Abstract
Background
The application of 21-gene assays in clinical practice is jeopardized by their cost and availability. This study aimed to predict the recurrence score (RS) of a 21-gene assay using MRI peritumoral radiomics in ER+/HER2- breast cancers.
Methods
154 and 39 patients with ER+/HER2- breast cancer from two centers were enrolled, who underwent 21-gene test and preoperative MRI. Patients from Center 1 were divided into training (n = 108) and internal validation (n = 46) cohorts, and patients from Center 2 were enrolled in the external validation cohort. Radiomics features were extracted from the tumoral, peritumoral and dilation volumes of interest with peritumoral ranges of 1 mm, 3 mm, 5 mm, 7 mm, and 9 mm. After feature selection, RS-prediction models were constructed using support vector machine method to distinguish high (RS ≥ 26) from low RS (RS < 26).
Results
As the thickness of the peritumor tissue increased, the AUC of models increased and then decreased, with the 3-mm model performing the best. Among all RS-prediction models, the 3 mm peritumoral model based on T2WI (T2-p3) achieved larger AUCs (0.70 and 0.69 in the internal and external validation cohorts, separately). The peritumoral-fusion model integrating intratumoral radiomic and imaging-clinicopathological features with the T2-p3 model, obtained greater AUCs (0.82 and 0.75 in the internal and external validation cohorts, separately).
Conclusions
MRI peritumoral radiomic data exhibits the potential to serve as a biomarker of recurrence risk in patients with ER+/HER2- breast cancer.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12880-026-02153-1.
Keywords: Artificial intelligence, Breast cancer, Machine learning, Magnetic resonance imaging
Background
With 2.3 million new cases occurring in 2020, breast cancer is the most common cancer worldwide [1], and rational treatment can help reduce recurrence and improve prognosis [2]. With the development of molecular techniques, breast cancer has been shown to be a highly heterogeneous tumor [3], which poses a challenge for accurately determining patient prognosis and developing a rational treatment plan. The expression of 21 selected genes (16 cancer-related genes and 5 reference genes) was detected by a reverse-transcriptase polymerase chain reaction followed by a defined algorithm, the Oncotype Dx 21-gene assay (Genomic Health, Redwood City, CA), and a recurrence score (RS) was generated [4]. Since 2007, the American Society of Clinical Oncology has recommended the use of the 21-gene test as a decision support tool to evaluate the need for postoperative adjuvant chemotherapy in patients with estrogen receptor-positive (ER+) and human epidermal growth factor receptor type 2-negative (HER2-) early breast cancer [5]. However, 21-gene testing is expensive (maximum tariff €4487.02 [6]) and time-consuming, making it difficult to use widely in clinical practice at this time. In addition, 21-gene testing is applied mainly after surgery, which is so late that it can only help guide postoperative adjuvant chemotherapy. If the RS can be obtained preoperatively and the recurrence risk can be accurately assessed before surgery, it may be possible to recommend that some high-risk patients undergo preoperative neoadjuvant therapy [7] to optimize the treatment regimen and improve patient outcomes.
MRI is an indispensable modality for breast imaging practice, and its indications in known malignancies are primarily for preoperative staging and evaluating the response to neoadjuvant therapy [8–10], both of which are relevant to patient outcomes. Traditionally, breast lesions have been analyzed and evaluated according to the Breast Imaging Reporting and Data System (BI-RADS) criteria using a limited number of subjective semantic words, and the accuracy of this assessment is highly dependent on the personal experience of the radiologist. With the rise and development of radiomics, medical image information has been deeply mined, and high-throughput objective features have been used to characterize lesions in detail. An important application scenario for radiomics is the combination of artificial intelligence techniques to characterize tumors or predict their biological behavior [11–13]. Since the tissue surrounding the tumor contains important biological information [14–17], the region of interest for radiomics includes not only the tumor but also the peritumoral area. Several studies have been published on the use of MRI radiomics for predicting the RS, and most of them have analyzed the correlation of intratumoral radiomic features with the RS [7, 18–20]. To the best of our knowledge, there is only one published study on peritumoral radiomics and RS, and it included a limited number of subjects (n = 62) and a limited peritumoral range, even if we exclude the fact only dynamic contrast-enhancement (DCE) sequences and no T2W images were involved [21]. There are still many gaps in knowledge regarding the feasibility of using MRI peritumoral radiomics to predict 21-gene RS in breast cancer patients.
Therefore, the objectives of this study were to extract peritumoral characteristics in different peritumoral areas on dynamic contrast-enhanced and T2W images and to construct prediction models for RS by combining intratumoral, traditional imaging and clinicopathological features.
Methods
Patients
The institutional review board approved this retrospective study (2203252-23), and informed consent was waived. The clinical data of patients who were diagnosed with invasive ER+/HER2- breast cancer and underwent presurgical breast MRI at Fudan University Shanghai Cancer Center (Center 1) between February 2017 and August 2017 and at Peking University Cancer Hospital & Institute (Center 2) between April 2021 and January 2024 were reviewed. Patients who underwent 21-gene testing, breast surgery and consistent MRI of T2W and DCE images (1 plain and 3–5 enhanced phases) at these two centers were screened for enrollment. The exclusion criteria for Center 1 were as follows: missing pathological results (n = 1), MRI performed after minimally invasive surgery (n = 1), poorly visualized lesion on MRI (n = 1), and poor image quality (n = 2). For Center 2, two patients with poorly visualized lesion on MRI were excluded.
The enrolled patients from Center 1 were then arranged in chronological order and categorized into training and internal validation cohorts at a ratio of 7:3, and those from Center 2 was designated as the external validation cohort.
21-gene test
All the women underwent a 21-gene test at Center 1 and Center 2, and an RS was obtained by detecting the expression of 21 genes and performing mathematical operations on the results [4]. The RS ranges from 0 to 100, and the greater the RS, the greater the 10-year recurrence risk and the greater the benefit of chemotherapy. In accordance with the 4.2025 edition of the National Comprehensive Cancer Network (NCCN) clinical practice guidelines for breast cancer (available at NCCN.org), this study adopted RS = 26 as the cutoff value, and the patients were categorized as having low (RS < 26) or high recurrence risk (RS ≥ 26).
MRI protocol
The MRI scans were completed using six different MRI scanners, and the MRI protocol included transverse T2W turbo spin‒echo or short time inversion recovery sequence and volume imaging for breast assessment or 3D gradient-echo volumetric interpolated breath-hold examination dynamic contrast enhancement. During the scan, the patients were placed in the prone position, and their breasts were naturally placed into 16-channel breast-specific coils. For DCE, plane scanning was performed before injection of contrast agent, followed by immediate intravenous injection of contrast agent (Magnevist, Bayer HealthCare Pharmaceuticals Inc., Wayne, USA) at a rate of 1.5-2.0 ml/s and at a dose of 0.1 mmol/kg as well as finally 20 mL of saline. A total of 3–5 enhanced phases were acquired for DCE. The detailed scan parameters are provided in Supplemental Materials(Appendix A in the supplementary material).
