Skip to content
Frontiers
Search

Mini Review

Frontiers in Imaging ·

Volume 5 - 2026 | https://doi.org/10.3389/fimag.2026.1870528

Integration of imaging with liquid biopsy using artificial intelligence for ultra-early detection of breast cancer

  • Shalaka Ramgir-Naidu1*corresponding author
  • Asmita Govekar1
  • Abhishek Ojha1
  • Meghavi Soni1
  1. Symbiosis Institute of Health Sciences, Symbiosis International (Deemed University), Pune, India
Article metrics0 viewsView details
PDFCiteMetrics

Abstract

Breast cancer early detection using liquid biopsy, circulating tumor DNA (ctDNA), artificial intelligence (AI), and multimodal fusion offers a promising but still emerging research strategy to overcome the limitations of conventional imaging. Breast cancer screening and molecular diagnostics remain constrained by the inability of mammography and magnetic resonance imaging to detect pre-invasive disease, alongside the low sensitivity and spatial ambiguity of ctDNA in early-stage settings. This mini-review summarizes the rationale and recent advances in AI-driven multimodal frameworks that integrate imaging phenotypes with blood-derived genotypic signals through feature-, decision-, and intermediate-level fusion strategies. Such approaches improve diagnostic sensitivity and specificity by capturing complementary biological and structural information, enabling earlier detection and longitudinal risk assessment. Despite this progress, clinical translation is hindered by data heterogeneity, the lack of standardized multimodal datasets, and limited prospective validation. This study highlights the emerging biology-first, imaging-informed framework. Despite recent progress, current multimodal approaches remain largely investigational and require robust prospective evidence before clinical deployment. It outlines key future directions, including federated learning, longitudinal modeling, and large-scale validation, to support the future evolution of scalable and equitable early-detection strategies.

Keywords

1 Introduction

Breast cancer is the most commonly diagnosed malignancy among women worldwide and a leading cause of cancer-related mortality. Survival is strongly stage-dependent. Five-year survival exceeds 99% for Stage I disease but falls to approximately 28% for Stage IV (Siegel et al., 2023). Early detection of malignant molecular changes before clinically visible lesions develop could improve intervention opportunities and reduce mortality, although robust implementation remains challenging.

Contemporary screening relies on mammography, supplemented by MRI or ultrasound in high-risk groups. These modalities detect anatomically manifest lesions but cannot capture pre-invasive molecular changes that precede visible tumor formation by months to years (Lone et al., 2022). Liquid biopsy analyses tumor-derived material in peripheral blood, including ctDNA, cell-free DNA (cfDNA) methylation signatures, circulating tumor cells (CTCs), microRNAs, and extracellular vesicles. It offers a biologically earlier detection window.

Despite advances in multimodal AI systems, emerging evidence is largely derived from retrospective studies, limited datasets, and proof-of-concept investigations. Prospective evaluations demonstrating improved screening outcomes or reduced mortality remain limited (Easaw et al., 2025). Accordingly, this review presents these technologies as evolving investigational strategies rather than established clinical tools.

Both approaches have significant standalone limitations. Imaging is phenotypic and misses pre-structural genomic events. Liquid biopsy is spatially blind and suffers from low sensitivity at early stages (Easaw et al., 2025; Zhu et al., 2026). AI provides a computational framework for integrating both modalities, learning cross-modal patterns that surpass the diagnostic ceiling of either approach alone (Yang et al., 2025). This review examines the rationale, integration strategies, algorithmic approaches, clinical perspectives, technical challenges, validation pathways, and future directions for AI-driven multimodal breast cancer detection.

2 Conceptual framework and rationale

2.1 Limitations of standalone imaging and liquid biopsy

Mammography loses sensitivity in dense breast tissue, present in 25–40% of screening-age women, because tumors and fibroglandular tissue appear similarly radio-opaque. Dense breasts confer a 1.6- to 2-fold elevated cancer risk while reducing detection reliability (Isautier et al., 2024). Approximately half of women undergoing annual mammography over a decade receive at least one false-positive result (Stroth, 2025). Supplemental MRI halved interval cancer rates in the DENSE (Dense Tissue and Early Breast Neoplasm Screening) trial, but carries lower specificity, high cost, and access barriers that preclude universal deployment (Bakker et al., 2019). A 2023 meta-analysis found MRI detected 25.7 cancers per 1,000 screenings vs. 4.3–4.8 per 1,000 for ultrasound and tomosynthesis, yet false-positive burden persisted across all modalities (Tran and Ray, 2023). Crucially, all imaging modalities remain phenotypic and cannot visualize pre-invasive somatic mutations, epigenetic reprogramming, or clonal expansion.

Liquid biopsy faces complementary constraints. ctDNA constitutes less than 0.01% of total cfDNA in early-stage tumors, creating a severe signal-to-noise problem. Fragments are cleared in as little as 30 min, so a single blood draw captures only a transient molecular snapshot (Martin-Alonso et al., 2024). No gold-standard assay exists for early-stage detection, and most clinical assays interrogate only one or two tumor DNA features, limiting specificity (Easaw et al., 2025; Zhu et al., 2026; Parums, 2025). In low-prevalence screening populations, even small specificity reductions generate many false positives, leading to unnecessary biopsies, patient anxiety, and increased costs. Clonal hematopoiesis of indeterminate potential (CHIP) can mimic early cancer signals, and liquid biopsy cannot localize a detected signal or characterize a lesion's radiological phenotype, making it a complement rather than a replacement for imaging (Kuligina et al., 2025).

