Abstract
Background
The use of multiparametric magnetic resonance imaging (MRI) in predicting lymphovascular invasion (LVI) in breast cancer has been well-documented in the literature. However, the majority of the related studies have primarily focused on intratumoral characteristics, overlooking the potential contribution of peritumoral features. The aim of this study was to evaluate the effectiveness of multiparametric MRI in predicting LVI by analyzing both intratumoral and peritumoral radiomics features and to assess the added value of incorporating both regions in LVI prediction.
Methods
A total of 366 patients underwent preoperative breast MRI from two centers and were divided into training (n=208), validation (n=70), and test (n=88) sets. Imaging features were extracted from intratumoral and peritumoral T2-weighted imaging, diffusion-weighted imaging, and dynamic contrast-enhanced MRI. Five models were developed for predicting LVI status based on logistic regression: the tumor area (TA) model, peritumoral area (PA) model, tumor-plus-peritumoral area (TPA) model, clinical model, and combined model. The combined model was created incorporating the highest radiomics score and clinical factors. Predictive efficacy was evaluated via the receiver operating characteristic (ROC) curve and area under the curve (AUC). The Shapley additive explanation (SHAP) method was used to rank the features and explain the final model.
Results
The performance of the TPA model was superior to that of the TA and PA models. A combined model was further developed via multivariable logistic regression, with the TPA radiomics score (radscore), MRI-assessed axillary lymph node (ALN) status, and peritumoral edema (PE) being incorporated. The combined model demonstrated good calibration and discrimination performance across the training, validation, and test datasets, with AUCs of 0.888 [95% confidence interval (CI): 0.841–0.934], 0.856 (95% CI: 0.769–0.943), and 0.853 (95% CI: 0.760–0.946), respectively. Furthermore, we conducted SHAP analysis to evaluate the contributions of TPA radscore, MRI-ALN status, and PE in LVI status prediction.
Conclusions
The combined model, incorporating clinical factors and intratumoral and peritumoral radscore, effectively predicts LVI and may potentially aid in tailored treatment planning.
Keywords: Breast cancer, lymphovascular invasion (LVI), machine learning, radiomics, Shapley additive explanation (SHAP)
Introduction
Breast cancer is the predominant malignant tumor among women, exerting a significant burden on women’s health across the globe (1). Lymphovascular invasion (LVI), characterized by the presence of tumor cells within the lymphatic and vascular systems, is a critical biomarker for indicating breast cancer aggressiveness and poor prognosis (2,3). Research has consistently shown that LVI is strongly associated with increased tumor invasiveness, recurrence, involvement of axillary lymph nodes (ALNs), and distant metastasis (4,5). However, the current methods for assessing LVI are fraught with limitations. Traditional assessments primarily depend on postoperative histopathological examination, which fails to offer preoperative predictions of LVI status (6).
Magnetic resonance imaging (MRI) plays a critical role in diagnosing and evaluating breast cancer, particularly in predicting LVI (7). It provides critical information about the morphological features and function of the tumor, which is essential for the early detection and determination of the invasiveness of breast cancer (8,9). Previous research has established a strong correlation of MRI indicators such as peritumoral edema (PE) and the status of ALNs with the presence of LVI (10). However, limitations exist in the morphological assessment of conventional MRI. For instance, detecting minimal or subtle LVI manifestations can be challenging, especially in smaller tumors or those at early stages (11). As an innovative approach, radiomics can quantitatively analyze medical images and extract an array of image features (12,13). In recent years, radiomics has demonstrated considerable value in oncology—particularly in breast cancer for noninvasive prediction and classification tasks. Various studies have demonstrated that combining radiomics features with machine learning classifiers significantly improves diagnostic accuracy. For instance, Wu et al. (14) constructed a radiomics framework combining intratumoral, peritumoral, and habitat features extracted from dynamic contrast-enhanced MRI (DCE-MRI) and applied six different machine learning classifiers to predict ALN metastasis in patients with breast cancer. Similarly, Zhang et al. (15) built eight radiomics-based machine learning models using T2-weighted imaging (T2WI) and DCE-MRI to predict LVI in breast cancer.
Previous radiomics research into LVI in breast cancer has predominantly concentrated on the DCE-MRI (7,16,17). Multiparametric MRI, by integrating DCE-MRI with other sequences, offers a comprehensive evaluation of the biological dynamics of breast cancer tumors (6,12). Notably, peritumoral factors, including their impact on ALN status, have also demonstrated predictive value for neoadjuvant chemotherapy (18,19). However, there are relatively few studies that have combined intratumoral and peritumoral features in multiparametric MRI to predict LVI status (17). This highlights the need for further exploration to optimize LVI prediction and improve clinical decision-making.
Shapley additive explanation (SHAP) provides a robust framework for enhancing the interpretability of machine learning models, addressing the longstanding “black-box” problem associated with their clinical adoption (20,21). By offering both global and local explanations, SHAP enables the ranking of feature importance and the visualization of contributions through intuitive plots, facilitating a clearer interpretation of complex radiomics-based predictions. Recent studies have demonstrated the effectiveness of SHAP in improving the reliability and clinical applicability of predictive models, such as those for molecular subtype classification and ALN status assessment (18,22).
The objective of our study was thus to predict LVI status in patients with breast cancer by integrating intratumoral and peritumoral radiomics features derived from multiparametric MRI. Furthermore, SHAP was applied to the combined model incorporating both clinical and intra- and peritumoral radiomics features to generate global insights into feature importance and local explanations for individual predictions, thereby enabling a comprehensive interpretation of each feature’s contribution to LVI prediction. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2685/rc).
Methods
Patients and histopathology collection
The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and approved by the Institutional Ethics Committee of Guangzhou First People’s Hospital (approval No. K-2023161-01) and The Third Affiliated Hospital of Sun Yat-Sen University (approval No. SL-RG2024-070-01). The requirement for informed consent was waived due to the retrospective nature of the analysis.