MR imaging features
MR images were reviewed and evaluated by two experienced breast radiologists (12 and 7 years of experience in breast radiology) according to the 2013 version of the BI-RADS criteria [22]. These two radiologists were aware of the pathological results of invasive breast cancer but did not know the value of the RS. If they did not agree on their assessment, a third senior radiologist (with more than 20 years of experience in breast imaging) reassessed the images and determined the final result. The detailed MR imaging features are shown in Supplemental Materials(Appendix B in the supplementary material).
Volume of interest (VOI) segmentation
Tumoral VOI segmentation
Two radiologists (with 7 and 3 years of experience) manually segmented the tumoral VOI by outlining all the lesions layer-by-layer on the first phase of postenhancement (CF) images and then made copies on the T2W images by Python. After checking the VOI one by one, manual adjustments were made if the VOI outlining was inaccurate or out of the breast range.
Segmentation of the peritumoral VOI and dilated VOI
Python 3.8 was used to automatically dilate the tumoral VOI contours outward to obtain the dilated VOIs, with dilation ranges of 1 mm, 3 mm, 5 mm, 7 mm, and 9 mm. The tumoral VOI was subtracted from the dilation VOI, and the peritumoral annular region could be obtained (peritumoral VOI). Each peritumoral VOI and dilated VOI was checked one by one, and the VOIs containing areas outside of the mammary range (e.g., the chest wall) were modified and adjusted manually.
Therefore, a total of 11 VOIs were obtained for each tumor on each sequence: the tumor VOI, 5 peritumoral VOIs (1–9 mm annular region around the tumor), and 5 dilated VOIs (tumor + 1–9 mm annular peritumoral region) (Fig. 1). VOI segmentation was performed via itk-SNAP (version 3.8.0, http://www.itksnap.org/) open source software and Python (version 3.8).
Fig. 1.
Volume of interest (VOI) segmentation was performed on the first phase of postenhancement (CF) imaging. (a) CF images showing a 1.9 cm tumor with an irregular margin in the outer quadrant of the left breast; (b) tumoral VOI; (c ~ g) peritumoral VOIs with ranges of 1 mm (p1), 3 mm (p3), 5 mm (p5), 7 mm (p7), and 9 mm (p9); (h ~ l) dilation VOIs included tumor sites and peritumoral regions with ranges of 1 mm (d1), 3 mm (d3), 5 mm (d5), 7 mm (d7), and 9 mm (d9)
Radiomic feature extraction and model construction
Schematic representation of machine learning framework implementation is shown in Fig. 2.
Fig. 2.
Schematic representation of machine learning framework implementation. Firstly, the volume of interest (VOI) was segmented on the first phase of postenhancement (CF) and T2W images. Secondly, after image preprocessing, radiomics features were extracted from original, Laplacian of Gaussian (LoG) and wavelet images. Then, the Pearson correlation coefficient selection method and the least absolute shrinkage and selection operator (LASSO)-based recursive feature elimination (RFE) method were used to reduce the overfitting between features. Finally, the linear support vector machine (SVM) method was used to construct a classifier. GLCM = gray level co-occurrence matrix. GLRLM = gray level run length matrix. GLSZM = gray level size zone matrix. GLDM = gray level dependence matrix. ROC = receiver operating characteristic
Image preprocessing
Since the MRI scans were completed under 6 different scanners with various scanning parameters, the standardization of image signal intensity and the unification of image spatial resolution were performed before extracting radiomic features. (1) The following formula was applied to standardize the image signal intensity.
 |
Inormalize is the standardized image signal intensity, I is the original image signal intensity, and 
and 
are the mean and standard deviation of I, respectively. (2) The spatial resolution of the CF images was resized to 0.9 × 0.9 × 2.2 mm, and the spatial resolution of the T2W images was resized to 1.2 × 1.2 × 7 mm by using the cubic spline interpolation algorithm.
Feature extraction
A total of 1046 radiomic features were extracted from each VOI in each sequence, including 100 original image features, 258 Laplacian of Gaussian (LoG) image features (δ set to 2, 3, 4) and 688 wavelet image features. The LoG and wavelet images were obtained by filtering the original image with LoG and wavelet filters. The bin width was set to 5..
Feature selection
Z score normalization was performed on the features in the training cohort, while features in the internal and external validation cohorts were normalized using the mean and standard deviation of the corresponding features in the training cohort. To reduce the overfitting between features, the Pearson correlation coefficient selection method and recursive feature elimination (RFE) method were used. First, the Pearson correlation coefficient between each pair of features was calculated, and feature pairs with a Pearson correlation coefficient > 0.7 were selected. Then, the average Pearson correlation coefficient between each feature in those pairs and all the other features was calculated, and the feature with the larger average Pearson correlation coefficient was eliminated. For RFE feature selection, the base estimator was the least absolute shrinkage and selection operator (LASSO) algorithm, and α was set to the default value of 1. The minimum and maximum numbers of selected features were set at 1% and 10% of the sample size, respectively, and the feature set with the highest model accuracy was selected to develop the model.
Model construction
Because the distribution of patients in the training cohort was not balanced between the low-risk and high-risk groups in this study, the patients in the training cohort were resampled using the synthetic minority oversampling technique (SMOTE) to improve the robustness of the model. Patients in the internal and external validation cohorts were not resampled. The linear support vector machine (SVM) method was used to construct a model differentiating recurrence risk. All the above data processing steps were performed in Python 3.8.
Statistical analysis
Continuous variables are summarized as the mean ± standard deviation or median (quartiles) according to the normality test, and categorical variables are described with frequencies and percentages. Continuous variables that conformed to a normal distribution were compared for differences between groups using Student’s t test or the Satterthwaite t test, and those that were not normally distributed were compared using the Mann‒Whitney test. The Chi-square test and Fisher’s exact test were performed for categorical variables. The predictive performance of the model was assessed by receiver operating characteristic (ROC) curves. Statistical analysis and data processing were performed with SPSS software (version 25.0) and Python (version 3.8), and P < 0.05 was considered to indicate statistical significance.
Results
Clinical and imaging characteristics
According to the chronological order and 7:3 ratio, patients from Center 1 who underwent preoperative MRI examination from February 2017 to June 2017 were classified as the training cohort, with a total of 108 patients (33 at low risk and 75 at high risk), and patients who underwent MRI from July 2017 to August 2017 were divided into the internal validation cohort, with a total of 46 patients (24 at low risk and 22 at high risk). The external validation cohort included 39 patients, with 13 in the low-risk group and 26 in the high-risk group.
