Breast cancer remains the single most commonly diagnosed cancer worldwide, with over 2.3 million new cases each year. Traditional screening relies heavily on mammography, which, while widely used, has limited sensitivity in women with dense breast tissue and frequently produces both false negatives (missed cancers) and false positives (unnecessary biopsies). MRI offers higher sensitivity at 94.6% compared to mammography's 54.5%, but that gain comes at the cost of reduced specificity, elevated expense, and limited availability. Ultrasound fills a complementary role, especially for dense breasts, but it is operator-dependent and carries modestly lower specificity when used for screening.
The multimodal imaging landscape: Modern breast imaging follows a stepwise, multimodal approach. Mammography or digital breast tomosynthesis (DBT) serves as the first-line screening tool, with ultrasound or MRI added based on individual risk factors or equivocal findings. DBT has been increasingly adopted because it decomposes the breast into thin slices, improving lesion conspicuity, lowering recall rates, and reducing interval cancers. Nuclear medicine techniques such as FDG PET/CT and molecular breast imaging are reserved for staging, therapy response assessment, and selected preoperative decisions rather than routine screening.
Biological complexity: Breast cancer is not a single disease. It encompasses hormone receptor-positive/HER2-negative, HER2-positive, and triple-negative subtypes, each with distinct growth patterns, metastatic behavior, and therapeutic responsiveness. Prognosis depends on grade, stage, and molecular biomarkers from histopathology and genomics. Risk itself is shaped by age, family history, germline variants (BRCA1/2), breast density, reproductive factors, and modifiable exposures such as obesity and alcohol use. The global burden is unevenly distributed, with later-stage diagnosis and limited access to quality imaging driving excess mortality in many settings.
Enter AI and deep learning: Recent advances in deep learning, particularly convolutional neural networks (CNNs), Vision Transformers (ViTs), and generative adversarial networks (GANs), have introduced transformative potential. AI-driven mammography has demonstrated superior diagnostic accuracy in dense breast tissue compared to traditional radiological assessment. Models trained on extensive datasets show reduced false-positive and false-negative rates. Beyond imaging, AI supports predictive modeling and personalized diagnosis by integrating genetic, histological, and clinical information. However, the authors stress that widespread clinical adoption still faces significant barriers: limited external validation, domain shifts across institutions, variable reporting standards, limited interpretability, and ethical and regulatory constraints.
CNNs and their evolution: Convolutional Neural Networks have fundamentally transformed medical image analysis. Key architectures such as AlexNet, VGGNet, InceptionNet, DenseNet, and ResNet have each addressed specific challenges. AlexNet and VGGNet laid the groundwork for deep feature extraction. InceptionNet introduced multi-scale feature extraction, allowing the network to capture patterns at different resolutions simultaneously. DenseNet introduced dense layer connections that promote efficient gradient flow and feature reuse, which proves especially valuable for detecting abnormalities in dense breast tissue. ResNet, through its skip connections, mitigates the vanishing gradient problem and enables training of deeper networks for high-dimensional datasets like DBT.
Transfer learning advantage: A critical enabler for medical AI is transfer learning, where models pre-trained on large general datasets (such as ImageNet) are fine-tuned for specific clinical tasks like breast lesion detection, risk stratification, and image classification. This approach significantly alleviates the chronic challenge of limited annotated medical datasets, which are both costly and time-consuming to curate. Despite this, CNNs still face challenges including architecture optimization for specific imaging tasks, class imbalance in medical datasets, and limited interpretability for clinical adoption.
Vision Transformers (ViTs) - a paradigm shift: ViTs represent a fundamentally different approach to image analysis. Instead of using convolutional filters, they divide images into patches and treat them as sequences, applying self-attention mechanisms to capture both local and global contextual information simultaneously. This is particularly powerful for breast cancer imaging, where tumors often exhibit complex morphological and spatial relationships spanning multiple regions. Hybrid models combining CNNs for local feature extraction with ViTs for long-range dependencies have demonstrated superior performance in challenging cases such as dense breast tissue and multifocal tumors.
