A Habitat Imaging-Based Radiomics and Deep Learning Fusion for Preoperative Prediction of Lymphovascular Invasion in Invasive Breast Cancer: A Multicenter Study

Abstract

Background

Lymphovascular invasion (LVI) is a critical prognostic factor in invasive breast cancer; however, reliable preoperative prediction remains challenging because of the lack of non-invasive and accurate assessment tools. Ultrasound-based radiomics and deep learning have shown promise, but conventional single-modality approaches often fail to capture intratumoral heterogeneity, thereby limiting their predictive performance and clinical interpretability.

Aims

To develop and validate a non-invasive ultrasound-based model that integrates habitat radiomic features with deep learning features for the preoperative prediction of LVI in invasive breast cancer and to evaluate whether this model could enhance the diagnostic performance of radiologists.

Study Design

Retrospective multicenter diagnostic study.

Methods

A total of 1,096 patients from three institutions were included. Tumors on routine B-mode ultrasound were partitioned into three habitats using unsupervised k-means clustering (k = 3). Habitat2, identified as a high-risk subregion associated with LVI, was used for selective radiomic feature extraction. In parallel, deep learning features from all habitat subregions were extracted using a pretrained ResNet50 model. Radiomic and deep learning features were fused at the feature level, and a LightGBM classifier was trained and evaluated in an internal validation cohort and two independent external cohorts. Model performance was compared with that of the Habitat2 radiomics-only and ResNet50-only models using the DeLong test and with the performance of unaided senior and junior radiologists. Biological interpretability was assessed through correlation analysis with CD31-stained microvessel density.

Results

The fusion model achieved areas under the receiver operating characteristic curve (AUCs) of 0.918 (training), 0.898 (internal validation), 0.890 (External Cohort 1), and 0.905 (External Cohort 2), significantly outperforming the Habitat2 radiomics-only and ResNet50-only models (DeLong test: fusion vs. Habitat2 radiomics, all p < 0.001; fusion vs. ResNet50, p < 0.001, 0.021, 0.029, and 0.026, respectively). Habitat-specific radiomic and deep learning features showed weak-to-moderate correlations with CD31-stained microvessel density, supporting the biological interpretability of the model. Compared with radiologists, the fusion model achieved higher diagnostic performance than unaided senior and junior radiologists in both external cohorts, and model assistance substantially improved the performance of junior radiologists (AUCs up to 0.865 and 0.905).

Conclusion

This interpretable habitat-deep learning fusion approach based on routine ultrasound provides accurate preoperative LVI prediction with robust generalizability across multiple centers and suggests the potential to enhance clinical decision-making by improving radiologist diagnostic performance, particularly for less-experienced radiologists.

INTRODUCTION

Breast cancer is one of the most common malignancies among women worldwide and a leading cause of cancer-related mortality.¹ In China, its incidence continues to increase, particularly among women living in urban areas.² Early diagnosis and timely treatment are essential for improving patient survival.³ Lymphovascular invasion (LVI), defined as the presence of tumor cells within adjacent lymphatic or blood vessels,⁴ has important clinical significance in breast cancer because it is a prerequisite for tumor dissemination and distant metastasis.⁴ LVI is associated with more aggressive clinicopathological characteristics, including larger tumor size, higher histological grade, advanced T stage, and axillary lymph node metastasis.⁵ More importantly, LVI is an independent adverse prognostic factor. Even in patients with lymph node-negative early-stage breast cancer, the presence of LVI is associated with higher rates of local recurrence and distant metastasis, as well as shorter progression-free survival and overall survival.⁵ Therefore, LVI status should be considered during staging and treatment planning because it indicates an increased risk of recurrence and may guide decisions regarding adjuvant radiotherapy and chemotherapy.² Accurate preoperative prediction of LVI is therefore of considerable value for disease staging, prognostic assessment, and individualized treatment planning.

Currently, LVI assessment relies primarily on postoperative histopathological examination, which is invasive, delayed, and subject to interobserver variability,⁶ thereby limiting its utility for preoperative risk stratification and early screening. Breast ultrasound is a widely used imaging modality for breast cancer because it is radiation-free, cost-effective, and enables real-time evaluation.^{7, 8} It can also assess axillary lymph node involvement and provide important information for clinical staging.⁷ However, conventional ultrasound has limited ability to visualize microvascular invasion and intratumoral heterogeneity, which may contribute to missed diagnoses.⁹

In recent years, deep learning, particularly convolutional neural networks (CNNs), has significantly advanced medical image analysis.¹⁰ CNNs can automatically learn high-level discriminative features from large-scale image datasets, overcoming the limitations of handcrafted features.¹⁰ In breast imaging, deep learning has been widely applied to tumor detection, benign-malignant classification, and image segmentation, achieving performance comparable to that of human experts.¹¹ Several CNN-based ultrasound models have demonstrated diagnostic performance comparable to that of experienced sonographers, with some outperforming human radiologists on specific performance metrics.¹¹ By capturing subtle texture patterns, margin characteristics, and peritumoral changes, deep learning has shown considerable potential for predicting tumor biological behavior.¹⁰ Its integration into clinical workflows may further improve diagnostic efficiency, reduce human error, and provide more objective quantitative assessments.¹¹

