC = breast cancer, HER2 = human epidermal growth factor receptor 2, NAC = neoadjuvant chemotherapy, pCR = pathologic complete response.
Figure 2:
Image shows the workflow of this study. DCA = decision curve analysis, GA = genetic algorithm, ICC = intraclass correlation coefficient, pCR = pathologic complete response, ROC = receiver operating characteristic, SHAP = Shapley additive explanations, S-Radiomics = synthetic radiomics model, T-Radiomics = traditional radiomics model.
Table 1:
Clinical Characteristics of the Patients in Different Datasets
| Characteristic | Training Set (n = 150) | Internal Test Set (n = 51) | External Test Set (n = 82) | I-SPY 2 Set (n = 64) | I-SPY 1 Set (n = 57) |
|---|---|---|---|---|---|
| Outcome | |||||
| Non-pCR | 83 (55.3) | 28 (54.9) | 43 (52.4) | 30 (46.8) | 32 (56.1) |
| pCR | 67 (44.7) | 23 (45.1) | 39 (47.6) | 34 (53.2) | 25 (43.9) |
| Age (y) | 51 ± 9 | 50 ± 9 | 50 ± 9 | 49 ± 11 | 48 ± 8 |
| Menstrual status | |||||
| Premenopausal | 65 (43.3) | 23 (45.1) | 40 (48.7) | NA | NA |
| Postmenopausal | 85 (56.7) | 28 (54.9) | 42 (51.3) | NA | NA |
| Tumor location | |||||
| Left | 80 (53.3) | 26 (50.9) | 39 (47.5) | NA | NA |
| Right | 70 (46.7) | 25 (49.1) | 43 (52.5) | NA | NA |
| ER status | |||||
| Negative | 65 (43.3) | 22 (43.1) | 45 (54.8) | NA | 33 (57.8) |
| Positive | 85 (56.7) | 29 (56.9) | 37 (45.2) | NA | 24 (42.2) |
| PR status | |||||
| Negative | 78 (52.0) | 24 (47.0) | 40 (48.7) | NA | 42 (73.6) |
| Positive | 72 (48.0) | 27 (53.0) | 42 (51.3) | NA | 15 (26.4) |
| Ki-67 status | |||||
| ≤20 | 28 (18.6) | 9 (17.6) | 20 (24.3) | NA | NA |
| >20 | 122 (81.4) | 42 (82.3) | 62 (75.7) | NA | NA |
| Clinical T stage | |||||
| 1 | 14 (9.4) | 6 (11.8) | 7 (8.5) | NA | NA |
| 2 | 87 (58.0) | 27 (52.9) | 40 (48.8) | NA | NA |
| 3 | 20 (13.3) | 7 (13.7) | 14 (17.1) | NA | NA |
| 4 | 29 (19.3) | 11 (21.2) | 21 (25.6) | NA | NA |
| Clinical N stage | |||||
| Negative | 99 (66.0) | 33 (64.7) | 49 (59.8) | NA | NA |
| Positive | 51 (34.0) | 18 (35.3) | 33 (40.2) | NA | NA |
Note.—Data are numbers with percentages in parentheses or means ± SDs. ER = estrogen receptor, NA = not available, pCR = pathologic complete response, PR = progesterone receptor.
Performance of the Three Models
Univariable and multivariable logistic regression analyses revealed significant clinical and pathologic factors for the clinical model. PR status, Ki-67 status, and clinical T stage were independent predictors (P < .05; Table 2). The clinical model achieved an AUC of 0.70 (95% CI: 0.64, 0.77) on the training set, an AUC of 0.77 (95% CI: 0.63, 0.91) on the internal test set, and an AUC of 0.66 (95% CI: 0.55, 0.78) on the external test set (Fig 3A). The T-Radiomics, which is based on 12 traditional radiomic features, achieved an AUC of 0.76 (95% CI: 0.69, 0.85) on the training set, an AUC of 0.78 (95% CI: 0.66, 0.91) on the internal test set, and an AUC of 0.73 (95% CI: 0.62, 0.84) on the external test set (Fig 3B). The S-Radiomics, which incorporated 13 traditional radiomic features, two topologic radiomic features, and three Hessian-based radiomic features, achieved an AUC of 0.84 (95% CI: 0.78, 0.91) on the training set, an AUC of 0.86 (95% CI: 0.75, 0.97) on the internal validation set, and an AUC of 0.84 (95% CI: 0.75, 0.92) on the external test set (Fig 3C). Additionally, the accuracy, specificity, sensitivity, positive predictive value, and negative predictive value results for each model are shown in Table 3, and the receiver operating characteristic curves are displayed in Figure 3A–3C.