A total of 366 female patients with invasive breast cancer who underwent surgery from April 2019 to May 2023 in center 1 (Guangzhou First People’s Hospital) and from July 2019 to May 2020 in center 2 (The Third Affiliated Hospital of Sun Yat-Sen University) were included in the study. The inclusion criteria were as follows: (I) MRI examination performed prior to the biopsy procedure; (II) confirmed LVI status via postoperative pathology; (III) no prior treatment or surgery for breast cancer; and (IV) a time interval between MRI examination and surgery of <2 weeks. Meanwhile, the exclusion criteria were as follows: (I) obvious artifacts on magnetic resonance (MR) images; (II) incomplete clinical characteristics (such as absence of menstrual status and immunohistochemistry); and (III) neoadjuvant therapy before MRI. Patients from center 1 were randomly divided into training (n=208) and validation (n=70) sets in a 3:1 ratio, while all patients from center 2 were assigned to the test set (n=88) for evaluation of the model’s generalizability. Figure 1 provides a flowchart of patient enrollment.
Figure 1.
Flowchart of patient enrollment. LVI, lymphovascular invasion; MRI, magnetic resonance imaging.
Clinicopathological information was collected for the enrolled patients from electronic medical records and included age at initial diagnosis, menopausal status, LVI status, histologic type, estrogen receptor (ER) status, progesterone receptor (PR) status, hormonal receptor (HR) status, human epidermal growth factor receptor type 2 (HER2) status, and Ki-67 index. LVI was considered positive if cancer cells were present in lymphatic or small blood vessels external to the tumor. The threshold for ER and PR positivity was set at 1%, and that for the Ki-67 index was 20%. HR-positive was defined as ER-positive or PR-positive. HER2-negative tumors were defined as an immunohistochemical staining score of 0 or 1+, whereas HER2-positive tumors were identified by a score of 3+. Chromogenic in situ hybridization was used to confirm HER2 amplification in cases scored 2+.
MRI acquisition and evaluation
MRI scans were conducted with two 1.5-T MRI systems (center 1: uMR 560, United Imaging, Shanghai, China; center 2: Achieva, Philips Healthcare, Best, the Netherlands), employed through dedicated bilateral breast coils and with patients in the prone position. The MRI protocols included fat-suppressed T2WI, DCE, and diffusion-weighted imaging (DWI) with b-values of 0 and 800 s/mm2. For DCE-MRI, both centers administered a gadolinium-based contrast agent (Gd-DTPA, Magnevist; Bayer, Leverkusen, Germany) at a dosage of 0.2 mL/kg body weight and an infusion rate of 1.5 mL/s. Center 1 acquired six DCE-MRI phases (1 precontrast and 5 postcontrast), while center 2 acquired seven phases (1 precontrast and 6 postcontrast). The detailed acquisition parameters of the two centers are summarized in Table S1.
To delineate MRI morphological characteristics, two radiologists who were blinded to the clinicopathological data visually assessed the MR images following the Breast Imaging Reporting and Data System guidelines. They evaluated the following MRI features of tumors: lesion type, mass shape, margin, background parenchymal enhancement (BPE), PE, multifocal lesions, time-signal intensity curve (TIC) pattern, MRI-ALN status, and largest diameter (LD). PE was considered present when fat-suppressed T2 showed that the signal around the tumor was higher than the signal of the breast tissue and almost equal to the signal of water (23). MRI-ALN metastasis was considered present when lymph nodes exhibited a round or irregular shape, loss of fatty hilum, cortical thickening ≥3 mm, or a short-axis diameter of ≥10 mm (24-26). The contrast enhancement pattern was classified based on the TIC into three conventional types: persistent (type I), plateau (type II), and washout (type III).
MRI segmentation and radiomics feature extraction
The data flowchart for this study is presented in Figure 2. To enhance feature robustness across datasets, images underwent resampling to a voxel spacing of 1 mm × 1 mm × 1 mm, followed by z-score normalization and intensity discretization with a fixed bin width of 25. Regions of interest (ROIs) encompassing the entire tumor region were manually delineated slice by slice along the lesion boundary on transverse images from the peak phase of DCE-MRI, T2WI, and DWI for each patient by a junior radiologist using ITK-SNAP 4.0 software (http://www.itksnap.org). In cases of multicentric lesions, the largest tumor lesion was designated as the target lesion. A subset of 30 patients was randomly selected for a second delineation by the senior radiologist to assess the intraclass correlation coefficients (ICCs). After lesion delineation, ROIs of the peritumoral area (PA) 4 mm around the tumor were obtained automatically (27,28). The segmented regions were denoted as the tumor area (TA), PA, and tumor-plus-peritumoral area (TPA). Image preprocessing and feature extraction processes were executed via the open-source Pyradiomics software, which facilitated the extraction of various features from all ROIs, including first-order features, texture features, shape features, and wavelet features.
Figure 2.
Flowchart of image processing, including image segmentation, feature extraction, feature reduction, and model evaluation. GA, genetic algorithm; ICC, intraclass correlation coefficient; PA, peritumoral area; SHAP, Shapley additive explanation; TA, tumor area; TPA, tumor-plus-peritumoral area.
Feature selection
We developed a three-step procedure to select robust radiomic features. Initially, the ICC was calculated to assess the stability of the radiomic features, and features exhibiting an ICC less than 0.75 were excluded. Subsequently, Pearson correlation analysis was employed to remove redundant features under a threshold of r>0.90. Finally, the genetic algorithm (GA) was used to identify the optimal subset of features (details are presented in Appendix 1).