Habitat imaging is an emerging radiomics approach that uses unsupervised clustering to partition heterogeneous tumors into habitat subregions with distinct imaging characteristics. These subregions reflect the spatial heterogeneity of the tumor microenvironment, including variations in blood supply, oxygen metabolism, and cell density.^{12, 13, 14} Compared with conventional radiomics, habitat imaging enables multidimensional and visually interpretable characterization of tumor heterogeneity. This heterogeneity has been closely associated with tumor proliferation, invasiveness, metastatic potential, and treatment sensitivity.¹² In breast imaging, habitat-based analysis has demonstrated promising clinical value. Multiparametric magnetic resonance imaging (MRI)-derived habitat features have significantly improved the prediction of neoadjuvant chemotherapy response and LVI status.⁴ More recently, ultrasound habitat radiomics has also shown encouraging results. In a study involving 945 patients, this approach identified three representative habitat patterns and effectively predicted the response to neoadjuvant chemotherapy, achieving an area under the receiver operating characteristic curve (AUC) of 0.74 in the test cohort.¹⁵ These findings suggest that applying habitat imaging to breast ultrasound may reveal deeper intratumoral heterogeneity and provide novel imaging biomarkers for assessing tumor malignancy and prognosis.

We hypothesized that a fusion model integrating habitat radiomic and deep learning features would provide superior predictive performance for LVI compared with single-modality models and significantly improve the diagnostic performance of radiologists. Therefore, this study aimed to integrate habitat radiomic features with deep learning features extracted from ultrasound habitat images to develop a model for the preoperative prediction of LVI status in invasive breast cancer. The model was further compared with the diagnostic performance of radiologists with different levels of experience to evaluate its predictive performance and potential clinical utility.

MATERIALS AND METHODS

Study design and patient population

This multicenter retrospective cohort study included 1,096 patients with pathologically confirmed invasive breast cancer from three medical centers: 706 patients from Center 1 (January 2016 to December 2024), 200 patients from Center 2 (January 2021 to December 2024), and 190 patients from Center 3 (January 2022 to December 2024). All patients underwent preoperative conventional breast ultrasound, and LVI status was determined postoperatively by histopathological examination. The study was approved by the Institutional Ethics Committees of all participating institutions: Sixth People’s Hospital Affiliated to Medical College of Shanghai Jiao Tong University [approval no. 2024-KT-117(K)], The Affiliated Jiangsu Shengze Hospital of Nanjing Medical University (approval no. 2024-KYCS-002), and Fudan University Shanghai Cancer Center (approval no. 2001213-23-NSFC). The study was also registered in the Chinese Clinical Trial Registry (ChiCTR2400090587). The requirement for informed consent was waived because of the retrospective study design.

The inclusion criteria were as follows: (1) histopathologically confirmed invasive breast carcinoma with documented LVI status; (2) preoperative breast ultrasound images of adequate quality; and (3) complete clinical and pathological data. The exclusion criteria were as follows: (1) receipt of neoadjuvant therapy before surgery; (2) incomplete imaging data or poor-quality ultrasound images; and (3) special types of malignant breast tumors, such as sarcoma or lymphoma. The patient selection flowchart is presented in Figure 1.

Figure 1.

The flowchart of patient selection.

LVI, lymphovascular invasion.

Image acquisition and preprocessing

All ultrasound examinations were performed by experienced radiologists using high-frequency linear-array transducers (7-15 MHz). Specifically, Center 1 used the Philips EPIQ 7, EPIQ 5, Affiniti 70, and iU Elite systems; the Mindray Resona R9; and the GE Healthcare LOGIQ E9 (USA). Center 2 used the Siemens Healthineers ACUSON S3000, Philips EPIQ 7, and Samsung P80 Plus systems. Center 3 used the Philips EPIQ 7 and Affiniti 70, the Mindray Resona R9, and the GE Healthcare LOGIQ E11 (USA). Images were acquired in both the transverse and longitudinal planes to ensure comprehensive tumor coverage. To minimize inter-scanner variability, acquisition parameters, including gain, dynamic range, and focal zones, were harmonized across centers, and image quality was centrally monitored by senior radiologists to ensure inter-site consistency. Digital Imaging and Communications in Medicine images were de-identified and underwent standardized preprocessing. To mitigate the impact of scanner-dependent grayscale variations on habitat construction, a rigorous grayscale intensity standardization protocol was implemented. First, global histogram matching was performed to align the overall intensity distributions across different scanners. Second, within each manually delineated tumor region of interest (ROI), pixel intensities were normalized using min-max scaling (rescaled to a range of 0–1), followed by Z-score standardization. Because habitat generation relies on k-means clustering based on relative grayscale similarity within the ROI, this intra-tumor normalization effectively minimized the influence of absolute grayscale differences resulting from variations in gain or time-gain compensation settings, ensuring that the identified habitats reflected true tissue heterogeneity rather than scanner-specific artifacts. Furthermore, pathological assessment of LVI followed a standardized protocol. All hematoxylin and eosin (H&E)-stained sections were independently reviewed by two senior pathologists who were blinded to the imaging findings. Any discrepancies were resolved by consensus with a third expert pathologist.