2.2 Phenotypic vs. genotypic information, integration need, and core hypothesis

Imaging captures macroscopic phenotypic features tissue shape, texture, vascularity, and architectural distortion reflecting downstream structural consequences of malignant transformation. Liquid biopsy provides direct genotypic evidence, ctDNA mutations, promoter methylation, copy number variations, and fragment length patterns that may precede detectable anatomical changes by months to years (Lone et al., 2022). Specific alterations, including PIK3CA mutations, BRCA1/2 methylation changes, TP53 mutations, and HER2 amplification, can be detectable in plasma at the pre-invasive stage when imaging remains uninformative. The relationship is bidirectional, certain MRI texture and kinetic features correlate with receptor status and tumor mutational burden, while ctDNA methylation signatures carry tissue-of-origin signals that can direct organ-specific imaging (Foser et al., 2024).

The central hypothesis is that joint AI analysis of imaging phenotypes and blood-borne genotypes may eventually improve sensitivity and specificity for earlier-stage detection compared with either modality alone, however, this remains to be validated in large prospective clinical studies, potentially supporting a more biology-informed and imaging-supported detection framework in future screening paradigms (Kuligina et al., 2025).

Combining both data streams can reduce false negatives from each modality's blind spots and false positives by requiring corroborating evidence across independent data streams (Martin-Alonso et al., 2024). In quantitative terms, standalone liquid biopsy assays for early-stage disease typically show lower sensitivity (approximately 30–70%) despite high specificity (>90%), as shown in Table 1, while unimodal imaging systems generally achieve AUC values of approximately 0.75–0.92 (Table 1), depending on the modality and population (Foser et al., 2024). In contrast, multimodal fusion methods integrating imaging and liquid biopsy data have reported relative AUC improvements of 10–20% and sensitivity gains of 15–25% (Table 1) in retrospective datasets, although these benefits vary across studies and depend on cohort size and data quality (Wang et al., 2024).

Table 1. Advanced comparative overview of imaging, liquid biopsy, and AI-based multimodal integration for ultra-early breast cancer detection.

S. no.DomainData typeKey role in early detectionStrengthsLimitationsAI methods commonly usedPerformance metrics (reported trends)Clinical utility/ application stagePractical/ regulatory/ ethical issuesIntegration readinessReferences
1Imaging (mammography/ DBT/MRI/US)Radiology phenotypes (spatial, structural)Detects morphological abnormalities such as masses, calcifications, and architectural distortionsWidely standardized screening tool; MRI shows highest sensitivity (~90–99% in high-risk populations); DBT reduces recall ratesReduced sensitivity in dense breasts (~30–48% drop); high false positives; lacks molecular-level detectionCNNs (ResNet, DenseNet), vision transformers (ViT), radiomics pipelinesAUC: ~0.75–0.92 depending on modality; sensitivity improves with AI-assisted detection (~5–15%)Primary screening and diagnostic imagingOverdiagnosis risk; variability in imaging protocols; regulatory approval required (FDA/CE); bias in training datasets (ethnicity, breast density)High (already integrated clinically with AI assistance)McKinney et al., 2020
2Liquid biopsy (ctDNA, cfDNA, CTCs, miRNA, EVs)Molecular biomarkers (non-spatial, circulating)Detects tumor-derived genomic, epigenetic, and transcriptomic alterations before radiological visibilityUltra-sensitive molecular detection; enables minimal residual disease (MRD) monitoring; non-invasive and repeatableExtremely low analyte abundance in early stages (< 0.01% VAF); lack of tumor localization; assay variabilityML classifiers (Random Forest, SVM), Bayesian models, deep sequencing models, transformers, LSTM (temporal trends)Sensitivity varies (~30–70% in early-stage); specificity often >90%; improves with multi-marker panelsEarly detection, recurrence monitoring, treatment response trackingStandardization issues; false positives from clonal hematopoiesis; regulatory uncertainty; cost and scalability concernsModerate (clinical adoption increasing but not universal)Wan et al., 2017
3Radiomics + genomics integrationCombined imaging + molecular featuresLinks radiological phenotype with genomic alterations (e.g., TP53, PIK3CA, HER2 status)Enables phenotype-genotype mapping; improves risk stratification and subtype prediction; enhances precision diagnosticsSmall matched datasets; poor reproducibility; domain shift between institutions; lack of standardized pipelinesMulti-view learning, canonical correlation analysis (CCA), feature fusion models, graph-based learningAUC improvement of ~5–10% over unimodal models; better subtype classification accuracy (~80–90%)Translational research; emerging precision oncology applicationsData harmonization challenges; interpretability issues; regulatory complexity for multi-modal diagnosticsLow–moderate (research-heavy, limited clinical deployment)Bera et al., 2022
4Early fusion (feature-level integration)Concatenated imaging + molecular embeddingsSimultaneous learning of cross-modal relationships at feature levelCaptures complex biological interactions early; relatively straightforward implementationCurse of dimensionality; risk of overfitting; modality dominance (imaging often outweighs molecular signals)CNN + MLP fusion, autoencoders, deep feature concatenation networksImproved classification accuracy (~5–12% vs. unimodal); unstable performance with small datasetsExperimental multimodal AI pipelinesRequires large annotated datasets; computationally intensive; interpretability concernsLow (experimental stage)Ngiam et al., 2011
5Intermediate fusion (shared latent space models)Learned joint embedding spaceAligns imaging and molecular data into unified representationBest balance between performance and interpretability; captures cross-modal interactions effectivelyRequires large, well-aligned datasets; computational complexity; training instabilityCross-attention transformers, variational autoencoders (VAEs), contrastive learning frameworksHighest reported gains (~10–20% improvement in multimodal tasks); improved generalizationAdvanced research and pilot clinical AI systemsData privacy concerns (multi-source data); high computational cost; reproducibility challengesModerate (emerging frontier)Suzuki and Nitanda, 2021
6Late fusion (decision-level integration)Independent modality predictionsCombines outputs from imaging and liquid biopsy modelsRobust to missing modalities; flexible clinical deployment; easier validationCannot capture biological interactions between modalities; limited synergy exploitationEnsemble learning, weighted averaging, stacking models, meta-learnersStable performance; modest improvement (~3–8% over single modality)Clinical decision-support systems (CDSS)Easier regulatory pathway but still requires validation; risk of conflicting predictionsModerate–high (clinically feasible)Wolpert, 1992
7Temporal multimodal modelingLongitudinal imaging + serial liquid biopsyTracks disease evolution and detects malignancy before clinical manifestationEnables ultra-early detection; captures tumor dynamics; useful for MRD and relapse predictionMissing or asynchronous time points; patient compliance; complex data alignmentLSTM, GRU, time-aware transformers, sequential deep learning modelsEarly detection sensitivity improved (~15–25% in longitudinal models); strong predictive value for recurrenceLongitudinal monitoring and surveillanceData incompleteness; ethical concerns in predictive diagnosis; high data storage requirementsLow–moderate (still developing clinically)Rajkomar et al., 2018
8Explainable AI (XAI in breast cancer)Model interpretability outputs (heatmaps, feature attribution)Enhances transparency and clinical trust in AI decisionsImproves clinician confidence; aids regulatory approval; identifies biologically relevant featuresMostly correlational; does not establish causality; risk of misleading explanationsGrad-CAM, SHAP, LIME, attention maps, saliency mappingNo direct performance gain; improves interpretability scores and clinical adoption likelihoodClinical validation and regulatory compliance layerRegulatory requirement for transparency; medico-legal accountability; interpretability vs. performance trade-offHigh (essential for deployment)Ribeiro et al., 2016