Table 2:
Univariable and Multivariable Logistic Regression Analyses of Clinical Characteristics for Predicting a pCR Based on the Training Set
| Characteristic | Univariable | Multivariable | ||
|---|---|---|---|---|
| OR (95% CI) | P Value | OR (95% CI) | P Value | |
| Age | 1.01 (0.97, 1.04) | .76 | 1.03 (0.97, 1.11) | .32 |
| Menstrual status | ||||
| Premenopausal | Ref | — | Ref | — |
| Postmenopausal | 1.00 (0.52, 1.92) | .99 | 0.66 (0.21, 2.09) | .48 |
| ER status | ||||
| Negative | Ref | — | Ref | — |
| Positive | 0.55 (0.28, 1.05) | .07 | 1.64 (0.56, 4.82) | .37 |
| PR status | ||||
| Negative | Ref | — | Ref | — |
| Positive | 0.41 (0.21, 0.79) | .01 | 0.27 (0.01, 0.77) | .01 |
| Ki-67 status | ||||
| Negative | Ref | — | Ref | — |
| Positive | 2.34 (0.96, 0.79) | .06 | 3.49 (1.19, 10.32) | .02 |
| Clinical T stage | ||||
| 1 | Ref | .05 | Ref | .02 |
| 2 | 6.56 (1.59, 27.07) | .01 | 11.69 (2.40, 56.87) | <.001 |
| 3 | 2.04 (0.81, 5.10) | .13 | 1.90 (0.71, 5.07) | .20 |
| 4 | 3.21 (0.97, 10.65) | .06 | 3.08 (0.85, 11.17) | .09 |
| Clinical N stage | ||||
| Negative | Ref | — | Ref | — |
| Positive | 1.87 (0.95, 3.71) | .07 | 1.61 (0.73, 3.53) | .24 |
| Location | ||||
| Left | Ref | — | Ref | — |
| Right | 1.20 (0.63, 2.30) | .57 | 1.32 (0.62, 2.82) | .47 |
Note.—ER = estrogen receptor, OR = odds ratio, pCR = pathologic complete response, PR = progesterone receptor, Ref = reference.
Figure 3:
Graphs show the performance of the models in predicting a pathologic complete response in patients with human epidermal growth factor receptor 2–positive breast cancer after neoadjuvant chemotherapy across the three datasets. (A–C) The receiver operating characteristic curves show the model performance on the (A) training, (B) internal test, and (C) external test sets. (D–F) Decision curve analysis reveals that within the clinical threshold probability range of 20%–60%, the synthetic radiomics model (S-Radiomics) exhibited greater net clinical benefits than both the traditional radiomics model (T-Radiomics) and clinical model across all datasets: (D) the training set (P = .02), (E) the internal test set (P = .19), and (F) the external test set (P = .15). AUC = area under the receiver operating characteristic curve.