Construction and validation of the different models
The radiomics score (radscore) was defined as a linear combination of selected radiomic features weighted by their corresponding logistic regression coefficients after feature selection with a GA. The clinical model was constructed from the most significant clinical factors as identified through univariate and multivariate logistic regression analyses. The combined model was established through the incorporation of both the highest radiomics signature and clinical factors via multivariate logistic regression. Five models were developed for predicting LVI status based on logistic regression: the TA model, PA model, TPA model, clinical model, and combined model (Table S2).
Model interpretability with SHAP
The SHAP method was applied to rank feature importance and interpret the final habitat model via the computation of both global and individual Shapley values, thereby enhancing model transparency and supporting its clinical applicability. SHAP was used to interpret the final logistic regression model by quantifying the contribution of each feature. The logistic regression model was implemented via the logistic regression class from the scikit-learn library (penalty =“l2”, C =1.0, solver = “lbfgs”, maximum iterations =100). The SHAP analysis was conducted with the “shap” Python package (Python Software Foundation, Wilmington, DE, USA), and summary plots were generated to visualize global feature importance and individual explanations.
Sample size evaluation
Based on the following inputs and hypotheses, a sample size of at least 44 patients with breast cancer (13 LVI-positive patients and 31 LVI-negative patients) was deemed necessary for the study: alternative hypothesis for the area under the receiver operating characteristic (ROC) curve (AUC), 0.800; AUC for the null hypothesis, 0.500; power, 80%; two-sided significance threshold, 0.05; and allocation ratio of sample sizes in the negative and positive groups, 7:3. Therefore, under the assumption that the true AUC was >0.800, sample sizes of 208 in training, 70 in the validation, and 88 in the test set were considered adequate to detect an AUC difference of 0.500 with 80% power (29).
Statistical analysis
Statistical analyses were performed with SPSS version 26.0 (IBM Corp., Armonk, NY, USA) and R software version 4.4.1 (The R Foundation for Statistical Computing). Continuous variables were compared between the LVI-positive and LVI-negative groups via independent sample t-tests, while categorical variables were analyzed with the chi-squared test. Variables that reached statistical significance (P<0.05) in the univariate analysis were fitted into the multivariate logistic regression analysis to estimate the odds ratio (OR) and 95% confidence interval (CI) for risk factors of LVI via the enter method. ROC curves were constructed, with optimal cutoff values being derived from the maximum Youden index. Subsequently, the AUC was calculated to assess the performance of the models. The DeLong test was employed for statistical comparisons, and the calibration of the models was evaluated with calibration curves, which demonstrated the agreement between predicted probabilities and actual outcomes. To evaluate the clinical utility of the models, decision curve analysis (DCA) was applied, with the net benefits across a range of threshold probabilities being calculated. All tests were two-tailed, and a P value of less than 0.05 was considered statistically significant.
Results
Patient characteristics
This investigation analyzed data from 366 patients. From center 1, 208 patients were placed into a training set and 70 patients into a validation set; from center 2, 88 patients were placed into a test set (Table 1). The ICC for interobserver agreement for MRI characteristics ranged from 0.705 to 0.980 (Table S3). In the univariate logistic regression of the training set, the significantly different characteristics between LVI-negative and LVI-positive patients included LD, MRI-ALN status, and PE (all P values <0.05). According to multivariable logistic regression, the predictive variables for LVI were MRI-ALN status (OR =2.670; 95% CI: 1.384–5.153; P=0.003) and PE (OR =2.303; 95% CI: 1.246–4.257; P=0.008) (Table 2). The clinical model was constructed from MRI-ALN status and PE.
Table 1. Clinical characteristics of the training, validation, and test sets in the study.
| Characteristics | Training set (n=208) | Validation set (n=70) | Test set (n=88) |
|---|---|---|---|
| LVI status | |||
| Negative | 137 (65.8) | 48 (68.6) | 59 (67.0) |
| Positive | 71 (34.2) | 22 (31.4) | 29 (33.0) |
| Age (years) | 54.73±10.43 | 53.74±11.04 | 48.77±11.11 |
| Menstrual status | |||
| Premenopausal | 71 (34.1) | 27 (38.6) | 53 (60.2) |
| Postmenopausal | 137 (65.9) | 43 (61.4) | 35 (39.8) |
| ER status | |||
| Negative | 44 (21.2) | 16 (22.9) | 34 (38.6) |
| Positive | 164 (78.8) | 54 (77.1) | 54 (61.4) |
| PR status | |||
| Negative | 57 (27.4) | 22 (31.4) | 40 (45.5) |
| Positive | 151 (72.6) | 48 (68.6) | 48 (54.5) |
| HER2 status | |||
| Negative | 158 (76.0) | 50 (71.4) | 57 (64.8) |
| Positive | 50 (24.0) | 20 (28.6) | 31 (35.2) |
| Ki-67 status | |||
| Negative | 58 (27.9) | 17 (24.3) | 20 (22.7) |
| Positive | 150 (72.1) | 53 (75.7) | 68 (77.3) |
| Location | |||
| Left | 105 (50.5) | 29 (41.4) | 49 (55.7) |
| Right | 103 (49.5) | 41 (58.6) | 39 (44.3) |
| LD (mm) | 21.93±9.03 | 25.37±12.35 | 27.00±12.69 |
| TIC | |||
| Persistent | 17 (8.2) | 6 (8.6) | 6 (6.8) |
| Plateau | 170 (81.7) | 51 (72.8) | 70 (79.6) |
| Washout | 21 (10.1) | 13 (18.6) | 12 (13.6) |
| MRI-ALN status | |||
| Absent | 153 (73.6) | 47 (67.1) | 55 (62.5) |
| Present | 55 (26.4) | 23 (32.9) | 33 (37.5) |
| PE | |||
| Absent | 114 (54.8) | 36 (51.4) | 48 (54.5) |
| Present | 94 (45.2) | 34 (48.6) | 40 (45.5) |
| Multifocal or multicentric | |||
| Absent | 175 (84.1) | 59 (84.3) | 77 (87.5) |
| Present | 33 (15.9) | 11 (15.7) | 11 (12.5) |
| Shape | |||
| Oval/round | 48 (23.0) | 6 (8.5) | 11 (12.5) |
| Irregular | 160 (77.0) | 64 (91.5) | 77 (87.5) |
| Margin | |||
| Circumscribed | 36 (17.3) | 6 (8.6) | 11 (12.5) |
| Noncircumscribed | 172 (82.7) | 64 (91.4) | 77 (87.5) |
| Internal enhancement | |||
| Homogeneous | 11 (5.3) | 6 (8.6) | 9 (10.2) |
| Heterogeneous | 193 (92.8) | 64 (91.4) | 71 (80.7) |
| Rim | 4 (1.9) | 0 | 8 (9.1) |
| FGT | |||
| Heterogeneously dense or dense | 96 (46.2) | 38 (54.3) | 35 (39.8) |
| Predominantly fatty or scattered | 112 (53.8) | 32 (45.7) | 53 (60.2) |
Data are presented as mean ± standard deviation or n (%). ALN, axillary lymph node; ER, estrogen receptor; FGT, fibroglandular tissue; HER2, human epidermal growth factor receptor type 2; LD, largest diameter; LVI, lymphovascular invasion; MRI, magnetic resonance imaging; PE, peritumoral edema; PR, progesterone receptor; TIC, time-signal intensity curve.
