Multimodal automated diagnosis of lymphovascular invasion in breast cancer on contrast-enhanced MRI: ResUNet++ segmentation and Transformer-based classification

Multimodal automated diagnosis of lymphovascular invasion in breast cancer on contrast-enhanced MRI: ResUNet++ segmentation and Transformer-based classification | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Multimodal automated diagnosis of lymphovascular invasion in breast cancer on contrast-enhanced MRI: ResUNet++ segmentation and Transformer-based classification Junyu Lin, Zichang Ma, Yuxi Tao, Yun Liang, Yuhan Wei, Huajin Liu, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7867527/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Objectives To develop and evaluate an automated, multimodal Transformer model for preoperative prediction of lymphovascular invasion (LVI) in invasive breast cancer using contrast-enhanced MRI. Materials and Methods A retrospective study analyzed 288 patients with pathologically confirmed invasive breast cancer who all underwent preoperative dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). The study included an internal cohort (238 patients) and an external cohort (50 patients). Tumor regions of interest (ROI) were manually delineated by radiologists and automatic tumor segmentation was performed using ResUNet++. The output results were then processed with 4mm boundary dilation, while radiomic features were extracted and radiologists assessed MRI features according to BI-RADS criteria. Single-modality and multi-modality models were constructed for comparison, with the multi-modal fusion network integrating enhanced images, radiomic features, and MRI features. Model differences were assessed using DeLong test, and interpretability analysis was performed using Grad-CAM and SHAP methods. Results Automated segmentation was robust (Dice 0.916 internal and 0.921 external). The two-stage multimodal classifier achieved the highest AUC, 0.873 internally and 0.845 externally, compared with the best single-modality Transformer at 0.801 internally and 0.762 externally. Conclusion Integrating automated MRI segmentation with Transformer-based multimodal learning enables reliable preoperative LVI prediction and shows promising cross-center generalizability for clinical translation. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Breast cancer is one of the most common forms of cancer among women worldwide, with approximately 90% of breast cancer-related deaths attributed to tumor metastasis ( 1 ). LVI is a critical stage in breast cancer metastasis, referring to the infiltration of tumor cells into the lymphatic or blood vessels surrounding invasive cancer. This is a key step in the metastatic cascade, where cancer cells invade blood and lymphatic vessels, serving as the primary pathway for cancer cells to spread to secondary sites( 2 ). Regardless of lymph node involvement, LVI is considered as an independent factor for poor prognosis in invasive breast cancer ( 3 ). Tumor cell invasion through lymphatic and blood vessels is the main route of vascular invasion in breast cancer, and this process is closely associated with lymph node metastasis, the development of distant metastases, and reduces disease-free survival (DFS) and overall survival (OS) rates ( 4 ). Studies have shown that patients with lymphovascular invasion, particularly those with Luminal B, triple-negative, and Her-2 overexpressing breast cancer subtypes, have worse OS and DFS outcomes( 5 ), also including recurrence-free survival (RFS) in early-stage breast cancer patients( 6 , 7 ). An increasing number of studies emphasize the importance of LVI status in determining surgical interventions( 8 ), providing essential guidance for adjuvant therapies and suggesting optimal surgical margins. Radiotherapy studies have shown that retrospective evaluations of pathological reports reveal that the extent of lymphovascular invasion determines different radiotherapy modalities( 9 ). Additionally, another study demonstrated that endocrine therapy alone is insufficient to prevent distant metastases in breast cancer patients with lymphovascular invasion( 10 ). Therefore, the detection of LVI is crucial for the diagnosis and prognostic assessment of breast cancer patients, particularly in guiding treatment decisions and evaluating metastatic risk. However, accurately and noninvasively predicting LVI status before surgery remains a challenge. The gold standard for diagnosing LVI is obtained through postoperative pathology in breast cancer patients, but it is critical to develop treatment plans in advance of surgery ( 11 ). Multiparametric Magnetic Resonance Imaging (MRI) offers excellent soft tissue contrast and high anatomical resolution, making it the most sensitive method for breast cancer detection among current clinical imaging techniques( 12 , 13 ). Several previous studies have indicated that certain MRI morphological features (MRI-MF) are associated with lymphovascular invasion ( 14 , 15 ), such as mass margins, adjacent vessel sign, peritumoral edema, tumor size, background parenchymal enhancement, apparent diffusion coefficient patterns, and contrast enhancement patterns. However, the subjectivity involved in feature evaluation sometimes leads to inconsistent research findings. To address the prediction inaccuracies caused by subjective errors in MRI imaging feature assessments, it is crucial to introduce additional methods to assist radiologists in interpreting MRI images for LVI prediction. With the rapid development of artificial intelligence (AI) technologies ( 16 , 17 ), their application in medical imaging analysis has opened new possibilities for preoperative LVI prediction. However, studies have shown that although thousands of radiomics features are available for radiological analysis( 18 ), the hierarchical nature of texture features may fail to fully capture the potential information embedded in images ( 19 ). In radiomics analysis, the quality of ROI delineation directly impacts the accuracy and stability of feature extraction. However, ROI delineation currently relies heavily on the experience and subjective judgment of radiologists( 20 ), which inevitably leads to interobserver variability( 21 ). Furthermore, previous studies have utilized deep learning to construct multimodal convolutional network models for LVI prediction, integrating radiomics features from multiple modalities( 22 ). These studies have demonstrated the ability of deep learning to extract higher-level abstract imaging features. Deep learning has shown significant potential in preoperative LVI prediction, offering strong support for personalized treatment and prognostic evaluation in breast cancer patients. Transformer models, with their self-attention mechanisms, have the ability to capture global features and complex logical relationships, making them highly capable of extracting deeper-level features. Comparing Transformer models with existing multimodal convolutional network models could provide new insights and improved performance for LVI prediction( 23 ). Therefore, the aim of this study is to develop a multimodal Transformer model enhanced by automated segmentation of breast MRI, integrating radiological and radiomics features. The study will evaluate whether this model can serve as a preoperative auxiliary tool for predicting lymphovascular invasion in patients with invasive breast cancer. Materials and Methods Participants: This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013) . This retrospective study was approved by the Medical Ethics Committee of the Fifth Affiliated Hospital of Sun Yat-sen University (Approval No. K243-1). Given the retrospective use of medical records and biospecimens obtained during routine clinical care, the requirement for individual informed consent was waived by the ethics committee. The study dataset was collected from the Fifth Affiliated Hospital of Sun Yat-sen University in December 2024, including 238 patients who underwent preoperative DCE-MRI examination and were pathologically confirmed to have invasive breast cancer after surgery between June 2015 and December 2024. Additionally, 50 similar patients were collected from the Sun Yat-sen Memorial Hospital of Sun Yat-sen University, as outlined in the study workflow (Fig. 1 ). Figure 2 shows the inclusion and exclusion criteria. Inclusion criteria included: 1) Primary breast lesion visible on magnetic resonance imaging; 2) Solid tumor present on magnetic resonance imaging; 3) Newly diagnosed invasive breast cancer confirmed by pathological assessment of surgical specimens. Patients were excluded if they had: 1) Biopsy of breast lesions prior to MRI examination; 2) Preoperative chemotherapy or chemoradiotherapy; 3) Poor quality magnetic resonance images. MRI Acquisition: The study utilized two 3T MRI scanners: the MAGNETOM Verio manufactured by Siemens and the SIGNA Pioneer manufactured by General Electric. The scanning sequences primarily included axial fat-suppressed T2-weighted imaging (T2WI), T1-weighted imaging (T1WI), and dynamic contrast-enhanced MRI (DCE-MRI).The imaging parameters for T2WI were as follows: inversion time = 230 msec, repetition time/echo time (TR/TE) = 3800/72 msec, flip angle = 80°, field of view (FOV) = 249–319 mm, acquisition matrix = 384 × 210, averages = 2, and slice thickness = 4 mm. For T1WI, the parameters were: TR/TE = 3.99/1.3 msec, flip angle = 9°, FOV = 233–340 mm, acquisition matrix = 320 × 256, averages = 2, and slice thickness = 1.5 mm. Diffusion-weighted imaging (DWI) parameters were: b-values = 50 s/mm², TR/TE = 5450/68 msec, FOV = 159–340 mm, acquisition matrix = 128 × 64, and slice thickness = 4 mm. The parameters for DCE-MRI were: TR/TE = 7.67/4.25 msec, flip angle = 12°, FOV = 360–360 mm, acquisition matrix = 320 × 288, averages = 0.7, and slice thickness = 2 mm. For DCE-MRI, gadopentetate dimeglumine was used as the contrast agent, with a flow rate of 2 mL/s and a dose of 0.2 mL/kg. After contrast injection, seven repeated DCE-MRI scans were performed using the same parameters as the standard sequence, with each scan lasting 90 seconds. All image sequences were acquired with the patient in the prone position using a dedicated breast coil to enhance visualization. MRI Morphological Feature Acquisition: After anonymization and randomization, MRI features were independently evaluated by two experienced radiologists (Zichang Ma and Junyu Lin) following the standards of the American College of Radiology Breast Imaging Reporting and Data System (BI-RADS, 2023 Aug 28) to ensure unbiased assessments( 24 ). The evaluated MRI features included time-signal-intensity curves classified as persistent, plateau, or wash-out patterns based on DCE-MRI kinetics; fibroglandular tissue density and background parenchymal enhancement graded according to BI-RADS; peritumoral edema, defined as a rim of water-equivalent high signal surrounding the lesion on fat-suppressed T2-weighted images, with simultaneous assessment of subcutaneous edema and skin thickening; intratumoral high T2 signal on fat-suppressed T2-WI; the rim sign on diffusion-weighted imaging, recorded as complete when the hyperintense margin covered at least 90% of the tumor circumference and incomplete otherwise; the adjacent-vessel sign and increased ipsilateral vascularity; internal enhancement patterns categorised as homogeneous, heterogeneous, or rim-enhancing; MRI-based axillary lymph-node status considered positive when the short-axis diameter was 10 mm or greater or when central necrosis was present; intratumoral necrosis; and tumor-margin morphology, classified as regular or irregular. All features were independently assessed by two blinded radiologists, and inter-observer agreement was quantified using the kappa statistic. Region of Interest Acquisition: The delineation of ROI: The preprocessed images were subsequently imported into ITK-SNAP (version 4.0; http://www.itksnap.org/)(25) . Three radiologists with extensive experience in breast imaging (Zichang Ma, Junyu Lin, Huajin Liu) manually delineated a 3D regions of interest that includes the tumor (and all its approximations) to ensure optimal accuracy and reliability. Radiomics feature extraction and feature selection: Radiomics Feature Extraction and Selection: The feature values of patients need to be normalized by calculating their z-scores using the formula (x-µ)/σ, where x represents the feature value, µ is the mean of the feature values for all patients, and σ is the corresponding standard deviation( 26 ). Feature extraction was performed using FAE software(Version 0.5.12, https://github.com/salan668/FAE)(27) , with the enhanced images resampled to achieve a uniform voxel size of 1×1×1 mm. The analysis yielded first-order and intensity histogram statistics, 2D and 3D shape descriptors, and texture features, including gray-level co-occurrence matrix, gray-level run-length matrix, neighboring gray-tone difference matrix, gray-level dependence matrix, and gray-level size zone matrix. A total of 1,690 features were extracted from the ROI in each patient’s DCE-MRI series. Subsequently, a correlation analysis was performed to eliminate redundant features with a correlation coefficient ≥ 0.9. Lasso regression with five-fold cross-validation was then applied to automatically determine the optimal regularization parameter and feature coefficients, retaining only those features whose coefficients were non-zero. Development of automated segmentation models: To achieve automatic segmentation on breast MRI, we employed the ResUNet + + segmentation network as illustrated in Fig. 3 . This employs an encoder-decoder framework. In the encoder section, the model begins with the Conv3-32 layer for initial feature extraction, followed by a sequence of downsampling modules that incorporate Squeeze_Excite_Block and ResidualConv layers. These modules progressively increase the number of channels in the feature maps while reducing spatial resolution to effectively capture multi-scale information. At the deepest level, the ASPP-512 module is integrated to perform multi-scale feature extraction, thereby enhancing the model's receptive field. In the decoder section, upsampling operations are carried out using AttentionBlock and Upsample modules. Concurrently, skip connections are utilized to fuse feature maps from corresponding encoder layers with those in the decoder, ensuring the preservation of rich spatial detail information. Following each upsampling stage, ResidualConv layers are applied to further process the fused features, ensuring the effectiveness and stability of feature representations. Finally, the model generates segmentation results through the ASPP-32 module and the Output Conv-32-1 layer, with the outputs being normalized using a Sigmoid activation function. We first performed standardization