Table 3:
Predictive Performance of the Different Models
| Model | AUC | ACC (%) | SEN (%) | SPE (%) | PPV (%) | NPV (%) | P Value |
|---|---|---|---|---|---|---|---|
| Training set (n = 151) | |||||||
| Clinical | 0.70 (0.64, 0.77) | 67.3 (67.0, 67.6) | 58.2 (46.4, 70.0) | 74.7 (65.3, 84.1) | 65.0 (52.9, 77.1) | 68.9 (59.3, 78.5) | <.01 |
| T-Radiomics | 0.76 (0.69, 0.85) | 70.7 (70.4, 70.9) | 80.6 (71.1, 90.1) | 62.7 (52.2, 73.1) | 63.5 (53.3, 73.8) | 80.0 (70.8, 89.7) | .04 |
| S-Radiomics | 0.84 (0.78, 0.91) | 80.0 (79.8, 80.2) | 83.6 (74.7, 92.5) | 77.1 (69.1, 86.1) | 74.7 (64.8, 84.5) | 85.3 (77.3, 93.3) | Ref |
| Internal test set (n = 50) | |||||||
| Clinical | 0.77 (0.63, 0.91) | 78.4 (77.8, 79.1) | 82.6 (67.1, 98.1) | 75.0 (59.0, 91.0) | 73.1 (56.0, 90.1) | 84.0 (69.6, 98.4) | .37 |
| T-Radiomics | 0.78 (0.66, 0.91) | 74.5 (73.8, 75.2) | 73.9 (56.0, 91.9) | 75.0 (59.0, 91.0) | 70.8 (52.6, 89.0) | 77.8 (62.1, 93.5) | .37 |
| S-Radiomics | 0.86 (0.75, 0.97) | 84.3 (83.8, 84.8) | 91.3 (79.8, 100) | 78.6 (63.4, 93.8) | 77.8 (62.1, 93.5) | 91.7 (80.6, 100) | Ref |
| External test set (n = 82) | |||||||
| Clinical | 0.66 (0.55, 0.78) | 64.6 (64.1, 65.2) | 69.2 (54.7, 83.7) | 60.5 (45.9, 75.1) | 61.4 (47.0, 75.8) | 68.4 (53.6, 83.2) | .02 |
| T-Radiomics | 0.73 (0.62, 0.84) | 70.7 (70.2, 71.2) | 74.4 (60.7, 88.1) | 67.4 (53.4, 81.4) | 67.4 (53.4, 81.4) | 74.4 (60.7, 88.1) | .11 |
| S-Radiomics | 0.84 (0.75, 0.92) | 78.0 (77.6, 78.5) | 82.1 (70.0, 94.1) | 74.4 (61.4, 87.5) | 74.4 (61.4, 87.5) | 82.1 (70.0, 94.1) | Ref |
| I-SPY 2 set (n = 64) | |||||||
| Clinical | NA | NA | NA | NA | NA | NA | NA |
| T-Radiomics | 0.77 (0.65, 0.89) | 71.9 (71.3, 72.5) | 76.5 (62.2, 90.7) | 66.7 (49.8, 83.5) | 72.2 (57.6, 86.9) | 71.4 (54.7, 88.2) | .52 |
| S-Radiomics | 0.80 (0.69, 91) | 76.6 (76.0, 77.1) | 76.5 (62.2, 90.7) | 76.7 (61.5, 91.8) | 78.8 (64.8, 92.7) | 74.2 (58.8, 89.6) | Ref |
| I-SPY 1 set (n = 57) | |||||||
| Clinical | NA | NA | NA | NA | NA | NA | NA |
| T-Radiomics | 0.76 (0.63, 0.89) | 75.4 (74.8, 76.1) | 68.0 (49.7, 86.3) | 81.2 (67.7, 94.8) | 73.9 (56.0, 91.9) | 76.5 (62.2, 90.7) | .26 |
| S-Radiomics | 0.83 (0.72, 0.94) | 78.9 (78.4, 79.5) | 80.0 (64.3, 90.7) | 78.1 (63.8, 92.4) | 74.1 (57.5, 90.6) | 83.3 (70.0, 96.7) | Ref |
Note.—Values in parentheses are 95% CIs. ACC = accuracy, AUC = area under the receiver operating characteristic curve, NA = not available, NPV = negative predictive value, PPV = positive predictive value, SEN = sensitivity, SPE = specificity, S-Radiomics = synthetic radiomics model, T-Radiomics = traditional radiomics model.
Model performance was also validated using public databases. On the I-SPY 1 set, the T-Radiomics and S-Radiomics achieved AUCs of 0.76 and 0.83, respectively. On the I-SPY 2 set, the T-Radiomics and S-Radiomics achieved AUCs of 0.77 and 0.80 respectively (Fig S1A and S1B).
In addition, decision curve analysis was conducted on the training, internal validation, and external test sets to assess clinical utility (Fig 3D–3F). The S-Radiomics consistently exhibited the greatest net benefit across a clinically rational threshold range (approximately 0.2–0.6), outperforming both the T-Radiomics and clinical model. Decision curve analysis was also performed on the I-SPY 1 and I-SPY 2 sets (Fig 1C and 1D).
