Noninvasive molecular subtyping of breast cancer using multimodal ultrasound spatiotemporal transformer Authors: Tao Chen, Qinghua Niu, Zichen An, Chenhui Wang, Zhihao Chen, Lianfang Du, Chao Jia, Chunxiao Li, Jiying Gu, Ting Sun, Yuxiang Wang, Fan Li, Hongming Shan Published: 2026 ABSTRACT Molecular subtyping is essential for guiding systemic therapy in breast cancer but currently requires invasive biopsy. Conventional B-mode ultrasound offers rich anatomical information, yet lacks the functional dynamics needed to capture the comprehensive biology of tumors. Here, we present the first multimodal ultrasound spatiotemporal transformer, MUST-Sub, which integrates paired B-mode morphological features with contrast-enhanced ultrasound (CEUS) hemodynamic patterns to classify Luminal, HER2-enriched, and triple-negative subtypes. Training on a retrospective development cohort, and validated on internal, prospective, and multicenter external cohorts, MUST-Sub achieved macro-average areas under the receiver operating characteristic curve (AUCs) of 0.94, 0.90, and 0.92, respectively, and Luminal versus non-Luminal AUCs of 0.92, 0.88, and 0.91, outperforming B-mode–only deep learning baselines. MUST-Sub also produced interpretable quantitative biomarkers derived from spatiotemporal attention: the morphology-associated biomarker showed inverse correlations with tumor size (Spearman ρ = [− 0.34, − 0.23]; all p < . 05), while the hemodynamics-associated biomarker correlated positively with tumor size (ρ = [0.24, 0.32]; all p < . 05) and Ki-67 proliferation index (ρ = [0.21, 0.24]; all p < . 05). These results suggest that multimodal ultrasound with spatiotemporal modeling can serve as a promising adjunctive approach for non-invasive pre-biopsy molecular phenotyping of breast cancer. Introduction Breast cancer is the most common malignancy among women worldwide1,2. Molecular subtyping that classifies breast tumors into Luminal, human epidermal growth factor receptor 2 (HER2)-enriched, and triple-negative is essential for guiding systemic therapy3–5. However, the diagnostic standard, core needle biopsy, is invasive and vulnerable to sampling error arising from intratumoral heterogeneity6, highlighting the need for a noninvasive global alternative. Ultrasound (US) represents a promising candidate due to its widespread use and capability to assess the entire tumor noninvasively7. However, single-modality US is limited in capturing the complete pathophysiological picture. Conventional B-mode US provides high-resolution anatomical delineation but offers limited insight into functional dynamics8,9. Contrast-enhanced US (CEUS) complements this limitation by dynamically visualizing tumor neovascularization and perfusion, capturing functional insights that have been directly associated with distinct molecular subtypes10,11. Despite their complementary biological relevance, US remains underutilized for quantitative tumor characterization due to substantial operator dependence and the difficulty of interpreting complex morphological and hemodynamic patterns12. Artificial intelligence (AI) offers an opportunity to overcome these challenges. However, most prior AI models13–16 face three major limitations: (1) they rely predominantly on B-mode imaging, overlooking the critical functional dimension available from CEUS; (2) they are constrained to label scarcity due to the cost of subtype annotation; and (3) they often operate as “black boxes”, providing limited interpretability and insufficient biological plausibility, which remain key barriers to clinical adoption. In this paper, we conducted a multicenter study to develop and validate MUST-Sub, a Multimodal Ultrasound Spatiotemporal Transformer model designed to predict molecular Subtypes of breast cancer from paired B-mode US and CEUS (Fig. 1). To our knowledge, MUST-Sub is the first approach to integrate B-mode morphological cues with CEUS hemodynamic patterns within a unified spatiotemporal transformer framework, while leveraging large-scale unlabeled data for self-supervised pretraining to improve generalization. The model was rigorously evaluated across internal, prospective, and multicenter external cohorts, demonstrating a consistent improvement of 9-11% macro-average areas under the receiver operating characteristic curve (AUC) and 6-8% Luminal versus non-Luminal AUC over the best B-mode-only models. Furthermore, we derive quantitative spatiotemporal biomarkers from the trained MUST-Sub and validate their associations with established pathological indicators, including tumor size and the Ki-67 proliferation index, providing biologically grounded evidence that supports model interpretability.Fig. 1Overview of the study design, datasets, and model development.a Clinical workflow for breast cancer diagnosis and surrogate molecular subtyping based on IHC/FISH. b Multicenter cohorts used in this study, including unlabeled and labeled paired B-mode US and CEUS data, and the internal, prospective, and external validation splits. c MUST-Sub development workflow, comprising self-supervised pretraining and supervised fine-tuning for subtype classification and biomarker derivation. The autoencoder compresses B-mode images and CEUS video sequences into compact latent representations, enabling tractable spatiotemporal modeling. US ultrasound, CEUS contrast-enhanced ultrasound, PR progesterone receptor, ER estrogen receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemistry, FISH fluorescence in situ hybridization. Overview of the study design, datasets, and model development. a Clinical workflow for breast cancer diagnosis and surrogate molecular subtyping based on IHC/FISH. b Multicenter cohorts used in this study, including unlabeled and labeled paired B-mode US and CEUS data, and the internal, prospective, and external validation splits. c MUST-Sub development workflow, comprising self-supervised pretraining and supervised fine-tuning for subtype classification and biomarker derivation. The autoencoder compresses B-mode images and CEUS video sequences into compact latent representations, enabling tractable spatiotemporal modeling. US ultrasound, CEUS contrast-enhanced ultrasound, PR progesterone receptor, ER estrogen receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemistry, FISH fluorescence in situ hybridization. Results