preprocessing on the masked regions to ensure that the image intensity mean and standard deviation conform to the distribution of the breast MRI dataset. Subsequently, the images were resampled to a uniform dimension of 56×112×112 to standardize data dimensions and enhance training efficacy. Data augmentation was applied to the training set, including random rotations (± 30 degrees), horizontal and vertical flips (each with a 50% probability), brightness adjustments (range 0.8–1.2), and contrast adjustments (range 0.8–1.2). The Dice loss function was employed to optimize the overlap of the segmentation results. The Adam optimizer was used, and segmentation performance was validated using the Dice coefficient on the test set. Model training was implemented using Python 3.11.0 and PyTorch 2.0 ( https://pytorch.org ). The training process was conducted on a workstation equipped with an NVIDIA RTX 4090 GPU, Intel Core i9 processor, and 64GB of memory to ensure efficient computational performance. Deep Learning Model Development: To develop a deep learning (DL) model capable of diagnosing LVI, we applied two architectures: the traditional convolutional neural network (CNN) model ResNet101 and the transformer-based Swin Transformer, as shown in Fig. 3 . The transformer model begins by employing Patch Partition to divide the input image into several non-overlapping small patches. These patches are then processed through Patch Merging, which progressively downsamples the feature maps while increasing the number of channels. In Stage 1, following Linear Embedding, the features are passed through a Swin Transformer Block for initial feature extraction and modeling. Subsequent stages, Stage 2 through Stage 4, each consist of Patch Merging and multiple Swin Transformer Blocks, which further extract and integrate higher-level features. This hierarchical design, combined with the window-based self-attention mechanism, effectively captures both local and global information. Finally, the feature maps undergo Global Average Pooling to aggregate into a global feature vector, which is then transformed by a Linear Layer to produce the final output. In the data preprocessing stage, the bounding boxes were resampled to 224×224 pixels to serve as inputs for both DL models. To address the issue of insufficient training samples, both models employed transfer learning. ResNet utilized pre-trained ResNet101 weights from torchvision, while Swin Transformer used the swin-tiny-patch4-window7-224 weights for subsequent training. The Swin Transformer source code can be found at https://github.com/microsoft/Swin-Transformer . For ResNet101, only the final fully connected layer was fine-tuned, with the remaining layers kept frozen to leverage the feature extraction capabilities of the pre-trained model. In contrast, Swin Transformer fine-tuned all layers to fully adapt to the data distribution of this study. The optimizer used was AdamW with a weight decay coefficient set to 0.05 to prevent overfitting by reducing the weight magnitude through L2 regularization. The learning rate scheduling strategy employed was StepLR with an initial learning rate of 0.0001. The loss function used was the weighted cross-entropy loss function. Similarly, model training was implemented using Python 3.11.0 and PyTorch 2.0 ( https://pytorch.org ). The training process was conducted on a workstation equipped with an NVIDIA RTX 4090 GPU, Intel Core i9 processor, and 64GB of memory. All classification models are evaluated for differences using the DeLong test. To thoroughly analyze the prediction mechanism of neural network models and their contribution to LVI diagnosis, we employed Gradient-weighted Class Activation Mapping (Grad-CAM) and SHAP(Shapley Additive Explanations) analysis methods( 28 , 29 ). This approach extracts gradient features from the final output of convolutional layers, calculates weighted feature maps, and generates heat maps to assist in understanding the model's decision-making basis. Results Patients' Characteristics: A total of 288 patients were included: 167 were randomly assigned to the training cohort, 71 to the internal test cohort, and 50 constituted an independent external test cohort. Baseline characteristics were comparable between the training and internal test cohorts (all P > 0.05; Table 1 ). Table 1 Comparison of Baseline Characteristics Between the Training and Test Sets Level Training Set Test Set Independent Test Set P-value LVI Negative 112(67.1%) 47(66.2%) 32(64.0%) 0.922 Positive 55(32.9%) 24(33.8%) 18(36.0%) TIC 1 69(41.3%) 24(33.8%) 14(28.0%) 0.392 2 58(34.7%) 28(39.4%) 24(48.0%) 3 40(24.0%) 19(26.8%) 12(24.0%) Peritumoral Edema Negative 93(55.7%) 38(53.5%) 26(52.0%) 0.883 Positive 74(44.3%) 33(46.5%) 24(48.0%) Adjacent Vessel Sign Negative 127(76.0%) 49(69.0%) 37(74.0%) 0.527 Positive 40(24.0%) 22(31.0%) 13(26.0%) Age Median (min, max) 53.00 (min 24.00, max 84.00) 49.00 (min 27.00, max 84.00) 45.50 (min 30.00, max 78.00) 0.025 Menopausal Status Premenopausal 77(46.1%) 37(52.1%) 32(64.0%) 0.082 Postmenopausal 90(53.9%) 34(47.9%) 18(36.0%) Location Right 84(50.3%) 35(49.3%) 23(46.0%) 0.867 Left 83(49.7%) 36(50.7%) 27(54.0%) Symmetry Negative 105(62.9%) 47(66.2%) 35(70.0%) 0.63 Positive 62(37.1%) 24(33.8%) 15(30.0%) BPE None/minimal 101(60.5%) 47(66.2%) 32(64.0%) 0.781 Mild 45(26.9%) 17(23.9%) 15(30.0%) Moderate 14(8.4%) 6(8.5%) 2(4.0%) Marked 7(4.2%) 1(1.4%) 1(2.0%) Subcutaneous Edema Negative 125(74.9%) 51(71.8%) 37(74.0%) 0.889 Positive 42(25.1%) 20(28.2%) 13(26.0%) High Signal Intensity within Tumor Negative 115(68.9%) 47(66.2%) 33(66.0%) 0.886 Positive 52(31.1%) 24(33.8%) 17(34.0%) Internal Enhancement Pattern Heterogeneous 11(6.6%) 5(7.0%) 3(6.0%) 0.928 Homogeneous 133(79.6%) 55(77.5%) 42(84.0%) Rim Enhancement 23(13.8%) 11(15.5%) 5(10.0%) Increased Ipsilateral Vascular Distribution Negative 72(43.1%) 29(40.8%) 22(44.0%) 0.93 Positive 95(56.9%) 42(59.2%) 28(56.0%) Axillary Lymph Node Status Negative 100(59.9%) 48(67.6%) 29(58.0%) 0.458 Positive 67(40.1%) 23(32.4%) 21(42.0%) Intratumoral Necrosis Negative 133(79.6%) 58(81.7%) 41(82.0%) 0.898 Positive 34(20.4%) 13(18.3%) 9(18.0%) Tumor-margin Morphology Negative 67(40.1%) 36(50.7%) 27(54.0%) 0.124 Positive 100(59.9%) 35(49.3%) 23(46.0%) DWI Margin Sign Positive 85(50.9%) 32(45.1%) 22(44.0%) 0.572 Negative 82(49.1%) 39(54.9%) 28(56.0%) MRI Morphological Feature We found that patients with LVI(+) were more likely to develop peripheral tumor edema (OR: 2.06; 95% confidence interval [CI] 1.04–4.06) and DWI margin signal (OR: 2.34; 95% CI 1.19–4.63) in the training set (Table 2 ). Table 2 Univariate and Multivariate Logistic Analysis for LVI of Breast Cancer. Variable Name Single-factor OR (95%CI) p-value Multifactorial OR (95%CI) p-value TIC Curve 0.86 (0.57, 1.30) 0.472 Peritumoral Edema 2.07 (1.08, 3.99) 0.029 2.06 (1.04, 4.06) 0.037 Adjacent Vessel Sign 0.61 (0.27, 1.36) 0.224 Age 1.01 (0.99, 1.04) 0.324 Menstrual Status 0.75 (0.39, 1.43) 0.384 Location 0.70 (0.36, 1.33) 0.273 Symmetry 1.07 (0.55, 2.08) 0.843 BPE 0.85 (0.56, 1.28) 0.426 Subcutaneous Edema 1.02 (0.49, 2.15) 0.949 High Signal Intensity within Tumor 1.12 (0.56, 2.23) 0.756 Internal Enhancement Pattern 1.15 (0.56, 2.38) 0.699 Increased Ipsilateral Vascular Distribution 1.35 (0.70, 2.62) 0.368 Axillary Lymph Node Status 1.74 (0.90, 3.34) 0.099 1.69 (0.85, 3.35) 0.131 Intratumoral Necrosis 0.56 (0.24, 1.34) 0.195 Tumor Margin 1.13 (0.58, 2.19) 0.720 DWI Margin Sign 2.16 (1.12, 4.18) 0.022 2.34 (1.19, 4.63) 0.014 Selected Radiomics Features: At the optimal α = 0.04977, a total of 15 radiomics features were selected. Ranked by SHAP importance from highest to lowest, they were: original_firstorder_90Percentile, lbp-3D-m1_glszm_LowGrayLevelZoneEmphasis, squareroot_glszm_GrayLevelVariance, square_glrlm_RunEntropy, logarithm_glszm_LowGrayLevelZoneEmphasis, original_shape_Flatness, lbp-3D-m2_glrlm_RunEntropy, lbp-3D-k_glcm_Correlation, wavelet-HLH_firstorder_Skewness, original_shape_Elongation, exponential_gldm_LargeDependenceHighGrayLevelEmphasis, lbp-3D-k_glszm_LowGrayLevelZoneEmphasis, original_shape_LeastAxisLength, squareroot_firstorder_Kurtosis, and logarithm_glrlm_RunEntropy.(Fig. 4 ) Segmentation performance: The segmentation performance of the model on the dataset is summarized in Table 3 . On Test1 (N = 71), the model achieved a Dice coefficient of 0.838 ± 0.050 (95% CI: 0.826–0.849), an IoU of 0.724 ± 0.073 (95% CI: 0.707–0.740), a Precision of 0.879 ± 0.056 (95% CI: 0.866–0.892), and a Recall of 0.804 ± 0.073 (95% CI: 0.787–0.822). On Test2 (N = 50), the model demonstrated improved performance with a Dice coefficient of 0.867 ± 0.044 (95% CI: 0.854–0.878), an IoU of 0.768 ± 0.065 (95% CI: 0.749–0.784), a Precision of 0.923 ± 0.046 (95% CI: 0.910–0.935), and a Recall of 0.820 ± 0.059 (95% CI: 0.803–0.835). Some of the segmentation results are shown in Fig. 5 . Table 3 Segmentation performance of the method on the dataset (macro: case-level mean ± SD with 95% CI) N Dice IoU Precision Recall Test1 71 0.838 ± 0.050 [0.826, 0.849] 0.724 ± 0.073 [0.707, 0.740] 0.879 ± 0.056 [0.866, 0.892] 0.804 ± 0.073 [0.787, 0.822] Test2 50 0.867 ± 0.044 [0.854, 0.878] 0.768 ± 0.065 [0.749, 0.784] 0.923 ± 0.046 [0.910, 0.935] 0.820 ± 0.059 [0.803, 0.835] Model Validation: For single-modality models, the logistic-regression model based on handcrafted imaging features achieved an AUC(Area Under Curve) of 0.641 (95% CI 0.549–0.721) in the internal test set and 0.707 (95% CI 0.580–0.821) in the independent external test set. The standalone convolutional neural network (CNN) obtained an AUC of 0.741 (95% CI 0.576–0.884) internally and 0.702 (95% CI 0.515–0.880) externally. The radiomics model yielded an AUC of 0.770 (95% CI 0.643–0.870) in the internal test set but only 0.571 (95% CI 0.400–0.742) in the external cohort. The pure Transformer model reached an AUC of 0.801 (95% CI 0.665–0.918) internally and 0.762 (95% CI 0.576–0.918) externally. For multimodal approaches, the ResNet-based multimodal network achieved an AUC of 0.855 (95% CI 0.737–0.947) in the internal test set and 0.802 (95% CI 0.623–0.935) in the external test set.The multimodal architecture that combines ResNet-based segmentation with Transformer-based classification delivered the best results, with an AUC of 0.873 (95% CI 0.758–0.967) internally and 0.845 (95% CI 0.687–0.964) externally. The multimodal Transformer model with automated segmentation outperformed all standalone models (Fig. 6 ).The Grad-CAM heat map visualizes the attention regions of the Transformer model; SHAP quantifies the contribution weights of different features to the prediction of LVI positive or negative (Fig. 7 ). Discussion In this study, we explored the relationship between LVI and invasive breast cancer and developed a multimodal model combining automatic segmentation, MRI morphological features, radiomics features, and a Transformer-based deep learning framework to assess the likelihood of LVI in breast cancer patients. The combined model demonstrated higher diagnostic efficiency in distinguishing LVI-positive and LVI-negative cases compared to single models. In previous studies, MRI morphological features have been shown to predict LVI status in breast cancer patients( 30 ). However, the reliability of these results may be affected by interobserver variability due to differences in radiologists' experience levels. In this study, we identified only peritumoral edema and the DWI edge sign as independent indicators of LVI, consistent with previous findings. Peritumoral edema may result from tumor cell invasion into lymphatic vessels, obstructing normal lymphatic flow and drainage, or from damage to the integrity of vascular walls, increasing vascular permeability and allowing fluid to leak into surrounding tissues. The DWI edge sign may be caused by lymphatic and vascular invasion, leading to increased local cell density, edema, and stromal reactions in surrounding tissues, which restrict the free diffusion of water molecules. Radiomics features provide quantitative information about tumor phenotypes that are often imperceptible to the naked eye. Accordingly, we extracted radiomics features from the tumors and, after feature selection, trained machine-learning models. SHAP analysis showed that model decisions were primarily driven by texture-intensity composite features, including the upper gray-level quantile (original_firstorder_90Percentile), clustering of low-gray-level zones (lbp-3D-m1_glszm_LowGrayLevelZoneEmphasis), and texture run entropy (square_root_glrlm_RunEntropy). These metrics are strongly associated with intratumoral heterogeneity, suggesting that lesions with more complex tissue composition and more extreme gray-level distributions are more likely to be classified as high risk. Notably, the morphological features Flatness and Elongation also ranked among the top contributors, indicating that flatter and more elongated lesions have a higher probability of a positive prediction, which may reflect irregular shapes arising from infiltrative growth. Overall, increased image heterogeneity together with irregular morphology jointly underpins the model’s high-risk determinations. Manual ROI delineation is time-consuming and highly subjective, with significant variability between radiologists, which can affect the reliability of subsequent analyses. In this study, we employed a ResUNet + + automatic segmentation model. On Test1 (N = 71) and Test2 (N = 50), the model achieved Dice coefficients of 0.838 (95% CI: 0.826–0.849) and 0.867 (95% CI: 0.854–0.878), with corresponding IoU values of 0.724 (95% CI: 0.707–0.740) and 0.768 (95% CI: 0.749–0.784), respectively. Precision consistently exceeded Recall on both datasets (Test1: 0.879 vs. 0.804; Test2: 0.923 vs. 0.820), indicating the model's conservative tendency with fewer false positives and stable generalization capability. The automatic pipeline substantially reduces radiologists’ workload and improves the consistency of ROI delineation, providing a reliable foundation for subsequent model training. Introducing automatic segmentation is a major strength of this study, as it alleviates the subjectivity of manual contouring and lays the groundwork for developing multimodal models. Nevertheless, performance remains to be improved for tumors with indistinct boundaries. To address this, we post-processed the outputs by dilating the ROI margin by 4 mm and standardizing to a square input box to cover potentially uncertain boundary regions (see Fig. 5 ). To address this issue, we processed the automatic segmentation results by expanding the ROI boundaries by 4mm and using square input boxes, which resolved the boundary blurring problem. We also tested other segmentation networks, and found no significant differences in results among different segmentation networks after applying this processing