Comparison of Prediction Models
To evaluate whether integrating topologic and Hessian-based radiomic features improves the predictive performance of traditional radiomics models, the IDI and net reclassification index metrics of the S-Radiomics and T-Radiomics were compared (Table 4). On the training and external test sets, the IDI values of the S-Radiomics and T-Radiomics were 0.13 (P < .001) and 0.17 (P = .01), respectively, indicating an improvement in the pCR prediction accuracy with the S-Radiomics. Although the IDI value was positive on the internal test set (IDI = 0.09), there was no evidence of a difference between the two models (P = .24). Additionally, compared with that of the T-Radiomics, the net reclassification index of the S-Radiomics was better on the training and external test sets (P < .05), suggesting that the S-Radiomics can more accurately classify patients into pCR and non-pCR groups.
Table 4:
Improvement of the S-Radiomics on Different Datasets Compared with the T-Radiomics
| Set | IDI (95% CI) | P Value* | NRI (95% CI) | P Value† |
|---|---|---|---|---|
| Training | 0.13 (0.06, 0.21) | <.001 | 0.28 (0.11, 0.45) | <.01 |
| Internal validation | 0.09 (−0.06, 0.24) | .24 | 0.21 (−0.12, 0.55) | .21 |
| External validation | 0.17 (0.04, 0.24) | .01 | 0.30 (0.06, 0.55) | .01 |
| I-SPY 2 | 0.13 (0.05, 0.21) | <.01 | 0.03 (−0.18, 0.24) | .76 |
| I-SPY 1 | 0.14 (0.01, 0.26) | .03 | 0.12 (−0.17, 0.40) | .42 |
Note.—IDI = integrated discrimination improvement, NRI = net reclassification index.
P value is for the improvement in the IDI.
P value is for the improvement in the NRI.
SHAP Analysis for Model Interpretation
SHAP analysis was performed to visualize the feature contributions in the S-Radiomics. The SHAP summary plot revealed that the topologic radiomic feature Gaussian_curvature_mean was the most influential in terms of model decision-making, followed by the traditional radiomic feature original_shape_flatness (Fig 4A). Additionally, Hessian-based vascular and shape features ranked among the five most important features.
Figure 4:
Shapley additive explanations (SHAP) analysis illustrates feature contributions in the synthetic radiomics model (S-Radiomics). (A) Summary plot ranking features by their mean absolute SHAP values, reflecting their overall contribution to the model. The feature categories are color-coded: Blue for topologic radiomics features, green for traditional radiomics features, and orange for Hessian-based features. (B) Swarm plot showing the distribution of SHAP values for each feature across all samples. Red represents high feature values, and blue represents low values, indicating both the direction and magnitude of the feature impact on the model predictions.
The SHAP swarm plot illustrates the direction and distribution of each feature’s contribution to the model’s predictions (Fig 4B). Higher Gaussian_curvature_mean values, reflecting greater surface curvature and more convex tumor morphology, were associated with a predicted pCR. In contrast, higher original_shape_flatness values were associated with a greater probability of predicting non-pCR. Moreover, shape- and vascular-related Hessian-based features positively contributed to pCR predictions.
Validation of Biologic Mechanisms in Model Prediction
On the basis of the S-Radiomics predictions, patients in the I-SPY 2 set were categorized into predicted pCR and predicted non-pCR groups. Gene set enrichment analysis was performed, and normalized enrichment scores were calculated for each group. The most significantly enriched pathway in the predicted non-pCR group was the upregulated expression of metabotropic glutamate receptors (Fig 5A). In contrast, pathways related to RNA degradation, protein synthesis, and modification were upregulated in the predicted pCR group (Fig 5B).
Figure 5:
Graphs show the genomic and immune profiling of the predicted pathologic complete response (pCR) and non-pCR groups based on the synthetic radiomics model (S-Radiomics). (A,B) Gene set enrichment analysis shows the enriched pathways in the (A) predicted non-pCR and (B) predicted pCR groups based on the 6698/I-SPY 2 dataset, ranked by the normalized enrichment score (NES). (C) Box and whiskers plots compare the infiltration levels of immune cell subsets between the two groups. Orange and blue represent the predicted non-pCR and predicted pCR groups, respectively. *P < .05; ** P < .01.
Analysis of immune cell infiltration revealed that both T lymphocytes and CD8+ T cells were abundantly infiltrated in both groups. However, compared with patients in the non-pCR group, the patients in the pCR group exhibited greater infiltration of most immune cells, suggesting that immune responses may play a crucial role in achieving a pCR (Fig 5C).