Summary of datasets Between October 2018 and November 2024, this multicenter study enrolled 1039 unlabeled and 1054 labeled paired B-mode US and CEUS examinations. Labeled breast cancer cases were categorized into three molecular subtypes—Luminal, HER2-enriched, and triple-negative—according to estrogen receptor (ER), progesterone receptor (PR), and HER2 status, as determined by immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) (Fig. 1a). Clinicopathological characteristics by subtype are summarized in Table 1. As expected, the subtype groups differed significantly in histologic type (p = 0.002), pathologic tumor size (p = 0.014), histologic grade (p < 0.001), PR status (p < 0.001), HER2 status (p < 0.001), Ki-67 index (p < 0.001), and perineural invasion (p = 0.026), reflecting the known biologic heterogeneity of breast cancer.Table 1Clinicopathologic characteristics stratified by IHC/FISH-defined molecular subtypeLuminalHER2-enrichedTriple-negative(N = 771)(N = 144)(N = 139)Lesion location (p = 0.597)Left412 (53.4%)80 (55.6%)69 (49.6%)Right359 (46.6%)64 (44.4%)70 (50.4%)Age (years) (p = 0.144)Age ≤50250 (32.4%)40 (27.8%)35 (25.2%)Age > 50521 (67.6%)104 (72.2%)104 (74.8%)Histologic type (p = 0.002)Ductal carcinoma644 (83.5%)131 (91.0%)129 (92.8%)Lobular carcinoma27 (3.5%)0 (0.0%)1 (0.7%)Others100 (13.0%)13 (9.0%)9 (6.5%)Pathologic size (mm) (p = 0.014)≤20405 (52.5%)54 (37.5%)63 (45.3%)> 20340 (44.1%)87 (60.4%)74 (53.2%)NA26 (3.4%)3 (2.1%)2 (1.5%)Histologic grade (p < 0.001)Grade I90 (11.7%)2 (1.4%)4 (2.9%)Grade II398 (51.6%)37 (25.7%)32 (23.0%)Grade III283 (36.7%)105 (72.9%)103 (74.1%)PR status (p < 0.001)Positive664 (86.1%)0 (0.0%)0 (0.0%)Negative107 (13.9%)144 (100.0%)139 (100.0%)HER2 status (p < 0.001)Positive121 (15.7%)144 (100.0%)0 (0.0%)Negative650 (84.3%)0 (0.0%)139 (100.0%)Ki-67 index (%) (p < 0.001)≤14218 (28.3%)10 (6.9%)11 (7.9%)> 14553 (71.7%)134 (93.1%)128 (92.1%)Lymphovascular invasion (p = 0.096)Present214 (27.8%)41 (28.5%)27 (19.4%)Absent557 (72.2%)103 (71.5%)112 (80.6%)Perineural invasion (p = 0.026)Present98 (12.7%)8 (5.6%)17 (12.2%)Absent673 (87.3%)136 (94.4%)122 (87.8%)Lymph-node metastasis (p = 0.680)0528 (68.5%)99 (68.8%)98 (70.5%)1-3153 (19.8%)33 (22.9%)28 (20.1%)> 390 (11.7%)12 (8.3%)13 (9.4%)Variables are reported as n (%) and compared using the Pearson χ2 test, or Fisher’s exact test when Cochran’s rule was violated. For Pathologic size, the “NA” category was excluded prior to testing. Pathologic size refers to the maximum tumor diameter.NA not available, PR progesterone receptor, HER2 human epidermal growth factor receptor 2. Table 1Clinicopathologic characteristics stratified by IHC/FISH-defined molecular subtypeLuminalHER2-enrichedTriple-negative(N = 771)(N = 144)(N = 139)Lesion location (p = 0.597)Left412 (53.4%)80 (55.6%)69 (49.6%)Right359 (46.6%)64 (44.4%)70 (50.4%)Age (years) (p = 0.144)Age ≤50250 (32.4%)40 (27.8%)35 (25.2%)Age > 50521 (67.6%)104 (72.2%)104 (74.8%)Histologic type (p = 0.002)Ductal carcinoma644 (83.5%)131 (91.0%)129 (92.8%)Lobular carcinoma27 (3.5%)0 (0.0%)1 (0.7%)Others100 (13.0%)13 (9.0%)9 (6.5%)Pathologic size (mm) (p = 0.014)≤20405 (52.5%)54 (37.5%)63 (45.3%)> 20340 (44.1%)87 (60.4%)74 (53.2%)NA26 (3.4%)3 (2.1%)2 (1.5%)Histologic grade (p < 0.001)Grade I90 (11.7%)2 (1.4%)4 (2.9%)Grade II398 (51.6%)37 (25.7%)32 (23.0%)Grade III283 (36.7%)105 (72.9%)103 (74.1%)PR status (p < 0.001)Positive664 (86.1%)0 (0.0%)0 (0.0%)Negative107 (13.9%)144 (100.0%)139 (100.0%)HER2 status (p < 0.001)Positive121 (15.7%)144 (100.0%)0 (0.0%)Negative650 (84.3%)0 (0.0%)139 (100.0%)Ki-67 index (%) (p < 0.001)≤14218 (28.3%)10 (6.9%)11 (7.9%)> 14553 (71.7%)134 (93.1%)128 (92.1%)Lymphovascular invasion (p = 0.096)Present214 (27.8%)41 (28.5%)27 (19.4%)Absent557 (72.2%)103 (71.5%)112 (80.6%)Perineural invasion (p = 0.026)Present98 (12.7%)8 (5.6%)17 (12.2%)Absent673 (87.3%)136 (94.4%)122 (87.8%)Lymph-node metastasis (p = 0.680)0528 (68.5%)99 (68.8%)98 (70.5%)1-3153 (19.8%)33 (22.9%)28 (20.1%)> 390 (11.7%)12 (8.3%)13 (9.4%)Variables are reported as n (%) and compared using the Pearson χ2 test, or Fisher’s exact test when Cochran’s rule was violated. For Pathologic size, the “NA” category was excluded prior to testing. Pathologic size refers to the maximum tumor diameter.NA not available, PR progesterone receptor, HER2 human epidermal growth factor receptor 2. Clinicopathologic characteristics stratified by IHC/FISH-defined molecular subtype Variables are reported as n (%) and compared using the Pearson χ2 test, or Fisher’s exact test when Cochran’s rule was violated. For Pathologic size, the “NA” category was excluded prior to testing. Pathologic size refers to the maximum tumor diameter. NA not available, PR progesterone receptor, HER2 human epidermal growth factor receptor 2. The labeled cohort was then divided into retrospective development, internal test, prospective validation, and multicenter external validation cohorts, as shown in Fig. 1b. The prospective cohort was used to assess temporal reproducibility in newly acquired data from the same institution, whereas the multicenter external cohort, comprising four hospitals using different ultrasound systems (Canon Aplio, Mindray Resona R9, and GE Logiq E9), was used to evaluate transferability across institutions and imaging platforms. Patient characteristics across all cohorts are presented in Supplementary Table 3. Overall comparisons among the five cohorts showed no significant differences in most clinicopathological variables, except for age (p = 0.044) and perineural invasion (p = 0.002). Pairwise analyses suggested that these differences were primarily driven by comparisons involving the prospective and external test cohorts. No significant pairwise differences were observed for the remaining variables, supporting broad comparability of patient characteristics across cohorts. Overview of MUST-Sub MUST-Sub is a multimodal ultrasound spatiotemporal transformer model for breast cancer subtyping that integrates B-mode US and CEUS through a two-stage training strategy of self-supervised pretraining followed by supervised fine-tuning (Fig. 1c). Spatiotemporal dependencies are modeled using transformer modules operating in the latent space of a variational autoencoder (VAE), which enables computationally tractable spatiotemporal modeling while aligning B-mode and