In the training cohort, mass margins were significantly different between high- and low-risk patients (P = 0.013), with the largest proportion of low-risk patients showing spiculated margins (65.6%) and the largest proportion of high-risk patients showing irregular margins (54.2%). In the internal validation cohort, the rate of progesterone receptor positivity was significantly greater in low-risk patients than in high-risk patients (95.8% vs. 72.7%, P = 0.043). Within the external validation cohort, the median Ki-67 index was significantly lower in the low-risk group (10%) than in the high-risk group (30%) (P < 0.001). There were no statistically significant differences in the other clinical or imaging characteristics (Table 1).
Table 1.
Characteristics of patients and imaging
| Characteristics | Training cohort (n = 108) | Internal validation cohort (n = 46) | External validation cohort (n = 39) | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Low recurrence risk (n = 33) | High recurrence risk (n = 75) | P value | Low recurrence risk (n = 24) | High recurrence risk (n = 22) | P value | Low recurrence risk (n = 13) | High recurrence risk (n = 26) | P value | |
| Age | 54.7 ± 9.4a | 51.1 ± 9.0a | 0.069 | 51.2 ± 8.2a | 53.0 ± 8.9a | 0.483 | 44.0 (40.5, 61.5)b | 51.5 (40.3, 58.3)b | 0.929 |
| Invasive malignancy size | 33 | 75 | 0.832 | 24 | 22 | 0.813 | 13 | 26 | 0.172 |
| pTI | 24 (72.7%) | 56 (74.7%) | 15 (62.5%) | 13 (59.1%) | 8 (61.5%) | 9 (34.6%) | |||
| pTII/III | 9 (27.3%) | 19 (25.3%) | 9 (37.5%) | 9 (40.9%) | 5 (38.5%) | 17 (65.4%) | |||
| Invasive malignancy grade | 33 | 75 | 1.000 | 24 | 22 | 0.336 | 13 | 26 | 0.538 |
| Grade I/II | 29 (87.9%) | 65 (86.7%) | 23 (95.8%) | 19 (86.4%) | 13 (100%) | 23 (88.5%) | |||
| Grade III | 4 (12.1%) | 10 (13.3%) | 1 (4.2%) | 3 (13.6%) | 0 (0) | 3 (11.5%) | |||
| Progesterone receptor positive | 33 (100%) | 68 (90.7%) | 0.098 | 23 (95.8%) | 16 (72.7%) | 0.043 | 13 (100%) | 24 (92.3%) | 0.544 |
| Ki76 (%) | 20.0 (10.0, 20.0)b | 20.0 (10.0, 30.0)b | 0.317 | 15.0 (11.3, 23.8)b | 20.0 (10.0, 30.0)b | 0.413 | 10.0 (10.0, 15.0)b | 30.0 (15.0, 50.0)b | < 0.001 |
| Positive lymph node | 6 (18.2%) | 9 (12.0%) | 0.580 | 3 (12.5%) | 8 (36.4%) | 0.058 | 3 (23.1%) | 6 (23.1%) | 1.000 |
| Amount of fibroglandular tissue | 33 | 75 | 0.215 | 24 | 22 | 0.391 | 13 | 26 | 1.000 |
| a. Almost entirely fat | 2 (6.1%) | 0 (0) | 1 (4.2%) | 0 (0) | 0 (0) | 1 (3.8) | |||
| b. Scattered fibroglandular tissue | 5 (15.2%) | 11 (14.7%) | 2 (8.3%) | 5 (22.7%) | 2 (15.4%) | 4 (15.4) | |||
| c. Heterogeneous fibroglandular tissue | 22 (66.7%) | 50 (66.7%) | 18 (75.0%) | 13 (59.1%) | 9 (69.2%) | 18 (69.2) | |||
| d. Extreme fibroglandular tissue | 4 (12.1%) | 14 (18.7%) | 3 (12.5%) | 4 (18.2%) | 2 (15.4%) | 3 (11.5%) | |||
| Degree of background parenchymal enhancement | 33 | 75 | 0.853 | 24 | 22 | 0.597 | 13 | 26 | 0.603 |
| Minimal | 19 (57.6%) | 47 (62.7%) | 13 (54.2%) | 10 (45.5%) | 8 (61.5%) | 19 (73.1%) | |||
| Mild | 11 (33.3%) | 21 (28.0%) | 7 (29.2%) | 9 (40.9%) | 3 (23.1%) | 5 (19.2%) | |||
| Moderate | 3 (9.1%) | 7 (9.3%) | 4 (16.7%) | 2 (9.1%) | 2 (15.4%) | 1 (3.8%) | |||
| Marked | 0 (0) | 0 (0) | 0 (0) | 1 (4.5%) | 0 (0) | 1 (3.8%) | |||
| Long diameter | 17.0 (12.5, 22.0)b | 17.0 (13.0, 21.0)b | 0.686 | 17.5 (14.3, 25.5)b | 19.6 ± 7.8a | 0.904 | 17.8 ± 8.3a | 22.0 ± 7.3a | 0.111 |
| Short diameter | 15.0 (10.0, 16.0)b | 12.0 (10.0, 17.0)b | 0.460 | 14.0 ± 5.7a | 14.3 ± 4.7a | 0.882 | 11.0 (9.0, 15.5)b | 14.0 (12.0, 19.3)b | 0.107 |
| MRI finding | 33 | 75 | 1.000 | 24 | 22 | 1.000 | 13 | 26 | 0.589 |
| Mass | 32 (97.0%) | 72 (96.0%) | 21 (87.5%) | 19 (86.4%) | 11 (84.6%) | 24 (92.3%) | |||
| NME | 1 (3.0%) | 3 (4.0%) | 3 (12.5%) | 3 (13.6%) | 2 (15.4%) | 2 (7.7%) | |||
| Mass shape | 32 | 72 | 0.424 | 21 | 19 | 0.385 | 11 | 24 | 0.368 |
| Round | 2 (6.3%) | 3 (4.2%) | 0 (0) | 0 (0) | 1 (9.1%) | 1 (4.2%) | |||
| Oval | 6 (18.8%) | 23 (31.9%) | 7 (33.3%) | 4 (21.1%) | 4 (36.3%) | 5 (20.8%) | |||
| Irregular | 24 (75.0%) | 46 (63.9%) | 14 (66.7%) | 15 (78.9%) | 6 (54.5%) | 18 (75.0%) | |||
| Mass margin | 32 | 72 | 0.013 | 21 | 19 | 0.210 | 11 | 24 | 0.633 |
| Irregular | 8 (25.0%) | 39 (54.2%) | 8 (38.1%) | 11 (57.9%) | 4 (36.4%) | 13 (54.2%) | |||
| Circumscribed | 3 (9.4%) | 3 (4.2%) | 0 (0) | 0 (0) | 0 (0) | 1 (4.2%) | |||
| Spiculated | 21 (65.6%) | 30 (41.7%) | 13 (61.9%) | 8 (42.1%) | 7 (63.6%) | 10 (41.7%) | |||
| Mass internal enhancement | 32 | 72 | 0.637 | 21 | 19 | 1.000 | 11 | 24 | 0.64 |
| Heterogeneous | 24 (75.0%) | 57 (79.2%) | 18 (85.7%) | 17 (89.5%) | 9 (81.8%) | 21 (87.5%) | |||
| Rim enhancement | 8 (25.0%) | 15 (20.8%) | 3 (14.3%) | 2 (10.5%) | 2 (18.2%) | 3 (12.5%) | |||
| Mass with NME | 2 (6.3%) | 3 (4.2%) | 0.642 | 2 (9.5%) | 4 (21.1%) | 0.398 | 2 (18.2%) | 3 (12.5%) | 1.000 |
