In recent years, breast cancer (BC) has maintained its position as a leading cause of global cancer incidence and mortality (1). To enable precise prognosis and treatment stratification, researchers have classified BC into four molecular subtypes based on gene expression profiles: Luminal A, Luminal B, HER2-enriched, and triple-negative breast cancer (TNBC) (2). This classification system originates from immunohistochemical evaluation of three biomarkers: estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor-2 (HER2) expression status. TNBC, defined by the absence of ER, PR, and HER2 expression (3), accounts for 15%-20% of all breast cancer cases. Compared with other subtypes, TNBC demonstrates shorter survival durations, higher mortality rates, and enhanced metastatic potential (4). Current international guidelines recommend BRCA1/2 genetic testing for high-risk TNBC patients (characterized by large tumor size, lymph node metastasis, high histological grade, elevated Ki-67 index, or young age). For BRCA1/2 mutation carriers, adjuvant therapy with the PARP inhibitor olaparib improves disease-free survival, while PD-L1-positive patients may benefit from combined immunotherapy using pembrolizumab (5). Postoperative maintenance therapy with oral capecitabine is recommended for intermediate-to-low risk TNBC patients (presenting with smaller tumor size, node-negative status, or low-grade histology) to mitigate recurrence risk (6). Given the substantial costs associated with BRCA testing and PD-L1 immunohistochemical analysis, there exists an urgent clinical need for novel cost-effective methodologies to achieve precise risk stratification and prognostic prediction in TNBC management.

Risk stratification-guided precision therapy enables optimized prognostic outcomes for TNBC patients while maintaining equilibrium between treatment intensity and quality of life. Current breast cancer diagnostic modalities primarily encompass mammography, ultrasonography, and magnetic resonance imaging (MRI) (7, 8). However, the high prevalence of dense breast tissue in Chinese female patients elevates false-positive rates in mammographic screening (9). Although MRI demonstrates superior diagnostic accuracy, its clinical application is constrained by substantial costs and time-intensive protocols. Ultrasonography has emerged as the primary screening modality for breast disorders due to its non-invasive nature, cost-effectiveness, rapid implementation, and demonstrated superiority over mammography and MRI in detection rates, diagnostic accuracy, and cost-benefit ratios (10, 11). Conventional screening approaches exhibit two critical limitations: (1) subjective image interpretation may yield false-positive results (inducing patient anxiety and overtreatment) or false-negative findings (leading to occult lesion misdiagnosis) (12); (2) screening-undetected tumors often demonstrate more aggressive biological behavior (13). Artificial intelligence (AI)-driven predictive models, developed through image segmentation, feature extraction, and quantitative analysis, enhance detection accuracy by identifying suspicious lesions in breast imaging (14, 15). Therefore, developing prognostic prediction models integrating AI with ultrasound radiomics data holds significant theoretical and clinical implications, potentially heralding groundbreaking advancements in TNBC therapeutic management.

Artificial intelligence (AI), a cognitive function-mimicking intelligent system, demonstrates autonomous learning and complex decision-making capabilities through deep learning algorithms and neural network architectures. Recent years have witnessed rapid advancement of this disruptive technology in medical applications, particularly achieving transformative progress in breast cancer radiomics research, driving paradigm shifts toward precision diagnostics and therapeutics. Intelligent imaging assistance systems automate lesion delineation and tumor burden quantification, reducing diagnostic processing time by 40%.Studies have demonstrated AI’s exceptional performance in reducing false-negative rates: A research team developed an AI model using 108,970 mammograms achieving an area under the curve (AUC) of 0.83, detecting interval cancers missed by radiologists while reducing physician workload by 30% (16). Cruz-Ramos et al. developed a novel computer-aided diagnosis (CAD) system integrating ultrasound and mammography, achieving 97.6% accuracy and 98% precision in differentiating breast mass malignancy, substantially optimizing diagnostic workflows (17). Prognostic prediction models incorporating therapeutic response parameters enable dynamic follow-up strategy adjustments, while personalized treatment recommendation systems powered by real-world data optimize regimen selection for patients with specific genetic profiles. A multicenter study of 387 TNBC patients revealed that multiparametric MRI radiomics models demonstrated superior neoadjuvant chemotherapy response prediction (AUCs=0.97, 0.90, 0.86) compared to clinical models, with enhanced accuracy when integrated with clinicopathological risk factors (18). Zhu et al. developed a longitudinal radiomics model incorporating multi-temporal and multi-regional MRI analysis to precisely predict axillary lymph node response to neoadjuvant chemotherapy (NAC), potentially reducing unnecessary lymph node dissection and improving quality of life (19). Li’s research team proposed a biologically guided deep learning approach capable of simultaneous prediction of tumor immune/stromal microenvironment status and therapeutic outcomes from medical imaging. Importantly, this model identified mismatch repair-deficient tumor subsets non-responsive to immunotherapy, potentially informing patient selection for combination therapies (20). Furthermore, they established a non-invasive methodology combining radiomics and deep learning analysis to predict tumor microenvironment (TME) status from radiographic images. A multi-institutional cohort study of 2,686 gastric cancer patients validated this TME assessment approach, enabling longitudinal monitoring of therapeutic responses (21).