method. Through systematic comparison of six models, this study demonstrated that multimodal frameworks can significantly enhance prediction performance. The ResNet multimodal network achieved AUCs of 0.855 (95% CI 0.737–0.947) and 0.802 (95% CI 0.623–0.935) in the internal and external test sets, respectively, while the two-stage model combining automatic residual segmentation output with Transformer classification performed even better, with AUCs climbing to 0.873 (95% CI 0.758–0.967) and 0.845 (95% CI 0.687–0.964), respectively. First, the self-attention mechanism enables the model to capture complex interdependencies between different imaging features that traditional machine learning approaches might overlook( 23 , 31 ). By integrating morphological, textural, and pharmacokinetic features through the Transformer architecture, the model can more comprehensively characterize tumor heterogeneity. Second, the robustness demonstrated in external validation proves the model's generalizability across different imaging protocols and patient populations, which is particularly notable given the known challenges in cross-institutional validation of radiomics models (AUC = 0.571). The modest performance decline from internal testing to external testing indicates good model robustness, although some performance degradation is expected due to differences in imaging parameters and patient characteristics. Compared with previous studies using traditional machine learning approaches for LVI prediction( 32 ), our Transformer model shows improvements in both discrimination ability and stability. The attention mechanism's ability to automatically weight feature importance may be the key reason why this approach outperforms traditional feature selection methods. We employed visualization techniques such as Grad-CAM and SHAP to intuitively depict the attention regions of the Transformer model and to provide preliminary estimates of the relative contributions of different features to predicting LVI-positive vs. LVI-negative status, thereby offering clinicians spatial localization of the model’s decision-making. However, due to postoperative specimen deformation, geometric distortions introduced during tissue processing, and misregistration between imaging examinations and pathological assessments across 2D/3D coordinate systems, it remains challenging to obtain a voxel-level, one-to-one image–pathology correspondence as a reference standard for gold-standard validation. Overall, the Transformer architecture not only leads in single-modality scenarios but further amplifies its advantages in automated segmentation-driven multi-modal approaches, providing a solution pathway that combines both performance and interpretability for non-invasive assessment of LVI. Limitation: There are several limitations to our study that should be considered. First, our radiomics features were extracted only from the second phase of DCE-MRI and did not utilize data from other MRI sequences such as T1WI, T2WI, DWI and ADC maps. Second, we did not extract radiomics features solely from the peritumoral region, although our deep learning approach covered the peritumoral area, it may still result in the loss of some multimodal information. Declarations Ethics approval and consent to participate: Ethical approval for this retrospective study was granted by the Medical Ethics Committee of the Fifth Affiliated Hospital of Sun Yat-sen University (Approval No. K243-1), in accordance with the Declaration of Helsinki (as revised in 2013). The ethics committee waived the requirement for individual informed consent given the retrospective nature of this study using routinely collected clinical data and biospecimens. Consent for publication: Not applicable. Availability of data and materials: The datasets generated and/or analysed during the current study are not publicly available due to patient privacy and institutional data protection policies but are available from the corresponding author on reasonable request. Competing interests: The authors declare that they have no competing interests. Funding: This work was supported by the National Natural Science Foundation of China under Grant Nos. 81801809 and 82371917, the Basic and Applied Basic Research Foundation of Guangdong Province under Grant No. 2020A1515010572, and the Zhuhai Basic and Applied Basic Research Project Foundation under Grant No. ZH22017003200001PWC. Authors' contributions: J.L. and Z.M. contributed equally to this work, performed data analysis and wrote the main manuscript text. Y.T. and Y.L. contributed to data interpretation. J.L., Y.W., and H.L. were responsible for data collection and organization. Y.Z. designed and supervised the overall study. All authors read and approved the final manuscript. Acknowledgements: Not applicable. References Siegel RL, Kratzer TB, Giaquinto AN, Sung H, Jemal A. Cancer statistics, 2025. CA Cancer J Clin. 2025;75(1):10-45. Null JL, Kim DJ, McCann JV, Pramoonjago P, Fox JW, Zeng J, et al. Periostin+ Stromal Cells Guide Lymphovascular Invasion by Cancer Cells. Cancer Res. 2023;83(13):2105-22. Liu G, Kong X, Dai Q, Cheng H, Wang J, Gao J, et al. Clinical Features and Prognoses of Patients With Breast Cancer Who Underwent Surgery. JAMA Netw Open. 2023;6(8):e2331078. Kennedy WR, Tricarico C, Gabani P, Weiner AA, Altman MB, Ochoa LL, et al. Predictors of Distant Metastases in Triple-Negative Breast Cancer Without Pathologic Complete Response After Neoadjuvant Chemotherapy. J Natl Compr Canc Netw. 2020;18(3):288-96. Rakha EA, Martin S, Lee AH, Morgan D, Pharoah PD, Hodi Z, et al. The prognostic significance of lymphovascular invasion in invasive breast carcinoma. Cancer. 2012;118(15):3670-80. Houvenaeghel G, Cohen M, Classe JM, Reyal F, Mazouni C, Chopin N, et al. Lymphovascular invasion has a significant prognostic impact in patients with early breast cancer, results from a large, national, multicenter, retrospective cohort study. ESMO Open. 2021;6(6):100316. Zhang S, Zhang D, Gong M, Wen L, Liao C, Zou L. High lymphatic vessel density and presence of lymphovascular invasion both predict poor prognosis in breast cancer. BMC Cancer. 2017;17(1):335. Chauhan H, Jiwa N, Nagarajan VR, Thiruchelvam P, Hogben K, Al-Mufti R, et al. Clinicopathological Predictors of Positive Resection Margins in Breast-Conserving Surgery. Ann Surg Oncol. 2024;31(6):3939-47. Strauss JB. Extensive Lymphovascular Invasion Warrants Extensive Radiation Therapy. Int J Radiat Oncol Biol Phys. 2020;107(2):234-5. Kurebayashi J, Kanomata N, Shimo T, Yamashita T, Aogi K, Nishimura R, et al. Marked lymphovascular invasion, progesterone receptor negativity, and high Ki67 labeling index predict poor outcome in breast cancer patients treated with endocrine therapy alone. Breast Cancer. 2014;21(2):214-22. Nitz UA, Gluz O, Christgen M, Grischke EM, Augustin D, Kuemmel S, et al. De-escalation strategies in HER2-positive early breast cancer (EBC): final analysis of the WSG-ADAPT HER2+/HR- phase II trial: efficacy, safety, and predictive markers for 12 weeks of neoadjuvant dual blockade with trastuzumab and pertuzumab ± weekly paclitaxel. Ann Oncol. 2017;28(11):2768-72. Mann RM, Cho N, Moy L. Breast MRI: State of the Art. Radiology. 2019;292(3):520-36. Chen Y, Wang L, Luo R, Liu H, Zhang Y, Wang D. Focal breast edema and breast edema score on T2-weighted images provides valuable biological information for invasive breast cancer. Insights Imaging. 2023;14(1):73. Cheon H, Kim HJ, Lee SM, Cho SH, Shin KM, Kim GC, et al. Preoperative MRI features associated with lymphovascular invasion in node-negative invasive breast cancer: A propensity-matched analysis. J Magn Reson Imaging. 2017;46(4):1037-44. Acciavatti RJ, Lee SH, Reig B, Moy L, Conant EF, Kontos D, et al. Beyond Breast Density: Risk Measures for Breast Cancer in Multiple Imaging Modalities. Radiology. 2023;306(3):e222575. Bera K, Braman N, Gupta A, Velcheti V, Madabhushi A. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol. 2022;19(2):132-46. Hao J, Liu S, Wang T, Han X, Gao A, Wang H, et al. Differentiation of malignant from benign soft tissue tumors using radiomics based on pharmacokinetic parameter maps obtained from dynamic contrast-enhanced magnetic resonance imaging data. Chinese Journal of Academic Radiology. 2024;7(3):219-28. Bi WL, Hosny A, Schabath MB, Giger ML, Birkbak NJ, Mehrtash A, et al. Artificial intelligence in cancer imaging: Clinical challenges and applications. CA Cancer J Clin. 2019;69(2):127-57. Tomaszewski MR, Gillies RJ. The Biological Meaning of Radiomic Features. Radiology. 2021;298(3):505-16. Zhang Y, Chan S, Park VY, Chang KT, Mehta S, Kim MJ, et al. Automatic Detection and Segmentation of Breast Cancer on MRI Using Mask R-CNN Trained on Non-Fat-Sat Images and Tested on Fat-Sat Images. Acad Radiol. 2022;29 Suppl 1(Suppl 1):S135-s44. Hagiwara A, Fujita S, Ohno Y, Aoki S. Variability and Standardization of Quantitative Imaging: Monoparametric to Multiparametric Quantification, Radiomics, and Artificial Intelligence. Invest Radiol. 2020;55(9):601-16. Yang X, Fan X, Lin S, Zhou Y, Liu H, Wang X, et al. 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(6):2238-49. Li J, Chen J, Tang Y, Wang C, Landman BA, Zhou SK. Transforming medical imaging with Transformers? A comparative review of key properties, current progresses, and future perspectives. Med Image Anal. 2023;85:102762. Magny SJ, Shikhman R, Keppke AL. Breast Imaging Reporting and Data System. StatPearls. Treasure Island (FL) ineligible companies. Disclosure: Rachel Shikhman declares no relevant financial relationships with ineligible companies. Disclosure: Ana Keppke declares no relevant financial relationships with ineligible companies.: StatPearls Publishing Copyright © 2025, StatPearls Publishing LLC.; 2025. Yushkevich PA, Piven J, Hazlett HC, Smith RG, Ho S, Gee JC, et al. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage. 2006;31(3):1116-28. Traverso A, Wee L, Dekker A, Gillies R. Repeatability and Reproducibility of Radiomic Features: A Systematic Review. Int J Radiat Oncol Biol Phys. 2018;102(4):1143-58. Song Y, Zhang J, Zhang YD, Hou Y, Yan X, Wang Y, et al. FeAture Explorer (FAE): A tool for developing and comparing radiomics models. PLoS One. 2020;15(8):e0237587. Chen Y, Xiao Z, Pan Y, Zhao L, Dai H, Wu Z, et al. Mask-Guided Vision Transformer for Few-Shot Learning. IEEE Trans Neural Netw Learn Syst. 2025;36(5):9636-47. Singh A, Sengupta S, Lakshminarayanan V. Explainable Deep Learning Models in Medical Image Analysis. J Imaging. 2020;6(6). Xu Z, Xie Y, Wu L, Chen M, Shi Z, Cui Y, et al. 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(5):1580-9. Sun W, Qin Z, Deng H, Wang J, Zhang Y, Zhang K, et al. Vicinity Vision Transformer. IEEE Trans Pattern Anal Mach Intell. 2023;45(10):12635-49. Zhang J, Wang G, Ren J, Yang Z, Li D, Cui Y, et al. 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(6):4079-89. Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterial.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 24 Feb, 2026 Reviewers agreed at journal 15 Feb, 2026 Reviewers agreed at journal 10 Feb, 2026 Reviewers invited by journal 11 Nov, 2025 Editor invited by journal 22 Oct, 2025 Editor assigned by journal 21 Oct, 2025 Submission checks completed at journal 21 Oct, 2025 First submitted to journal 15 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7867527","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":543284348,"identity":"f20e09ca-8694-4667-bf93-0a54b3fdbbff","order_by":0,"name":"Junyu Lin","email":"","orcid":"","institution":"Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Junyu","middleName":"","lastName":"Lin","suffix":""},{"id":543284349,"identity":"765c1336-c1ea-4cf9-a307-103aa8eb9dfa","order_by":1,"name":"Zichang Ma","email":"","orcid":"","institution":"Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Zichang","middleName":"","lastName":"Ma","suffix":""},{"id":543284350,"identity":"7796faa2-f0b2-4ad6-8795-a2764ee4067e","order_by":2,"name":"Yuxi Tao","email":"","orcid":"","institution":"Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Yuxi","middleName":"","lastName":"Tao","suffix":""},{"id":543284351,"identity":"a5d16675-8139-4b7e-829b-1286e4b30d46","order_by":3,"name":"Yun Liang","email":"","orcid":"","institution":"Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Yun","middleName":"","lastName":"Liang","suffix":""},{"id":543284352,"identity":"92e89031-6a62-405d-a504-e4455dba9936","order_by":4,"name":"Yuhan Wei","email":"","orcid":"","institution":"Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Yuhan","middleName":"","lastName":"Wei","suffix":""},{"id":543284353,"identity":"40be5382-f856-448a-a8d8-38df23c3a741","order_by":5,"name":"Huajin Liu","email":"","orcid":"","institution":"Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Huajin","middleName":"","lastName":"Liu","suffix":""},{"id":543284354,"identity":"ddaf9b5e-5350-4b6f-92e3-615b83c52f45","order_by":6,"name":"Yaqin Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDElEQVRIiWNgGAWjYDACCRBhwCDHwMDYeABFkJAWY6CWBlK0MDAkNgAJ4rTIz+4x/PCm4E762vbDQFv+HLY3OMB88DYPg10eLi2Mc84YS84xeJa77UxiwwHGtsOJGw6wJVvzMCQX49LCLJFjIM1jcDh32wGQlobDCQYHeMykeRgOgJ2KDbBJ5Bj/BmpJNzv/EOYw/m94tfBI5JiBbEkwuwG0hYHtMOOGAzxseLVISKSVWc4xOGy47QbQlsS29MSZh9mMgSLJOLXIz0jefOPNn8PyZufTHz748Mfanu9488MbbyrscGphYOAwYOCBsRMYmoEhAmIZ4FQPBOwPEFoYGOrwKR0Fo2AUjIIRCgCDrl0Xp1jZTAAAAABJRU5ErkJggg==","orcid":"","institution":"Sun Yat-sen University","correspondingAuthor":true,"prefix":"","firstName":"Yaqin","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2025-10-15 11:23:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7867527/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7867527/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96491037,"identity":"28cf1cf2-622d-4df8-b7a4-0c4b9cad6e3c","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"jpeg","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1479869,"visible":true,"origin":"","legend":"","description":"","filename":"figure1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/3bbc137f8b030432b86e0bf5.jpeg"},{"id":96491035,"identity":"545934ed-f62a-45d5-bc48-736d146aacb0","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"jpeg","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":785867,"visible":true,"origin":"","legend":"","description":"","filename":"figure2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/4ecbf7a8739146e7a14edce1.jpeg"},{"id":96603986,"identity":"6a0dcd98-c76b-44f0-8b4e-5bdf0eebae48","added_by":"auto","created_at":"2025-11-24 09:12:20","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2239989,"visible":true,"origin":"","legend":"","description":"","filename":"lvimanuscriptbmcv2.1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/ff760e14f9dfd6fcacae8bdd.docx"},{"id":96603545,"identity":"8aea558f-df8d-462f-a64f-5710c4fed63e","added_by":"auto","created_at":"2025-11-24 