Tumor ROIs were manually delineated using Insight Toolkit-SNAP by two radiologists, each with more than 5 years of experience in breast ultrasound. Both radiologists were blinded to the pathological results and manually traced the tumor boundaries pixel by pixel to include the entire lesion while excluding the surrounding normal tissue. Any discrepancies were resolved by consensus, and the final ROI masks were used for subsequent analyses.

Construction of habitat imaging

For each tumor ROI, unsupervised subregion partitioning was performed using a habitat imaging approach. Specifically, k-means clustering was applied to segment each tumor into multiple subregions based on grayscale intensity similarity within the ROI. The Calinski-Harabasz (CH) index was used to evaluate clustering performance across different numbers of clusters and determine the optimal number of habitats. In the training cohort, the CH index indicated an optimal elbow point at three clusters; therefore, each ROI was divided into three internal subregions representing distinct tumor habitats.

Radiomic features were then extracted separately from each habitat subregion using an open-source Python-based radiomics toolkit (e.g., PyRadiomics). The extracted features included first-order, shape, and texture features, such as gray-level co-occurrence matrix and gray-level run-length matrix descriptors. Features from the three habitat subregions were subsequently concatenated to form a high-dimensional feature set.

To reduce dimensionality and minimize overfitting, a two-step feature selection strategy was applied in the training cohort. Although both methods perform feature selection, they serve complementary roles in the analysis pipeline. First, minimum redundancy maximum relevance (mRMR) was used as a filter method to retain features that were most relevant to LVI while minimizing redundancy among them. This step is particularly important because radiomic features are often highly collinear. Applying mRMR before least absolute shrinkage and selection operator (LASSO) regression effectively removes redundant and highly correlated features, thereby stabilizing the subsequent regularization process and improving computational efficiency. Second, LASSO logistic regression (LR) with 10-fold cross-validation was used to identify the most predictive features and construct a radiomic signature. This two-stage strategy was preferred over embedded feature selection alone (e.g., LASSO only) because it reduces the likelihood of LASSO arbitrarily selecting among highly correlated features, thereby yielding a more robust, parsimonious, and generalizable model.

Following feature extraction, the resulting high-dimensional feature matrix was rigorously screened to ensure data integrity. Radiomic or deep learning features containing missing or infinite values in the training cohort were identified and excluded from further analysis. In addition, features with near-zero variance were removed to improve data stability. Consequently, the final feature matrix used for the mRMR and LASSO feature selection pipeline was complete and contained no missing values requiring statistical imputation. This data-cleaning protocol was applied exclusively to the training cohort, and the same set of valid features was subsequently extracted and evaluated in the validation and external cohorts.

Deep learning model construction

Deep feature learning for ultrasound habitat subregions was performed using the ResNet50 architecture, which was originally developed for ImageNet-scale image classification. The model was trained in a supervised framework to distinguish LVI-positive from LVI-negative breast cancers using categorical cross-entropy loss and the Adam optimizer with a batch size of 32. Training was conducted for 100 epochs, with early stopping and learning-rate decay applied to minimize overfitting. The model weights that achieved the best performance on the internal validation cohort were retained.

For feature extraction, the trained ResNet50 model was truncated at the penultimate layer, and the output of the global average pooling layer was used as the deep representation of each habitat image, yielding a 2048-dimensional feature vector.

Fusion model construction

The selected habitat-based radiomic features and the 2048-dimensional deep features extracted by ResNet50 were concatenated at the sample level to generate a fused feature vector. Because these features differed in origin and scale, Z-score normalization was applied before classifier training. To strictly prevent data leakage, the mean and standard deviation (SD) used for Z-score transformation were calculated exclusively from the training cohort. These training-derived parameters were then directly applied to normalize the internal validation and external cohorts. Similarly, the specific radiomic and deep learning features selected using mRMR and LASSO in the training cohort were fixed and directly extracted from the validation and test cohorts without further feature evaluation or selection. The standardized fused features were subsequently used to develop a binary classifier for predicting LVI status in breast cancer. To ensure the reproducibility of the findings, a fixed global random seed (random_state = 42) was used for all stochastic processes, including data splitting, cross-validation, and model initialization. The hyperparameters of the LightGBM classifier were optimized using a grid search combined with 10-fold cross-validation exclusively in the training cohort. The search space included num_leaves (20-100), max_depth (3-10), learning_rate (0.01-0.2), and n_estimators (100-1000). The optimal hyperparameter combination was selected based on the highest mean AUC across the 10 folds. During model training, an early stopping strategy with a patience of 20 rounds was applied, with performance on the internal validation cohort used as the monitoring criterion to minimize overfitting. The final model, incorporating the optimized hyperparameters and early-stopped weights, was then fixed and evaluated on the external cohorts without any further tuning.