3 Data Integration strategies

3.1 Multimodal datasets: definition and assembly

A multimodal dataset combines at least two fundamentally distinct data types from the same patients at matched time points (Steyaert et al., 2023). The imaging component may comprise mammography DICOM (Digital Imaging and Communications in Medicine) files, dynamic contrast-enhanced MRI sequences, ultrasound volumes, or digital breast tomosynthesis (Foser et al., 2024). The molecular component may include ctDNA variant allele frequencies, cfDNA methylation beta-values, CTC counts, miRNA profiles, or multi-analyte protein panels. Co-registration with ground-truth outcome labels is critical for supervised learning, as shown in Figure 1.

AI-driven multimodal fusion pipeline for early breast cancer detection. Imaging and liquid biopsy signals are encoded by modality-specific neural networks. Their fused representations generate a unified risk score for early detection, stratification, and longitudinal monitoring. DCE-MRI, dynamic contrast-enhanced MRI; US, ultrasonography; ctDNA, circulating tumor DNA; cfDNA, cell-free DNA; CTC, circulating tumor cells; miRNA, microRNA, CNN, convolutional neural networks; LSTM, long short-term memory; MLP, multilayer perceptron; VAEs, variational autoencoders; MRD, minimal residual disease; XAI, explainable AI.
Figure 1. AI-driven multimodal fusion pipeline for early breast cancer detection. Imaging and liquid biopsy signals are encoded by modality-specific neural networks. Their fused representations generate a unified risk score for early detection, stratification, and longitudinal monitoring. DCE-MRI, dynamic contrast-enhanced MRI; US, ultrasonography; ctDNA, circulating tumor DNA; cfDNA, cell-free DNA; CTC, circulating tumor cells; miRNA, microRNA, CNN, convolutional neural networks; LSTM, long short-term memory; MLP, multilayer perceptron; VAEs, variational autoencoders; MRD, minimal residual disease; XAI, explainable AI.

The Cancer Genome Atlas (TCGA) has provided foundational matched datasets for cancer AI development (Wang et al., 2024), and prospective trials including DETECT, STRIVE, and NHS-Galleri are generating matched datasets for screening populations (Zhu et al., 2026). Assembly challenges are substantial, imaging data require standardized DICOM preprocessing pipelines while molecular data require modality-specific normalization. Most current multimodal imaging–liquid biopsy datasets are limited in size, compiled retrospectively, and institution-specific, often including only a few 100 matched cases (Steyaert et al., 2023). This limits statistical power, increases the risk of selection bias, and remains a major barrier to reproducibility and external validation.

3.2 Feature-level and decision-level fusion

Feature-level fusion extracts modality-specific representations independently and presents a joint vector to a shared downstream classifier. Its principal advantage is enabling the model to learn cross-modal interactions simultaneously, potentially uncovering correlations invisible to either modality alone. It is computationally straightforward, requiring only a single downstream model (Waqas et al., 2024). However, when imaging and molecular features differ substantially in scale and dimensionality, one modality can dominate the joint feature space. This increases overfitting risk in small matched datasets, where high-capacity models may learn dataset-specific noise rather than biologically generalizable patterns (Nikolaou et al., 2025). Reported performance gains in certain fusion studies should therefore be interpreted cautiously, especially without external validation.

Decision-level fusion trains separate unimodal models and combines their output risk scores via majority voting, Bayesian averaging, or learned ensemble weights. Its key advantage is modularity: each sub-model can be developed independently, and the system remains functional when one modality is unavailable (Nikolaou et al., 2025). Decision-level fusion yields AUC increases of 3–8% over unimodal baselines (Table 1), while feature-level fusion shows larger gains of approximately 5–12% but with greater variability. The largest improvements come from intermediate, attention-based architectures, with AUC increases of 10–20% in retrospective benchmarks (Table 1), though prospective validation remains lacking (Isautier et al., 2024). Emerging consensus favors hybrid intermediate architectures where modality-specific encoders project into a shared latent space fused through cross-attention mechanisms as offering the best balance between interaction learning and robustness to missing data (Yang et al., 2025).