09:09:56","extension":"jpeg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1348337,"visible":true,"origin":"","legend":"","description":"","filename":"figure3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/8ad5188b896f65e1374c7d65.jpeg"},{"id":96491048,"identity":"a236bb67-623f-455a-937c-d057a45b457e","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1045870,"visible":true,"origin":"","legend":"","description":"","filename":"figure4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/d06bc50e314df9b3a635a1f3.jpeg"},{"id":96491043,"identity":"c42edf3c-b5a1-4420-b024-aca03163e497","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":610784,"visible":true,"origin":"","legend":"","description":"","filename":"figure5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/b12cade6f8cd96f0e5af4906.jpeg"},{"id":96491052,"identity":"7db7c533-e9af-450a-93ea-fa1bcda34063","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1176332,"visible":true,"origin":"","legend":"","description":"","filename":"figure6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/4aef218a5d333ebf665ba105.jpeg"},{"id":96602987,"identity":"3cc1d52a-70aa-48d4-a92d-14547e7cc956","added_by":"auto","created_at":"2025-11-24 09:05:47","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":983643,"visible":true,"origin":"","legend":"","description":"","filename":"figure7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/1c80c524524722146121d9b2.jpeg"},{"id":96491054,"identity":"45841dae-7a2d-44a1-a8d3-db3d9bb5aefd","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"json","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8361,"visible":true,"origin":"","legend":"","description":"","filename":"06172c7dfb254793a5447ec2466c9d3d.json","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/f16a9ee06153d27e27a2bec0.json"},{"id":96491063,"identity":"61b8629f-4050-4c4a-bc53-86abc5e6d892","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"docx","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":317525,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/2617b4bd22b09d385c622792.docx"},{"id":96491049,"identity":"2dbcf1c0-7794-4060-9717-2253ee201553","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"xml","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":125252,"visible":true,"origin":"","legend":"","description":"","filename":"06172c7dfb254793a5447ec2466c9d3d1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/d25c2f9d59ba835527fbeaf0.xml"},{"id":96491061,"identity":"e21b1855-c5f6-479c-a7ab-6c7c8ff87fed","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"jpeg","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1479869,"visible":true,"origin":"","legend":"","description":"","filename":"figure1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/a351e7ea448f24736f187764.jpeg"},{"id":96602961,"identity":"5b83d7d1-bed8-43d1-b38c-2b62ca651bef","added_by":"auto","created_at":"2025-11-24 09:05:33","extension":"jpeg","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":785867,"visible":true,"origin":"","legend":"","description":"","filename":"figure2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/a0b533af57be200481c17cf4.jpeg"},{"id":96602974,"identity":"d2075e7e-4078-4458-b57f-cf4a87703355","added_by":"auto","created_at":"2025-11-24 09:05:40","extension":"jpeg","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1348337,"visible":true,"origin":"","legend":"","description":"","filename":"figure3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/fc7351f4fdc94622b2e84564.jpeg"},{"id":96491065,"identity":"50e735f8-995f-4c7c-973d-0b6979e5c330","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"jpeg","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1045870,"visible":true,"origin":"","legend":"","description":"","filename":"figure4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/5c13a6789454210d64352a7b.jpeg"},{"id":96603034,"identity":"555bf7dc-7d43-4084-a037-e5a76b6c23b8","added_by":"auto","created_at":"2025-11-24 09:06:17","extension":"jpeg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":610784,"visible":true,"origin":"","legend":"","description":"","filename":"figure5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/1040a9968c4a1eba08a7cb18.jpeg"},{"id":96603793,"identity":"44008e8d-acdf-4704-9e46-f059c39d6528","added_by":"auto","created_at":"2025-11-24 09:11:34","extension":"jpeg","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1176332,"visible":true,"origin":"","legend":"","description":"","filename":"figure6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/4564bab7e67d88a92f615e1d.jpeg"},{"id":96491079,"identity":"269bd292-1437-4041-9b62-8d0911e72f8b","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"jpeg","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":983643,"visible":true,"origin":"","legend":"","description":"","filename":"figure7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/8210671a2bd9a58bfaed5da0.jpeg"},{"id":96491056,"identity":"6fbc32aa-54fa-4839-b164-525c629605ad","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":148623,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/d89c591e7050ff1c518d3d49.png"},{"id":96491053,"identity":"baf1b811-7a5c-4b85-9a84-9f3a512c403d","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":56640,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/c543f8dbd46ec26741392562.png"},{"id":96603646,"identity":"4b9a2184-f84d-419e-85be-76f9c476fb85","added_by":"auto","created_at":"2025-11-24 09:10:51","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":213623,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/78d2cb5df216555994b7bc6c.png"},{"id":96491059,"identity":"f055442e-2e92-4e55-ad24-4b6ade648670","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1063032,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/99eb332a2d84ae6897963fad.png"},{"id":96604048,"identity":"586414ef-0b64-459e-a194-5f977ede0fc3","added_by":"auto","created_at":"2025-11-24 09:12:32","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":80359,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/d5f52fa785a678cf7ed8f7ec.png"},{"id":96603575,"identity":"c8815d62-27e4-46a6-a83d-0bc9b4ef07fa","added_by":"auto","created_at":"2025-11-24 09:10:17","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":78319,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/9e8405d5448acd3f272a3d45.png"},{"id":96603771,"identity":"7423777e-83b3-456c-ac85-63c528f1fdc2","added_by":"auto","created_at":"2025-11-24 09:11:29","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":509257,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/9344a09b6639bf7b31e848ae.png"},{"id":96603629,"identity":"c5bcbfa4-cb2a-41fd-ae7d-4a772f84fa57","added_by":"auto","created_at":"2025-11-24 09:10:46","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":431934,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/38f039f280db6b902c625a52.png"},{"id":96491064,"identity":"4f9b57ab-3c85-42f7-8a76-a59f887b4bd2","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":549200,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/4576af77321b26a90e21f8c0.png"},{"id":96604020,"identity":"dd3d52b9-56e2-411d-ace3-e52d28158476","added_by":"auto","created_at":"2025-11-24 09:12:27","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":390159,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/755aa428936500aef076eea4.png"},{"id":96491075,"identity":"9b88941d-1189-42f4-bc13-9c4304ffbe1c","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":327958,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/42dd6a2326bae2b32f9ff8a3.png"},{"id":96603559,"identity":"e0558840-2f22-46bc-b9e7-832fce772440","added_by":"auto","created_at":"2025-11-24 09:10:11","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":166668,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/800be47eff3c3e7654d0bc28.png"},{"id":96491078,"identity":"131d3647-9a14-4843-ad0d-1e85fa59d467","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":317762,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/45e276ef8eeb1d05bcaca7d1.png"},{"id":96491073,"identity":"661c1135-b0ac-47ce-94c4-2bed876302c5","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":461096,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure7.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/19043b9fcae1ca3d08d4b9a4.png"},{"id":96491081,"identity":"5692d30b-0a3d-4119-b6bc-1e64f9808d61","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":37690,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/c619e0c502173feec23cfa5a.png"},{"id":96491067,"identity":"f20b4b08-48e1-46fa-8823-e349c35e92b9","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":27094,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/436a9548f9d40353cceab838.png"},{"id":96491070,"identity":"a748d412-cfbd-46f5-9cde-08d054012bc0","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":53335,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/e31f254ebfa385d2b471fcac.png"},{"id":96491069,"identity":"7a7820ed-f0c9-4d05-9876-a4c490a96f30","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":247836,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/886e407a6133bc59199b1809.png"},{"id":96491068,"identity":"436c6865-e0c1-42c9-85fc-eac9d3f4cf05","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17430,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/465ddd676246337bc672d40a.png"},{"id":96491083,"identity":"36bff31d-10e9-4eb7-b5eb-2da04ec5e462","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":37,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":23690,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/75111f3c72b4e38777c7d125.png"},{"id":96491074,"identity":"a2b4d8c5-e6c3-48c5-af25-afa25f2a1648","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"png","order_by":38,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":60922,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/0472e2d086e7ab4cfe86548e.png"},{"id":96603883,"identity":"03a10fa5-8361-49dd-9bbc-452e7ea5a23d","added_by":"auto","created_at":"2025-11-24 09:11:56","extension":"xml","order_by":39,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":123251,"visible":true,"origin":"","legend":"","description":"","filename":"06172c7dfb254793a5447ec2466c9d3d1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/7cc4ae7fbe7a1c86dfa53f1c.xml"},{"id":96491077,"identity":"82ea2c9c-6239-4ac4-8416-c9b97de9c55d","added_by":"auto","created_at":"2025-11-21 17:49:45","extension":"html","order_by":40,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":131852,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/19a5501344d274be650354fc.html"},{"id":96603191,"identity":"a876a137-2c10-4ec9-9901-838451d1c0f8","added_by":"auto","created_at":"2025-11-24 09:07:24","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1479869,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of the entire research design.\u003c/p\u003e","description":"","filename":"figure1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/a73ebc945a53ee290778192d.jpeg"},{"id":96603566,"identity":"0d2f72fe-cd26-406c-a787-92535592d327","added_by":"auto","created_at":"2025-11-24 09:10:15","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":785867,"visible":true,"origin":"","legend":"\u003cp\u003eThe inclusion and exclusion criteria.\u003c/p\u003e","description":"","filename":"figure2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/333b0187ac9b87cda0f33406.jpeg"},{"id":96491039,"identity":"754af806-463d-4b24-8765-2dbbc3d4d461","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1348337,"visible":true,"origin":"","legend":"\u003cp\u003eThe structure of deep learning. (a:ResUnet, b:Swin Transformer)\u003c/p\u003e","description":"","filename":"figure3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/a3b2edcf9470a8b7dbf68687.jpeg"},{"id":96602914,"identity":"5e2e88ea-b1bf-493e-8ece-06b22f91bd06","added_by":"auto","created_at":"2025-11-24 09:04:40","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1045870,"visible":true,"origin":"","legend":"\u003cp\u003eImportance of radiomic features.\u003c/p\u003e","description":"","filename":"figure4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/f5416381800ae2123a8c0663.jpeg"},{"id":96491041,"identity":"414fcfa8-4284-4a7f-a76b-9a7f784d4f54","added_by":"auto","created_at":"2025-11-21 17:49:44","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":610784,"visible":true,"origin":"","legend":"\u003cp\u003eResUNet++ segmentation and post-processing.\u003c/p\u003e","description":"","filename":"figure5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/1d639062b791b5d3e64000b5.jpeg"},{"id":96603848,"identity":"feeadde6-a930-458f-a427-08a23a72fdcd","added_by":"auto","created_at":"2025-11-24 09:11:47","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1176332,"visible":true,"origin":"","legend":"\u003cp\u003eThe performance evaluation of different models.\u003c/p\u003e","description":"","filename":"figure6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/3cd061c264cf845f5e510574.jpeg"},{"id":96603838,"identity":"311e368a-1233-4fd2-9005-b89acd792e36","added_by":"auto","created_at":"2025-11-24 09:11:44","extension":"jpeg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":983643,"visible":true,"origin":"","legend":"\u003cp\u003eModel interpretability analysis. (a), (c) are LVI-positive patients, (b), (d) are LVI-negative patients. (a), (b) show model attention region heatmaps based on Grad-CAM method; (c), (d) display feature contribution distribution for LVI prediction based on SHAP method. Color bars represent activation intensity (top) and SHAP values (bottom).\u003c/p\u003e","description":"","filename":"figure7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/1c73b019fa261d6559c49c73.jpeg"},{"id":96709042,"identity":"e9bafdaa-438d-44ef-9c27-a282066c4408","added_by":"auto","created_at":"2025-11-25 10:07:17","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":8434527,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/db4361bc-98bd-4bd8-bed7-022c21578d4c.pdf"},{"id":96603581,"identity":"6d5fc724-cc85-4e1c-967a-ba768c0ca3d1","added_by":"auto","created_at":"2025-11-24 09:10:20","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":317525,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-7867527/v1/3e4e81d95005ebfe8e92e493.