Notably, the center variable was intentionally excluded as a predictor to prevent site-specific learning and improve the model’s generalizability to new institutions. Instead, center-related confounding and potential domain shift were addressed through three strategies: (1) harmonization of ultrasound acquisition parameters and centralized quality control across participating institutions; (2) confirmation of balanced baseline clinical and pathological characteristics, including LVI prevalence, across all cohorts; and (3) rigorous evaluation using two completely independent external cohorts (Centers 2 and 3) without any model retraining or hyperparameter tuning, thereby empirically demonstrating the model’s robustness to domain shift.

Reader study and evaluation of model-assisted diagnosis

To evaluate the clinical utility of the fusion model, two radiologists with different levels of experience (one junior and one senior) independently assessed LVI status in the external cohorts from Centers 2 and 3 while blinded to the pathological findings. Their assessments were directly compared with the predictions generated by the fusion model. To further evaluate the model’s clinical applicability, the same junior radiologist independently re-evaluated the same cases with assistance from the fusion model while remaining blinded to the pathological findings. Diagnostic performance before and after model assistance, including the AUC and accuracy, was compared to determine the extent to which model assistance improved diagnostic performance, particularly for the less-experienced reader.

To minimize potential recall, learning, and order effects, a washout period of at least 4 weeks was implemented between the unaided and AI-assisted reading sessions. In addition, the case order was completely re-randomized for the second reading session, and the junior radiologist remained blinded to both the previous assessments and the pathological ground truth throughout the evaluation.

Statistical analysis

All machine learning pipelines, including image preprocessing, radiomic feature extraction, deep learning feature extraction, and model training, were implemented in Python (version 3.8.10). The primary libraries included PyTorch (version 1.10.0) for the ResNet50 architecture, PyRadiomics (version 3.0.1) for radiomic feature extraction, and LightGBM (version 3.3.2) for classifier development. Statistical analyses, including the DeLong test for AUC comparison, McNemar’s test, and data visualization, were performed using R (version 4.1.3) and IBM SPSS Statistics (version 26.0, IBM Corp., Armonk, NY, USA).

Continuous variables were assessed for normality using the Shapiro-Wilk test. Normally distributed continuous variables are presented as the mean ± SD and were compared using the independent-samples t-test or one-way analysis of variance. Non-normally distributed continuous variables are presented as the median with interquartile range and were compared using the Mann-Whitney U test or Kruskal-Wallis test. Categorical variables are presented as frequencies and percentages and were compared using the chi-square test or Fisher’s exact test.

To strictly prevent data leakage, model development and evaluation followed a sequential, isolated pipeline. Feature selection (mRMR and LASSO), data normalization (Z-score transformation), and hyperparameter optimization were performed exclusively in the training cohort. The mean and SD calculated from the training cohort were subsequently used to normalize the internal validation and external cohorts. Likewise, only the features selected in the training cohort were extracted and applied to the validation and test cohorts. The internal validation cohort was used solely to monitor early stopping during model training, whereas the two independent external cohorts were reserved exclusively for final performance evaluation. The predictive performance of the models and radiologists was evaluated using the AUC with 95% confidence intervals (CIs). AUCs were compared using the non-parametric DeLong test. Sensitivity, specificity, and accuracy were compared between the fusion model and the radiologists, as well as between unaided and model-assisted interpretations, using McNemar’s test. All statistical tests were two-sided, and exact p values were reported to three decimal places, except when p < 0.001. A p value < 0.05 was considered statistically significant. To assess the agreement between the predicted probabilities and the observed outcomes, calibration curves were generated for the model in all cohorts.

RESULTS

Patient characteristics

The overall study design and analytical pipeline are summarized in Figure 2. This multicenter retrospective study included 1,096 patients with pathologically confirmed invasive breast cancer from three institutions and analyzed routine B-mode ultrasound images. The dataset comprised an internal training cohort (n = 494), an internal validation cohort (n = 212), and two external cohorts (Cohort 1: n = 200; Cohort 2: n = 190). LVI status was determined by histopathological evaluation of H&E-stained sections, supplemented by CD31 immunohistochemistry. Overall, 44.34% (n = 486) of patients were LVI-positive, whereas 55.66% (n = 610) were LVI-negative. To formally assess potential center-specific heterogeneity, clinical characteristics were compared across the participating institutions. As shown in Table 1, no statistically significant differences were observed in age (p = 0.401) or tumor diameter (p = 0.090) among the cohorts. Importantly, the prevalence of LVI was highly consistent across all cohorts: 43.93% (217/494) in the training cohort, 44.81% (95/212) in the internal validation cohort, 45.00% (90/200) in External Cohort 1, and 44.21% (84/190) in External Cohort 2 (p = 0.993). The comparable prevalence of LVI and balanced baseline characteristics across cohorts suggest the absence of clinically meaningful demographic or disease-related heterogeneity among the participating centers. Furthermore, harmonization of ultrasound acquisition parameters and centralized pathological review helped minimize technical and assessment-related variability, thereby supporting the robustness of the multicenter validation (Table 1).