Prognostic Analysis
In the center 1 cohort, recurrence-free survival analysis based on the S-Radiomics revealed differences in prognosis between the predicted pCR and non-pCR groups. The Kaplan-Meier curves revealed that, compared with the high-risk group, the low-risk group had a greater recurrence-free survival probability (P = .02, log-rank test; Fig 6). The median follow-up time for this cohort was 33.9 months (IQR, 19.8–50.0 months). Among the 78 included patients, 13 (16.7%) experienced recurrence during the follow-up period ending March 1, 2026, and the remaining 65 (83.3%) did not (Fig 6).
Figure 6:
Graph shows the Kaplan-Meier survival analysis based on the synthetic radiomics model risk stratification in the center 1 cohort. Patients were stratified into predicted non–pathologic complete response (pCR; orange) and predicted pCR (blue) groups. Compared with the high-risk group, the low-risk group had significantly better survival outcomes (P = .02). SY = Guangzhou First People’s Hospital.
Discussion
In this study, the incremental value of an S-Radiomics integrating traditional, topologic, and Hessian-based features for pCR prediction in HER2-positive breast cancer was evaluated, along with its biologic interpretability and prognostic value. The S-Radiomics significantly outperformed the T-Radiomics in the training set (AUC, 0.84 vs 0.76; P = .04), with significant improvements in both integrated discrimination improvement (P < .001) and net reclassification index (P < .01). Furthermore, predicted responders exhibited significantly longer disease-free survival than nonresponders (P = .02).
Previous studies using MRI-based traditional radiomic features to predict a pCR in patients with HER2-positive breast cancer have reported AUCs between 0.58 and 0.75 (9). In our study, the T-Radiomics performed better than existing models (AUC, 0.73–0.78), and the proposed S-Radiomics exhibited better predictive ability (AUC, 0.80–0.86). The IDI and net reclassification index analyses confirmed that the addition of topologic and Hessian-based features increased the predictive accuracy and reduced misclassification. Topologic features quantify global and local geometric properties such as curvature and spatial compactness (11,12), whereas Hessian-based features capture structural patterns related to vascular complexity (13,20). These descriptors complemented traditional features by capturing morphologic and vascular complexities not well characterized by texture or intensity features alone, improving the model’s capacity to stratify heterogeneous HER2-positive tumors.
To understand the underlying predictive mechanism, SHAP analysis was performed to assess instance-level additive feature contributions. Gaussian_curvature_mean, a topologic descriptor of surface curvature, was identified as the most influential variable, followed by flatness, a traditional shape metric. Higher Gaussian_curvature_mean and lower flatness values, reflecting a more convex, compact, and less elongated tumor morphology, were associated with pCR. These features delineate a spheroid and spatially confined phenotype, which has been associated with favorable treatment responses in previous studies (21,22). The incorporation of topologic descriptors extends the ability of radiomics models to represent morphologic characteristics, enabling geometrically sensitive quantification of tumor structures and explaining the improved performance of the S-Radiomics.
In addition, a vascular-related Hessian-based feature ranked among the top five predictors, suggesting that internal architectural complexity, particularly vessel-like patterns, contributes meaningfully to response prediction. HER2-positive tumors are frequently characterized by rich and heterogeneous vasculature (23,24), which may promote both morphologic compactness and intratumoral perfusion, thereby facilitating more effective drug delivery. By capturing vascular structures, Hessian-derived features offer functional insights into the tumor microenvironment. Combined with shape descriptors, they provide biologically relevant information linking vascular complexity to therapeutic responses, delineating a treatment-sensitive phenotype characterized by spatial regularity and perfusion efficiency.
Transcriptomic and immunologic profiling reinforced the biologic relevance of our imaging features. Immune infiltration was significantly higher in patients achieving pCR, potentially driven by chemotherapy-induced tumor immunogenicity and enhanced vasculature that facilitated drug delivery and immune cell migration (22,25). Gene set enrichment analysis of the subgroups predicted by the S-Radiomics revealed additional biologic differences. The pCR group exhibited upregulated pathways related to protein repair and RNA degradation, which may improve the clearance of aberrant proteins and transcriptional errors, thereby limiting proliferation, resistance, and immune evasion (26,27). In contrast, non-pCR tumors exhibit elevated expression of metabotropic glutamate receptors, which may increase adaptation to microenvironmental stress and promote metabolic reprogramming, contributing to resistance (28,29).
Survival analysis based on the S-Radiomics was conducted to evaluate its potential utility in clinical management. Predicted nonresponders had shorter disease-free survival than predicted responders (P = .02), consistent with previous findings (30,31). These results suggest that patients with a lower predicted probability of achieving a pCR may require treatment modification and more intensive follow-up.