CEUS representations in a unified latent space. Specifically, an autoregressive temporal transformer models CEUS dynamics and a diffusion-based spatial transformer models B-mode morphology. During self-supervised pretraining, both unlabeled and labeled development data are used to learn transferable representations of B-mode morphological patterns and CEUS enhancement dynamics. During supervised fine-tuning, cross-attention fuses CEUS temporal information with B-mode spatial context, and the resulting representation is classified using a multilayer perceptron. Model attention provides intrinsic interpretability, which we summarize using two attention-derived biomarkers: a morphology-associated biomarker, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}$$\end{document}Bm, from spatial attention on B-mode US and a hemodynamics-associated biomarker, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{h}$$\end{document}Bh, from temporal attention on CEUS (see Methods for formal definitions). MUST-Sub achieves robust performance across validation cohorts We assessed MUST-Sub on the internal, prospective, and multicenter external validation cohorts by assigning each case to the class with the highest predicted probability (Table 2). MUST-Sub demonstrated strong and consistent discrimination across cohorts, with macro-average sensitivity ranging from 0.73 to 0.90 and macro-average specificity from 0.87 to 0.93. Notably, the model achieved excellent specificity for triple-negative disease (0.98–0.99 across cohorts), indicating a low false-positive rate for this aggressive subtype. In the multicenter external cohort, which provided a stringent test of cross-site generalization, MUST-Sub maintained high specificity (0.92) and positive predictive value (0.98) for Luminal tumors, while achieving considerable sensitivity for HER2-enriched (1.00) and triple-negative (0.85) tumors. The low positive predictive value (PPV) for HER2-enriched in the prospective cohort (0.39) is consistent with the low prevalence of this subtype in that cohort (10.1%, 15/149), as PPV is mathematically dependent on prevalence. For illustration, under hypothetical prevalences of 20% and 50%, the expected PPV would increase to 0.58 and 0.85, respectively. The high negative predictive value (NPV) for this subtype (0.96) indicates that negative predictions were usually correct in this cohort. Misclassification of HER2-enriched or triple-negative tumors as Luminal remained infrequent across cohorts (HER2-enriched: 0/13 in the external cohort; triple-negative: 4/22, 3/18, and 2/13 across the internal, prospective, and external cohorts, respectively). Confusion matrices and t-distributed stochastic neighbor embedding (t-SNE)17 visualizations for each cohort are provided in Supplementary Fig. 1.Table 2Performance of MUST-Sub in subtype classification across the validation cohortsSensitivitySpecificityPPVNPV(95% CI)(95% CI)(95% CI)(95% CI)Internal test setLuminal0.9380.8040.9220.841(0.892-0.981)(0.681–0.909)(0.868–0.965)(0.731–0.945)HER2-enriched0.7920.9630.7920.963(0.613–0.950)(0.930–0.993)(0.625–0.955)(0.927–0.992)Triple-negative0.8180.9850.9000.971(0.631–0.960)(0.962–1.000)(0.750–1.000)(0.938–0.993)Macro-average0.8490.9180.8710.925(0.763–0.923)(0.874–0.955)(0.799–0.937)(0.874–0.955)Prospective cohortLuminal0.8620.7580.9260.610(0.797–0.924)(0.607–0.898)(0.874–0.972)(0.452–0.757)HER2-enriched0.6670.8810.3850.959(0.400–0.909)(0.824–0.933)(0.191–0.579)(0.921–0.992)Triple-negative0.6670.9770.8000.955(0.450–0.870)(0.948–1.000)(0.583–1.000)(0.918–0.985)Macro-average0.7320.8720.7030.841(0.620–0.839)(0.819–0.922)(0.608–0.797)(0.799–0.892)Multicenter external cohortLuminal0.8440.9230.9760.615(0.766–0.911)(0.810–1.000)(0.940–1.000)(0.455–0.763)HER2-enriched1.0000.8810.5001.000(1.000–1.000)(0.813–0.936)(0.300–0.692)(1.000–1.000)Triple-negative0.8460.9820.8460.982(0.615–1.000)(0.954–1.000)(0.625–1.000)(0.954–1.000)Macro-average0.8970.9280.7740.866(0.814–0.958)(0.884–0.964)(0.673–0.863)(0.824–0.904)Metrics were derived by assigning each case to the class with the highest predicted probability (i.e., argmax rule) and are presented with 95% confidence intervals (CI).PPV positive predictive value, NPV negative predictive value. Table 2Performance of MUST-Sub in subtype classification across the validation cohortsSensitivitySpecificityPPVNPV(95% CI)(95% CI)(95% CI)(95% CI)Internal test setLuminal0.9380.8040.9220.841(0.892-0.981)(0.681–0.909)(0.868–0.965)(0.731–0.945)HER2-enriched0.7920.9630.7920.963(0.613–0.950)(0.930–0.993)(0.625–0.955)(0.927–0.992)Triple-negative0.8180.9850.9000.971(0.631–0.960)(0.962–1.000)(0.750–1.000)(0.938–0.993)Macro-average0.8490.9180.8710.925(0.763–0.923)(0.874–0.955)(0.799–0.937)(0.874–0.955)Prospective cohortLuminal0.8620.7580.9260.610(0.797–0.924)(0.607–0.898)(0.874–0.972)(0.452–0.757)HER2-enriched0.6670.8810.3850.959(0.400–0.909)(0.824–0.933)(0.191–0.579)(0.921–0.992)Triple-negative0.6670.9770.8000.955(0.450–0.870)(0.948–1.000)(0.583–1.000)(0.918–0.985)Macro-average0.7320.8720.7030.841(0.620–0.839)(0.819–0.922)(0.608–0.797)(0.799–0.892)Multicenter external cohortLuminal0.8440.9230.9760.615(0.766–0.911)(0.810–1.000)(0.940–1.000)(0.455–0.763)HER2-enriched1.0000.8810.5001.000(1.000–1.000)(0.813–0.936)(0.300–0.692)(1.000–1.000)Triple-negative0.8460.9820.8460.982(0.615–1.000)(0.954–1.000)(0.625–1.000)(0.954–1.000)Macro-average0.8970.9280.7740.866(0.814–0.958)(0.884–0.964)(0.673–0.863)(0.824–0.904)Metrics were derived by assigning each case to the class with the highest predicted probability (i.e., argmax rule) and are presented with 95% confidence intervals (CI).PPV positive predictive value, NPV negative predictive value. Performance of MUST-Sub in subtype classification across the validation cohorts Metrics were derived by assigning each case to the class with the highest predicted probability (i.e., argmax rule) and are presented with 95% confidence intervals (CI). PPV positive predictive value, NPV negative predictive value. As discrimination metrics such as AUC do not indicate whether predicted probabilities are well aligned with observed event frequencies, we further evaluated the calibration of MUST-Sub across all three cohorts using subtype-wise calibration curves and Brier scores (Supplementary Fig. 5). Brier scores were low for HER2-enriched (0.049, 0.063, 0.054) and