Figure 2.

Schematic diagram of the study design and process.

ROI, region of interest.

TABLE 1. Patient Characteristics in the Training Cohort, Internal Test Cohort, External Validation Cohort 1, and Validation Cohort 2.

Feature	Internal training cohort (n = 494)	Internal validation cohort (n = 212)	External test cohort 1 (n = 200)	External test cohort 2 (n = 190)	p value -
Age	49.17 ± 13.57	48.24 ± 13.49	49.89 ± 12.13	47.89 ± 13.36	0.401
Diameter	18.23 ± 6.61	17.58 ± 7.32	18.78 ± 7.05	19.21 ± 6.95	0.090
LVI status	-	-	-	-	0.993

Performance of habitat radiomics and its biological relevance

Based on tumor ROIs identified on routine B-mode ultrasound, unsupervised k-means clustering (k = 3) partitioned each tumor into three distinct habitat subregions (Figure 3a). A total of 107 radiomic features were extracted from each habitat, including first-order, shape, and higher-order texture features (Figure 3b).

Figure 3.

K-means unsupervised clustering and diagnostic performance of each habitat subregion. (a) Calinski-Harabasz (CH) score curve for determining the optimal number of clusters, showing an optimal peak at three clusters. (b) Three-dimensional visualization of the K-means unsupervised clustering results, displaying the spatial distribution and voxel proportions of the three habitat subregions (Habitat1: 29.01%, Habitat2: 56.71%, Habitat3: 14.28%). (c) AUC (train vs. test) across all models and habitats, where a shorter line indicates higher stability.

AUC, area under the curve.

We evaluated the performance of LightGBM, LR, support vector machine (SVM), k-nearest neighbors (KNN), and ExtraTrees classifiers across four feature sets (Habitat1-3, Habitat1, Habitat2, and Habitat3) using training and testing AUCs to assess model generalizability. Most classifiers achieved high training AUCs but showed substantial declines in testing AUCs, indicating varying degrees of minimize overfitting. For the Habitat1-3 feature set, KNN and SVM achieved training AUCs of 0.851 and 0.810, respectively, whereas their testing AUCs decreased to 0.613 and 0.659. In contrast, LightGBM demonstrated greater stability, with its AUC declining only from 0.724 to 0.676. A similar pattern was observed for Habitat3, where KNN and SVM achieved training AUCs of 0.810 and 0.705 but testing AUCs of 0.627 and 0.586, respectively, whereas LightGBM maintained relatively stable performance (0.666 to 0.653) (Figure 3c).

Habitat2 was identified as the optimal feature set for LightGBM, with training and testing AUCs of 0.722 and 0.714, respectively, resulting in the smallest performance gap (0.008). The testing AUCs of the other classifiers were 0.705 for ExtraTrees, 0.715 for KNN, 0.748 for SVM, and 0.769 for LR, all of which exhibited larger discrepancies between the training and testing AUCs (e.g., ExtraTrees: 0.800 to 0.705; LR: 0.659 to 0.769) (Figure 3c). Accordingly, Habitat2 was designated as the high-risk habitat subregion, and LightGBM was selected for subsequent analyses because it provided the best balance between discriminative performance and generalizability.

Development and performance of deep learning and fusion models

Deep features were extracted from all habitat images using an ImageNet-pretrained ResNet50 model. These features were subsequently fused with the radiomic features derived from Habitat2, and a LightGBM classifier was used to construct the final fusion model (ResNet50 + Habitat2).

The Habitat2-only radiomic model demonstrated moderate predictive performance, with AUCs of 0.722 (95% CI: 0.676-0.769) in the training cohort, 0.715 (95% CI: 0.644-0.786) in the internal validation cohort, 0.706 (95% CI: 0.632-0.779) in External Cohort 1, and 0.714 (95% CI: 0.639-0.789) in External Cohort 2. The ResNet50-only model achieved substantially better performance, with AUCs of 0.890 (95% CI: 0.863-0.918), 0.864 (95% CI: 0.813-0.914), 0.855 (95% CI: 0.802-0.909), and 0.840 (95% CI: 0.782-0.898), respectively (Figure 4a-d, Table 2).

Figure 4.

Diagnostic performance of Habitat2, Deep Learning (ResNet50), and the Fusion Model in the training cohort, internal test cohort, external validation cohort 1, and validation cohort 2, along with DeLong test results. (a-d) ROC curves for the training, internal validation, external test cohort 1, and external test cohort 2, respectively. (e-h) DeLong test results comparing the fusion model with Habitat2 and ResNet50 in the respective cohorts.

AUC, area under the curve; CI, confidence interval; ROC, receiver operating characteristic.

TABLE 2. Diagnostic Performance of Habitat2, Deep Learning (ResNet50), and the Fusion Model in the Training Cohort, Internal Test Cohort, External Validation Cohort 1, and Validation Cohort 2.