Additionally, multivariable logistic regression analyses revealed lower clinical T stage and PR negativity as significant predictors of a pCR, which is consistent with previous findings (32,33). A lower T stage reflects a lower tumor burden and limited local invasion, increasing pCR likelihood (32). Compared with the hormone receptor–negative and HER2-positive subtype, hormone receptor–positive and HER2-positive breast cancer exhibits greater molecular heterogeneity (33). PR status is a key determinant of response, with PR-negative tumors more likely to achieve pCR (34). This finding aligns with those reported in the literature indicating that elevated Ki-67 expression reflects increased tumor proliferation, which confers increased sensitivity to neoadjuvant chemotherapy (35).
This study has several limitations. First, as a retrospective multicenter study, minor protocol variations may have existed; however, all patients whose data were used for model development received standardized HER2-targeted neoadjuvant therapy, minimizing interinstitutional variability. Second, the genomic and immune associations remain correlative, and experimental confirmation is needed. Third, tumor segmentation was performed manually, which is time-consuming and introduces observer-dependent variability.
In conclusion, our study demonstrates the potential of a synthetic radiomics model for predicting pCR in patients with HER2-positive breast cancer, offering added value beyond traditional radiomics models. The model is predictive, interpretable, and biologically relevant, with utility in stratifying long-term prognosis. Incorporating clinicopathologic features further reinforced its potential for guiding personalized clinical decision-making.
Supplemental Files
Acknowledgments
Acknowledgments
The authors thank the radiology technicians and research coordinators at Guangzhou First People’s Hospital, Henan Cancer Hospital, and Liuzhou People’s Hospital for their assistance in image acquisition and data collection. We also express our gratitude to The Cancer Imaging Archive (TCIA) database for providing access to the relevant imaging data.
S.C. and Y.C. contributed equally to this work.
Q.Z. and Y.G. are co–senior authors.
Funding: Supported by the National Natural Science Foundation of China (nos. 82302314, 62271448, and 81901711), Basic and Applied Basic Research Foundation of Guangdong Province (no. 2024A1515010653), Medical Scientific Research Foundation of Guangdong Province (nos. A2023073 and A2024116), Science and Technology Projects in Guangzhou (nos. 2024A03J1030, 2025A03T4163, and 2025A03T4162), and the Special Fund for the Construction of High-level Key Clinical Specialty (Medical Imaging) in Guangzhou, Guangzhou Key Laboratory of Molecular Imaging and Clinical Translational Medicine.
Abbreviations:
- AUC
- area under the receiver operating characteristic curve
- DCE
- dynamic contrast-enhanced
- HER2
- human epidermal growth factor receptor 2
- IDI
- integrated discrimination improvement
- NAC
- neoadjuvant chemotherapy
- pCR
- pathologic complete response
- PR
- progesterone receptor
- SHAP
- Shapley additive explanations
- S-Radiomics
- synthetic radiomics model
- T-Radiomics
- traditional radiomics model
Disclosures of conflicts of interest
Please see ICMJE form(s) for author conflicts of interest. These have been provided as supplemental materials.