3.3 Temporal alignment and high-dimensional challenges

Both data streams are more informative when modeled longitudinally. Longitudinal imaging tracks changes in lesion size, morphology, and radiomic feature trajectories, identifying subtle progressive trends that are individually sub-threshold but collectively detectable (Luo et al., 2024). Serial liquid biopsy sampling through monthly or quarterly blood draws adds a temporal molecular dimension: a rising ctDNA mutant allele fraction may signal clonal expansion before any imaging-detectable lesion forms. ctDNA monitoring can predict breast cancer relapse with a lead time of up to 3 years (Zhu et al., 2026). The core challenge is the asynchrony between imaging cycles and blood draw intervals, requiring careful co-registration onto a common time axis.

Combining imaging and genomic data creates a high-dimensional joint feature space. The curse of dimensionality can cause models to memorize noise rather than learn generalizable signals, particularly when matched datasets contain only a few 100 cases (Luo et al., 2024). This challenge is further exacerbated by the imbalance between highly dimensional imaging and genomic feature spaces (Steyaert et al., 2023). As a result, multimodal fusion models are prone to overfitting, overly optimistic internal validation results, and poor cross-institutional reproducibility. Moreover, many published studies rely on retrospective datasets without independent external validation, limiting confidence in their clinical robustness (Martin-Alonso et al., 2024). One pan-cancer survival study had to discard over 99% of variables before late-fusion ensembling (Waqas et al., 2024). Feature selection methods, autoencoders, variational autoencoders (VAEs), and canonical correlation analysis (CCA) address this problem, while deep learning methods including VAEs learn non-linear shared representations that capture complex cross-modal associations (Yang et al., 2025; Nakach et al., 2024), as shown in Figure 1. Significant inter-study heterogeneity from differences in imaging acquisition protocols and sequencing platforms further hampers direct model comparison and remains a major barrier to reproducibility (Luo et al., 2024). The lack of standardized benchmarking frameworks further hampers direct comparison between multimodal models and remains a major barrier to reproducibility (Yang et al., 2025).

4 AI algorithmic approaches

The AI techniques covered in this section span a range of clinical maturity, from deployed imaging algorithms to early-stage experimental multimodal architectures. Their current level of evidence is presented in context.

4.1 Data types, CNNs, and sequential models

Each data type favors specific AI architectural inductive biases. Medical images are spatially correlated pixel arrays, favoring CNNs (convolutional neural networks) or vision transformers. Genomic sequence data and temporal ctDNA measurements carry long-range sequential dependencies, favoring LSTM (long short-term memory) networks or attention-based transformers. Tabular molecular data, such as protein panels and CTC counts, are efficiently processed by fully connected networks or gradient-boosted trees (Yang et al., 2025; Waqas et al., 2024).

CNNs learn hierarchical spatial feature representations. Early layers detect edges and textures; intermediate layers capture lesion margins; deeper layers encode mass shape, heterogeneity, and enhancement kinetics. CNN-based transfer learning using backbone architectures such as ResNet, AlexNet, and DenseNet is integrated into clinically deployed imaging AI systems (Zhu et al., 2023). Hybrid CNN-transformer models, including TransBreastNet and MammoFormer, have achieved up to 13% performance improvement over individual architectures (Brahmareddy and Selvan, 2025; Peter et al., 2025). Image-only deep learning systems perform approximately 5–15% better than standard radiologist assessment (Table 1). Combined with molecular features, multimodal systems can increase AUC by approximately 5–20% and sensitivity by 10–25% over unimodal baselines (Table 1), though gains depend strongly on dataset size, modality alignment, and external validation rigor (Waqas et al., 2024).

RNNs (Recurrent Neural Networks), particularly LSTM networks and gated recurrent units (GRUs), model dependencies across variable-length sequences and maintain memory of earlier inputs, directly addressing the temporal structure of longitudinal biomarker monitoring (Zhu et al., 2023). CNN-LSTM hybrid architectures have achieved up to 99.90% classification accuracy on standardized benchmarks (Kaddes et al., 2025), though such results may reflect dataset homogeneity or overfitting rather than true clinical performance. Transformer architectures use self-attention to model long-range dependencies without the vanishing gradient problem of classical RNNs. A multimodal co-attention transformer for survival prediction using whole-slide images and genomic sequences substantially outperformed single-modality alternatives (Waqas et al., 2024).

4.2 Multimodal architectures, shared latent space, explainability, and transfer learning

Multimodal deep learning architectures integrate imaging and molecular data through designs ranging from simple embedding concatenation to complex cross-attention and graph-based interaction models. Attention-based and hybrid architectures have been identified as the most performant and clinically relevant approaches (Nakach et al., 2024). They require large, well-aligned multimodal datasets that remain scarce in breast cancer research. Representative models include SurvPGC, which integrates pathology images, genomic data, and clinical records through a cross-attention module (Hou et al., 2025), and DeepClinMed-PGM, which fuses molecular data, pathological slides, and clinical information for disease-free survival prediction (Wang et al., 2024). GNNs are gaining traction for encoding inter-modality relationships as graph structures (Waqas et al., 2024). Key AI models are summarized in Table 1.

In intermediate fusion architectures, modality-specific encoders project inputs onto a common lower-dimensional manifold where imaging and molecular representations are geometrically aligned (Ing et al., 2025). This shared latent space enables cross-modal biomarker discovery, imputation of missing modalities, and uncertainty quantification through probabilistic sampling (Foser et al., 2024). Explainable AI (XAI) tools are essential for clinical adoption: Grad-CAM highlights spatial imaging regions driving predictions, SHAP quantifies the contribution of each molecular feature, and LIME provides locally faithful approximations around individual predictions (Yang et al., 2025). Transfer learning and federated learning address dataset scarcity, enabling model training across institutions without centralizing patient data (Lilhore et al., 2025; Lipkova et al., 2022).

5 Clinical and biological perspectives

The technologies covered in this section range from novel translational biomarkers (cfDNA, ctDNA) to experimental applications in early cancer diagnosis. Their clinical maturity varies widely and should be interpreted accordingly.