-	Model name	Accuracy	AUC 95% CI	Sensitivity	Specificity	Precision	F1
Internal training cohort	Habitat2	0.727	0.722 (0.6758-0.7688)	0.507	0.899	0.797	0.620
	ResNet50	0.814	0.89 (0.8628-0.9176)	0.714	0.892	0.838	0.771
	Fusion model	0.862	0.918 (0.8939-0.9429)	0.802	0.91	0.874	0.837
Internal validation cohort	Habitat2	0.703	0.715 (0.6439-0.7856)	0.6	0.786	0.695	0.644
	ResNet50	0.802	0.864 (0.8134-0.9144)	0.842	0.769	0.748	0.792
	Fusion model	0.858	0.898 (0.8553-0.9399)	0.779	0.923	0.892	0.831
External test cohort 1	Habitat2	0.695	0.706 (0.6318-0.7793)	0.589	0.782	0.688	0.635
	ResNet50	0.790	0.855 (0.8021-0.9088)	0.900	0.700	0.711	0.794
	Fusion model	0.85	0.89 (0.8446-0.9352)	0.767	0.918	0.885	0.821
External test cohort 2 -	Habitat2	0.737	0.714 (0.6393-0.7886)	0.476	0.943	0.870	0.615
	ResNet50	0.821	0.840 (0.7822-0.8977)	0.762	0.868	0.821	0.790
	Fusion model	0.868	0.905 (0.8618-0.9475)	0.750	0.962	0.940	0.834

AUC, area under the curve; CI, confidence interval.

The fusion model consistently outperformed both single-modality models, achieving AUCs of 0.918 (95% CI: 0.894-0.943) in the training cohort, 0.898 (95% CI: 0.855-0.940) in the internal validation cohort, 0.890 (95% CI: 0.845-0.935) in External Cohort 1, and 0.905 (95% CI: 0.862-0.948) in External Cohort 2 (fusion vs. Habitat2: all p < 0.001; fusion vs. ResNet50: p = 0.000, 0.021, 0.029, and 0.026; DeLong test) (Figure 4e-h). Across all cohorts, the fusion model also demonstrated high specificity (0.910-0.962), sensitivity (0.750-0.802), precision (0.874-0.940), and F1-score (0.821-0.837). The optimal probability cut-off, determined using the Youden index, ranged from 0.492 to 0.521.

Model calibration

The calibration curves demonstrated good agreement between the predicted probabilities generated by the fusion model and the observed frequencies of LVI in the training, internal validation, and both external cohorts. In all four cohorts, the calibration curves closely followed the ideal diagonal reference line (Supplementary Figure 1), indicating good calibration, reliable probability estimation, and potential clinical utility for decision-support system. Nevertheless, slight deviations from the ideal reference line were observed in some intermediate- and high-probability ranges, suggesting that the predicted probabilities were not perfectly calibrated across all risk intervals.

Model interpretability: SHapley Additive exPlanations analysis and heatmap of feature correlations with CD31 expression

To improve the interpretability of the multimodal fusion model, SHapley Additive exPlanations (SHAP) analysis was performed. Figure 5a presents the global SHAP summary plot for the test cohort. Among all features, habitatDL1897 and original_shape_MeshVolume_h2 were identified as the most influential predictors.

Figure 5.

(a) Feature importance ranking based on SHapley Additive exPlanations (SHAP), represented by the mean absolute SHAP value for each feature. (b) and (c) SHAP force plots for Patient A (lymphovascular invasion-positive) and Patient B (with recurrence), respectively. Colors indicate the direction of each feature’s contribution to the model prediction (red for positive contribution, blue for negative contribution), with bar length reflecting the magnitude of contribution.

In a representative patient with LVI-positive breast cancer, the model generated a high-risk score of 0.8, which was primarily driven by original_shape_MeshVolume_h2 (feature value = 1.6282) and habitatDL1462 (feature value = 1.5308) (Figure 5b). In contrast, for a representative LVI-negative patient, the model assigned a very low-risk score (-0.93), largely because a smaller tumor volume, reflected by original_shape_MeshVolume_h2 (feature value = -1.2149), substantially reduced the predicted risk (Figure 5c).

To investigate the biological relevance of the identified imaging features, correlations between CD31 expression and habitat radiomic and deep learning-derived features were analyzed. Habitat2 radiomic texture features were negatively correlated with CD31 expression, including original_glszm_ZoneEntropy_h2 (ρ = -0.24) and original_glszm_ZonePercentage_h2 (ρ = -0.21) (Supplementary Figure 2a). In contrast, deep learning-derived habitat features exhibited a bidirectional correlation pattern: habitatDL678 (ρ = 0.22) and habitatDL751 (ρ = 0.20) were positively correlated with CD31 expression, whereas habitatDL1521 (ρ = -0.21) was negatively correlated (Supplementary Figure 2b). These findings support the potential of the fusion model as a non-invasive surrogate for assessing tumor angiogenic status.