References
- 1. Loibl S , Gianni L . HER2-positive breast cancer . Lancet 2017. ; 389 ( 10087 ): 2415 – 2429 . [DOI] [PubMed] [Google Scholar]
- 2. van der Voort A , van der Hoogt KJJ , Wessels R , et al . Diffusion-weighted imaging in addition to contrast-enhanced MRI in identifying complete response in HER2-positive breast cancer . Eur Radiol 2024. ; 34 ( 12 ): 7994 – 8004 . [Published correction appears in Eur Radiol 2025;35(4):2357.] [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Seidman A , Hudis C , Pierri MK , et al . Cardiac dysfunction in the trastuzumab clinical trials experience . J Clin Oncol 2002. ; 20 ( 5 ): 1215 – 1221 . [DOI] [PubMed] [Google Scholar]
- 4. Fan M , Wang K , Pan D , et al . Radiomic analysis reveals diverse prognostic and molecular insights into the response of breast cancer to neoadjuvant chemotherapy: a multicohort study . J Transl Med 2024. ; 22 ( 1 ): 637 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Liu W , Yang Y , Wang X , et al . A Comprehensive Model Outperformed the Single Radiomics Model in Noninvasively Predicting the HER2 Status in Patients with Breast Cancer . Acad Radiol 2025. ; 32 ( 1 ): 24 – 36 . [DOI] [PubMed] [Google Scholar]
- 6. Portnow LH , Kochkodan-Self JM , Maduram A , et al . Multimodality Imaging Review of HER2-positive Breast Cancer and Response to Neoadjuvant Chemotherapy . RadioGraphics 2023. ; 43 ( 2 ): e220103 . [DOI] [PubMed] [Google Scholar]
- 7. Elias SG , Adams A , Wisner DJ , et al . Imaging features of HER2 overexpression in breast cancer: a systematic review and meta-analysis . Cancer Epidemiol Biomarkers Prev 2014. ; 23 ( 8 ): 1464 – 1483 . [DOI] [PubMed] [Google Scholar]
- 8. Zhang X , Chen X , Fu Y , Zhou H , Lin Y . Study on heterogeneity of vascularity and cellularity via multiparametric MRI habitat imaging in breast cancer . BMC Med Imaging 2025. ; 25 ( 1 ): 159 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Liu Z , Li Z , Qu J , et al . Radiomics of Multiparametric MRI for Pretreatment Prediction of Pathologic Complete Response to Neoadjuvant Chemotherapy in Breast Cancer: A Multicenter Study . Clin Cancer Res 2019. ; 25 ( 12 ): 3538 – 3547 . [DOI] [PubMed] [Google Scholar]
- 10. Huang Y , Cao Y , Chen H , et al . Quantifying tumor morphological complexity based on pretreatment MRI fractal analysis for predicting pathologic complete response and survival in breast cancer: a retrospective, multicenter study . Breast Cancer Res 2025. ; 27 ( 1 ): 86 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Ismail M , Hill V , Statsevych V , et al . Shape Features of the Lesion Habitat to Differentiate Brain Tumor Progression from Pseudoprogression on Routine Multiparametric MRI: A Multisite Study . AJNR Am J Neuroradiol 2018. ; 39 ( 12 ): 2187 – 2193 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Somasundaram E , Litzler A , Wadhwa R , Barker-Clarke R , Scott J . Persistent homology of tumor CT scans is associated with survival in lung cancer . Med Phys 2021. ; 48 ( 11 ): 7043 – 7051 . [DOI] [PubMed] [Google Scholar]
- 13. Xie T , Gong J , Zhao Q , et al . Development and validation of peritumoral vascular and intratumoral radiomics to predict pathologic complete responses to neoadjuvant chemotherapy in patients with triple-negative breast cancer . BMC Med Imaging 2024. ; 24 ( 1 ): 136 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Chen S , Zheng B , Tang W , et al . The longitudinal changes in multiparametric MRI during neoadjuvant chemotherapy can predict treatment response early in patients with HER2-positive breast cancer . Eur J Radiol 2024. ; 178 : 111656 . [DOI] [PubMed] [Google Scholar]
- 15. Hylton NM , Gatsonis CA , Rosen MA , et al. ; ACRIN 6657 Trial Team and I-SPY 1 TRIAL Investigators . Neoadjuvant Chemotherapy for Breast Cancer: Functional Tumor Volume by MR Imaging Predicts Recurrence-free Survival-Results from the ACRIN 6657/CALGB 150007 I-SPY 1 TRIAL . Radiology 2016. ; 279 ( 1 ): 44 – 55 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Saha A , Harowicz MR , Grimm LJ , et al . A machine learning approach to radiogenomics of breast cancer: a study of 922 subjects and 529 DCE-MRI features . Br J Cancer 2018. ; 119 ( 4 ): 508 – 516 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Pennisi A , Kieber-Emmons T , Makhoul I , Hutchins L . Relevance of Pathological Complete Response after Neoadjuvant Therapy for Breast Cancer . Breast Cancer (Auckl) 2016. ; 10 : 103 – 106 