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Multimodal automated diagnosis of lymphovascular invasion in breast cancer on contrast-enhanced MRI: ResUNet++ segmentation and Transformer-based classification","fulltext":[{"header":"Introduction","content":"\u003cp\u003eBreast cancer is one of the most common forms of cancer among women worldwide, with approximately 90% of breast cancer-related deaths attributed to tumor metastasis (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). LVI is a critical stage in breast cancer metastasis, referring to the infiltration of tumor cells into the lymphatic or blood vessels surrounding invasive cancer. This is a key step in the metastatic cascade, where cancer cells invade blood and lymphatic vessels, serving as the primary pathway for cancer cells to spread to secondary sites(\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). Regardless of lymph node involvement, LVI is considered as an independent factor for poor prognosis in invasive breast cancer (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). Tumor cell invasion through lymphatic and blood vessels is the main route of vascular invasion in breast cancer, and this process is closely associated with lymph node metastasis, the development of distant metastases, and reduces disease-free survival (DFS) and overall survival (OS) rates (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e). Studies have shown that patients with lymphovascular invasion, particularly those with Luminal B, triple-negative, and Her-2 overexpressing breast cancer subtypes, have worse OS and DFS outcomes(\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e), also including recurrence-free survival (RFS) in early-stage breast cancer patients(\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eAn increasing number of studies emphasize the importance of LVI status in determining surgical interventions(\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e), providing essential guidance for adjuvant therapies and suggesting optimal surgical margins. Radiotherapy studies have shown that retrospective evaluations of pathological reports reveal that the extent of lymphovascular invasion determines different radiotherapy modalities(\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e). Additionally, another study demonstrated that endocrine therapy alone is insufficient to prevent distant metastases in breast cancer patients with lymphovascular invasion(\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e). Therefore, the detection of LVI is crucial for the diagnosis and prognostic assessment of breast cancer patients, particularly in guiding treatment decisions and evaluating metastatic risk. However, accurately and noninvasively predicting LVI status before surgery remains a challenge. The gold standard for diagnosing LVI is obtained through postoperative pathology in breast cancer patients, but it is critical to develop treatment plans in advance of surgery (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eMultiparametric Magnetic Resonance Imaging (MRI) offers excellent soft tissue contrast and high anatomical resolution, making it the most sensitive method for breast cancer detection among current clinical imaging techniques(\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Several previous studies have indicated that certain MRI morphological features (MRI-MF) are associated with lymphovascular invasion (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e), such as mass margins, adjacent vessel sign, peritumoral edema, tumor size, background parenchymal enhancement, apparent diffusion coefficient patterns, and contrast enhancement patterns. However, the subjectivity involved in feature evaluation sometimes leads to inconsistent research findings. To address the prediction inaccuracies caused by subjective errors in MRI imaging feature assessments, it is crucial to introduce additional methods to assist radiologists in interpreting MRI images for LVI prediction.\u003c/p\u003e\u003cp\u003eWith the rapid development of artificial intelligence (AI) technologies (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e), their application in medical imaging analysis has opened new possibilities for preoperative LVI prediction. However, studies have shown that although thousands of radiomics features are available for radiological analysis(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e), the hierarchical nature of texture features may fail to fully capture the potential information embedded in images (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e). In radiomics analysis, the quality of ROI delineation directly impacts the accuracy and stability of feature extraction. However, ROI delineation currently relies heavily on the experience and subjective judgment of radiologists(\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e), which inevitably leads to interobserver variability(\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eFurthermore, previous studies have utilized deep learning to construct multimodal convolutional network models for LVI prediction, integrating radiomics features from multiple modalities(\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). These studies have demonstrated the ability of deep learning to extract higher-level abstract imaging features. Deep learning has shown significant potential in preoperative LVI prediction, offering strong support for personalized treatment and prognostic evaluation in breast cancer patients. Transformer models, with their self-attention mechanisms, have the ability to capture global features and complex logical relationships, making them highly capable of extracting deeper-level features. Comparing Transformer models with existing multimodal convolutional network models could provide new insights and improved performance for LVI prediction(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eTherefore, the aim of this study is to develop a multimodal Transformer model enhanced by automated segmentation of breast MRI, integrating radiological and radiomics features. The study will evaluate whether this model can serve as a preoperative auxiliary tool for predicting lymphovascular invasion in patients with invasive breast cancer.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eParticipants:\u003c/h2\u003e\u003cp\u003eThis study was conducted in accordance with the \u003cb\u003eDeclaration of Helsinki (as revised in 2013)\u003c/b\u003e. This retrospective study was approved by the Medical Ethics Committee of the Fifth Affiliated Hospital of Sun Yat-sen University (Approval No. K243-1). Given the retrospective use of medical records and biospecimens obtained during routine clinical care, the requirement for individual informed consent was waived by the ethics committee. The study dataset was collected from the Fifth Affiliated Hospital of Sun Yat-sen University in December 2024, including 238 patients who underwent preoperative DCE-MRI examination and were pathologically confirmed to have invasive breast cancer after surgery between June 2015 and December 2024. Additionally, 50 similar patients were collected from the Sun Yat-sen Memorial Hospital of Sun Yat-sen University, as outlined in the study workflow (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the inclusion and exclusion criteria. Inclusion criteria included: 1) Primary breast lesion visible on magnetic resonance imaging; 2) Solid tumor present on magnetic resonance imaging; 3) Newly diagnosed invasive breast cancer confirmed by pathological assessment of surgical specimens. Patients were excluded if they had: 1) Biopsy of breast lesions prior to MRI examination; 2) Preoperative chemotherapy or chemoradiotherapy; 3) Poor quality magnetic resonance images.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eMRI Acquisition:\u003c/h3\u003e\n\u003cp\u003eThe study utilized two 3T MRI scanners: the MAGNETOM Verio manufactured by Siemens and the SIGNA Pioneer manufactured by General Electric. The scanning sequences primarily included axial fat-suppressed T2-weighted imaging (T2WI), T1-weighted imaging (T1WI), and dynamic contrast-enhanced MRI (DCE-MRI).The imaging parameters for T2WI were as follows: inversion time\u0026thinsp;=\u0026thinsp;230 msec, repetition time/echo time (TR/TE)\u0026thinsp;=\u0026thinsp;3800/72 msec, flip angle\u0026thinsp;=\u0026thinsp;80\u0026deg;, field of view (FOV)\u0026thinsp;=\u0026thinsp;249\u0026ndash;319 mm, acquisition matrix\u0026thinsp;=\u0026thinsp;384 \u0026times; 210, averages\u0026thinsp;=\u0026thinsp;2, and slice thickness\u0026thinsp;=\u0026thinsp;4 mm. For T1WI, the parameters were: TR/TE\u0026thinsp;=\u0026thinsp;3.99/1.3 msec, flip angle\u0026thinsp;=\u0026thinsp;9\u0026deg;, FOV\u0026thinsp;=\u0026thinsp;233\u0026ndash;340 mm, acquisition matrix\u0026thinsp;=\u0026thinsp;320 \u0026times; 256, averages\u0026thinsp;=\u0026thinsp;2, and slice thickness\u0026thinsp;=\u0026thinsp;1.5 mm. Diffusion-weighted imaging (DWI) parameters were: b-values\u0026thinsp;=\u0026thinsp;50 s/mm\u0026sup2;, TR/TE\u0026thinsp;=\u0026thinsp;5450/68 msec, FOV\u0026thinsp;=\u0026thinsp;159\u0026ndash;340 mm, acquisition matrix\u0026thinsp;=\u0026thinsp;128 \u0026times; 64, and slice thickness\u0026thinsp;=\u0026thinsp;4 mm. The parameters for DCE-MRI were: TR/TE\u0026thinsp;=\u0026thinsp;7.67/4.25 msec, flip angle\u0026thinsp;=\u0026thinsp;12\u0026deg;, FOV\u0026thinsp;=\u0026thinsp;360\u0026ndash;360 mm, acquisition matrix\u0026thinsp;=\u0026thinsp;320 \u0026times; 288, averages\u0026thinsp;=\u0026thinsp;0.7, and slice thickness\u0026thinsp;=\u0026thinsp;2 mm. For DCE-MRI, gadopentetate dimeglumine was used as the contrast agent, with a flow rate of 2 mL/s and a dose of 0.2 mL/kg. After contrast injection, seven repeated DCE-MRI scans were performed using the same parameters as the standard sequence, with each scan lasting 90 seconds. All image sequences were acquired with the patient in the prone position using a dedicated breast coil to enhance visualization.\u003c/p\u003e\n\u003ch3\u003eMRI Morphological Feature Acquisition:\u003c/h3\u003e\n\u003cp\u003eAfter anonymization and randomization, MRI features were independently evaluated by two experienced radiologists (Zichang Ma and Junyu Lin) following the standards of the American College of Radiology Breast Imaging Reporting and Data System (BI-RADS, 2023 Aug 28) to ensure unbiased assessments(\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). The evaluated MRI features included time-signal-intensity curves classified as persistent, plateau, or wash-out patterns based on DCE-MRI kinetics; fibroglandular tissue density and background parenchymal enhancement graded according to BI-RADS; peritumoral edema, defined as a rim of water-equivalent high signal surrounding the lesion on fat-suppressed T2-weighted images, with simultaneous assessment of subcutaneous edema and skin thickening; intratumoral high T2 signal on fat-suppressed T2-WI; the rim sign on diffusion-weighted imaging, recorded as complete when the hyperintense margin covered at least 90% of the tumor circumference and incomplete otherwise; the adjacent-vessel sign and increased ipsilateral vascularity; internal enhancement patterns categorised as homogeneous, heterogeneous, or rim-enhancing; MRI-based axillary lymph-node status considered positive when the short-axis diameter was 10 mm or greater or when central necrosis was present; intratumoral necrosis; and tumor-margin morphology, classified as regular or irregular. All features were independently assessed by two blinded radiologists, and inter-observer agreement was quantified using the kappa statistic.\u003c/p\u003e\n\u003ch3\u003eRegion of Interest Acquisition:\u003c/h3\u003e\n\u003cp\u003eThe delineation of ROI: The preprocessed images were subsequently imported into ITK-SNAP (version 4.0; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.itksnap.org/)(25)\u003c/span\u003e\u003cspan address=\"http://www.itksnap.org/)(25)\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Three radiologists with extensive experience in breast imaging (Zichang Ma, Junyu Lin, Huajin Liu) manually delineated a 3D regions of interest that includes the tumor (and all its approximations) to ensure optimal accuracy and reliability.\u003c/p\u003e\n\u003ch3\u003eRadiomics feature extraction and feature selection:\u003c/h3\u003e\n\u003cp\u003eRadiomics Feature Extraction and Selection: The feature values of patients need to be normalized by calculating their z-scores using the formula (x-\u0026micro;)/σ, where x represents the feature value, \u0026micro; is the mean of the feature values for all patients, and σ is the corresponding standard deviation(\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e). Feature extraction was performed using FAE software(Version 0.5.12, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/salan668/FAE)(27)\u003c/span\u003e\u003cspan address=\"https://github.com/salan668/FAE)(27)\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e, with the enhanced images resampled to achieve a uniform voxel size of 1\u0026times;1\u0026times;1 mm. The analysis yielded first-order and intensity histogram statistics, 2D and 3D shape descriptors, and texture features, including gray-level co-occurrence matrix, gray-level run-length matrix, neighboring gray-tone difference matrix, gray-level dependence matrix, and gray-level size zone matrix. A total of 1,690 features were extracted from the ROI in each patient\u0026rsquo;s DCE-MRI series. Subsequently, a correlation analysis was performed to eliminate redundant features with a correlation coefficient\u0026thinsp;\u0026ge;\u0026thinsp;0.9. Lasso regression with five-fold cross-validation was then applied to automatically determine the optimal regularization parameter and feature coefficients, retaining only those features whose coefficients were non-zero.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eDevelopment of automated segmentation models:\u003c/h2\u003e\u003cp\u003eTo achieve automatic segmentation on breast MRI, we employed the ResUNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;segmentation network as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. This employs an encoder-decoder framework. In the encoder section, the model begins with the Conv3-32 layer for initial feature extraction, followed by a sequence of downsampling modules that incorporate Squeeze_Excite_Block and ResidualConv layers. These modules progressively increase the number of channels in the feature maps while reducing spatial resolution to effectively capture multi-scale information. At the deepest level, the ASPP-512 module is integrated to perform multi-scale feature extraction, thereby enhancing the model's receptive field. In the decoder section, upsampling operations are carried out using AttentionBlock and Upsample modules. Concurrently, skip connections are utilized to fuse feature maps from corresponding encoder layers with those in the decoder, ensuring the preservation of rich spatial detail information. Following each upsampling stage, ResidualConv layers are applied to further process the fused features, ensuring the effectiveness and stability of feature representations. Finally, the model generates segmentation results through the ASPP-32 module and the Output Conv-32-1 layer, with the outputs being normalized using a Sigmoid activation function.