Comparison with radiologists’ performance

In both external cohorts, the fusion model (AUC = 0.890 and 0.905, respectively) outperformed both the junior radiologist without model assistance (AUC = 0.675 and 0.670) and the senior radiologist (AUC = 0.770 and 0.768) (all p < 0.001, DeLong test) (Supplementary Figure 3, Table 3).

TABLE 3. Diagnostic Performance of Radiologists with and without the Fusion Model in External Validation Cohort 1 and Validation Cohort 2.

-	Model name	Accuracy	AUC 95% CI	Sensitivity	Specificity	Precision	F1
Radiologist with vs. without fusion model on external test cohort 1	Fusion model	0.85	0.89 (0.8446-0.9352)	0.767	0.918	0.885	0.821
	Junior without fusion model	0.68	0.675 (0.6093-0.7402)	0.622	0.727	0.651	0.636
	Senior without fusion model	0.78	0.770 (0.7116-0.8278)	0.667	0.873	0.811	0.732
	Junior with fusion model	0.90	0.896 (0.8528-0.9391)	0.856	0.936	0.917	0.885
Radiologist with vs. without fusion model on external test cohort 2	Fusion model	0.868	0.905 (0.8618-0.9475)	0.750	0.962	0.940	0.834
	Junior without fusion model	0.684	0.670 (0.6039-0.7361)	0.548	0.792	0.676	0.605
	Senior without fusion model	0.774	0.768 (0.7066-0.8284)	0.714	0.821	0.759	0.736
	Junior with fusion model	0.905	0.902 (0.8581-0.9449)	0.869	0.934	0.912	0.890

Importantly, with assistance from the fusion model, the junior radiologist’s diagnostic performance improved substantially, with AUCs increasing to 0.896 in External Cohort 1 and 0.902 in External Cohort 2. This performance exceeded that of the unaided senior radiologist and closely approached that of the standalone fusion model (Figure 6, Table 3). Diagnostic performance also increased from 0.680-0.684 without assistance to 0.900-0.905 with model assistance. These findings highlight the value of the fusion model not only as a diagnostic tool but also as a clinical decision-support system that has the potential to narrow the performance gap between less-experienced and more-experienced radiologists.

Figure 6.

Radar chart visualization of the diagnostic performance of individual models and radiologists with/without the fusion model.

AUC, area under the curve.

DISCUSSION

In this multicenter study, we developed a feature-level fusion model that combined ultrasound-based habitat radiomics with deep learning (ResNet50) for the preoperative, non-invasive prediction of LVI in invasive breast cancer. The ResNet50 + Habitat2_LightGBM classifier demonstrated strong predictive performance across the training, internal validation, and two external cohorts, achieving AUCs of 0.918, 0.898, 0.890, and 0.905, respectively. These values exceeded those of both the habitat radiomics model (AUCs: 0.706-0.722) and the standalone ResNet50 model (AUCs: 0.840-0.890). Habitat2, identified as the high-risk habitat, showed weak-to-moderate correlations with CD31 expression. Specifically, original_glszm_ZoneEntropy_h2 (ρ = -0.24), original_glszm_ZonePercentage_h2 (ρ = -0.21), original_gldm_SmallDependenceEmphasis_h2 (ρ = 0.18), original_glrlm_LongRunEmphasis_h2 (ρ = 0.18), habitatDL678 (r = 0.22), habitatDL751 (r = 0.20), and habitatDL1521 (r = -0.21) were significantly correlated with CD31 expression, supporting the biological interpretability of the proposed model.

In both external cohorts, the fusion model outperformed both unaided senior and junior radiologists in terms of AUC and diagnostic performance. With model assistance, the junior radiologist achieved accuracies of 0.900 and 0.905 and AUCs of 0.896 and 0.902, exceeding the performance of the unaided senior radiologist. These findings support the potential value of the fusion model as a clinical decision-support tool, particularly in primary care or resource-limited settings.^{7, 8}

Most previously reported preoperative LVI prediction models have been based on dynamic contrast-enhanced-MRI or multiparametric MRI, with externally validated AUCs generally ranging from 0.85 to 0.90. However, these approaches are limited by the need for intravenous contrast agents, higher examination costs, longer acquisition times, and contraindications in patients with impaired renal function.^{16, 17, 18, 19, 20, 21, 22} For example, Yang et al.,¹⁶ Liang et al.,¹⁷ Chen et al.,¹⁸ Zheng et al.,¹⁹ and Zhang et al.²⁰ reported MRI-based models with AUCs of 0.872, 0.892, 0.894, 0.901, and 0.885, respectively. Although these MRI-based models demonstrated excellent predictive performance, they rely on contrast enhancement and are approximately 5-10 times more expensive than conventional ultrasound examinations.