5.1 ctDNA biology, biomarker kinetics, and pre-radiological detection

ctDNA shedding is highly variable across tumor types, stages, and individual patient factors. In asymptomatic screening, high specificity is crucial. In low-prevalence settings, false-positive ctDNA results can lead to unnecessary tests, biopsies, distress, and overtreatment (Wang et al., 2024). High-proliferative tumors such as TNBC (triple-negative breast cancer) and HER2-positive subtypes shed more ctDNA due to rapid cell turnover. Indolent hormone receptor-positive tumors release minimal ctDNA, reducing sensitivity (Lipsyc-Sharf et al., 2025). The short half-life of ctDNA (1–2 h) complicates detection, as hepatic and renal clearance rapidly removes circulating fragments (Davis et al., 2024). Tumor DNA accounts for less than 0.01% of total cfDNA in Stage I and II patients (Martin-Alonso et al., 2024). Ultrasensitive assays detecting one part per million ctDNA are under development and may provide a 15-month lead time before clinical relapse (Elliott et al., 2025; McKelvey et al., 2024). Pharmacological priming agents that transiently inhibit ctDNA clearance and noise reduction strategies targeting CHIP are being explored to improve sensitivity and specificity (Martin-Alonso et al., 2024). DCE-MRI and PET-CT patterns are linked to ctDNA shedding dynamics, and radiomic features reflecting tumor vascularity correlate with underlying molecular biology (Abbosh et al., 2023). The concept of pre-radiological detection proposes that molecular alterations may become detectable before structural imaging abnormalities emerge; however, the clinical reliability and reproducibility of this approach remain under investigation (Bakker et al., 2019). If validated prospectively, pre-radiological detection could have important clinical implications for earlier intervention and risk stratification (Tran and Ray, 2023). If prospectively validated, pre-radiological detection of molecular alterations before structural imaging abnormalities emerge could have important implications for earlier intervention and risk stratification (Siegel et al., 2023; Zhu et al., 2026).

6 Technical and logistical challenges

Translating AI-driven multimodal platforms into clinical practice faces substantial barriers. Variability in imaging techniques including differences in mammography kilovoltage, detector types, MRI field strength, and contrast dosing leads to inconsistent radiomic outputs susceptible to vendor-specific artifacts and deep learning domain shift (Schopf et al., 2024; Ali et al., 2025). Liquid biopsy is complicated by pre-analytical inconsistencies in tube types, centrifugation delays, and freeze-thaw cycles, as well as diverse analytical platforms including ddPCR (droplet-based digital PCR), NGS (Next-Generation Sequencing) panels, and methylome assays that resist direct comparison (Panet et al., 2024). Exosome isolation inconsistencies persist despite the MISEV2023 (Minimal Information for Studies of Extracellular Vesicles) guidelines (Welsh et al., 2024).

Inter-institutional variability in demographics, scanners, and clinical workflows exacerbates harmonization challenges. A lack of matched longitudinal datasets limits the ability to train models capturing tumor evolution over time (Yang et al., 2025). Annotation noise arises from inter-radiologist variability, long follow-up periods required for ctDNA ground-truth labels, false positives from CHIP, and imaging artifacts in dense breast tissue. Because early breast cancer is rare in asymptomatic populations, even small false-positive rates can create substantial clinical burden. Balancing sensitivity with sufficiently high specificity remains a key unresolved challenge (Davis et al., 2024). Overfitting in high-dimensional multimodal feature spaces, reliance on retrospective single-center datasets, and limited external validation further reduce confidence in real-world application (Elliott et al., 2025). These challenges highlight the urgent need for harmonized benchmarks, standardized pre-analytical protocols, and well-structured prospective trials (Ali et al., 2025; Haghighat et al., 2025).

7 Validation and clinical implementation

Current evidence derives predominantly from retrospective, single-site studies insufficient to demonstrate clinical utility or improved patient outcomes (Haghighat et al., 2025). Robust prospective trials are essential: they should enroll large diverse cohorts, evaluate mortality reductions, and incorporate adaptive designs (Panet et al., 2024). Platform frameworks such as I-SPY 2 and MRD trials including DARE, TRAK-ER, and TREAT ctDNA serve as models. Priority metrics include sensitivity, PPV, NPV, AUC, lead time to detection, and biopsy referral rates. In low-prevalence populations, specificity and PPV are especially critical, and prospective studies must also assess unnecessary biopsies, patient anxiety, and cost-effectiveness (Keam, 2026). Currently, no multimodal AI platform integrating breast imaging and liquid biopsy has been successfully incorporated into clinical workflows for population-level early breast cancer screening (Schopf et al., 2024).

Among the technologies discussed in this review, only AI-assisted imaging systems have received regulatory approval for clinical use (Abbosh et al., 2023). The FDA granted 510(k) clearance to nine AI mammography tools between 2017 and 2021, and Prognosia Breast achieved Breakthrough Device Designation in 2025. Liquid biopsies such as Guardant360 CDx are approved only for metastatic ESR1 detection in the US and Japan, with no endorsements for early screening (Schopf et al., 2024; Keam, 2026). Inter-study comparison remains challenging due to differences in datasets, outcome definitions, and evaluation protocols, despite reported performance improvements. Consequently, reported performance gains in multimodal systems should be interpreted as approximate ranges rather than directly comparable benchmarks (Davis et al., 2024). Multimodal imaging liquid biopsy AI systems remain experimental. Clinical integration will additionally require standardized phlebotomy and imaging protocols, streamlined data fusion pipelines, and explainable AI to foster clinician trust. Persistent barriers include high costs, geographic inequities, and disparities in access to next-generation sequencing among underrepresented populations (Davis et al., 2024; Ali et al., 2025).