| NME distribution | 1 | 3 | NA | 3 | 3 | 1.000 | 2 | 2 | 1.000 |
| Segmental | 1 (100%) | 3 (100%) | 2 (66.7%) | 1 (33.3%) | 0 (0) | 1 (50.0%) | |||
| Regional or multiple regions | 0 (0) | 0 (0) | 1 (33.3%) | 2 (66.7%) | 2 (100%) | 1 (50.0%) | |||
| NME internal enhancement | 1 | 3 | 1.000 | 3 | 3 | 1.000 | 2 | 2 | NA |
| Heterogeneous | 0 (0) | 1 (33.3%) | 2 (66.7%) | 3 (100%) | 2 (100%) | 2 (100%) | |||
| Clumped or clustered ring | 1 (100%) | 2 (66.7%) | 1 (33.3%) | 0 (0) | 0 (0) | 0 (0) | |||
| Initial phase of kinetic curve | 33 | 74 | 0.688 | 24 | 22 | 0.063 | 13 | 26 | 0.12 |
| Fast | 22 (66.7%) | 48 (64.9%) | 11 (45.8%) | 4 (18.2%) | 7 (54.8%) | 20 (76.9%) | |||
| Medium | 10 (30.3%) | 25 (33.8%) | 13 (54.2%) | 17 (77.3%) | 4 (30.8%) | 6 (23.1%) | |||
| Slow | 1 (3.0%) | 1 (1.4%) | 0 (0) | 1 (4.5%) | 2 (15.4%) | 0 (0) | |||
| Delayed phase of kinetic curve | 33 | 74 | 0.428 | 24 | 22 | 0.608 | 13 | 26 | 0.518 |
| Persistent | 2 (6.1%) | 2 (2.7%) | 0 (0) | 0 (0) | 2 (15.4%) | 1 (3.8%) | |||
| Plateau type | 7 (21.2%) | 23 (31.1%) | 4 (16.7%) | 6 (27.3%) | 4 (30.8%) | 9 (34.6%) | |||
| Washout | 24 (72.7%) | 49 (66.2%) | 20 (83.3%) | 16 (72.7%) | 7 (54.8%) | 16 (61.5%) | |||
| Assessment categories | 33 | 75 | 0.968 | 24 | 22 | 0.769 | 13 | 26 | 0.459 |
| 4 A | 1 (3.0%) | 3 (4.0%) | 0 (0) | 0 (0) | 0 (0) | 1 (3.8%) | |||
| 4B | 9 (27.3%) | 19 (25.3%) | 4 (16.7%) | 5 (22.7%) | 4 (30.8%) | 4 (15.4%) | |||
| 4 C | 17 (51.5%) | 37 (49.3%) | 16 (66.7%) | 12 (54.5%) | 7 (53.8%) | 12 (46.2%) | |||
| 5 | 6 (18.2%) | 16 (21.3%) | 4 (16.7%) | 5 (22.7%) | 2 (15.4%) | 9 (34.6%) | |||
aData are means ± standard deviation, and bdata are median with interquartile range in parentheses. Bold italic values mean statistically significant. NME = non-mass enhancement
Performance of the peritumoral model
The performances of the 10 peritumoral models are shown in Fig. 3a; Table 2, and the results are as follows:
Fig. 3.
(a) The area under the receiver operating characteristic curve (AUC) of peritumoral models with peritumoral regions of 1 mm (p1), 3 mm (p3), 5 mm (p5), 7 mm (p7), and 9 mm (p9) in the internal and external validation cohorts for distinguishing high-risk patients (RS ≥ 26) from low-risk patients (RS < 26). (b) Features selected by the 3 mm peritumoral model on T2W images (T2-p3)
Table 2.
Performance of peritumoral models
| Peritumoral model |
Training Cohort | Internal validation cohort | External validation cohort | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| AUC (95% CI) | Accuracy | Sensitivity | Specificity | AUC (95% CI) | Accuracy | Sensitivity | Specificity | AUC (95% CI) | Accuracy | Sensitivity | Specificity | |
| CF-p1 | 0.68 (0.60–0.77) | 0.65 | 0.71 | 0.60 | 0.66 (0.49–0.81) | 0.61 | 0.86 | 0.38 | 0.66 (0.47–0.83) | 0.69 | 0.85 | 0.38 |
| CF-p3 | 0.95 (0.90–0.98) | 0.87 | 0.84 | 0.89 | 0.68 (0.50–0.84) | 0.61 | 0.77 | 0.46 | 0.68 (0.49–0.84) | 0.69 | 0.69 | 0.69 |
| CF-p5 | 0.81 (0.74–0.88) | 0.77 | 0.85 | 0.69 | 0.66 (0.48–0.80) | 0.61 | 0.77 | 0.46 | 0.63 (0.43–0.83) | 0.64 | 0.69 | 0.54 |
| CF-p7 | 0.93 (0.89–0.97) | 0.87 | 0.83 | 0.91 | 0.65 (0.49–0.81) | 0.65 | 0.68 | 0.63 | 0.62 (0.42–0.82) | 0.67 | 0.65 | 0.69 |
| CF-p9 | 0.94 (0.89–0.98) | 0.84 | 0.75 | 0.93 | 0.64 (0.45–0.80) | 0.63 | 0.64 | 0.63 | 0.62 (0.42–0.81) | 0.64 | 0.77 | 0.38 |
| T2-p1 | 0.73 (0.65–0.81) | 0.69 | 0.53 | 0.84 | 0.67 (0.50–0.81) | 0.63 | 0.68 | 0.58 | 0.67 (0.44–0.82) | 0.62 | 0.58 | 0.69 |
| T2-p3 | 0.83 (0.75–0.89) | 0.76 | 0.73 | 0.79 | 0.70 (0.52–0.85) | 0.78 | 0.91 | 0.67 | 0.69 (0.49–0.86) | 0.72 | 0.81 | 0.54 |
| T2-p5 | 0.83 (0.76–0.90) | 0.78 | 0.76 | 0.80 | 0.66 (0.47–0.81) | 0.67 | 0.64 | 0.71 | 0.63 (0.41–0.83) | 0.64 | 0.69 | 0.54 |
| T2-p7 | 0.89 (0.83–0.94) | 0.78 | 0.76 | 0.80 | 0.65 (0.47–0.81) | 0.65 | 0.68 | 0.63 | 0.63 (0.43–0.80) | 0.59 | 0.58 | 0.62 |
| T2-p9 | 0.84 (0.77–0.90) | 0.75 | 0.68 | 0.83 | 0.65 (0.48–0.80) | 0.61 | 0.64 | 0.58 | 0.62 (0.43–0.79) | 0.64 | 0.73 | 0.46 |
CF-p1 ～ CF-p9: peritumoral models based on features extracted from 1 mm (p1), 3 mm (p3), 5 mm (p5), 7 mm (p7), 9 mm (p9) peritumoral ranges on the first-enhanced (CF) phases of dynamic contrast-enhancement (DCE) images
T2-p1 ～ T2-p9: peritumoral models based on features extracted from 1 mm (p1), 3 mm (p3), 5 mm (p5), 7 mm (p7), 9 mm (p9) peritumoral ranges on the T2W images
AUC = area under the curve, CI = confidence interval
(1) As the peritumoral range gradually increased, the areas under the ROC (AUCs) of both the CF- and T2WI-based models in the internal and external validation cohorts exhibited similar trends: the AUC of the 3 mm peritumoral model was greater than that of the 1 mm peritumoral model, and the performance of the peritumoral model with a peritumoral range of 3 mm to 9 mm gradually decreased or stabilized (Fig. 3a).