ViT performance highlights: The review presents a comprehensive table (Table 1) of ViT-based studies from 2023 to 2024. Standout results include: a fine-tuned ViT achieving 99.99% accuracy on the BreakHis histopathology dataset; a Compact Convolutional Transformer (CCT) reaching 99.92% accuracy in mammography classification; a wavelet-based ViT attaining an AUC of 0.984 in dense breast ultrasound classification; and a ViT with majority voting achieving 98.2% accuracy in MRI classification. Self-supervised learning further enhances ViTs by allowing pre-training on vast unlabeled medical image datasets, which is a critical advantage given the scarcity of labeled data in breast cancer diagnostics.
GANs solving data scarcity: Generative Adversarial Networks have emerged as powerful solutions for two persistent problems in breast cancer AI research: data scarcity and class imbalance. By generating synthetic mammograms and histopathological images, GANs enrich training datasets so that models generalize better across diverse patient populations. This is particularly critical for rare breast cancer subtypes where real-world data is often insufficient. GANs have demonstrated efficacy across multiple applications: super-resolution imaging that enhances low-quality images to improve visibility of diagnostic markers; cross-modality synthesis (converting between MRI and CT); image reconstruction and denoising; and radiation dose reduction.
Quantified GAN impact: The review catalogues specific GAN contributions in Table 2. DCGAN (Deep Convolutional GAN) synthesizes mammograms that enhance dataset diversity and clustering consistency. Conditional GANs (cGANs) improve classification by directly addressing class imbalance, producing a 3.9% improvement in classification accuracy. Pix2Pix GANs augment mammography data and improve VGG-16 and ResNet-50 accuracy. Synthetic ultrasound images generated by GANs boosted CNN classification accuracy by 15.3%. TVAE and CTGAN achieved a combined diagnostic accuracy of 96.66%. WA-SRGAN enhanced histopathology images to achieve an AUC of 0.826 for tumor grading.
Multi-task learning and explainability: The integration of advanced architectures has produced significant improvements through multi-task learning (MTL), where models perform classification, segmentation, and tumor grading concurrently. This integrated approach reduces the need for separate analyses and provides a holistic assessment of breast abnormalities. Explainable AI (XAI) techniques, including LIME, SHAP, and Grad-CAM, are becoming essential for addressing the "black box" nature of deep learning models and fostering trust among clinicians. These methods provide visual explanations for model predictions, enhancing transparency in AI-assisted diagnosis.
Federated learning and personalization: Federated learning is emerging as a solution to data-sharing barriers while maintaining patient privacy. Multiple institutions can collaboratively train models by sharing only model updates rather than raw data. Meanwhile, research is increasingly focusing on personalized AI models that incorporate patient-specific factors such as genetic predisposition, hormonal status, and breast density. Self-supervised learning is expected to further reduce reliance on large annotated datasets, paving the way for scalable AI solutions in breast cancer diagnostics.
Mammography and DBT: AI algorithms have surpassed radiologists in sensitivity and area under the curve (AUC) values for mammography interpretation. In digital breast tomosynthesis, AI reduced recall rates by 2 to 27% and reading times by up to 53%. AI-enhanced mammography addresses the core limitation of standard mammography in dense tissue by identifying subtle markers such as microcalcifications and architectural distortions with greater sensitivity. In DBT specifically, AI decomposes the 3D breast structure into thin slices, significantly improving the visibility of lesions hidden in traditional 2D mammography. Incorporating AI into DBT has yielded sensitivity improvements of up to 15% while simultaneously reducing recall and false-positive rates.
Ultrasound and elastography: Deep learning models such as U-Net and fully convolutional networks (FCNs) are automating breast lesion segmentation, accurately delineating lesion boundaries, assessing internal characteristics, and classifying malignancy likelihood. AI-based shear-wave and strain elastography techniques have demonstrated sensitivity of 80% or higher in distinguishing benign from malignant lesions, potentially reducing unnecessary biopsies. A novel mobile AI solution for real-time breast ultrasound analysis achieved 100% sensitivity for malignancy detection with an AUC of 0.835 to 0.850. These tools are especially valuable for women with dense breasts, where mammography alone is insufficient.