Table 2. Univariate and multivariate logistic regression of the clinical characteristics in the training set.
| Variants | Univariate | Multivariate | |||
|---|---|---|---|---|---|
| OR (95% CI) | P value | OR (95% CI) | P value | ||
| Age | 0.978 (0.951–1.005) | 0.115 | – | – | |
| Menstrual status | 0.930 (0.509–1.699) | 0.814 | – | – | |
| ER status | 1.103 (0.540–2.254) | 0.787 | – | – | |
| PR status | 1.142 (0.594–2.193) | 0.691 | – | – | |
| HER2 status | 1.104 (0.567–2.147) | 0.771 | – | – | |
| Ki-67 status | 1.882 (0.947–3.743) | 0.071 | – | – | |
| Location | 0.831 (0.468–1.476) | 0.528 | – | – | |
| LD (mm) | 1.033 (1.000–1.066) | 0.047 | 1.020 (0.986–1.055) | 0.246 | |
| TIC | 1.364 (0.694–2.681) | 0.368 | – | – | |
| MRI-ALN status | 2.653 (1.405–5.008) | 0.003 | 2.670 (1.384–5.153) | 0.003 | |
| Peritumoral edema | 2.367 (1.318–4.252) | 0.004 | 2.303 (1.246–4.257) | 0.008 | |
| Multifocal or multicentric | 1.771 (0.832–3.769) | 0.138 | – | – | |
| Shape | 1.531 (0.750–3.125) | 0.242 | – | – | |
| Margin | 2.349 (0.970–5.684) | 0.058 | – | – | |
| Internal enhancement | 0.620 (0.213–1.814) | 0.380 | – | – | |
| FGT | 0.787 (0.442–1.403) | 0.417 | – | – | |
ALN, axillary lymph node; CI, confidence interval; ER, estrogen receptor; FGT, fibroglandular tissue; HER2, human epidermal growth factor receptor type 2; LD, largest diameter; MRI, magnetic resonance imaging; OR, odds ratio; PE, peritumoral edema; PR, progesterone receptor; TIC, time-signal intensity curve.
Radiomics feature selection and model construction
From intratumoral regions, 1,130 radiomics features were extracted separately from the DCE, DWI, and T2WI sequences, with the same number of features extracted from the peritumoral region. Features with ICCs greater than 0.75 were initially selected, and then those with correlation coefficients exceeding 0.90 were removed, from each pair. GA was applied to further screen the radiomics features for final selection (Figure S1). The TPA model, composed of features from the intra- and peritumoral regions, yielded higher AUCs than did the TA or PA model in all datasets (Figure 3). The distributions of the TPA radscore for each patient in the training, validation, and test datasets are shown in Figure S2. Finally, the TPA radscore, MRI-ALN status, and PE were integrated to construct the combined model.
Figure 3.
ROC curves of the TA, PA, TPA, clinical, and combined models for predicting LVI status in the training (A), validation (B), and test (C) datasets. The TPA model achieved higher AUCs than did the TA and PA models across all datasets (AUCs of 0.878, 0.825, and 0.823, respectively). AUC, area under the curve; LVI, lymphovascular invasion; PA, peritumoral area; ROC, receiver operating characteristic; TA, tumor area; TPA, tumor-plus-peritumoral area.
Comparison of the different models
According to the DeLong test, the combined model exhibited significantly better ability in predicting LVI status than did the clinical model in the training set (AUC: 0.888 vs. 0.665; P<0.001), validation set (AUC: 0.856 vs. 0.635; P=0.001), and test set (AUC: 0.853 vs. 0.686; P=0.001) (Table 3). Calibration curves showed good agreement between actual probabilities and the probabilities estimated by the models (Figure S3). The DCA indicated that the combined model could provide a higher net benefit than could the clinical model in predicting LVI status in all datasets (Figure 4).