triple-negative subtypes (0.057, 0.066, 0.075), indicating reasonably good calibration, whereas Luminal subtypes showed higher scores (0.105, 0.164, 0.142), likely reflecting greater heterogeneity within this broader category across cohorts, as it comprises both Luminal A- and Luminal B-like tumors. Calibration curves showed a consistent directional asymmetry, with underestimation of Luminal probabilities and overestimation of HER2-enriched and triple-negative probabilities. Per-subtype one-vs-rest analyses showed that MUST-Sub achieved sensitivity > 0.80 for HER2-enriched and triple-negative subtypes at a low operating threshold (pt = 0.2; Supplementary Fig. 7), at the cost of lower specificity. Multi-class threshold tuning18 further suggested that additional threshold optimization offered only limited benefit: the optimal threshold vector τ⋆ = (0.28, 0.44, 0.28) was close to the default setting (1/3, 1/3, 1/3), with only marginal performance changes across cohorts (internal: 0.860 → 0.867; prospective: 0.703 → 0.716; external: 0.806 → 0.794). Decision curve analysis showed positive net benefit relative to treat-all and treat-none strategies over broad threshold ranges across the three cohorts (Supplementary Fig. 6), supporting an exploratory assessment of potential clinical utility. MUST-Sub outperforms B-mode–only state-of-the-art models To quantify the added value of CEUS, we compared MUST-Sub with representative state-of-the-art deep learning models trained using B-mode ultrasound alone: ResNet-1819, a widely used convolutional neural network; MedViT20, a vision transformer adapted for medical imaging; and MedMamba21, a state-space model. Across all cohorts, MUST-Sub consistently outperformed these B-mode–only baselines (Fig. 2), achieving macro-average AUCs of 0.94 (internal), 0.90 (prospective), and 0.92 (external), compared with 0.82/0.81/0.80 for ResNet-18, 0.79/0.78/0.80 for MedViT, and 0.83/0.81/0.82 for MedMamba. Pairwise DeLong tests22 showed that these macro-average AUC differences were significant in all nine uncorrected comparisons, six of which remained significant after Bonferroni correction23 (Supplementary Table 4). Improvements were observed across all three subtypes. For the clinically relevant binary task of distinguishing Luminal from non-Luminal tumors, MUST-Sub achieved AUCs of 0.92, 0.88, and 0.91 on the internal, prospective, and external cohorts, respectively, exceeding all B-mode–only models in every cohort.Fig. 2Receiver operating characteristic (ROC) curves comparing MUST-Sub with B-mode–only baseline models across validation cohorts.a Internal (n=159), b prospective (n=149), and c multicenter external (n=122). Curves are shown for one-vs-rest classification of each subtype and for the macro-average. CI confidence interval. Receiver operating characteristic (ROC) curves comparing MUST-Sub with B-mode–only baseline models across validation cohorts. a Internal (n=159), b prospective (n=149), and c multicenter external (n=122). Curves are shown for one-vs-rest classification of each subtype and for the macro-average. CI confidence interval. Model-derived spatiotemporal biomarkers are associated with pathological features Beyond subtype prediction, MUST-Sub provides interpretability by producing quantitative biomarkers derived from its spatiotemporal attention weights (Fig. 3, Supplementary Figs. 2–4).Fig. 3Spatiotemporal biomarkers derived from MUST-Sub and their associations with pathological variables.a Examples of spatial attention maps from the internal cohort. The B-mode image is overlaid with a segmentation contour generated by the Segment Anything Model (SAM)55 prompted by the attention heatmap. Arrows indicate salient morphological cues. b Subtype-specific perfusion kinetics on CEUS in the internal cohort. Curves show mean time-intensity profiles with standard deviation. Asterisks denote time points with the highest temporal attention weights. c, d Correlations of the morphological biomarker \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}$$\end{document}Bm and the hemodynamic biomarker \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{h}$$\end{document}Bh with tumor size and the Ki-67 index, respectively, across validation cohorts, assessed using Spearman’s rank correlation coefficient. Shaded areas indicate two-dimensional kernel density estimation (KDE), with darker regions representing higher sample density. Spatiotemporal biomarkers derived from MUST-Sub and their associations with pathological variables. a Examples of spatial attention maps from the internal cohort. The B-mode image is overlaid with a segmentation contour generated by the Segment Anything Model (SAM)55 prompted by the attention heatmap. Arrows indicate salient morphological cues. b Subtype-specific perfusion kinetics on CEUS in the internal cohort. Curves show mean time-intensity profiles with standard deviation. Asterisks denote time points with the highest temporal attention weights. c, d Correlations of the morphological biomarker \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}$$\end{document}Bm and the hemodynamic biomarker \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{h}$$\end{document}Bh with tumor size and the Ki-67 index, respectively, across validation cohorts, assessed using Spearman’s rank correlation coefficient. Shaded areas indicate two-dimensional kernel density estimation (KDE), with darker regions representing higher sample density. As shown in Fig. 3a and Supplementary Fig. 2, the spatial attention module highlights diagnostically relevant morphological cues on B-mode US, including spiculated margins, a feature associated with malignancy and commonly observed in invasive ductal carcinoma24, and angular contours suggestive of tumor infiltration25. Across cohorts (Fig. 3c, Supplementary Fig. 3), the morphology-associated biomarker \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}$$\end{document}Bm was inversely correlated with tumor size (internal: ρ = − 0.259, p = 0.001; prospective: ρ = − 0.336, p < 0.001; external: ρ = − 0.255, p = 0.014), suggesting that the model places greater morphology-related attention on boundary characteristics in smaller lesions. Consistent with a predominantly morphology-related interpretation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}$$\end{document}Bm