8 Future directions and innovations

Future research must prioritize large-scale, prospective, multicenter trials validating multimodal AI systems across diverse populations, including different ethnicities, ages, and breast density subtypes. Standardized protocols for serial liquid biopsy sampling, synchronized with high-resolution imaging, will be essential (Davis et al., 2024). Federated learning enables decentralized model training, but challenges remain, algorithmic bias, privacy governance, and medico-legal accountability for AI-driven decisions must all be resolved. Ethical concerns also arise from predictive anxiety when early molecular signals are detected without clear clinical actionability (Elliott et al., 2025). Transparent governance, explainable AI outputs, and legal frameworks defining responsibility among developers, clinicians, and institutions are required (Lipkova et al., 2022; Ali et al., 2025).

Personalized early-detection models integrating germline genetics, breast density, and hormonal history alongside imaging and molecular signals will improve sensitivity in high-risk subgroups. Longitudinal monitoring frameworks that continuously update risk scores using serial ctDNA trajectories and interval imaging represent the next generation of surveillance platforms (Abbosh et al., 2023). Expanding liquid biopsy beyond ctDNA to include CTCs, exosomes, and proteomics will provide more comprehensive molecular characterization (Welsh et al., 2024; Hao et al., 2025). Point-of-care integration using microfluidic assays and portable AI-enabled ultrasound offers the potential for single-visit personalized risk scoring (Ali et al., 2025). Surface-enhanced Raman spectroscopy combined with AI classification demonstrated high diagnostic performance for plasma exosomes, illustrating the feasibility of rapid molecular screening at point of care (Shin et al., 2023). Scaling these innovations globally will require multimodal data fusion, federated learning, and validation across diverse populations to ensure equitable benefits (Yang et al., 2025; Lipsyc-Sharf et al., 2025).

9 Conclusion

The integration of breast imaging and liquid biopsy, enabled by AI, represents a scientifically compelling research area with the potential to improve early cancer detection. Conventional imaging depends on detectable structural alterations and cannot identify preclinical disease. Liquid biopsy suffers from suboptimal sensitivity, variable specificity, and a lack of spatial resolution. Combined, these modalities yield complementary phenotypic and genotypic information that mitigates each other's shortcomings.

AI-based fusion techniques have shown promising results in retrospective studies, though clinical utility is not yet established. Realizing this potential requires standardized data platforms, rigorous prospective validation, and equitable implementation. Because breast cancer incidence is low in screened populations, very high specificity is essential to minimize false-positive molecular findings and avoid unnecessary procedures. Large prospective multicenter trials, standardized multimodal datasets, and regulatory validation are required before these strategies can be considered for routine clinical practice.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Mingquan Lin, University of Minnesota Twin Cities, United States

Reviewed by: Luka Bulić, St. Catherine Specialty Hospital, Croatia

Statements

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Funding

The author(s) declared that financial support was received for this work and/or its publication.

Author contributions

SR-N: Conceptualization, Methodology, Resources, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. AG: Writing – original draft, Writing – review & editing. AO: Writing – original draft, Writing – review & editing. MS: Writing – original draft.