Table 3. Performance of the different models for assessing LVI in invasive breast cancer.
| Data set | Model | AUC (95% CI) | ACC | SEN | SPE | P value |
|---|---|---|---|---|---|---|
| Training set | TA | 0.812 (0.754–0.870) | 0.755 | 0.718 | 0.774 | 0.013 |
| PA | 0.830 (0.772–0.887) | 0.803 | 0.746 | 0.832 | 0.046 | |
| TPA | 0.878 (0.828–0.927) | 0.832 | 0.831 | 0.832 | 0.344 | |
| Clinical | 0.665 (0.591–0.739) | 0.591 | 0.775 | 0.496 | <0.001 | |
| Combined | 0.888 (0.841–0.934) | 0.822 | 0.845 | 0.810 | Ref. | |
| Validation set | TA | 0.752 (0.629–0.874) | 0.743 | 0.682 | 0.771 | 0.163 |
| PA | 0.757 (0.641–0.873) | 0.729 | 0.773 | 0.708 | 0.097 | |
| TPA | 0.825 (0.726–0.924) | 0.757 | 0.864 | 0.708 | 0.214 | |
| Clinical | 0.635 (0.497–0.774) | 0.671 | 0.500 | 0.750 | 0.001 | |
| Combined | 0.856 (0.769–0.943) | 0.757 | 0.909 | 0.688 | Ref. | |
| Test set | TA | 0.766 (0.664–0.867) | 0.659 | 0.897 | 0.542 | 0.192 |
| PA | 0.704 (0.583–0.824) | 0.693 | 0.759 | 0.661 | 0.051 | |
| TPA | 0.823 (0.726–0.921) | 0.795 | 0.862 | 0.763 | 0.203 | |
| Clinical | 0.686 (0.571–0.802) | 0.670 | 0.690 | 0.661 | 0.001 | |
| Combined | 0.853 (0.760–0.946) | 0.852 | 0.793 | 0.881 | Ref |
ACC, accuracy; AUC, area under the curve; CI, confidence interval; LVI, lymphovascular invasion; PA, peritumoral area; Ref., reference; SEN, sensitivity; SPE, specificity; TA, tumor area; TPA, tumor-plus-peritumoral area.
Figure 4.
Decision curve analyses of the TA, PA, TPA, clinical, and combined models in the training (A), validation (B), and test (C) datasets. The x-axis and y-axis represent the threshold probability and net benefit, respectively. The purple, red, green, orange, and blue lines correspond to decisions made according to the TA, PA, TPA, clinical, and combined models, respectively. The black line represents the strategy of treating all patients as LVI-positive, while the gray line represents treating none of them. The combined model demonstrated higher clinical net benefit across thresholds as compared to the clinical model in all datasets. LVI, lymphovascular invasion; PA, peritumoral area; TA, tumor area; TPA, tumor-plus-peritumoral area.
Model explanation
The SHAP method was employed to interpret the output of the combined model by quantifying each variable’s contribution to the prediction. As shown in the SHAP summary plots in Figure 5A,5B, the contributions of the features to the combined model were evaluated via the mean SHAP values and are displayed in descending order. In addition, local explanation was used to analyze how the model made predictions for individual cases.
Figure 5.
Global model explanation by the SHAP method. (A) SHAP summary bar plot. (B) SHAP summary dot plot. The probability of LVI development increases with the higher SHAP value of a given feature. ALN, axillary lymph node; LVI, lymphovascular invasion; MRI, magnetic resonance imaging; PE, peritumoral edema; radscore, radiomics score; SHAP, Shapley additive explanation.
Figure 6A,6B illustrate two patients with breast cancer with different LVI statuses postoperation. Furthermore, an interactive SHAP force plot of the training set is provided in Figure 6C (https://pretend58.github.io/LVI-SHAP-FORCE/LVI-shap.html). This plot enables dynamic exploration of individual feature contributions.
Figure 6.
Local model explanation by the SHAP method. (A,B) SHAP waterfall plots of two examples of different LVI statuses. (a1,b1) DCE example; (a2,b2) T2WI example. (A) A representative patient with breast cancer and MRI-ALN positivity in whom the probability of LVI-negative status was 87.9% according to the combined model. (B) A representative patient with breast cancer with MRI-ALN negativity in whom the probability of LVI-positive status was 85.4% according to the combined model. (C) SHAP force plot for the training set. The x-axis represents individual patients as ordered by model output value. The y-axis indicates the SHAP value and each feature’s contribution to the model output. The red regions are features that contribute positively to the predicted probability of LVI positivity, whereas blue regions are features that contribute negatively. ALN, axillary lymph node; DCE, dynamic contrast-enhanced; LVI, lymphovascular invasion; MRI, magnetic resonance imaging; PE, peritumoral edema; SHAP, Shapley additive explanation; T2WI, T2-weighted imaging.
Discussion
The findings from this study indicate that the TPA constructed using multiparametric radiomics features from the combined intratumoral and peritumoral areas exhibited excellent performance across the training, validation, and test datasets. TPA features may have greater value in preoperative tumor evaluation and help inform the formulation of patient treatment strategies and clinical decisions. Furthermore, the combined model significantly outperformed the clinical model in predicting LVI.
Previous studies on LVI have primarily focused solely on evaluating LVI status using radiomics derived from DCE-MRI (16,30,31). However, multiparametric MRI radiomics offers a more comprehensive perspective on tumor biological behavior, enabling more precise assessment of LVI status (6,12). This approach enhances predictive capability by integrating a diversity sequence features. Features derived from multiparametric MRI leverage the strengths of each sequence, providing a richer and more detailed depiction of tumor characteristics (7,16). In addition, the development of LVI involves complex interactions among tumor cells and their surrounding microenvironment, including the degradation of the extracellular matrix, direct cell-to-cell contact, and the regulation of various cytokines and chemical signals (32,33). Previous MRI-based radiomics studies on breast cancer have suggested that PA with a 4-mm radial delineation distance from the tumor provides valuable information for diagnosing malignant lesions and predicting programmed cell death ligand-1 status (27,34). In line with this, our TPA model, which combines intratumoral and peritumoral radiomics features, outperformed the models based solely on TA or PA in predicting LVI status. Although no statistically significant difference was observed between the TPA model and the TA or PA models in a few of the datasets, the TPA model generally exhibited greater predictive ability, highlighting its potential advantage in comprehensive feature integration.