\u003c/p\u003e\u003cp\u003eWe first performed standardization preprocessing on the masked regions to ensure that the image intensity mean and standard deviation conform to the distribution of the breast MRI dataset. Subsequently, the images were resampled to a uniform dimension of 56\u0026times;112\u0026times;112 to standardize data dimensions and enhance training efficacy. Data augmentation was applied to the training set, including random rotations (\u0026plusmn;\u0026thinsp;30 degrees), horizontal and vertical flips (each with a 50% probability), brightness adjustments (range 0.8\u0026ndash;1.2), and contrast adjustments (range 0.8\u0026ndash;1.2). The Dice loss function was employed to optimize the overlap of the segmentation results. The Adam optimizer was used, and segmentation performance was validated using the Dice coefficient on the test set. Model training was implemented using Python 3.11.0 and PyTorch 2.0 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pytorch.org\u003c/span\u003e\u003cspan address=\"https://pytorch.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The training process was conducted on a workstation equipped with an NVIDIA RTX 4090 GPU, Intel Core i9 processor, and 64GB of memory to ensure efficient computational performance.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eDeep Learning Model Development:\u003c/h3\u003e\n\u003cp\u003eTo develop a deep learning (DL) model capable of diagnosing LVI, we applied two architectures: the traditional convolutional neural network (CNN) model ResNet101 and the transformer-based Swin Transformer, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The transformer model begins by employing Patch Partition to divide the input image into several non-overlapping small patches. These patches are then processed through Patch Merging, which progressively downsamples the feature maps while increasing the number of channels. In Stage 1, following Linear Embedding, the features are passed through a Swin Transformer Block for initial feature extraction and modeling. Subsequent stages, Stage 2 through Stage 4, each consist of Patch Merging and multiple Swin Transformer Blocks, which further extract and integrate higher-level features. This hierarchical design, combined with the window-based self-attention mechanism, effectively captures both local and global information. Finally, the feature maps undergo Global Average Pooling to aggregate into a global feature vector, which is then transformed by a Linear Layer to produce the final output. In the data preprocessing stage, the bounding boxes were resampled to 224\u0026times;224 pixels to serve as inputs for both DL models.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo address the issue of insufficient training samples, both models employed transfer learning. ResNet utilized pre-trained ResNet101 weights from torchvision, while Swin Transformer used the swin-tiny-patch4-window7-224 weights for subsequent training. The Swin Transformer source code can be found at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/microsoft/Swin-Transformer\u003c/span\u003e\u003cspan address=\"https://github.com/microsoft/Swin-Transformer\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. For ResNet101, only the final fully connected layer was fine-tuned, with the remaining layers kept frozen to leverage the feature extraction capabilities of the pre-trained model. In contrast, Swin Transformer fine-tuned all layers to fully adapt to the data distribution of this study. The optimizer used was AdamW with a weight decay coefficient set to 0.05 to prevent overfitting by reducing the weight magnitude through L2 regularization. The learning rate scheduling strategy employed was StepLR with an initial learning rate of 0.0001. The loss function used was the weighted cross-entropy loss function. Similarly, model training was implemented using Python 3.11.0 and PyTorch 2.0 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pytorch.org\u003c/span\u003e\u003cspan address=\"https://pytorch.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The training process was conducted on a workstation equipped with an NVIDIA RTX 4090 GPU, Intel Core i9 processor, and 64GB of memory.\u003c/p\u003e\u003cp\u003eAll classification models are evaluated for differences using the DeLong test. To thoroughly analyze the prediction mechanism of neural network models and their contribution to LVI diagnosis, we employed Gradient-weighted Class Activation Mapping (Grad-CAM) and SHAP(Shapley Additive Explanations) analysis methods(\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e). This approach extracts gradient features from the final output of convolutional layers, calculates weighted feature maps, and generates heat maps to assist in understanding the model's decision-making basis.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003ePatients' Characteristics:\u003c/h2\u003e\u003cp\u003eA total of 288 patients were included: 167 were randomly assigned to the training cohort, 71 to the internal test cohort, and 50 constituted an independent external test cohort. Baseline characteristics were comparable between the training and internal test cohorts (all P\u0026thinsp;\u0026gt;\u0026thinsp;0.05; Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of Baseline Characteristics Between the Training and Test Sets\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLevel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTraining Set\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest Set\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eIndependent Test Set\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eP-value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLVI\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e112(67.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e47(66.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e32(64.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.922\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e55(32.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e24(33.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e18(36.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTIC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e69(41.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e24(33.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e14(28.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.392\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e58(34.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e28(39.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e24(48.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e40(24.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e19(26.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e12(24.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePeritumoral Edema\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e93(55.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e38(53.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e26(52.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.883\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e74(44.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e33(46.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e24(48.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAdjacent Vessel Sign\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e127(76.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e49(69.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e37(74.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.527\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e40(24.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e22(31.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e13(26.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMedian (min, max)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e53.00 (min 24.00, max 84.00)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e49.00 (min 27.00, max 84.00)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e45.50 (min 30.00, max 78.00)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.025\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMenopausal Status\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePremenopausal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e77(46.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e37(52.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e32(64.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.082\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePostmenopausal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e90(53.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e34(47.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e18(36.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLocation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRight\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e84(50.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e35(49.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e23(46.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.867\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLeft\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e83(49.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e36(50.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e27(54.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSymmetry\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e105(62.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e47(66.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e35(70.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.63\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e62(37.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e24(33.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e15(30.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBPE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNone/minimal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e101(60.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e47(66.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e32(64.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.781\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMild\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e45(26.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e17(23.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e15(30.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eModerate\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e14(8.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e6(8.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2(4.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMarked\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e7(4.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1(1.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1(2.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSubcutaneous Edema\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e125(74.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e51(71.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e37(74.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.889\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e42(25.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e20(28.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e13(26.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHigh Signal Intensity within Tumor\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e115(68.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e47(66.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e33(66.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.886\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e52(31.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e24(33.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e17(34.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInternal Enhancement Pattern\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHeterogeneous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e11(6.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e5(7.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3(6.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.928\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHomogeneous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e133(79.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e55(77.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e42(84.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRim Enhancement\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e23(13.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e11(15.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e5(10.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIncreased Ipsilateral Vascular Distribution\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e72(43.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e29(40.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e22(44.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e95(56.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e42(59.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e28(56.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAxillary Lymph Node Status\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e100(59.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e48(67.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e29(58.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.458\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e67(40.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e23(32.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e21(42.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIntratumoral Necrosis\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e133(79.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e58(81.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e41(82.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.898\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e34(20.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e13(18.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e9(18.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTumor-margin Morphology\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e67(40.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e36(50.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e27(54.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.124\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e100(59.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e35(49.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e23(46.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDWI Margin Sign\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e85(50.