In summary, AI systems leveraging big data analytics and deep learning technologies demonstrate clinical utility in tumor screening, diagnosis, treatment planning, and prognostic prediction, thereby informing personalized therapeutic decision-making. Future advancements in AI technology, supported by multicenter studies with large-scale cohorts, will enable multimodal fusion algorithms to integrate imaging, pathology, omics, and electronic health record data for comprehensive multidimensional analysis. This integration promises to establish a closed-loop management system spanning precision diagnosis to intelligent follow-up throughout breast cancer care pathways (22). While ultrasound radiomics has demonstrated potential in breast cancer research, existing studies in this field have predominantly focused on the general breast cancer population, with relatively scarce research targeting the TNBC subtype characterized by poor prognosis and aggressive behavior. Furthermore, prior investigations have primarily concentrated on diagnostic tasks (23–25)—such as distinguishing benign from malignant lesions or conducting molecular subtyping—while prognostic prediction of recurrence and metastasis remains limited and often lacks longitudinal follow-up data. Most critically, few studies have closely integrated imaging biomarkers with long-term clinical outcomes following standardized treatment regimens. Therefore, this study aims to develop a predictive model by analyzing preoperative breast ultrasound imaging features from three distinct TNBC cohorts: patients maintaining five-year recurrence-free survival after surgery, those developing locoregional recurrence, and those with distant metastasis. The model is designed to evaluate prognosis in TNBC cases receiving standardized treatment, enabling precise risk stratification to inform personalized surveillance protocols and therapeutic decision-making. This approach ultimately seeks to optimize clinical outcomes by enhancing survival rates and improving the quality of life for TNBC patients.

TNBC remains a formidable challenge in oncological research due to its aggressive biological behavior, early metastatic propensity, and limited therapeutic options. Epidemiological studies indicate TNBC constitutes 15%-20% of all breast carcinomas, predominantly affecting premenopausal women under 40 years of age. Compared with other molecular subtypes, TNBC demonstrates significantly shorter median survival (13.3 months post-metastasis) and higher mortality rates, with 46% of patients developing distant metastases (27). The characteristic metastatic pattern predominantly involves visceral organs and central nervous system, with peak distant recurrence occurring during the third post-diagnostic year (28). Notably, TNBC exhibits accelerated disease progression with median recurrence-free interval of 19–40 months versus 35–67 months in non-TNBC cases, and carries 75% mortality risk within 3 months of recurrence (29). The absence of estrogen/progesterone receptors and HER2 expression renders TNBC refractory to endocrine/targeted therapies, establishing chemotherapy as the therapeutic cornerstone (30). NAC has emerged as standard care for locally advanced TNBC, achieving triple objectives: tumor downstaging, surgical optimization, and in vivo chemosensitivity assessment. NAC enables pathological complete response (pCR) rates of 30%-60% in TNBC, significantly improving breast conservation feasibility (from 15% to 45%) and reducing locoregional recurrence risk by 38% (31). Patients attaining pCR demonstrate superior 5-year disease-free survival (DFS: 85% vs 55%) and overall survival (OS: 90% vs 60%) compared with non-pCR counterparts (32). Our findings demonstrate NAC-treated patients had 42% lower 5-year recurrence risk than non-NAC counterparts (HR = 0.58, 95% CI: 0.43-0.78), underscoring the imperative for multicenter prospective validation to optimize NAC response prediction models (Table 1).