(2) Among the 10 peritumoral models evaluated in both internal and external validation cohorts, the highest AUC was for the 3 mm peritumoral model on T2WI (T2-p3), with AUC, accuracy, sensitivity, and specificity values of 0.70 (95% CI: 0.52–0.85) and 0.69 (95% CI: 0.49–0.86), 0.78 and 0.72, 0.91 and 0.81, and 0.67 and 0.54, respectively, incorporating 4 LoG features (Fig. 3b, Supplemental Materials (Appendix C in the supplementary material)).
(3) The optimal peritumoral model (T2-p3) was selected and fused with intratumoral and clinicopathologic features to construct a peritumoral fusion model. In the internal and external validation cohorts, the AUCs of the peritumoral fusion model was 0.82 and 0.75, with accuracy, sensitivity, and specificity values of 0.72 and 0.74, 0.91 and 0.81, and 0.54 and 0.62, respectively, which were greater than those of the optimal peritumoral model (T2-p3) and the intratumor + clinical imaging model (CF + T2 + Clinical) (Fig. 4). The peritumoral fusion model included two features from 3 mm peritumoral tissue on T2WI (T2-p3-log-sigma-4-0-mm-3D_glcm_Correlation, T2-p3-wavelet-HLH_glcm_ClusterShade), one intratumoral feature on CF (CF-wavelet-LLH_glcm_ClusterShade), and 3 clinical imaging features (mass margin, non-mass enhancement internal enhancement and age) (Supplemental Materials(Appendix C in the supplementary material)).
Fig. 4.
Receiver operating characteristic curve of the optimal peritumoral model (T2-p3), the intratumor + clinical-imaging model (CF + T2 + Clinical-imaging) and the peritumoral fusion model
Performance of the dilation model
The predictive performance of 10 dilation models is summarized in Fig. 5a; Table 3, and the results are summarized as follows:
Fig. 5.
(a) The area under the receiver operating characteristic curve (AUC) of dilation models with ranges of 1 mm (d1), 3 mm (d3), 5 mm (d5), 7 mm (d7), and 9 mm (d9) for distinguishing high-risk patients (RS ≥ 26) from low-risk patients (RS < 26) in the internal and external validation cohorts. (b) Features selected by the 3 mm dilation model on T2W images (T2-d3)
Table 3.
Performance of dilation models
| Dilation model | Training Cohort | Internal validation cohort | External validation cohort | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| AUC (95% CI) | Accuracy | Sensitivity | Specificity | AUC (95% CI) | Accuracy | Sensitivity | Specificity | AUC (95% CI) | Accuracy | Sensitivity | Specificity | |
| CF-d1 | 0.89 (0.82–0.94) | 0.81 | 0.75 | 0.88 | 0.70 (0.54–0.83) | 0.59 | 0.91 | 0.29 | 0.70 (0.52–0.85) | 0.64 | 0.65 | 0.62 |
| CF-d3 | 0.87 (0.80–0.93) | 0.82 | 0.84 | 0.80 | 0.72 (0.57–0.86) | 0.63 | 0.86 | 0.42 | 0.70 (0.52–0.87) | 0.69 | 0.65 | 0.77 |
| CF-d5 | 0.82 (0.75–0.88) | 0.75 | 0.63 | 0.88 | 0.70 (0.54–0.84) | 0.59 | 0.55 | 0.63 | 0.68 (0.45–0.87) | 0.79 | 0.96 | 0.46 |
| CF-d7 | 0.73 (0.65–0.80) | 0.67 | 0.57 | 0.77 | 0.67 (0.51–0.82) | 0.59 | 0.45 | 0.71 | 0.65 (0.46 0.83) | 0.67 | 0.69 | 0.62 |
| CF-d9 | 0.92 (0.86–0.96) | 0.83 | 0.81 | 0.85 | 0.66 (0.51–0.83) | 0.65 | 0.64 | 0.67 | 0.64 (0.45–0.83) | 0.69 | 0.85 | 0.38 |
| T2-d1 | 0.77 (0.70–0.85) | 0.70 | 0.67 | 0.73 | 0.71 (0.55–0.86) | 0.72 | 0.82 | 0.63 | 0.70 (0.50–0.86) | 0.67 | 0.88 | 0.23 |
| T2-d3 | 0.91 (0.85–0.96) | 0.87 | 0.79 | 0.95 | 0.73 (0.58–0.88) | 0.70 | 0.86 | 0.54 | 0.72 (0.52–0.88) | 0.67 | 0.69 | 0.62 |
| T2-d5 | 0.89 (0.83–0.94) | 0.84 | 0.84 | 0.84 | 0.71 (0.54–0.85) | 0.67 | 0.59 | 0.75 | 0.70 (0.51–0.87) | 0.59 | 0.54 | 0.69 |
| T2-d7 | 0.72 (0.64–0.80) | 0.66 | 0.75 | 0.57 | 0.68 (0.51–0.83) | 0.61 | 0.59 | 0.63 | 0.66 (0.46–0.85) | 0.74 | 0.88 | 0.46 |
| T2-d9 | 0.80 (0.72–0.86) | 0.64 | 0.69 | 0.59 | 0.66 (0.48–0.82) | 0.63 | 0.73 | 0.54 | 0.66 (0.45–0.85) | 0.59 | 0.62 | 0.54 |
CF-d1 ～ CF-d9: dilation models based on features extracted from 1 mm (d1), 3 mm (d3), 5 mm (d5), 7 mm (d7), 9 mm (d9) dilation ranges on the first-enhanced (CF) phases of dynamic contrast-enhancement (DCE) images
T2-d1 ～ T2-d9: dilation models based on features extracted from 1 mm (d1), 3 mm (d3), 5 mm (d5), 7 mm (d7), 9 mm (d9) dilation ranges on the T2W images.