PE typically indicates lymphatic obstruction and alterations in local blood circulation, suggesting the potential for tumor invasion into or in close proximity to the lymphovascular system (35,36). Furthermore, the characteristics of ALNs are essential in assessing LVI (36). The diagnostic performance of the combined model, particularly when integrated with TPA radscore and clinical factors, consistently surpassed that of the clinical model across all datasets (P<0.05). Compared to previous studies (30,36,37), our study achieved an improvement in predicting LVI through use of intratumoral and peritumoral multiparametric MRI.
Given the challenge clinicians face in accepting prediction models that lack direct explainability and interpretability, the SHAP method was employed to elucidate the workings of this model (38). SHAP-derived force maps provided detailed, patient-specific assessments, enhancing transparency in model interpretation. Moreover, SHAP’s visualization capabilities improved the clarity and depth of feature impact analysis, including their interactive effects. This approach enhances understanding and facilitates the model’s practical utility for clinicians. The high SHAP contribution of the TPA radscore suggests that the combined intra- and peritumoral radiomic patterns may reflect tumor heterogeneity and microenvironmental interactions associated with LVI status.
This study involved several limitations which should be noted. First, we employed a retrospective design, and further larger-scale, multicenter prospective trials are needed to validate and reinforce these findings. Second, the use of a 4-mm peritumoral region in this study was based on findings reported in the majority of the related literature. However, we did not further examine or determine the optimal size of the peritumoral region for generating the ROI. Finally, the scope was limited to invasive breast cancer and did not include ductal carcinoma in situ (DCIS), which may restrict the generalizability of the results. Future studies should incorporate DCIS to enhance diagnostic accuracy.
Conclusions
The TPA model demonstrated superior performance as compared to the TA model and PA model. By integrating the TPA radscore with clinical factors, the combined model may serve as a biomarker for predicting LVI. The results suggest that the combined model could aid clinicians in identifying patients with LVI, thereby facilitating the development of personalized treatment plans.
Supplementary
The article’s supplementary files as
Acknowledgments
None.
Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and was approved by the Institutional Ethics Committee of Guangzhou First People’s Hospital (Approval No. K-2023161-01) and the Third Affiliated Hospital of Sun Yat-Sen University (Approval No. SL-RG2024-070-01). The requirement for informed consent was waived due to the retrospective nature of the analysis.
Footnotes
Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2685/rc
Funding: This study was supported by (I) the National Natural Science Foundation of China (Nos. 82302314 and 81901711); (II) Basic and Applied Basic Research Foundation of Guangdong Province (Nos. 2022A1515110792 to W.T., 2023A1515220097 to Y.G., and 2024A1515010653 to Y.G.); (III) Medical Scientific Research Foundation of Guangdong Province (No. A2023073 to Q.K.); (IV) Science and Technology Projects in Guangzhou (Nos. 2024A03J1030 to W.T., 2025A03J4163 to Y.G., and 2025A03J4162 to W.L.); (V) Guangzhou First People’s Hospital Frontier Medical Technology Project (QY-C04 to Y.G.); (VI) 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.
Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2685/coif). W.T. reports that this research was funded by the National Natural Science Foundation of China (No. 82302314), Basic and Applied Basic Research Foundation of Guangdong Province (No. 2022A1515110792) and Science and Technology Projects in Guangzhou (No. 2024A03J1030). Y.F. is currently an employee of Shukun Technology Co., Ltd. Q.K. reports that this research was funded by the Medical Scientific Research Foundation of Guangdong Province (No. A2023073). Y.G. reports that this research was funded by the National Natural Science Foundation of China (No. 81901711), Basic and Applied Basic Research Foundation of Guangdong Province (Nos. 2023A1515220097 and 2024A1515010653), Science and Technology Projects in Guangzhou (No. 2025A03J4163), Guangzhou First People’s Hospital Frontier Medical Technology Project (QY-C04), the Special Fund for the Construction of High-level Key Clinical Specialty (Medical Imaging) in Guangzhou, and Guangzhou Key Laboratory of Molecular Imaging and Clinical Translational Medicine. W.L. reports that this research was funded by the Science and Technology Projects in Guangzhou (No. 2025A03J4162). The other authors have no conflicts of interest to declare.
Data Sharing Statement
Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2685/dss
References
- 1.Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. 10.3322/caac.21834 [DOI] [PubMed] [Google Scholar]
- 2.Britto AV, Schenka AA, Moraes-Schenka NG, Alvarenga M, Shinzato JY, Vassallo J, Ward LS. 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. 10.1186/1471-2407-9-109 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Rakha EA, Martin S, Lee AH, Morgan D, Pharoah PD, Hodi Z, Macmillan D, Ellis IO. The prognostic significance of lymphovascular invasion in invasive breast carcinoma. Cancer 2012;118:3670-80. 10.1002/cncr.26711 [DOI] [PubMed] [Google Scholar]