Prognostic evaluation and therapeutic response prediction constitute critical components of precision oncology. While conventional approaches integrate clinical parameters with TNM staging system (categorizing tumor size [T], nodal involvement [N], and distant metastasis [M]), persistent inaccuracies in prognostic stratification continue to challenge clinical decision-making. This investigation innovatively combines ultrasound radiomics with artificial intelligence, employing an optimized U-Net architecture to achieve automated tumor segmentation (Dice coefficient=0.89 ± 0.05) that minimizes interobserver variability. From 1,034 extracted quantitative features encompassing first-order statistics, morphometric parameters, and textural patterns, LASSO regression identified 13 recurrence-associated biomarkers and 7 metastasis-specific signatures. The integration of LASSO feature selection with support vector machine (SVM) classification effectively addressed high-dimensional data challenges, achieving dimensional reduction while preserving prognostic information. These radiomic biomarkers capture intrinsic tumor heterogeneity and marginal infiltration patterns correlating with TNBC’s aggressive phenotype. For instance, elevated gray-level co-occurrence matrix (GLCM) correlation values (range: 0.68-0.92) may indicate compact cellular architecture and enhanced invasive potential. The developed predictive model demonstrated robust risk stratification capability (C-index=0.83), enabling data-driven personalization of surveillance protocols and therapeutic regimens. Per NCCN guidelines, high-risk cohorts (characterized by tumor size >5cm, nodal involvement, high histological grade, Ki-67≥30%, or age <40) are recommended dose-dense anthracycline-taxane regimens (e.g., AC-T: doxorubicin/cyclophosphamide followed by paclitaxel) with platinum adjuncts to enhance pCR rates (ΔpCR=18.7%). Molecular profiling including PD-L1 expression and BRCA1/2 mutation analysis guides targeted therapy: pembrolizumab augmentation in PD-L1+ cases improves 3-year DFS by 32.6% (HR = 0.54), while olaparib maintenance in BRCA mutation carriers reduces recurrence risk by 42% (HR = 0.58).For intermediate-low risk patients (tumor ≤2cm, node-negative, low-grade), neoadjuvant chemotherapy serves primarily as in vivo chemosensitivity testing. Postoperative regimens may de-escalate to TC (docetaxel/cyclophosphamide ×4 cycles) or AC ×4 cycles, while ultra-low risk cases (T1aN0, Ki-67<15%) could potentially omit adjuvant chemotherapy with 5-year DFS exceeding 92% (33). Current barriers to BRCA testing adoption (cost: $3,000-5,000 per assay) highlight the cost-effectiveness advantage of our AI-powered stratification system. This radiomics model enables real-time therapeutic adaptation through: Dynamic regimen optimization based on recurrence probability;Risk-adapted surveillance intervals;Precision escalation/de-escalation of treatment intensity. Compared to conventional parameters, the model’s multidimensional tumor characterization improves high-risk detection rates, facilitating early intervention to enhance 5-year OS. Clinical implementation protocols suggest:Low-risk: Extended 6–12 month imaging intervals;High-risk: Intensive 3-month monitoring with ultrasound and serum biomarkers (CA15-3, CEA).The model’s exceptional metastatic pattern discrimination (training AUC = 0.9458, validation AUC = 0.8983) informs radiation field design and systemic therapy prioritization, reducing overtreatment rates while maintaining equivalent locoregional control.

AI technology is progressively enabling end-to-end integration across the oncological continuum, encompassing cancer screening, diagnosis, therapeutic intervention, and prognostic prediction. By leveraging deep learning architectures to decipher intrinsic correlations within multi-omics data, AI facilitates the establishment of more objective and multidimensional clinical decision-making frameworks. While current capabilities preclude complete replacement of molecular pathological assessments, the technology demonstrates considerable potential in mutational mechanism elucidation, clinical trial optimization, and precision patient stratification for targeted therapies. Through continuous algorithmic refinement and advanced multi-modal data fusion techniques, AI is poised to emerge as a pivotal catalyst in advancing precision oncology paradigms.

This study provides a foundation for future advancements in dynamic prognostic monitoring. An AI-powered continuous risk assessment system could be developed to analyze real-time ultrasound imaging alterations during follow-up surveillance (e.g., emerging microcalcifications or abnormal vascular patterns), enabling iterative updates of radiomics risk scores (RadScore). When integrated with circulating tumor DNA (ctDNA) load monitoring, this approach may establish an imaging-liquid biopsy synergy for early recurrence detection, empowering clinicians to dynamically adapt therapeutic strategies based on evolving disease profiles. Further exploration could focus on elucidating biological mechanisms by correlating radiomic features with genomic/transcriptomic data to decode molecular drivers of triple-negative breast cancer (TNBC) metastasis. Such radiogenomic integration may uncover novel therapeutic targets (e.g., DNA repair pathway vulnerabilities, immune checkpoint regulators) for precision treatment. Additionally, multidimensional prognostic models could be constructed by harmonizing mammographic patterns, dynamic contrast-enhanced MRI parameters, ctDNA mutation profiles, and whole-slide digital pathology images (WSI). To facilitate clinical implementation, user-friendly AI software compatible with hospital information systems (HIS) could be designed to deliver real-time prognostic assessments during routine clinical workflows.

This study developed an innovative multi-task prediction model based on ultrasound radiomics and artificial intelligence technology to assess the risk and patterns of postoperative recurrence and metastasis in triple-negative breast cancer (TNBC), representing a pioneering exploration in AI-driven personalized precision therapy for TNBC. The results demonstrate that ultrasound radiomic features effectively characterize the biological heterogeneity of TNBC, with the constructed AI model exhibiting outstanding performance in both predicting recurrence/metastasis risk and differentiating metastasis patterns. This research not only validates the technical advantages of AI in deciphering complex medical imaging features and uncovering prognostic correlations, but also proposes an innovative framework integrating multimodal data.

While achieving these results, we fully acknowledge the limitations of this study: class imbalance in the dataset, single-center retrospective design, lack of external validation, and insufficient exploration of biological mechanisms. These factors somewhat constrain the model’s clinical applicability and translational potential. Future research will focus on optimizing model performance through multicenter collaborations, integration of multimodal data, and rigorous external validation. We plan to develop an imaging-clinical integrated prediction system and enhance predictive accuracy through multi-omics data integration to accelerate clinical translation.

The deep integration of artificial intelligence and radiomics is reshaping the paradigm of cancer prognosis assessment. This study not only confirms the value of ultrasound imaging in TNBC management but also establishes a theoretical foundation for AI-driven diagnostic-therapeutic models. With continuing technological advancements, AI is poised to become the core engine of precision oncology, ultimately improving patients’ quality of life and clinical outcomes.