showed no significant association with the Ki-67 proliferation index in any cohort (all p > 0.05). As shown in Fig. 3b and Supplementary Fig. 2, the temporal attention module captures subtype-associated perfusion dynamics on CEUS. Triple-negative tumors exhibited rapid wash-in and wash-out kinetics, consistent with aggressive vascular behavior26, whereas Luminal tumors showed more gradual and sustained enhancement, and HER2-enriched tumors displayed intermediate kinetics. The hemodynamics-associated biomarker \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{h}$$\end{document}Bh was positively correlated with tumor size (internal: ρ = 0.245, p = 0.002; prospective: ρ = 0.323, p < 0.001; external: ρ = 0.237, p = 0.009) and Ki-67 proliferative index (internal: ρ = 0.209, p = 0.008; prospective: ρ = 0.212, p = 0.012; external: ρ = 0.239, p = 0.009; Fig. 3d, Supplementary Fig. 4). These results suggest that greater temporal dispersion in CEUS attention, as captured by MUST-Sub, is associated with larger tumors and higher proliferative activity. Qualitative error analysis To gain deeper insight into the model’s decision-making process and failure modes, we performed a qualitative analysis of representative cases from the prospective and external validation cohorts. Representative cases illustrated how the incorporation of CEUS provides complementary functional information that corrects errors made by B-mode–only models (Fig. 4a–c). For instance, in Fig. 4a, a pathologically confirmed HER2-enriched tumor was misclassified as Luminal by B-mode–only baselines. On B-mode US, the lesion appeared as an irregular mass with microlobulated margins, a morphology that can resemble Luminal tumors. In contrast, MUST-Sub focused on the rapid wash-in pattern on CEUS, consistent with HER2-enriched disease, resulting in the correct prediction.Fig. 4Representative case studies from the prospective and external cohorts.a HER2-enriched (prospective): B-mode–only baselines misclassified as Luminal, whereas MUST-Sub correctly predicted HER2-enriched. b Luminal (external): B-mode–only baselines misclassified as triple-negative, whereas MUST-Sub was correct. c Triple-negative (external): B-mode–only baselines misclassified as Luminal, whereas MUST-Sub was correct. d HER2-enriched (prospective): all models misclassified as Luminal. e Triple-negative (prospective): all models misclassified as Luminal. f Luminal (external): all models misclassified as triple-negative. Each panel shows the B-mode image, CEUS sequence, and the corresponding spatial and temporal attention maps. Representative case studies from the prospective and external cohorts. a HER2-enriched (prospective): B-mode–only baselines misclassified as Luminal, whereas MUST-Sub correctly predicted HER2-enriched. b Luminal (external): B-mode–only baselines misclassified as triple-negative, whereas MUST-Sub was correct. c Triple-negative (external): B-mode–only baselines misclassified as Luminal, whereas MUST-Sub was correct. d HER2-enriched (prospective): all models misclassified as Luminal. e Triple-negative (prospective): all models misclassified as Luminal. f Luminal (external): all models misclassified as triple-negative. Each panel shows the B-mode image, CEUS sequence, and the corresponding spatial and temporal attention maps. We also present cases that were misclassified by both B-mode–only models and MUST-Sub (Fig. 4d–f). In Fig. 4d, MUST-Sub misclassified a HER2-enriched tumor as Luminal. The lesion showed a well-defined margin on B-mode US and relatively smooth enhancement kinetics on CEUS, jointly resembling a Luminal-like imaging phenotype. These examples highlight the persistent challenge of distinguishing subtypes when tumors exhibit overlapping or hybrid imaging signatures. Discussion This multicenter study developed and validated MUST-Sub, a deep learning model that advances ultrasound-based breast cancer subtyping through three main contributions: (1) multimodal integration of B-mode US and CEUS using spatiotemporal transformers, (2) improved label efficiency and generalization via self-supervised pretraining on unlabeled paired data; and (3) enhanced interpretability through attention-derived spatiotemporal biomarkers that provide quantitative summaries of morphology- and hemodynamics-related model focus. AI-based approaches for breast cancer subtyping have been explored across multiple data modalities. Histopathology-based methods25,27–29 are constrained by invasiveness and sampling bias arising from intratumoral heterogeneity. MRI-based approaches30–36 provide improved soft-tissue contrast but are limited by cost and accessibility. Ultrasound offers a practical imaging modality, and recent deep learning models have shown promise for subtype prediction15. Multimodal strategies combining ultrasound with mammography8 and histopathology9 have further improved performance over single-modality methods. However, most existing ultrasound-based approaches primarily rely on static anatomical information and do not explicitly model dynamic functional signals, such as tumor vascular perfusion, that may provide complementary biological information relevant to subtype characterization. CEUS is well-suited to fill this gap because it is clinically accessible and captures tumor perfusion dynamics in real time. CEUS has been successfully applied to a range of diagnostic tasks, including characterization of peripheral pulmonary lesions37, hepatocellular carcinoma38, and liver lesions39. Nevertheless, its integration with AI for breast cancer subtyping has remained largely unexplored. By combining B-mode US with CEUS, MUST-Sub achieved robust performance across internal, prospective, and multicenter external validation cohorts. The consistency of performance across prospective and multicenter external validation provides evidence for the model’s robustness under temporal and cross-site variation. For the clinically relevant task of distinguishing Luminal from non-Luminal disease, MUST-Sub achieved AUCs of 0.92, 0.88, and 0.91, respectively. Notably, these results were comparable to approaches that incorporate hematoxylin–eosin-stained whole-slide images9 (AUCs = 0.93 and 0.90), supporting the premise that CEUS-derived functional patterns provide clinically meaningful information for subtype characterization. In