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e32(45.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e22(44.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.572\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e82(49.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e39(54.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e28(56.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eMRI Morphological Feature\u003c/h2\u003e\u003cp\u003eWe found that patients with LVI(+) were more likely to develop peripheral tumor edema (OR: 2.06; 95% confidence interval [CI] 1.04\u0026ndash;4.06) and DWI margin signal (OR: 2.34; 95% CI 1.19\u0026ndash;4.63) in the training set (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eUnivariate and Multivariate Logistic Analysis for LVI of Breast Cancer.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVariable Name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-factor OR (95%CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ep-value\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMultifactorial OR (95%CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ep-value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTIC Curve\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.86 (0.57, 1.30)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.472\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003ePeritumoral Edema\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e2.07 (1.08, 3.99)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.029\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e2.06 (1.04, 4.06)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.037\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAdjacent Vessel Sign\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.61 (0.27, 1.36)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.224\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.01 (0.99, 1.04)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.324\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMenstrual Status\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.75 (0.39, 1.43)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.384\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLocation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.70 (0.36, 1.33)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.273\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSymmetry\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.07 (0.55, 2.08)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.843\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBPE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.85 (0.56, 1.28)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.426\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSubcutaneous Edema\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.02 (0.49, 2.15)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.949\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHigh Signal Intensity within Tumor\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.12 (0.56, 2.23)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.756\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInternal Enhancement Pattern\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.15 (0.56, 2.38)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.699\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIncreased Ipsilateral Vascular Distribution\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.35 (0.70, 2.62)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.368\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAxillary Lymph Node Status\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.74 (0.90, 3.34)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.099\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.69 (0.85, 3.35)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.131\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIntratumoral Necrosis\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.56 (0.24, 1.34)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.195\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTumor Margin\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.13 (0.58, 2.19)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.720\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDWI Margin Sign\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e2.16 (1.12, 4.18)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.022\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e2.34 (1.19, 4.63)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.014\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eSelected Radiomics Features:\u003c/h2\u003e\u003cp\u003eAt the optimal α\u0026thinsp;=\u0026thinsp;0.04977, a total of 15 radiomics features were selected. Ranked by SHAP importance from highest to lowest, they were: original_firstorder_90Percentile, lbp-3D-m1_glszm_LowGrayLevelZoneEmphasis, squareroot_glszm_GrayLevelVariance, square_glrlm_RunEntropy, logarithm_glszm_LowGrayLevelZoneEmphasis, original_shape_Flatness, lbp-3D-m2_glrlm_RunEntropy, lbp-3D-k_glcm_Correlation, wavelet-HLH_firstorder_Skewness, original_shape_Elongation, exponential_gldm_LargeDependenceHighGrayLevelEmphasis, lbp-3D-k_glszm_LowGrayLevelZoneEmphasis, original_shape_LeastAxisLength, squareroot_firstorder_Kurtosis, and logarithm_glrlm_RunEntropy.(Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e)\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003eSegmentation performance:\u003c/h2\u003e\u003cp\u003eThe segmentation performance of the model on the dataset is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. On Test1 (N\u0026thinsp;=\u0026thinsp;71), the model achieved a Dice coefficient of 0.838\u0026thinsp;\u0026plusmn;\u0026thinsp;0.050 (95% CI: 0.826\u0026ndash;0.849), an IoU of 0.724\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073 (95% CI: 0.707\u0026ndash;0.740), a Precision of 0.879\u0026thinsp;\u0026plusmn;\u0026thinsp;0.056 (95% CI: 0.866\u0026ndash;0.892), and a Recall of 0.804\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073 (95% CI: 0.787\u0026ndash;0.822). On Test2 (N\u0026thinsp;=\u0026thinsp;50), the model demonstrated improved performance with a Dice coefficient of 0.867\u0026thinsp;\u0026plusmn;\u0026thinsp;0.044 (95% CI: 0.854\u0026ndash;0.878), an IoU of 0.768\u0026thinsp;\u0026plusmn;\u0026thinsp;0.065 (95% CI: 0.749\u0026ndash;0.784), a Precision of 0.923\u0026thinsp;\u0026plusmn;\u0026thinsp;0.046 (95% CI: 0.910\u0026ndash;0.935), and a Recall of 0.820\u0026thinsp;\u0026plusmn;\u0026thinsp;0.059 (95% CI: 0.803\u0026ndash;0.835). Some of the segmentation results are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eSegmentation performance of the method on the dataset (macro: case-level mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD with 95% CI)\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDice\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eIoU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTest1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e71\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.838\u0026thinsp;\u0026plusmn;\u0026thinsp;0.050\u003c/p\u003e\u003cp\u003e[0.826, 0.849]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.724\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073 [0.707, 0.740]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.879\u0026thinsp;\u0026plusmn;\u0026thinsp;0.056 [0.866, 0.892]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.804\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073 [0.787, 0.822]\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTest2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.867\u0026thinsp;\u0026plusmn;\u0026thinsp;0.044 [0.854, 0.878]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.768\u0026thinsp;\u0026plusmn;\u0026thinsp;0.065 [0.749, 0.784]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.923\u0026thinsp;\u0026plusmn;\u0026thinsp;0.046 [0.910, 0.935]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.820\u0026thinsp;\u0026plusmn;\u0026thinsp;0.059 [0.803, 0.835]\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eModel Validation:\u003c/h2\u003e\u003cp\u003eFor single-modality models, the logistic-regression model based on handcrafted imaging features achieved an AUC(Area Under Curve) of 0.641 (95% CI 0.549\u0026ndash;0.721) in the internal test set and 0.707 (95% CI 0.580\u0026ndash;0.821) in the independent external test set. The standalone convolutional neural network (CNN) obtained an AUC of 0.741 (95% CI 0.576\u0026ndash;0.884) internally and 0.702 (95% CI 0.515\u0026ndash;0.880) externally. The radiomics model yielded an AUC of 0.770 (95% CI 0.643\u0026ndash;0.870) in the internal test set but only 0.571 (95% CI 0.400\u0026ndash;0.742) in the external cohort. The pure Transformer model reached an AUC of 0.801 (95% CI 0.665\u0026ndash;0.918) internally and 0.762 (95% CI 0.576\u0026ndash;0.918) externally.\u003c/p\u003e\u003cp\u003eFor multimodal approaches, the ResNet-based multimodal network achieved an AUC of 0.855 (95% CI 0.737\u0026ndash;0.947) in the internal test set and 0.802 (95% CI 0.623\u0026ndash;0.935) in the external test set.The multimodal architecture that combines ResNet-based segmentation with Transformer-based classification delivered the best results, with an AUC of 0.873 (95% CI 0.758\u0026ndash;0.967) internally and 0.845 (95% CI 0.687\u0026ndash;0.964) externally. The multimodal Transformer model with automated segmentation outperformed all standalone models (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e).The Grad-CAM heat map visualizes the attention regions of the Transformer model; SHAP quantifies the contribution weights of different features to the prediction of LVI positive or negative (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we explored the relationship between LVI and invasive breast cancer and developed a multimodal model combining automatic segmentation, MRI morphological features, radiomics features, and a Transformer-based deep learning framework to assess the likelihood of LVI in breast cancer patients. The combined model demonstrated higher diagnostic efficiency in distinguishing LVI-positive and LVI-negative cases compared to single models.\u003c/p\u003e\u003cp\u003eIn previous studies, MRI morphological features have been shown to predict LVI status in breast cancer patients(\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e). However, the reliability of these results may be affected by interobserver variability due to differences in radiologists' experience levels. In this study, we identified only peritumoral edema and the DWI edge sign as independent indicators of LVI, consistent with previous findings. Peritumoral edema may result from tumor cell invasion into lymphatic vessels, obstructing normal lymphatic flow and drainage, or from damage to the integrity of vascular walls, increasing vascular permeability and allowing fluid to leak into surrounding tissues. The DWI edge sign may be caused by lymphatic and vascular invasion, leading to increased local cell density, edema, and stromal reactions in surrounding tissues, which restrict the free diffusion of water molecules.\u003c/p\u003e\u003cp\u003eRadiomics features provide quantitative information about tumor phenotypes that are often imperceptible to the naked eye. Accordingly, we extracted radiomics features from the tumors and, after feature selection, trained machine-learning models. SHAP analysis showed that model decisions were primarily driven by texture-intensity composite features, including the upper gray-level quantile (original_firstorder_90Percentile), clustering of low-gray-level zones (lbp-3D-m1_glszm_LowGrayLevelZoneEmphasis), and texture run entropy (square_root_glrlm_RunEntropy). These metrics are strongly associated with intratumoral heterogeneity, suggesting that lesions with more complex tissue composition and more extreme gray-level distributions are more likely to be classified as high risk. Notably, the morphological features Flatness and Elongation also ranked among the top contributors, indicating that flatter and more elongated lesions have a higher probability of a positive prediction, which may reflect irregular shapes arising from infiltrative growth. Overall, increased image heterogeneity together with irregular morphology jointly underpins the model\u0026rsquo;s high-risk determinations.\u003c/p\u003e\u003cp\u003eManual ROI delineation is time-consuming and highly subjective, with significant variability between radiologists, which can affect the reliability of subsequent analyses. In this study, we employed a ResUNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;automatic segmentation model. On Test1 (N\u0026thinsp;=\u0026thinsp;71) and Test2 (N\u0026thinsp;=\u0026thinsp;50), the model achieved Dice coefficients of 0.838 (95% CI: 0.826\u0026ndash;0.849) and 0.867 (95% CI: 0.854\u0026ndash;0.878), with corresponding IoU values of 0.724 (95% CI: 0.707\u0026ndash;0.740) and 0.768 (95% CI: 0.749\u0026ndash;0.784), respectively. Precision consistently exceeded Recall on both datasets (Test1: 0.879 vs. 0.804; Test2: 0.923 vs. 0.820), indicating the model's conservative tendency with fewer false positives and stable generalization capability. The automatic pipeline substantially reduces radiologists\u0026rsquo; workload and improves the consistency of ROI delineation, providing a reliable foundation for subsequent model training. Introducing automatic segmentation is a major strength of this study, as it alleviates the subjectivity of manual contouring and lays the groundwork for developing multimodal models. Nevertheless, performance remains to be improved for tumors with indistinct boundaries. To address this, we post-processed the outputs by dilating the ROI margin by 4 mm and standardizing to a square input box to cover potentially uncertain boundary regions (see Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). To address this issue, we processed the automatic segmentation results by expanding the ROI boundaries by 4mm and using square input boxes, which resolved the boundary blurring problem. We also tested other segmentation networks, and found no significant differences in results among different segmentation networks after applying this processing method.