- 4.Curigliano G, Burstein HJ, Winer EP, Gnant M, Dubsky P, Loibl S, et al. De-escalating and escalating treatments for early-stage breast cancer: the St. Gallen International Expert Consensus Conference on the Primary Therapy of Early Breast Cancer 2017. Ann Oncol 2017;28:1700-12. 10.1093/annonc/mdx308 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Wang C, Wu Y, Shao J, Liu D, Li W. Clinicopathological variables influencing overall survival, recurrence and post-recurrence survival in resected stage I non-small-cell lung cancer. BMC Cancer 2020;20:150. 10.1186/s12885-020-6621-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Zhang J, Wang G, Ren J, Yang Z, Li D, Cui Y, Yang X. Multiparametric MRI-based radiomics nomogram for preoperative prediction of lymphovascular invasion and clinical outcomes in patients with breast invasive ductal carcinoma. Eur Radiol 2022;32:4079-89. 10.1007/s00330-021-08504-6 [DOI] [PubMed] [Google Scholar]
- 7.Yang X, Fan X, Lin S, Zhou Y, Liu H, Wang X, Zuo Z, Zeng Y. Assessment of Lymphovascular Invasion in Breast Cancer Using a Combined MRI Morphological Features, Radiomics, and Deep Learning Approach Based on Dynamic Contrast-Enhanced MRI. J Magn Reson Imaging 2024;59:2238-49. 10.1002/jmri.29060 [DOI] [PubMed] [Google Scholar]
- 8.Chen S, Sui Y, Ding S, Chen C, Liu C, Zhong Z, Liang Y, Kong Q, Tang W, Guo Y. A simple and convenient model combining multiparametric MRI and clinical features to predict tumour-infiltrating lymphocytes in breast cancer. Clin Radiol 2023;78:e1065-74. 10.1016/j.crad.2023.08.029 [DOI] [PubMed] [Google Scholar]
- 9.Chen S, Du Y, Hu Y, Ling R, Huang D, Xiang J, Liang Y, Wei X, Tang W, Guo Y. Preoperative MRI of breast squamous cell carcinoma: diagnostic value of distinguishing between two subtypes. Clin Radiol 2022;77:e321-8. 10.1016/j.crad.2021.12.018 [DOI] [PubMed] [Google Scholar]
- 10.Zhang C, Liang Z, Feng Y, Xiong Y, Manwa C, Zhou Q. Risk Factors for Lymphovascular Invasion in Invasive Ductal Carcinoma Based on Clinical and Preoperative Breast MRI Features: a Retrospective Study. Acad Radiol 2023;30:1620-7. 10.1016/j.acra.2022.10.029 [DOI] [PubMed] [Google Scholar]
- 11.Şendur HN, Şendur AB. MRI-Based Radiomics May Provide More In-depth Information Regarding Lymphovascular Invasion Status in Patients with Breast Cancer. Acad Radiol 2023;30:2710-1. 10.1016/j.acra.2023.07.014 [DOI] [PubMed] [Google Scholar]
- 12.Jiang Y, Zeng Y, Zuo Z, Yang X, Liu H, Zhou Y, Fan X. Leveraging multimodal MRI-based radiomics analysis with diverse machine learning models to evaluate lymphovascular invasion in clinically node-negative breast cancer. Heliyon 2024;10:e23916. 10.1016/j.heliyon.2023.e23916 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Lv L, Zhang Z, Zhang D, Chen Q, Liu Y, Qiu Y, Fu W, Yin X, Chen X. Machine-learning radiomics to predict bone marrow metastasis of neuroblastoma using magnetic resonance imaging. Cancer Innov 2023;2:405-15. 10.1002/cai2.92 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Wu PQ, Guo FL, Wang J, Gao Y, Feng ST, Chen SL, Ma J, Liu YB. Development and validation of a dynamic contrast-enhanced magnetic resonance imaging-based habitat and peritumoral radiomic model to predict axillary lymph node metastasis in patients with breast cancer: a retrospective study. Quant Imaging Med Surg 2024;14:8211-26. 10.21037/qims-24-558 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Zhang C, Zhong M, Liang Z, Zhou J, Wang K, Bu J. MRI-based radiomic and machine learning for prediction of lymphovascular invasion status in breast cancer. BMC Med Imaging 2024;24:322. 10.1186/s12880-024-01501-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Zheng H, Jian L, Li L, Liu W, Chen W. Delta-Radiomics Based on Dynamic Contrast-Enhanced MRI for Predicting Lymphovascular Invasion in Invasive Breast Cancer. Acad Radiol 2024;31:1762-72. 10.1016/j.acra.2023.11.017 [DOI] [PubMed] [Google Scholar]
- 17.Jiang W, Meng R, Cheng Y, Wang H, Han T, Qu N, Yu T, Hou Y, Xu S. Intra- and Peritumoral Based Radiomics for Assessment of Lymphovascular Invasion in Invasive Breast Cancer. J Magn Reson Imaging 2024;59:613-25. 10.1002/jmri.28776 [DOI] [PubMed] [Google Scholar]
- 18.Liu Z, Hong M, Li X, Lin L, Tan X, Liu Y. Predicting axillary lymph node metastasis in breast cancer patients: A radiomics-based multicenter approach with interpretability analysis. Eur J Radiol 2024;176:111522. 10.1016/j.ejrad.2024.111522 [DOI] [PubMed] [Google Scholar]
- 19.Zheng G, Peng J, Shu Z, Jin H, Han L, Yuan Z, Qin X, Hou J, He X, Gong X. Predicting pathological complete response to neoadjuvant chemotherapy in breast cancer patients: use of MRI radiomics data from three regions with multiple machine learning algorithms. J Cancer Res Clin Oncol 2024;150:147. 10.1007/s00432-024-05680-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Zhang R, Hong M, Cai H, Liang Y, Chen X, Liu Z, Wu M, Zhou C, Bao C, Wang H, Yang S, Hu Q. Predicting the pathological invasiveness in patients with a solitary pulmonary nodule via Shapley additive explanations interpretation of a tree-based machine learning radiomics model: a multicenter study. Quant Imaging Med Surg 2023;13:7828-41. 10.21037/qims-23-615 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Xue T, Zhu L, Tao Y, Ye X, Yu H. Development and validation of an interpretable delta radiomics-based model for predicting invasive ground-glass nodules in lung adenocarcinoma: a retrospective cohort study. Quant Imaging Med Surg 2024;14:4086-97. 10.21037/qims-23-1711 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Huang Y, Wang X, Cao Y, Li M, Li L, Chen H, Tang S, Lan X, Jiang F, Zhang J. Multiparametric MRI model to predict molecular subtypes of breast cancer using Shapley additive explanations interpretability analysis. Diagn Interv Imaging 2024;105:191-205. 10.1016/j.diii.2024.01.004 [DOI] [PubMed] [Google Scholar]