AUC = area under the curve, CI = confidence interval
(1) With the gradual increase in the dilation range, the performances of the CF-based model and T2WI-based model showed the same trend. From the 1 mm to 3 mm dilation region, the model efficacy showed an increasing or equal trend and reached a maximum in the 3 mm dilation region, and from the 3 mm to 9 mm dilation region, the model efficacy showed a decreasing or equal trend (Fig. 5a).
(2) Among the 10 dilation models, the one constructed on T2W images within the 3 mm dilation region (T2-d3) had relatively greater efficacy in the internal and external validation cohorts, with AUC, accuracy, sensitivity, and specificity values of 0.73 (95% CI: 0.58–0.88) and 0.72 (95% CI: 0.52–0.88), 0.70 and 0.67, 0.86 and 0.69, and 0.54 and 0.62, respectively. A total of 9 radiomic features were selected for the T2-d3 model, including 8 wavelet features and one LoG feature (Fig. 5b, Supplemental Materials (Appendix D in the supplementary material)).
(3) Incorporating clinical imaging features into the optimal dilation model resulted in a dilation fusion model with an AUC of 0.75 and 0.74 in the internal and external validation cohorts respectively, which was significantly greater than that of the clinical imaging model (AUCs = 0.60 and 0.59) (Fig. 6). The dilation fusion model incorporated nine T2-d3 features (T2-d3-wavelet-LLH_firstorder_Mean, T2-d3-wavelet-HLL_firstorder_Kurtosis, T2-d3-wavelet-HLL_firstorder_Skewness, T2-d3-wavelet-HLH_firstorder_Kurtosis, T2-d3-wavelet-HLH_firstorder_Skewness, T2-d3-wavelet-HHL_firstorder_Mean, T2-d3-wavelet-HHH_glcm_Idn, T2-d3-wavelet-LLL_glcm_Idn and T2-d3-log-sigma-2-0-mm-3D_glcm_Idn) and the margin of the mass (Supplemental Materials (Appendix D in the supplementary material)).
Fig. 6.
Receiver operating characteristic curve of the optimal dilation model (T2-d3), the clinical-imaging model and the dilation-fusion model
Comparison of the peritumoral model and dilation model
A comparison of the peritumoral model and the dilation model showed that given the same sequence and the same peritumoral range, the dilation model performed better than the corresponding peritumoral model (Table 2 vs. Table 3; Fig. 3a vs. Fig. 5a).
Discussion
The NCCN guidelines recommend the use of 21-gene tests to help predict recurrence risk and assist in the clinical development of postoperative treatment strategies for early-stage breast cancer. However, the 21-gene test is expensive, time-consuming, and not currently available preoperatively. In this study, multiple RS prediction models were developed using peritumoral radiomics features and machine learning techniques, which demonstrated the potential value of peritumoral radiomics in predicting recurrence risk. In addition, the present study showed that the peritumoral region may influence the predictive performance, and the model incorporating 3 mm of peritumoral information performed best. This study also suggested the integration of peritumoral features, internal tumor features, and clinical imaging features to comprehensively characterize lesions to improve their predictive ability.
The present study showed that models constructed with CF and T2W peritumoral features were potentially valuable for predicting recurrence risk, with AUCs ranging from 0.62 to 0.70 in the validation cohorts. Chiacchiaretta et al. predicted the RS with a maximum AUC of 0.76 using a partial least square model constructed with peritumoral and intratumor radiomic features on early and late DCE images [21], and their study showed that the peritumoral features on DCE images could be able to reflect recurrence risk, which is consistent with our study. Some studies have shown that the features extracted from peritumoral tissue via DCE could be helpful in predicting sentinel lymph node metastasis [23] and assessing the response to neoadjuvant therapy in breast cancer patients [17, 24]. This may be because, on the one hand, the peritumoral environment secretes a large number of growth factors and cytokines that induce hypoxia and angiogenesis, and these physiopathological processes play important roles in tumor development, progression or metastasis [25]. On the other hand, peritumoral lymphovascular infiltration is also associated with tumor aggressiveness, recurrence and metastasis [26]– [27]. The features of peritumoral DCE imaging could characterize the vascularity of peritumoral tissues and therefore can reflect tumor biological behavior and patient prognosis to some extent. To the best of our knowledge, only one recently published study has investigated the association between T2W peritumoral characteristics and RS [28], and its findings are largely consistent with our results. It has been shown that tumors with peritumoral edema on T2W images may be larger in size, have a greater Ki-67 index, exhibit more neovascularization or angiolymphatic invasion, and thus may be more invasive and have a poorer prognosis [16, 29–32], which is somewhat supported by the present study.
This study not only promoted the value of the peritumoral radiomics but also further analyzed the difference in value for different peritumoral ranges. The predictive effectiveness showed an overall trend of increasing and then decreasing as the peritumoral range increased, a trend that was consistent across both the peritumoral and dilation models based on CF and T2WI sequences. This suggests that the peritumoral range may affect the predictive performance and that 3 mm of peritumoral tissue is relatively important for predicting patient prognosis, indicating that the vascularity, peritumoral edema, tissue microstructure, and pathophysiological process within a 3 mm peritumoral environment vary to a relatively substantial degree among different patient subgroups. We hypothesize that this trend may be because the 1 mm peritumoral region is relatively small compared to the 3 mm peritumoral region and contains limited information that affects the model’s effectiveness, whereas the 5 mm ~ 9 mm peritumoral regions contain “normal” tissue farther from the tumor, reducing interlesion variability and thus affecting the model’s effectiveness. In predicting the RS, Chiacchiaretta et al. [21] only extracted features within 4 mm of surrounding tissues (2 mm of inward erosion and 2 mm of outward dilation of the tumor) and did not further compare the value of peritumoral tissue larger than 4 mm. When investigating the value of peritumoral tissue in predicting breast cancer prognosis, many previous studies included only one peritumoral area [25, 33–35], and the peritumoral areas used in different studies were not completely consistent. Therefore, the design protocol of the present study was more in-depth and refined in terms of exploring the peritumoral area.