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Ghasemi A , Hashtarkhani S , Schwartz DL , Shaban-Nejad A . Explainable artificial intelligence in breast cancer detection and risk prediction: A systematic scoping review . Cancer Innov 2024. ; 3 ( 5 ): e136 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Li X , Zhang N , Hu C , et al . CT-based radiomics signature of visceral adipose tissue for prediction of disease progression in patients with Crohn’s disease: A multicentre cohort study . EClinicalMedicine 2023. ; 56 : 101805 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Le QC , Arimura H , Ninomiya K , Kabata Y . Radiomic features based on Hessian index for prediction of prognosis in head-and-neck cancer patients . Sci Rep 2020. ; 10 ( 1 ): 21301 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Li W , Newitt DC , Gibbs J , et al . Predicting breast cancer response to neoadjuvant treatment using multi-feature MRI: results from the I-SPY 2 TRIAL . NPJ Breast Cancer 2020. ; 6 ( 1 ): 63 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Lin P , Wan WJ , Kang T , et al . Molecular hallmarks of breast multiparametric magnetic resonance imaging during neoadjuvant chemotherapy . Radiol Med 2023. ; 128 ( 2 ): 171 – 183 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Han L , Zhu Y , Liu Z , et al . Radiomic nomogram for prediction of axillary lymph node metastasis in breast cancer . Eur Radiol 2019. ; 29 ( 7 ): 3820 – 3829 . [DOI] [PubMed] [Google Scholar]
- 24. Haodong G , Jianguo Z , Pylypenko D , et al . Ultrafast dynamic contrast-enhanced breast MRI with quantitative perfusion parameters in differentiating breast cancer: a study focusing on triple-negative and HER2 positive breast cancer . Front Oncol 2024. ; 14 : 1457918 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Rinnerthaler G , Egle D , Bartsch R , et al . Neoadjuvant atezolizumab in combination with dual HER2 blockade plus epirubicin in women with early HER2-positive breast cancer: the randomized phase 2 ABCSG-52/ATHENE trial . Nat Cancer 2025. ; 6 ( 1 ): 41 – 50 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Bastola P , Oien DB , Cooley M , Chien J . Emerging Cancer Therapeutic Targets in Protein Homeostasis . AAPS J 2018. ; 20 ( 6 ): 94 . [DOI] [PubMed] [Google Scholar]
- 27. Cui Z , Cong M , Yin S , et al . Role of protein degradation systems in colorectal cancer . Cell Death Discov 2024. ; 10 ( 1 ): 141 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Nandi I , Ji L , Smith HW , et al . Targeting fatty acid oxidation enhances response to HER2-targeted therapy . Nat Commun 2024. ; 15 ( 1 ): 6587 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Gandhi N , Das GM . Metabolic Reprogramming in Breast Cancer and Its Therapeutic Implications . Cells 2019. ; 8 ( 2 ): 89 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Takada M , Toi M . Neoadjuvant treatment for HER2-positive breast cancer . Chin Clin Oncol 2020. ; 9 ( 3 ): 32 . [DOI] [PubMed] [Google Scholar]
- 31. Takada M , Ishiguro H , Nagai S , et al . Survival of HER2-positive primary breast cancer patients treated by neoadjuvant chemotherapy plus trastuzumab: a multicenter retrospective observational study (JBCRG-C03 study) . Breast Cancer Res Treat 2014. ; 145 ( 1 ): 143 – 153 . [DOI] [PubMed] [Google Scholar]
- 32. van Mackelenbergh MT , Loibl S , Untch M , et al. ; CTNeoBC project . Pathologic Complete Response and Individual Patient Prognosis After Neoadjuvant Chemotherapy Plus Anti-Human Epidermal Growth Factor Receptor 2 Therapy of Human Epidermal Growth Factor Receptor 2-Positive Early Breast Cancer . J Clin Oncol 2023. ; 41 ( 16 ): 2998 – 3008 . [DOI] [PubMed] [Google Scholar]
- 33. Dieci MV , Miglietta F , Griguolo G , Guarneri V . Biomarkers for HER2-positive metastatic breast cancer: Beyond hormone receptors . Cancer Treat Rev 2020. ; 88 : 102064 . [DOI] [PubMed] [Google Scholar]
- 34. Jung YJ , Lee S , Kang SK , et al . Clinicopathological Factors Predicting Pathological Complete Response to Neoadjuvant Anti-HER2 Therapy in HER2-Positive Breast Cancer . Oncology 2025. ; 103 ( 5 ): 351 – 359 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35. Zheng G , Hou J , Shu Z , et al . Prediction of neoadjuvant chemotherapy pathological complete response for breast cancer based on radiomics nomogram of intratumoral and derived tissue . BMC Med Imaging 2024. ; 24 ( 1 ): 22 . [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.