\u003c/p\u003e\u003cp\u003eThrough systematic comparison of six models, this study demonstrated that multimodal frameworks can significantly enhance prediction performance. The ResNet multimodal network achieved AUCs of 0.855 (95% CI 0.737\u0026ndash;0.947) and 0.802 (95% CI 0.623\u0026ndash;0.935) in the internal and external test sets, respectively, while the two-stage model combining automatic residual segmentation output with Transformer classification performed even better, with AUCs climbing to 0.873 (95% CI 0.758\u0026ndash;0.967) and 0.845 (95% CI 0.687\u0026ndash;0.964), respectively. First, the self-attention mechanism enables the model to capture complex interdependencies between different imaging features that traditional machine learning approaches might overlook(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e). By integrating morphological, textural, and pharmacokinetic features through the Transformer architecture, the model can more comprehensively characterize tumor heterogeneity. Second, the robustness demonstrated in external validation proves the model's generalizability across different imaging protocols and patient populations, which is particularly notable given the known challenges in cross-institutional validation of radiomics models (AUC\u0026thinsp;=\u0026thinsp;0.571). The modest performance decline from internal testing to external testing indicates good model robustness, although some performance degradation is expected due to differences in imaging parameters and patient characteristics. Compared with previous studies using traditional machine learning approaches for LVI prediction(\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e), our Transformer model shows improvements in both discrimination ability and stability. The attention mechanism's ability to automatically weight feature importance may be the key reason why this approach outperforms traditional feature selection methods. We employed visualization techniques such as Grad-CAM and SHAP to intuitively depict the attention regions of the Transformer model and to provide preliminary estimates of the relative contributions of different features to predicting LVI-positive vs. LVI-negative status, thereby offering clinicians spatial localization of the model\u0026rsquo;s decision-making. However, due to postoperative specimen deformation, geometric distortions introduced during tissue processing, and misregistration between imaging examinations and pathological assessments across 2D/3D coordinate systems, it remains challenging to obtain a voxel-level, one-to-one image\u0026ndash;pathology correspondence as a reference standard for gold-standard validation.\u003c/p\u003e\u003cp\u003eOverall, the Transformer architecture not only leads in single-modality scenarios but further amplifies its advantages in automated segmentation-driven multi-modal approaches, providing a solution pathway that combines both performance and interpretability for non-invasive assessment of LVI.\u003c/p\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003eLimitation:\u003c/h2\u003e\u003cp\u003eThere are several limitations to our study that should be considered. First, our radiomics features were extracted only from the second phase of DCE-MRI and did not utilize data from other MRI sequences such as T1WI, T2WI, DWI and ADC maps. Second, we did not extract radiomics features solely from the peritumoral region, although our deep learning approach covered the peritumoral area, it may still result in the loss of some multimodal information.\u003c/p\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEthical approval for this retrospective study was granted by the Medical Ethics Committee of the Fifth Affiliated Hospital of Sun Yat-sen University (Approval No. K243-1), in accordance with the Declaration of Helsinki (as revised in 2013). The ethics committee waived the requirement for individual informed consent given the retrospective nature of this study using routinely collected clinical data and biospecimens.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated and/or analysed during the current study are not publicly available due to patient privacy and institutional data protection policies but are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the National Natural Science Foundation of China under Grant Nos. 81801809 and 82371917, the Basic and Applied Basic Research Foundation of Guangdong Province under Grant No. 2020A1515010572, and the Zhuhai Basic and Applied Basic Research Project Foundation under Grant No. ZH22017003200001PWC.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJ.L. and Z.M. contributed equally to this work, performed data analysis and wrote the main manuscript text. Y.T. and Y.L. contributed to data interpretation. J.L., Y.W., and H.L. were responsible for data collection and organization. Y.Z. designed and supervised the overall study. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSiegel RL, Kratzer TB, Giaquinto AN, Sung H, Jemal A. Cancer statistics, 2025. CA Cancer J Clin. 2025;75(1):10-45.\u003c/li\u003e\n\u003cli\u003eNull JL, Kim DJ, McCann JV, Pramoonjago P, Fox JW, Zeng J, et al. Periostin+ Stromal Cells Guide Lymphovascular Invasion by Cancer Cells. Cancer Res. 2023;83(13):2105-22.\u003c/li\u003e\n\u003cli\u003eLiu G, Kong X, Dai Q, Cheng H, Wang J, Gao J, et al. Clinical Features and Prognoses of Patients With Breast Cancer Who Underwent Surgery. JAMA Netw Open. 2023;6(8):e2331078.\u003c/li\u003e\n\u003cli\u003eKennedy WR, Tricarico C, Gabani P, Weiner AA, Altman MB, Ochoa LL, et al. Predictors of Distant Metastases in Triple-Negative Breast Cancer Without Pathologic Complete Response After Neoadjuvant Chemotherapy. J Natl Compr Canc Netw. 2020;18(3):288-96.\u003c/li\u003e\n\u003cli\u003eRakha EA, Martin S, Lee AH, Morgan D, Pharoah PD, Hodi Z, et al. The prognostic significance of lymphovascular invasion in invasive breast carcinoma. Cancer. 2012;118(15):3670-80.\u003c/li\u003e\n\u003cli\u003eHouvenaeghel G, Cohen M, Classe JM, Reyal F, Mazouni C, Chopin N, et al. Lymphovascular invasion has a significant prognostic impact in patients with early breast cancer, results from a large, national, multicenter, retrospective cohort study. ESMO Open. 2021;6(6):100316.\u003c/li\u003e\n\u003cli\u003eZhang S, Zhang D, Gong M, Wen L, Liao C, Zou L. High lymphatic vessel density and presence of lymphovascular invasion both predict poor prognosis in breast cancer. BMC Cancer. 2017;17(1):335.\u003c/li\u003e\n\u003cli\u003eChauhan H, Jiwa N, Nagarajan VR, Thiruchelvam P, Hogben K, Al-Mufti R, et al. Clinicopathological Predictors of Positive Resection Margins in Breast-Conserving Surgery. Ann Surg Oncol. 2024;31(6):3939-47.\u003c/li\u003e\n\u003cli\u003eStrauss JB. Extensive Lymphovascular Invasion Warrants Extensive Radiation Therapy. Int J Radiat Oncol Biol Phys. 2020;107(2):234-5.\u003c/li\u003e\n\u003cli\u003eKurebayashi J, Kanomata N, Shimo T, Yamashita T, Aogi K, Nishimura R, et al. Marked lymphovascular invasion, progesterone receptor negativity, and high Ki67 labeling index predict poor outcome in breast cancer patients treated with endocrine therapy alone. Breast Cancer. 2014;21(2):214-22.\u003c/li\u003e\n\u003cli\u003eNitz UA, Gluz O, Christgen M, Grischke EM, Augustin D, Kuemmel S, et al. De-escalation strategies in HER2-positive early breast cancer (EBC): final analysis of the WSG-ADAPT HER2+/HR- phase II trial: efficacy, safety, and predictive markers for 12\u0026thinsp;weeks of neoadjuvant dual blockade with trastuzumab and pertuzumab \u0026plusmn; weekly paclitaxel. Ann Oncol. 2017;28(11):2768-72.\u003c/li\u003e\n\u003cli\u003eMann RM, Cho N, Moy L. Breast MRI: State of the Art. Radiology. 2019;292(3):520-36.\u003c/li\u003e\n\u003cli\u003eChen Y, Wang L, Luo R, Liu H, Zhang Y, Wang D. Focal breast edema and breast edema score on T2-weighted images provides valuable biological information for invasive breast cancer. Insights Imaging. 2023;14(1):73.\u003c/li\u003e\n\u003cli\u003eCheon H, Kim HJ, Lee SM, Cho SH, Shin KM, Kim GC, et al. Preoperative MRI features associated with lymphovascular invasion in node-negative invasive breast cancer: A propensity-matched analysis. J Magn Reson Imaging. 2017;46(4):1037-44.\u003c/li\u003e\n\u003cli\u003eAcciavatti RJ, Lee SH, Reig B, Moy L, Conant EF, Kontos D, et al. Beyond Breast Density: Risk Measures for Breast Cancer in Multiple Imaging Modalities. Radiology. 2023;306(3):e222575.\u003c/li\u003e\n\u003cli\u003eBera K, Braman N, Gupta A, Velcheti V, Madabhushi A. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol. 2022;19(2):132-46.\u003c/li\u003e\n\u003cli\u003eHao J, Liu S, Wang T, Han X, Gao A, Wang H, et al. Differentiation of malignant from benign soft tissue tumors using radiomics based on pharmacokinetic parameter maps obtained from dynamic contrast-enhanced magnetic resonance imaging data. Chinese Journal of Academic Radiology. 2024;7(3):219-28.\u003c/li\u003e\n\u003cli\u003eBi WL, Hosny A, Schabath MB, Giger ML, Birkbak NJ, Mehrtash A, et al. Artificial intelligence in cancer imaging: Clinical challenges and applications. CA Cancer J Clin. 2019;69(2):127-57.\u003c/li\u003e\n\u003cli\u003eTomaszewski MR, Gillies RJ. The Biological Meaning of Radiomic Features. Radiology. 2021;298(3):505-16.\u003c/li\u003e\n\u003cli\u003eZhang Y, Chan S, Park VY, Chang KT, Mehta S, Kim MJ, et al. Automatic Detection and Segmentation of Breast Cancer on MRI Using Mask R-CNN Trained on Non-Fat-Sat Images and Tested on Fat-Sat Images. Acad Radiol. 2022;29 Suppl 1(Suppl 1):S135-s44.\u003c/li\u003e\n\u003cli\u003eHagiwara A, Fujita S, Ohno Y, Aoki S. Variability and Standardization of Quantitative Imaging: Monoparametric to Multiparametric Quantification, Radiomics, and Artificial Intelligence. Invest Radiol. 2020;55(9):601-16.\u003c/li\u003e\n\u003cli\u003eYang X, Fan X, Lin S, Zhou Y, Liu H, Wang X, et al. 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(6):2238-49.\u003c/li\u003e\n\u003cli\u003eLi J, Chen J, Tang Y, Wang C, Landman BA, Zhou SK. Transforming medical imaging with Transformers? A comparative review of key properties, current progresses, and future perspectives. Med Image Anal. 2023;85:102762.\u003c/li\u003e\n\u003cli\u003eMagny SJ, Shikhman R, Keppke AL. Breast Imaging Reporting and Data System. StatPearls. Treasure Island (FL) ineligible companies. Disclosure: Rachel Shikhman declares no relevant financial relationships with ineligible companies. Disclosure: Ana Keppke declares no relevant financial relationships with ineligible companies.: StatPearls Publishing Copyright \u0026copy; 2025, StatPearls Publishing LLC.; 2025.\u003c/li\u003e\n\u003cli\u003eYushkevich PA, Piven J, Hazlett HC, Smith RG, Ho S, Gee JC, et al. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage. 2006;31(3):1116-28.\u003c/li\u003e\n\u003cli\u003eTraverso A, Wee L, Dekker A, Gillies R. Repeatability and Reproducibility of Radiomic Features: A Systematic Review. Int J Radiat Oncol Biol Phys. 2018;102(4):1143-58.\u003c/li\u003e\n\u003cli\u003eSong Y, Zhang J, Zhang YD, Hou Y, Yan X, Wang Y, et al. FeAture Explorer (FAE): A tool for developing and comparing radiomics models. PLoS One. 2020;15(8):e0237587.\u003c/li\u003e\n\u003cli\u003eChen Y, Xiao Z, Pan Y, Zhao L, Dai H, Wu Z, et al. Mask-Guided Vision Transformer for Few-Shot Learning. IEEE Trans Neural Netw Learn Syst. 2025;36(5):9636-47.\u003c/li\u003e\n\u003cli\u003eSingh A, Sengupta S, Lakshminarayanan V. Explainable Deep Learning Models in Medical Image Analysis. J Imaging. 2020;6(6).\u003c/li\u003e\n\u003cli\u003eXu Z, Xie Y, Wu L, Chen M, Shi Z, Cui Y, et al. 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(5):1580-9.\u003c/li\u003e\n\u003cli\u003eSun W, Qin Z, Deng H, Wang J, Zhang Y, Zhang K, et al. Vicinity Vision Transformer. IEEE Trans Pattern Anal Mach Intell. 2023;45(10):12635-49.\u003c/li\u003e\n\u003cli\u003eZhang J, Wang G, Ren J, Yang Z, Li D, Cui Y, et al. 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(6):4079-89.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7867527/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7867527/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eObjectives\u003c/h2\u003e\u003cp\u003eTo develop and evaluate an automated, multimodal Transformer model for preoperative prediction of lymphovascular invasion (LVI) in invasive breast cancer using contrast-enhanced MRI.\u003c/p\u003e\u003ch2\u003eMaterials and Methods\u003c/h2\u003e\u003cp\u003eA retrospective study analyzed 288 patients with pathologically confirmed invasive breast cancer who all underwent preoperative dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). The study included an internal cohort (238 patients) and an external cohort (50 patients). Tumor regions of interest (ROI) were manually delineated by radiologists and automatic tumor segmentation was performed using ResUNet++. The output results were then processed with 4mm boundary dilation, while radiomic features were extracted and radiologists assessed MRI features according to BI-RADS criteria. Single-modality and multi-modality models were constructed for comparison, with the multi-modal fusion network integrating enhanced images, radiomic features, and MRI features. Model differences were assessed using DeLong test, and interpretability analysis was performed using Grad-CAM and SHAP methods.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eAutomated segmentation was robust (Dice 0.916 internal and 0.921 external). The two-stage multimodal classifier achieved the highest AUC, 0.873 internally and 0.845 externally, compared with the best single-modality Transformer at 0.801 internally and 0.762 externally.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e\u003cp\u003eIntegrating automated MRI segmentation with Transformer-based multimodal learning enables reliable preoperative LVI prediction and shows promising cross-center generalizability for clinical translation.\u003c/p\u003e","manuscriptTitle":"Multimodal automated diagnosis of lymphovascular invasion in breast cancer on contrast-enhanced MRI: ResUNet++ segmentation and Transformer-based classification","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-21 17:49:39","doi":"10.21203/rs.3.rs-7867527/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-02-24T15:07:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"197180435241364464129271136973481630700","date":"2026-02-15T14:27:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"59776834724556268687359984025916671616","date":"2026-02-10T13:51:29+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-11-11T11:09:46+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-10-22T23:49:58+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-21T12:41:11+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-21T12:41:08+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2025-10-15T11:18:22+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"1362ff6b-a9b8-43a0-9e12-abb368af298c","owner":[],"postedDate":"November 21st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-11-21T17:49:39+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-21 17:49:39","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7867527","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7867527","identity":"rs-7867527","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}