When intratumor and clinicopathologic features were included in the optimal peritumoral model or clinicopathologic information was included in the optimal dilation model, the efficacy of the fusion model was greater than that of the prefusion model, suggesting that there may be complementary information on these features. A comparison of the peritumoral and dilation models showed that the efficacy of the dilation model was greater than or equal to that of the corresponding peritumoral model with the same sequence and peritumoral region. This may be because the dilated VOI range covered not only the peri-tumor VOI region but also the internal tumor tissue, providing more comprehensive information and therefore greater predictive efficacy. This further confirmed that there may be information complementarity between internal tumor characteristics and peritumor characteristics. There is a large body of literature supporting this viewpoint in the prediction of breast cancer outcomes. Li et al. extracted pretreatment DCE radiomics features from 448 patients to construct a model to predict the response to neoadjuvant therapy, and their results showed that the combined intratumoral + peritumoral classifier (AUC = 0.92) outperformed both intratumoral (AUC = 0.89) and peritumoral (AUC = 0.78) models [36]. Jiang et al. conducted a study to determine whether a radiomics approach could predict lymphovascular invasion and showed that combining intratumoral and peritumoral features improved the model’s performance on DCE, DWI, or combined sequences [35].
There are several limitations to this study. First, although two different methods (the standardization of image signal intensity and the unification of image spatial resolution) were used to minimize the effect of different scanning parameters on the images and to maximize the robustness of the constructed models, related image differences cannot be completely avoided. Research on the effect of scanning parameters on model construction is needed. Second, although this retrospective study incorporated patients from two independent centers, the sample size remained limited, especially compared with the large number of radiomic features, possibly limiting the effectiveness of the machine learning model to some extent. Finally, although the peritumoral region was automatically generated by Python, the tumor VOI was still manually outlined, which involved a large amount of work. Although manual segmentation of the VOI tends to achieve high reproducibility, interobserver analysis was not performed in this study. In the future, fully automated or semiautomated VOI segmentation methods, together with automatic correction techniques, will help radiologists substantially improve the efficiency of VOI segmentation.
Conclusions
In conclusion, peritumoral radiomic features on breast MR images are potentially valuable for assessing recurrence risk in ER+/HER2- breast cancer patients. The peritumoral range may affect prediction performance, with the best performance occurring in the 3 mm peritumoral area. We recommend a comprehensive analysis of intratumoral, peritumoral, and clinicopathological information when assessing a patient’s recurrence risk.
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements
The authors thank Caixia Fu (MR Application Development, Siemens Shenzhen Magnetic Resonance Ltd., Shenzhen, China) for her suggestions on image preprocessing.
Abbreviations
- ADC
Apparent diffusion coefficient
- AUC
Area under the curve
- BI-RADS
Breast Imaging Reporting and Data System
- CF
First phase of postenhancement
- DCE
Dynamic contrast-enhancement
- DCIS
Ductal carcinoma in situ
- ER
Estrogen receptor
- GLCM
Gray level co-occurrence matrix
- GLDM
Gray level dependence matrix
- GLRLM
Gray level run length matrix
- GLSZM
Gray level size zone matrix
- HER2
Human epidermal growth factor receptor type 2
- LASSO
Least absolute shrinkage and selection operator
- LoG
Laplacian of Gaussian
- NCCN
National Comprehensive Cancer Network
- RFE
Recursive feature elimination
- ROC
Receiver operating characteristic curve
- RS
Recurrence score
- SMOTE
Synthetic minority oversampling technique
- SVM
Support vector machine
- VOI
Volume of interest
Author contributions
Conception and design: YC, LY, YS, WT, YG and WP; data collection and aggregation: YC, LY, YH, WT, QX, LZ, RL, and QW; technical support: YC, LY, LX, LZ, RL and TX; verification of the underlying data: YC, LY, YH, LX, QX, TX, YS, YG, WP and WT; development of methodology: YC, LY, YH, LX, QX, YS, WT, YG and WP; data analysis and interpretation: YC, LY, YH, LX, TX, YS, WT, YG and WP; original draft writing: YC, YH, QX and LX; approval of final version manuscript: all authors.
Funding
This study has received funding by National Natural Science Foundation of China (NSFC 82071878), Shanghai Science and Technology Innovation Action Plan Medical Innovation Research Project (21Y11910200), and Shanghai Anticancer Association EAGLET PROJECT (SACA-CY21C04).
Data availability
Due to the privacy of patients, the clinical data related to patients and MRI images cannot be available for public access but can be obtained from the corresponding author on reasonable request.
Declarations
Ethics approval and consent to participate
Our study complied with the Declaration of Helsinki. This study was approved by the Ethics Committee of Fudan University Shanghai Cancer Center (2203252-23), and informed consent was waived owing to the retrospective nature of this study.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Yang Chen, Liang You and Yan Huang contributed equally to this work and are co-first authors for this study.
Contributor Information
Yingshi Sun, Email: sys27@163.com.
Wei Tang, Email: tangwei105@163.com.
Yajia Gu, Email: guyajia@shca.org.cn.
Weijun Peng, Email: weijun_2020@163.com.