References

  1. Abbosh C., Frankell A. M., Harrison T., Kisistok J., Garnett A., Johnson L.. (2023). Tracking early lung cancer metastatic dissemination in TRACERx using ctDNA. Nature 616, 553 – 562. doi: doi:10.1038/s41586-023-05776-4
  2. Ali A., Alghamdi M., Marzuki S. S., Tengku Din T. A., Yamin M. S., Alrashidi M.. (2025). Exploring AI approaches for breast cancer detection and diagnosis: a review article. Breast Cancer Targets Therapy 17, 927 – 947. doi: doi:10.2147/BCTT.S550307
  3. Bakker M. F., de Lange S. V., Pijnappel R. M., Mann R. M., Peeters P. H., Monninkhof E. M.. (2019). Supplemental MRI screening for women with extremely dense breast tissue. New Engl. J. Med. 381, 2091 – 2102. doi: doi:10.1056/NEJMoa1903986
  4. Bera K., Braman N., Gupta A., Velcheti V., Madabhushi A. (2022). Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat. Rev. Clin. Oncol. 19, 132 – 146. doi: doi:10.1038/s41571-021-00560-7
  5. Brahmareddy A., Selvan M. P. (2025). TransBreastNet a CNN transformer hybrid deep learning framework for breast cancer subtype classification and temporal lesion progression analysis. Sci. Rep. 15: 35106. doi: doi:10.1038/s41598-025-19173-6
  6. Davis A. A., Bucheit L. A., Zhang N., Pretz C., Strickler J. H., Chachoua A.. (2024). Utilization of circulating tumor DNA (ctDNA) testing by race and ethnicity in more than 135,000 patients. doi: doi:10.1200/JCO.2024.42.16_suppl.e13702
  7. Easaw S., Hsu J., Steuerwald N., Heeke A. L. (2025). Liquid clues: tracking early-stage breast cancer with ctDNA a mini review. Front. Oncol. 15: 1634859. doi: doi:10.3389/fonc.2025.1634859
  8. Elliott M. J., Howarth K., Main S., Fuentes Antrás J., Echelard P., Dou A.. (2025). Ultrasensitive detection and monitoring of circulating tumor DNA using structural variants in early-stage breast cancer. Clin. Cancer Res. 31, 1520 – 1532. doi: doi:10.1158/1078-0432.CCR-24-3472
  9. Foser S., Maiese K., Digumarthy S. R., Puig-Butille J. A., Rebhan C. (2024). Looking to the future of early detection in cancer: liquid biopsies, imaging, and artificial intelligence. Clin. Chem. 70, 27 – 32. doi: doi:10.1093/clinchem/hvad196
  10. Haghighat F., Nemati Z., Rambodrad A., Negareshifard P., Jafari E. (2025). Multimodal deep learning and data fusion in precision breast oncology: clinical Applications, fusion Strategies, and future directions. InfoSci. Trends 2, 81 – 115. doi: doi:10.61882/ist.202502.10.06
  11. Hao X., Liu Z., Ma F., Li T., Liu C., Wang N.. (2025). Exosome-based liquid biopsy in early screening and diagnosis of cancers. Dose Response 23: 15593258251344480. doi: doi:10.1177/15593258251344480
  12. Hou J., Zhang R., Xie Y., Li C., Qin W. (2025). Multimodal deep learning for cancer prognosis prediction with clinical information prompts integration. NPJ Digital Med. 9: 76. doi: doi:10.1038/s41746-025-02257-y
  13. Ing A., Andrades A., Cosenza M. R., Korbel J. O. (2025). Integrating multimodal cancer data using deep latent variable path modelling. Nat. Machine Intell. 7, 1053 – 1075. doi: doi:10.1038/s42256-025-01052-4
  14. Isautier J. M., Houssami N., Hadlow C., Marinovich M. L., Hope S., Zackrisson S.. (2024). Clinical guidelines for the management of mammographic density: a systematic review of breast screening guidelines worldwide. JNCI Cancer Spectrum 8: pkae103. doi: doi:10.1093/jncics/pkae103
  15. Kaddes M., Ayid Y. M., Elshewey A. M., Fouad Y. (2025). Breast cancer classification based on hybrid CNN with LSTM model. Sci. Rep. 15: 4409. doi: doi:10.1038/s41598-025-88459-6
  16. Keam S. J. (2026). Imlunestrant: first approval. Drugs 86, 135 – 141. doi: doi:10.1007/s40265-025-02266-x
  17. Kuligina E. S., Yanus G. A., Imyanitov E. N. (2025). Improvement of the sensitivity of circulating tumor DNA-based liquid biopsy: current approaches and future perspectives. Explor. Target. Anti-tumor Therapy 6: 1002333. doi: doi:10.37349/etat.2025.1002333
  18. Lilhore U. K., Sharma Y. K., Shukla B. K., Vadlamudi M. N., Simaiya S., Alroobaea R.. (2025). Hybrid convolutional neural network and bi-LSTM model with EfficientNet-B0 for high-accuracy breast cancer detection and classification. Sci. Rep. 15: 12082. doi: doi:10.1038/s41598-025-95311-4
  19. Lipkova J., Chen R. J., Chen B., Lu M. Y., Barbieri M., Shao D.. (2022). Artificial intelligence for multimodal data integration in oncology. Cancer Cell 40: 1095 – 1110. doi: doi:10.1016/j.ccell.2022.09.012
  20. Lipsyc-Sharf M., Aushev V., Rivero-Hinojosa S., Scalise C. B., Green R., Ortiz J. B.. (2025). 37P Genetic ancestry and tumor mutations influence circulating tumor DNA (ctDNA) detection rates in breast cancer: a large real-world study. ESMO Open 10: 104591. doi: doi:10.1016/j.esmoop.2025.104591
  21. Lone S. N., Nisar S., Masoodi T., Singh M., Rizwan A., Hashem S.. (2022). Liquid biopsy: a step closer to transforming diagnosis, prognosis and future of cancer treatments. Mol. Cancer 21: 79. doi: doi:10.1186/s12943-022-01543-7
  22. Luo L., Wang X., Lin Y., Ma X., Tan A., Chan R.. (2024). Deep learning in breast cancer imaging: a decade of progress and future directions. IEEE Rev. Biomed. Eng. 18, 130 – 151. doi: doi:10.1109/RBME.2024.3357877
  23. Martin-Alonso C., Tabrizi S., Xiong K., Blewett T., Sridhar S., Crnjac A.. (2024). Priming agents transiently reduce the clearance of cell-free DNA to improve liquid biopsies. Science 383: eadf2341. doi: doi:10.1126/science.adf2341
  24. McKelvey B. A., Andrews H. S., Baehner F. L., Chen J., Espenschied C. R., Fabrizio D.. (2024). Advancing evidence generation for circulating tumor DNA: lessons learned from a Multi-Assay study of baseline circulating tumor DNA levels across cancer types and stages. Diagnostics 14: 912. doi: doi:10.3390/diagnostics14090912
  25. McKinney S. M., Sieniek M., Godbole V., Godwin J., Antropova N., Ashrafian H.. (2020). International evaluation of an AI system for breast cancer screening. Nature 577, 89 – 94. doi: doi:10.1038/s41586-019-1799-6