addition, our study assembled 1039 unlabeled and 1054 labeled paired B-mode/CEUS cases across multiple centers. To our knowledge, this constitutes the largest paired multimodal ultrasound dataset reported to date for breast cancer subtyping, providing a strong basis for both model development and cross-site validation. A key challenge in medical AI is the limited availability of subtype labels relative to the volume of routine imaging. Self-supervised learning leverages unlabeled data to learn transferable representations, improving robustness and generalization40,41. Building on this paradigm, MUST-Sub employs a two-stage training strategy that exploits the abundance of clinically acquired paired ultrasound examinations lacking subtype annotations. Pretraining on unlabeled data substantially improves cross-center generalization, increasing macro-average AUC from 0.87 to 0.92 on the external validation cohort (Supplementary Table 1). This result highlights a pathway for deploying ultrasound AI at scale by harnessing the large reservoir of unlabeled clinical data. Beyond predictive performance, MUST-Sub improves interpretability by deriving quantitative spatiotemporal biomarkers associated with pathological features. The morphology-associated biomarker \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}$$\end{document}Bm, derived from spatial attention, emphasizes discriminative anatomical patterns such as spiculated margins and angular contours and showed inverse correlations with tumor size across cohorts (Spearman ρ = [− 0.34, − 0.23]; all p < 0.05). This finding suggests that the model may place greater morphology-related attention on boundary characteristics in smaller lesions. In contrast, the hemodynamics-associated biomarker \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{h}$$\end{document}Bh, derived from temporal attention on CEUS, was positively correlated with both tumor size (ρ = [0.24, 0.32]; all p < 0.05) and Ki-67 proliferation index (ρ = [0.21, 0.24]; all p < 0.05), consistent with the association between aggressive tumor biology and abnormal neovascular perfusion patterns. Together, these biomarkers provide cohort-consistent quantitative summaries of morphology- and hemodynamics-related model attention and suggest associations between imaging phenotypes and pathological indicators beyond categorical subtype labels. We emphasize that these findings should be interpreted as exploratory: the modest effect sizes (ρ ≈ 0.2–0.3) reflect the inherently indirect relationship between macroscopic imaging features and molecular or cellular indices, and further validation against long-term clinical outcomes such as recurrence-free survival is needed to establish their clinical utility. We further included a secondary benchmark against recent large vision-language models (VLMs), which have shown strong general-purpose visual understanding and encouraging performance in several medical imaging applications42,43. This analysis was intended to assess whether such general-purpose foundation models could serve as a competitive alternative for molecular subtyping without task-specific training. As detailed in Supplementary Note 4, we prompted these models to perform three-class subtype classification using either B-mode US only or paired B-mode US and CEUS sequences. However, their performance on the multicenter external cohort was substantially lower than that of dedicated discriminative models (Supplementary Table 1). Notably, incorporating CEUS did not improve VLM performance and, in some cases, led to further degradation, plausibly because contemporary VLMs have limited exposure to CEUS data and lack explicit mechanisms for modeling contrast-enhancement kinetics. Qualitatively, the VLMs frequently defaulted to predicting the triple-negative subtype for a large fraction of cases, suggesting a systematic bias toward mapping “high-risk” visual patterns (e.g., irregular margins or heterogeneous echotexture) to the most aggressive category. This behavior likely reflects several limitations of current VLMs for fine-grained biomedical phenotyping: (1) subtype-related signals in ultrasound are subtle and indirect, requiring sensitivity to weak, distributed cues rather than coarse semantic descriptors; (2) CEUS interpretation relies on temporally calibrated perfusion dynamics that are not naturally represented in VLMs trained predominantly on still images and text; and (3) prompt-based inference is typically uncalibrated and may amplify prior biases when class boundaries are overlapping and clinically nuanced. These results suggest that off-the-shelf VLMs, used without task-specific adaptation, are not yet sufficient for ultrasound-based molecular subtyping, underscoring the need for specialized architectures and training strategies tailored to spatiotemporal perfusion modeling. Several limitations should be acknowledged. First, classification of HER2-enriched tumors remained comparatively challenging, potentially reflecting both recognized histological heterogeneity44 and the lower prevalence of this subtype in our cohorts. Second, although external validation included patients from four additional institutions across three cities, the external cohort size remained modest, and all cohorts were recruited from Chinese institutions; breast cancer molecular subtype distribution and tumor biology may vary across ethnic populations1,45, so international external validation is needed before broader generalization can be claimed. Third, MUST-Sub did not incorporate clinical variables (e.g., age, menopausal status, or treatment history), which may further improve performance and calibration. Fourth, the attention-derived biomarkers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}$$\end{document}Bm and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{h}$$\end{document}Bh show consistent but modest correlations with pathological indices and should be regarded as proof-of-concept rather than validated clinical biomarkers at this stage. Fifth, the relationship between model-derived spatiotemporal biomarkers and standard BI-RADS descriptors remains to be systematically characterized; a more granular mapping would require fine-grained per-descriptor annotations that are beyond the scope of the current study. Sixth, although the training cohort included