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(3):209–49. [DOI] [PubMed] [Google Scholar]
- 2.Kamangar F, Dores GM, Anderson WF. Patterns of cancer incidence, mortality, and prevalence across five continents: defining priorities to reduce cancer disparities in different geographic regions of the world. J Clin Oncol. 2006;24(14):2137–50. [DOI] [PubMed] [Google Scholar]
- 3.Yersal O, Barutca S. Biological subtypes of breast cancer: prognostic and therapeutic implications. World J Clin Oncol. 2014;5(3):412–24. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Paik S, Shak S, Tang G, et al. A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer. New Engl J Med. 2004;351(27):2817–26. [DOI] [PubMed] [Google Scholar]
- 5.Harris L, Fritsche H, Mennel R, et al. American society of clinical oncology 2007 update of recommendations for the use of tumor markers in breast cancer. J Clin Oncol. 2007;25(33):5287–312. [DOI] [PubMed] [Google Scholar]
- 6.de Jongh FE, Efe R, Herrmann KH, Spoorendonk JA. Cost and clinical benefits associated with oncotype DX(R) test in patients with Early-Stage HR+/HER2- Node-Negative breast cancer in the Netherlands. Int J Breast Cancer. 2022;2022:5909724. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Romeo V, Cuocolo R, Sanduzzi L, et al. MRI radiomics and machine learning for the prediction of oncotype Dx recurrence score in invasive breast cancer. Cancers (Basel). 2023;15(6):1840. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Mann RM, Cho N, Moy L, Breast MRI. State of the Art. Radiology. 2019;292(3):520–36. [DOI] [PubMed] [Google Scholar]
- 9.Mann RM, Kuhl CK, Kinkel K, Boetes C. Breast MRI: guidelines from the European society of breast imaging. Eur Radiol. 2008;18(7):1307–18. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Sardanelli F, Boetes C, Borisch B, et al. Magnetic resonance imaging of the breast: recommendations from the EUSOMA working group. Eur J Cancer. 2010;46(8):1296–316. [DOI] [PubMed] [Google Scholar]
- 11.Bhinder B, Gilvary C, Madhukar NS, Elemento O. Artificial intelligence in cancer research and precision medicine. Cancer Discov. 2021;11(4):900–15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Bi WL, Hosny A, Schabath MB, et al. Artificial intelligence in cancer imaging: clinical challenges and applications. CA Cancer J Clin. 2019;69(2):127–57. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Davey MG, Davey MS, Boland MR, et al. Radiomic differentiation of breast cancer molecular subtypes using pre-operative breast imaging - A systematic review and meta-analysis. Eur J Radiol. 2021;144:109996. [DOI] [PubMed] [Google Scholar]
- 14.Santucci D, Faiella E, Gravina M, et al. CNN-Based approaches with different tumor bounding options for lymph node status prediction in breast DCE-MRI. Cancers (Basel). 2022;14:19. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Musall BC, Adrada BE, Candelaria RP, et al. Quantitative apparent diffusion coefficients from peritumoral regions as early predictors of response to neoadjuvant systemic therapy in Triple-Negative breast cancer. J Magn Reson Imaging. 2022;56(6):1901–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Kwon BR, Shin SU, Kim SY, et al. Microcalcifications and peritumoral edema predict survival outcome in luminal breast cancer treated with neoadjuvant chemotherapy. Radiology. 2022;304(2):310–9. [DOI] [PubMed] [Google Scholar]
- 17.Braman N, Prasanna P, Whitney J, et al. Association of peritumoral radiomics with tumor biology and pathologic response to preoperative targeted therapy for HER2 (ERBB2)-Positive breast cancer. JAMA Netw Open. 2019;2(4):e192561. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Fan M, Cui Y, You C, et al. Radiogenomic signatures of oncotype DX recurrence score enable prediction of survival in Estrogen Receptor-Positive breast cancer: A multicohort study. Radiology. 2022;302(3):516–24. [DOI] [PubMed] [Google Scholar]
- 19.Kim JY, Kim JJ, Hwangbo L, et al. Diffusion-weighted MRI of Estrogen receptor-positive, HER2-negative, node-negative breast cancer: association between intratumoral heterogeneity and recurrence risk. Eur Radiol. 2020;30(1):66–76. [DOI] [PubMed] [Google Scholar]
- 20.Jacobs MA, Umbricht CB, Parekh VS, et al. Integrated multiparametric radiomics and informatics system for characterizing breast tumor characteristics with the oncotypedx gene assay. Cancers (Basel). 2020;12(10):2772. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Chiacchiaretta P, Mastrodicasa D, Chiarelli AM et al. MRI-Based radiomics approach predicts tumor recurrence in ER + /HER2 - Early breast cancer patients. J Digit Imaging. 2023;36(3):1071-1080. [DOI] [PMC free article] [PubMed]
- 22.Morris EA, Comstock CE, Lee CH, et al. ACR BI-RADS® magnetic resonance imaging. ACR BI-RADS® atlas, breast imaging reporting and data system. Reston, VA: American College of Radiology; 2013. [Google Scholar]
- 23.Liu C, Ding J, Spuhler K, et al. Preoperative prediction of Sentinel lymph node metastasis in breast cancer by radiomic signatures from dynamic contrast-enhanced MRI. J Magn Reson Imaging. 2019;49(1):131–40. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Braman NM, Etesami M, Prasanna P, et al. 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(1):57. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Han X, Cao W, Wu L, Liang C. Radiomics assessment of the tumor immune microenvironment to predict outcomes in breast cancer. Front Immunol. 2021;12:773581. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Lauria R, Perrone F, Carlomagno C, et al. The prognostic value of lymphatic and blood vessel invasion in operable breast cancer. Cancer. 1995;76(10):1772–8. [DOI] [PubMed] [Google Scholar]
- 27.Britto AV, Schenka AA, Moraes-Schenka NG, et al. Immunostaining with D2-40 improves evaluation of lymphovascular invasion, but May not predict Sentinel lymph node status in early breast cancer. BMC Cancer. 2009;9:109. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Chen Y, Xie L, Tang W, et al. MRI radiomics signatures of 21-Gene recurrence score for predicting survival in ER+/HER2- breast cancer. Cancer Med. 2025;14(17):e71172. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Bae MS, Shin SU, Ryu HS, et al. Pretreatment MR imaging features of Triple-Negative breast cancer: association with response to neoadjuvant chemotherapy and Recurrence-Free survival. Radiology. 2016;281(2):392–400. [DOI] [PubMed] [Google Scholar]
- 30.Baltzer PA, Yang F, Dietzel M, et al. Sensitivity and specificity of unilateral edema on T2w-TSE sequences in MR-Mammography considering 974 histologically verified lesions. Breast J. 2010;16(3):233–9. [DOI] [PubMed] [Google Scholar]
- 31.Cheon H, Kim HJ, Kim TH, et al. Invasive breast cancer: prognostic value of peritumoral edema identified at preoperative MR imaging. Radiology. 2018;287(1):68–75. [DOI] [PubMed] [Google Scholar]
- 32.Harada TL, Uematsu T, Nakashima K, et al. Evaluation of breast edema findings at T2-weighted breast MRI is useful for diagnosing occult inflammatory breast cancer and can predict prognosis after neoadjuvant chemotherapy. Radiology. 2021;299(1):53–62. [DOI] [PubMed] [Google Scholar]
- 33.Jiang T, Song J, Wang X et al. Intratumoral and peritumoral analysis of Mammography, tomosynthesis, and multiparametric MRI for predicting Ki-67 level in breast cancer: a radiomics-based study. Mol Imag Biol. 2022;24(4):550-9. [DOI] [PubMed]
- 34.Caballo M, Sanderink WBG, Han L, et al. Four-Dimensional machine learning radiomics for the pretreatment assessment of breast cancer pathologic complete response to neoadjuvant chemotherapy in dynamic Contrast-Enhanced MRI. J Magn Reson Imaging. 2023;57(1):97–110. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Jiang W, Meng R, Cheng Y, et al. Intra- and peritumoral based radiomics for assessment of lymphovascular invasion in invasive breast cancer. J Magn Reson Imaging. 2024;59(2):613–25. [DOI] [PubMed] [Google Scholar]
- 36.Li C, Lu N, He Z, et al. A noninvasive tool based on magnetic resonance imaging radiomics for the preoperative prediction of pathological complete response to neoadjuvant chemotherapy in breast cancer. Ann Surg Oncol. 2022;29(12):7685–93. [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.
Supplementary Materials
Data Availability Statement
Due to the privacy of patients, the clinical data related to patients and MRI images cannot be available for public access but can be obtained from the corresponding author on reasonable request.