Previous ultrasound-based studies have also demonstrated the potential of radiomics. Xu et al.²³ reported a grayscale ultrasound radiomics nomogram with a validation AUC of approximately 0.85, significantly outperforming a model based solely on conventional ultrasound features (AUC ≈ 0.71). A subsequent multicenter study integrating B-mode ultrasound with color Doppler flow imaging achieved an external AUC of approximately 0.91 and a sensitivity of 86.9%.²⁴ Li et al.²⁵ proposed a deep learning radiomics nomogram that combined B-mode and contrast-enhanced ultrasound, achieving an external AUC of 0.914. However, this approach still depended on contrast agents. In contrast, our study achieved comparable predictive performance using only conventional B-mode ultrasound, thereby maximizing clinical accessibility by eliminating the need for contrast agents, avoiding radiation exposure, reducing examination costs, enabling bedside application, and maintaining strong generalizability, with external AUCs of 0.890 and 0.905.

Conventional radiomics typically extracts global, intratumoral, or peritumoral features but often overlooks the spatial heterogeneity of intratumoral subregions associated with microvasculature, hypoxia, and necrosis, which are key drivers of LVI.^{5, 12} In this study, we applied unsupervised k-means clustering (k = 3) to partition each tumor into three habitat subregions and identified Habitat2 as a high-risk habitat, characterized by a vascular index of 0.42 ± 0.11, which was significantly higher than those of the other subregions (p < 0.001). The features derived from Habitat2 were strongly correlated with CD31-based microvessel density, providing, for the first time in ultrasound imaging, a biologically meaningful characterization of LVI-associated microvascular invasion. Previous habitat imaging studies have primarily focused on MRI for predicting axillary nodal burden or response to neoadjuvant chemotherapy in breast cancer.^{26, 27} To our knowledge, this is the first study to apply habitat imaging to LVI prediction and to demonstrate that feature-level fusion of habitat radiomics and deep learning improves the AUC by 0.03-0.15 compared with single-modality models, highlighting the value of habitat-based features for capturing tumor heterogeneity associated with microvascular invasion.

This study has several limitations. First, its retrospective design may have introduced selection bias, and prospective validation in multiethnic cohorts is warranted. Second, habitat construction was based on manually delineated ROIs from a single static two-dimensional B-mode ultrasound image. Because ultrasound is inherently operator-dependent and dynamic, reliance on a single imaging plane may not fully capture the three-dimensional (3D) spatial heterogeneity of the entire lesion. Consequently, variability in habitat delineation may affect the robustness of the extracted radiomic features, as grayscale intensity distributions can differ across imaging planes. Although strict image quality control and acquisition parameter harmonization were implemented to minimize these effects, future studies incorporating 3D ultrasound volumes, cine ultrasound imaging, or shear-wave elastography, together with automated 3D segmentation, may further refine habitat delineation and improve the robustness and biological representativeness of the extracted features. Third, the relationship between high-risk habitats and molecular subtypes (e.g., triple-negative versus luminal tumors) has not been fully investigated and warrants dedicated subgroup analyses. Fourth, the reader study used a sequential design in which the same junior radiologist interpreted the cases first without and then with model assistance. Although a 4-week washout period and case re-randomization were implemented to reduce recall and learning effects, this design may still overestimate the incremental benefit of model assistance compared with a true first-read scenario. Therefore, the observed improvement should be interpreted as reflecting the model’s potential to modify and refine diagnostic decisions rather than providing an unbiased estimate of real-world first-read model-assisted performance. Future prospective studies using a randomized, counterbalanced crossover design with multiple radiologists and adequate washout periods are needed to provide more robust validation of the model’s clinical utility. Fifth, habitat generation was based on grayscale intensity similarity, which, in B-mode ultrasound, may be influenced by acquisition-related factors such as gain, time-gain compensation, and scanner-specific post-processing algorithms. Although rigorous grayscale intensity standardization, including global histogram matching and intra-tumor min-max scaling followed by Z-score normalization, was implemented to minimize these effects, some residual inter-scanner variability may still have remained. This variability could introduce subtle noise into the habitat delineation process. Nevertheless, the robust and consistent performance of the model across different ultrasound vendors in the external cohorts suggests that the extracted radiomic and deep learning features capture fundamental textural and structural patterns that are relatively resilient to such residual grayscale variations.

Furthermore, by training the model exclusively on data from Center 1 and validating it on two independent external cohorts (Centers 2 and 3) without including center identity as a predictor, we effectively minimized site-specific bias. The consistently strong performance across independent institutions demonstrates the model’s robustness, transportability, and generalizability in real-world clinical settings.

In summary, we developed a biologically interpretable habitat-deep learning fusion model based on routine breast ultrasound for the non-invasive preoperative prediction of LVI in invasive breast cancer. The proposed model has the potential to facilitate preoperative risk stratification, support clinical decision-making regarding sentinel lymph node biopsy, and reduce unnecessary overtreatment.

Supplementary Materials

Supplementary Figures

https://d2v96fxpocvxx.cloudfront.net/beb8919b-f013-4ea1-b1c8-40332e840fe1/content-images/ea1ea57c-5af4-4052-ab49-516f203a8d49.pdf