examinations from two experienced radiologists and external validation spanned four institutions with different scanner models, residual operator-dependent variability in CEUS acquisition cannot be fully eliminated. Future work will prioritize international multicenter prospective validation with standardized acquisition protocols and more diverse imaging equipment, as well as prospective comparison with experienced breast radiologists to assess clinical added value. Longer-term directions include multimodal fusion with clinical metadata, fine-grained mapping of model-derived biomarkers to BI-RADS descriptors, and longitudinal modeling to support treatment-response monitoring. In conclusion, MUST-Sub advances ultrasound from qualitative lesion assessment toward quantitative molecular phenotyping by combining multimodal ultrasound with spatiotemporal modeling. Its robust performance across multiple validation cohorts, together with interpretable attention-derived biomarkers, supports its potential as an adjunctive, non-invasive pre-biopsy tool for preliminary molecular phenotyping and triage, particularly in settings where access to biopsy or IHC may be delayed. Histopathological confirmation remains necessary before any treatment decision. Methods Data collection This multicenter study was approved by the Institutional Review Board of Shanghai General Hospital (Approval No. [2023] Institutional Ethics Quick Review 383). Written informed consent was obtained from all participants before CEUS. Consecutive patients were recruited from five hospitals between October 2018 and November 2024. All participants underwent multimodal ultrasound, including B-mode US and CEUS. Patients who did not subsequently undergo surgery were assigned to the unlabeled cohort, which comprised 1039 cases from Hospital A (i.e., Shanghai General Hospital). For patients who subsequently underwent surgery, postoperative histopathology served as the reference standard for molecular subtyping, constituting the labeled cohort. Inclusion criteria for the labeled cohort were: (i) female patients aged > 18 years; (ii) histopathologically confirmed primary breast cancer; (iii) paired B-mode US and CEUS performed prior to any biopsy or treatment; and (iv) complete IHC and/or FISH results for ER, PR, HER2, and Ki-67. Exclusion criteria were: (i) pregnancy or lactation; (ii) prior biopsy, neoadjuvant therapy, or surgery directed at the index lesion; (iii) non-primary breast malignancy; and (iv) inadequate image quality. The patient enrollment process is illustrated in Fig. 5.Fig. 5Patient enrollment flowchart across the five participating hospitals.The labeled cohort was split into a retrospective cohort (Hospital A, October 2018–December 2023), a prospective validation cohort (Hospital A, January–November 2024), and a multicenter external validation cohort (Hospitals B–E). Patient enrollment flowchart across the five participating hospitals. The labeled cohort was split into a retrospective cohort (Hospital A, October 2018–December 2023), a prospective validation cohort (Hospital A, January–November 2024), and a multicenter external validation cohort (Hospitals B–E). After applying the criteria, 783 labeled patients from Hospital A enrolled before 2024 were assigned to the retrospective cohort and randomly split into 547 training cases (70%), 77 internal validation cases (10%), and 159 internal test cases (20%). Additionally, 149 labeled patients from Hospital A enrolled in 2024 and 122 labeled patients from Hospital B–E (i.e., Shanghai Chest Hospital, Shidong Hospital, Qingdao Municipal Hospital, and Shanxi Province Cancer Hospital) served as the prospective and multicenter external validation cohorts, respectively (Fig. 1b). Internal cohorts used Canon Aplio 500 or i900 systems exclusively, whereas the external cohort included additional systems: Mindray Resona R9 at Shanghai Chest Hospital and Shidong Hospital, and GE Logiq E9 at Qingdao Municipal Hospital and Shanxi Province Cancer Hospital. These differences introduced inter-institutional heterogeneity in equipment and acquisition conditions within the external validation cohort. Image acquisition protocol Ultrasound examinations were performed by board-certified breast radiologists. At Hospital A, examinations were performed by two board-certified breast radiologists each with over 10 years of experience, thereby incorporating natural inter-operator variability into the training data. All scanners were equipped with linear-array transducers (5–14 MHz; 14-5L). All examinations followed a standardized protocol. Patients were positioned supine with the ipsilateral arm elevated. Bilateral breasts and axillae were systematically scanned in radial and anti-radial planes. The index lesion was assessed from multiple angles, and the plane with the maximum lesion diameter was selected for B-mode imaging. CEUS was subsequently performed on the plane showing the most prominent vascularity. SonoVue® (Bracco, Italy; reconstituted with 5 mL of 0.9% saline) was administered as an intravenous bolus via a cubital vein, followed by a 5 mL saline flush. Cine recording started immediately after injection and continued for 60 s in dual-screen mode. For lesions larger than the probe field of view (~ 4–5 cm), a pre-scan using B-mode and Color Doppler was performed to identify the most actively perfused cross-section, typically the thickest, most solid, and hypervascular area while avoiding regions of central necrosis. The probe was then kept stationary over this representative area during CEUS recording. For non-mass lesions, ROI delineation was based on the spatial extent of abnormal contrast enhancement during the peak arterial phase rather than B-mode margins. Histopathological assessment All patients in the labeled cohort underwent surgical excision. IHC was performed for ER, PR, HER2, and Ki-67. ER and PR positivity were defined as ≥1% nuclear staining in tumor cells. HER2 status was assessed by IHC, with FISH performed for equivocal cases: HER2(-) was considered negative and HER2(3+) positive; HER2(+) or HER2(2+) prompted FISH testing, with gene amplification classified as positive and non-amplification as negative. The Ki-67 index was dichotomized at 14%. Surrogate molecular subtypes were assigned according to the 2013 St. Gallen International Expert