  26. Nakach F. Z., Idri A., Goceri E. (2024). A comprehensive investigation of multimodal deep learning fusion strategies for breast cancer classification. Artif. Intell. Rev. 57: 327. doi: doi:10.1007/s10462-024-10984-z
  27. Ngiam J., Khosla A., Kim M., Nam J., Lee H., Ng A. Y.. (2011). Multimodal deep learning. InIcml 11, 689 – 696.
  28. Nikolaou N., Salazar D., RaviPrakash H., Gonçalves M., Mulla R., Burlutskiy N.. (2025). A machine learning approach for multimodal data fusion for survival prediction in cancer patients. NPJ Precision Oncol. 9: 128. doi: doi:10.1038/s41698-025-00917-6
  29. Panet F., Papakonstantinou A., Borrell M., Vivancos J., Vivancos A., Oliveira M.. (2024). Use of ctDNA in early breast cancer: analytical validity and clinical potential. NPJ Breast Cancer 10: 50. doi: doi:10.1038/s41523-024-00653-3
  30. Parums D. V. (2025). A review of circulating tumor DNA (ctDNA) and the liquid biopsy in cancer diagnosis, screening, and monitoring treatment response. Med. Sci. Monitor Int. Med. J. Exp. Clin. Res. 31: e949300. doi: doi:10.12659/MSM.949300
  31. Peter O. O., Emakporuena D., Tunde B. D., Abdulkarim M., Umar A. B. (2025). Transformer-based explainable deep learning for breast cancer detection in mammography: the mammoformer framework. arXiv [Preprint] arXiv:2508.06137. doi: doi:10.11648/j.ajcst.20250802.16
  32. Rajkomar A., Oren E., Chen K., Dai A. M., Hajaj N., Hardt M.. (2018). Scalable and accurate deep learning with electronic health records. NPJ Digital Med. 1: 18. doi: doi:10.1038/s41746-018-0029-1
  33. Ribeiro M. T., Singh S., Guestrin C. (2016). Model-agnostic interpretability of machine learning. arXiv [Preprint] arXiv:1606.05386. doi: doi:10.48550/arXiv.1606.05386
  34. Schopf C. M., Ramwala O. A., Lowry K. P., Hofvind S., Marinovich M. L., Houssami N.. (2024). Artificial intelligence-driven mammography-based future breast cancer risk prediction: a systematic review. J. Am. Coll. Radiol. 21, 319 – 328. doi: doi:10.1016/j.jacr.2023.10.018
  35. Shin H., Choi B. H., Shim O., Kim J., Park Y., Cho S. K.. (2023). Single test-based diagnosis of multiple cancer types using Exosome-SERS-AI for early stage cancers. Nat. Commun. 14: 1644. doi: doi:10.1038/s41467-023-37403-1
  36. Siegel R. L., Miller K. D., Wagle N. S., Jemal A. (2023). Cancer statistics, 2023. CA Cancer J. Clin. 73, 17 – 48. doi: doi:10.3322/caac.21763
  37. Steyaert S., Pizurica M., Nagaraj D., Khandelwal P., Hernandez-Boussard T., Gentles A. J.. (2023). Multimodal data fusion for cancer biomarker discovery with deep learning. Nat. Machine Intell. 5, 351 – 362. doi: doi:10.1038/s42256-023-00633-5
  38. Stroth K. (2025). Revolutionizing breast cancer screening: the impact of artificial intelligence on early cancer detection in women (Doctoral dissertation). Northeastern University, Boston, MA, UnIte States.
  39. Suzuki T., Nitanda A. (2021). Deep learning is adaptive to intrinsic dimensionality of model smoothness in anisotropic Besov space. Adv. Neural Inf. Process. Syst. 34, 3609 – 3621.
  40. Tran E., Ray K. (2023). Meta-analysis of supplemental breast cancer screening modalities in women with dense breasts and negative mammography. Radiol. Imag. Cancer 5: e239012. doi: doi:10.1148/rycan.239012
  41. Wan J. C., Massie C., Garcia-Corbacho J., Mouliere F., Brenton J. D., Caldas C.. (2017). Liquid biopsies come of age: towards implementation of circulating tumour DNA. Nat. Rev. Cancer 17, 223 – 238. doi: doi:10.1038/nrc.2017.7
  42. Wang Z., Lin R., Li Y., Zeng J., Chen Y., Ouyang W.. (2024). Deep learning-based multi-modal data integration enhancing breast cancer disease- free survival prediction. Precision Clin. Med. 7: pbae012. doi: doi:10.1093/pcmedi/pbae012
  43. Waqas A., Tripathi A., Ramachandran R. P., Stewart P. A., Rasool G. (2024). Multimodal data integration for oncology in the era of deep neural networks: a review. Front. Artif. Intell. 7: 1408843. doi: doi:10.3389/frai.2024.1408843
  44. Welsh J. A., Goberdhan D. C., O'Driscoll L., Buzas E. I., Blenkiron C., Bussolati B.. (2024). Minimal information for studies of extracellular vesicles (MISEV2023): From basic to advanced approaches. J. Extracell. Vesicles 13: e12404. doi: doi:10.1002/jev2.12404
  45. Wolpert D. H. (1992). Stacked generalization. Neural Netw. 5, 241 – 259. doi: doi:10.1016/S0893-6080(05)80023-1
  46. Yang H., Yang M., Chen J., Yao G., Zou Q., Jia L.. (2025). Multimodal deep learning approaches for precision oncology: a comprehensive review. Briefings Bioinformat. 26: bbae699. doi: doi:10.1093/bib/bbae699
  47. Zhu H., Li Z., Xie K., Kassim S. H., Cao C., Huang K.. (2026). Liquid biopsy in early screening of cancers: emerging technologies and new prospects. Biomedicines 14: 158. doi: doi:10.3390/biomedicines14010158
  48. Zhu Z., Wang S. H., Zhang Y. D. (2023). A survey of convolutional neural network in breast cancer. Comput. Model. Eng. Sci. 136: 2127. doi: doi:10.32604/cmes.2023.025484

Citation

Ramgir-Naidu, S., et al. (2026). Integration of imaging with liquid biopsy using artificial intelligence for ultra-early detection of breast cancer. Frontiers in Imaging, Volume 5. https://doi.org/
Received
Accepted
Published
Volume
5 - 2026

Copyright

Copyright © 2026 Ramgir-Naidu, Govekar, Ojha and Soni.

License: https://creativecommons.org/licenses/by/4.0/

*Correspondence: Shalaka Ramgir-Naidu

Article metrics

Views
0
Downloads
0
Citations
0
Shares
0

Related articles

PerspectiveThe hidden value of uncertainty in medical segmentationSophie Fischer, Irina VoiculescuHypothesis and TheorySingle-image performance alone does not establish longitudinal technical performance: a paired-error framework for artificial intelligence-enabled quantitative imagingHunter Colson, Manuel ValdebranCorrectionCorrection: Aluminum impairs cellular ultrastructure and bone microarchitecture in newborn ratsMara Rubia Marques, Anderson Camargo Moreira, Iara Frangiotti Mantovani, Pedro Vale de Azevedo Brito, Isabela Cristina Gomes de Souza Nascimento, Celso Peres Fernandes, Fernanda Cristina Alcantara dos Santos