- 23.Yu X, Xiang J, Zhang Q, Chen S, Tang W, Li X, Sui Y, Liu W, Kong Q, Guo Y. Triple-negative breast cancer: predictive model of early recurrence based on MRI features. Clin Radiol 2023;78:e798-807. 10.1016/j.crad.2023.07.008 [DOI] [PubMed] [Google Scholar]
- 24.Han L, Zhu Y, Liu Z, Yu T, He C, Jiang W, Kan Y, Dong D, Tian J, Luo Y. Radiomic nomogram for prediction of axillary lymph node metastasis in breast cancer. Eur Radiol 2019;29:3820-9. 10.1007/s00330-018-5981-2 [DOI] [PubMed] [Google Scholar]
- 25.Zhao M, Wu Q, Guo L, Zhou L, Fu K. Magnetic resonance imaging features for predicting axillary lymph node metastasis in patients with breast cancer. Eur J Radiol 2020;129:109093. 10.1016/j.ejrad.2020.109093 [DOI] [PubMed] [Google Scholar]
- 26.Xu Z, Ding Y, Zhao K, Han C, Shi Z, Cui Y, Liu C, Lin H, Pan X, Li P, Chen M, Wang H, Deng X, Liang C, Xie Y, Liu Z. MRI characteristics of breast edema for assessing axillary lymph node burden in early-stage breast cancer: a retrospective bicentric study. Eur Radiol 2022;32:8213-25. 10.1007/s00330-022-08896-z [DOI] [PubMed] [Google Scholar]
- 27.Wu Z, Lin Q, Wang H, Chen J, Wang G, Fu G, Li L, Bian T. Intratumoral and Peritumoral Radiomics Based on Preoperative MRI for Evaluation of Programmed Cell Death Ligand-1 Expression in Breast Cancer. J Magn Reson Imaging 2024;60:588-99. 10.1002/jmri.29109 [DOI] [PubMed] [Google Scholar]
- 28.Xu H, Liu J, Chen Z, Wang C, Liu Y, Wang M, Zhou P, Luo H, Ren J. Intratumoral and peritumoral radiomics based on dynamic contrast-enhanced MRI for preoperative prediction of intraductal component in invasive breast cancer. Eur Radiol 2022;32:4845-56. 10.1007/s00330-022-08539-3 [DOI] [PubMed] [Google Scholar]
- 29.Li X, Zhang N, Hu C, Lin Y, Li J, Li Z, 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. 10.1016/j.eclinm.2022.101805 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Liu Z, Feng B, Li C, Chen Y, Chen Q, Li X, Guan J, Chen X, Cui E, Li R, Li Z, Long W. Preoperative prediction of lymphovascular invasion in invasive breast cancer with dynamic contrast-enhanced-MRI-based radiomics. J Magn Reson Imaging 2019;50:847-57. 10.1002/jmri.26688 [DOI] [PubMed] [Google Scholar]
- 31.Lai T, Chen X, Yang Z, Huang R, Liao Y, Chen X, Dai Z. Quantitative parameters of dynamic contrast-enhanced magnetic resonance imaging to predict lymphovascular invasion and survival outcome in breast cancer. Cancer Imaging 2022;22:61. 10.1186/s40644-022-00499-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Kariri YA, Aleskandarany MA, Joseph C, Kurozumi S, Mohammed OJ, Toss MS, Green AR, Rakha EA. Molecular Complexity of Lymphovascular Invasion: The Role of Cell Migration in Breast Cancer as a Prototype. Pathobiology 2020;87:218-31. 10.1159/000508337 [DOI] [PubMed] [Google Scholar]
- 33.Ribelles N, Santonja A, Pajares B, Llácer C, Alba E. The seed and soil hypothesis revisited: current state of knowledge of inherited genes on prognosis in breast cancer. Cancer Treat Rev 2014;40:293-9. 10.1016/j.ctrv.2013.09.010 [DOI] [PubMed] [Google Scholar]
- 34.Zhou J, Zhang Y, Chang KT, Lee KE, Wang O, Li J, Lin Y, Pan Z, Chang P, Chow D, Wang M, Su MY. Diagnosis of Benign and Malignant Breast Lesions on DCE-MRI by Using Radiomics and Deep Learning With Consideration of Peritumor Tissue. J Magn Reson Imaging 2020;51:798-809. 10.1002/jmri.26981 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Li X, Luo K, Zhang N, Chen W, Li B, Lu Z, Chen Y, Wu K. Prediction of Lymphovascular invasion status in breast cancer based on magnetic resonance imaging radiomics features. Magn Reson Imaging 2024;109:91-5. 10.1016/j.mri.2024.03.008 [DOI] [PubMed] [Google Scholar]
- 36.Xu Z, Xie Y, Wu L, Chen M, Shi Z, Cui Y, Han C, Lin H, Liu Y, Li P, Chen X, Ding Y, Liu Z. Using Machine Learning Methods to Assess Lymphovascular Invasion and Survival in Breast Cancer: Performance of Combining Preoperative Clinical and MRI Characteristics. J Magn Reson Imaging 2023;58:1580-9. 10.1002/jmri.28647 [DOI] [PubMed] [Google Scholar]
- 37.Kayadibi Y, Kocak B, Ucar N, Akan YN, Yildirim E, Bektas S. MRI Radiomics of Breast Cancer: Machine Learning-Based Prediction of Lymphovascular Invasion Status. Acad Radiol 2022;29 Suppl 1:S126-34. [DOI] [PubMed] [Google Scholar]
- 38.Hu J, Xu J, Li M, Jiang Z, Mao J, Feng L, Miao K, Li H, Chen J, Bai Z, Li X, Lu G, Li Y. Identification and validation of an explainable prediction model of acute kidney injury with prognostic implications in critically ill children: a prospective multicenter cohort study. EClinicalMedicine 2024;68:102409. 10.1016/j.eclinm.2023.102409 [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
The article’s supplementary files as
Data Availability Statement
Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2685/dss