Consensus: Luminal (ER- and/or PR-positive, any HER2 status), HER2-enriched (HER2-positive with ER- and PR-negative), and triple-negative (ER-, PR-, and HER2-negative). MUST-Sub MUST-Sub is a multimodal ultrasound spatiotemporal transformer model for breast cancer subtyping. It integrates B-mode and CEUS information through a two-stage training strategy: self-supervised pretraining to learn transferable multimodal representations, followed by supervised fine-tuning for subtype classification (Fig. 1c). The detailed network architectures and training protocol are provided in Supplementary Notes 1 and 2, respectively. During pretraining, both unlabeled data and labeled development data are used to learn general ultrasound representations spanning morphology in B-mode US and perfusion dynamics in CEUS. Spatial and temporal dependencies are modeled using transformer modules within a VAE latent space. The VAE compresses B-mode images from 3 × 256 × 256 to 4 × 32 × 32 and CEUS video sequences from 65 × 3 × 256 × 256 to 17 × 4 × 32 × 32 (compression factor 183.5 ×), making transformer-based modeling computationally tractable on the compressed latent representation. The latent space provides a shared representational space for B-mode morphological features and CEUS perfusion patterns, facilitating multimodal fusion. The reconstruction objective may also suppress part of the ultrasound-specific noise (e.g., speckle artifacts and operator-dependent variations), although this denoising effect is a beneficial byproduct rather than the primary design goal. The temporal transformer follows an autoregressive formulation46 to predict the next frame conditioned on preceding frames, whereas the spatial transformer adopts a diffusion-based strategy47 to progressively reconstruct an image from a masked input. The pretrained model can generate plausible B-mode images and CEUS sequences (Supplementary Note 3). Ablation results supporting the contribution of each component are provided in Supplementary Table 2. Subsequently, fine-tuning is performed using labeled development data to learn subtype-discriminative cues. Cross-attention in the spatiotemporal transformers is used to fuse CEUS temporal dynamics with B-mode US spatial contexts, and the fused representation is mapped to subtype probabilities using a multilayer perceptron classifier. To mitigate overfitting, we applied dropout (rate=0.1) throughout the transformer layers and MLP classifier, Adam optimizer with weight decay=0.01, and early stopping monitored on the internal validation set. Data augmentation included spatial perturbations (random horizontal and vertical flips, random rotation ± 90∘, scaling 0.8–1.2 ×, translation ± 10%, elastic deformation), temporal perturbations (random shift of the perfusion window by ± 5 seconds), and intensity perturbations (brightness/contrast adjustment ± 10%). Self-supervised pretraining on 1039 unlabeled samples further provides a regularized initialization, reducing reliance on labeled data. The subtype distribution in our cohorts is broadly consistent with the clinical predominance of Luminal breast cancers48. To reduce bias toward the majority class, we applied mini-batch resampling to maintain balanced subtype representation within each batch and replaced standard cross-entropy with focal loss49, which down-weights well-classified examples and directs the model’s attention toward harder minority-class cases. The data augmentation and self-supervised pretraining described above also help by increasing the effective diversity of minority-class samples and providing a class-agnostic initialization. The attention weights from the fine-tuned model provide intrinsic interpretability by highlighting spatiotemporal patterns supporting the prediction. To quantify the morphological patterns captured in B-mode US and the hemodynamic patterns captured in CEUS, we derive two attention-based biomarkers. The morphology-associated biomarker\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}$$\end{document}Bm measures intensity-weighted spatial attention over the B-mode image IB:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}=\frac{1}{H\times W}\mathop{\sum }\limits_{i=1}^{W}\mathop{\sum }\limits_{j=1}^{H}{{\boldsymbol{A}}}_{s}^{ij}\cdot {{\boldsymbol{I}}}_{\,{\rm{B}}}^{ij},$$\end{document}Bm=1H×W∑i=1W∑j=1HAsij⋅IBij,where As denotes the spatial attention map and H and W are image height and width. The hemodynamics-associated biomarker\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{h}$$\end{document}Bh quantifies temporal dispersion of attention weights across the CEUS sequence:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{h}=\sqrt{\frac{1}{T-1}{\sum }_{t=1}^{T}{({w}_{t}-\bar{w})}^{2}},$$\end{document}Bh=1T−1∑t=1T(wt−w¯)2,where T is the sequence length, wt is the temporal attention weight at time t, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{w}$$\end{document}w¯ is the mean attention weight over the sequence. Statistical analysis Statistical analysis was performed in Python using SciPy (version 1.15.3)50. Categorical variables were compared between groups using the Pearson χ2 test51, or Fisher’s exact test when Cochran’s rule was not satisfied52. Continuous biomarkers (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{h}$$\end{document}Bh and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathcal{B}}}_{m}$$\end{document}Bm) were compared across clinicopathological subgroups using the Mann–Whitney U test for two-group comparisons and the Kruskal–Wallis test for comparisons involving more than two groups (Supplementary Figs. 3 and 4). Associations between the biomarkers and pathological variables (tumor size and Ki-67 index) were assessed using Spearman’s rank correlation coefficient53 (Fig. 3c, d). The performance metrics of the prediction model are reported with 95% confidence intervals derived from nonparametric bootstrap resampling (1000 iterations) at the patient level54. Pairwise comparisons of macro-average AUC between MUST-Sub and baseline models were performed using the DeLong test22, with Bonferroni correction applied to the nine primary comparisons. All statistical tests were two-sided, and a p-value of less than 0.05 was considered statistically significant. Supplementary information Supplementary Information Supplementary Information