MRI and DCE-MRI: AI-enhanced radiomics applied to dynamic contrast-enhanced MRI (DCE-MRI) has shown promise in predicting molecular subtypes of invasive ductal breast cancer with high accuracy, particularly in distinguishing triple-negative and HER2-overexpressed subtypes. Radiomics integrates quantitative imaging features with clinical, histopathological, and genomic information to enable more personalized patient management. Deep learning models applied to breast MRI have achieved AUC values up to 0.98 for tasks including cancer diagnosis, molecular classification, chemotherapy response prediction, and lymph node involvement prediction.
Histopathology and digital pathology: AI-powered systems analyze whole-slide images to identify invasive tumors, detect lymph node metastases, and evaluate hormonal status with high precision. CNNs achieve expert-level performance in processing complex visual data from tissue specimens. AI tools can predict genetic alterations, identify prognostic biomarkers, and assess tumor-infiltrating lymphocytes. Cross-staining inference has achieved 89.6% accuracy on H&E-stained datasets and 80.5% on fluorescent-stained datasets, demonstrating potential for expanding AI across different staining technologies. These advances contribute to improved patient stratification and reduced delays between diagnosis and prognosis determination.
Comprehensive risk profiling: AI-based tools now integrate multi-omics data (genomics, transcriptomics, epigenomics, proteomics), imaging biomarkers, and clinical information to develop comprehensive patient risk profiles. These profiles enable personalized screening strategies and treatment planning. AI algorithms have demonstrated promising results in automating diagnosis, segmenting relevant data, and predicting tumor responses to neoadjuvant chemotherapy. Machine learning and deep learning methods show average accuracy rates of 90 to 96% for predicting treatment response, prognosis, and patient survival.
Real-time patient monitoring: AI has enhanced the accuracy of mammographic imaging, tomosynthesis, and other diagnostic modalities while simultaneously improving pathology workflows for identifying invasive tumors, lymph node metastasis, and hormonal status evaluation. A meta-analysis of AI algorithms for tumor metastasis detection reported a pooled sensitivity of 82% and specificity of 84%, with an AUC of 0.90. In abdominopelvic malignancies, AI-based radiomics models outperformed radiologists in lymph node metastasis detection, achieving AUCs of 0.895 for radiomics and 0.912 for deep learning compared to 0.774 for radiologists alone.
Wearable devices and continuous surveillance: Wearable devices integrated with AI further enhance monitoring by continuously collecting data on vital signs, activity levels, and physiological metrics. These data streams are analyzed to predict potential complications before clinical symptoms manifest. This proactive approach fosters a healthcare environment where interventions can be deployed early, reducing the burden on healthcare systems by prioritizing high-risk patients for more intensive surveillance. Circulating tumor DNA (ctDNA) analysis is also emerging as a complementary tool to imaging, offering potential advantages in the early detection of recurrence or progression.
Federated learning for broader validation: Federated learning enables multiple institutions to collaboratively train algorithms while keeping sensitive data localized within each institution. By sharing only model updates instead of raw data, this approach mitigates privacy concerns and reduces breach risks. FL has been successfully applied in medical imaging, COVID-19 research, and smart healthcare systems. Despite challenges such as potential privacy leakage from adversarial attacks, emerging privacy-preserving techniques continue to improve. Overall, federated learning offers substantial potential for improving the generalizability and external validity of AI models across heterogeneous multi-center data.
Tailored therapy through AI: Adaptive AI treatment plans are redefining cancer care by integrating pharmacogenomic, proteomic, and tumor evolution data to identify the most effective interventions for each patient. AI algorithms, including support vector machines (SVM), random forests (RF), and neural networks, have demonstrated high accuracy in predicting treatment outcomes, drug interactions, and resistance patterns. These techniques analyze complex datasets by combining patient-specific information with drug characteristics to optimize therapeutic decisions. AI applications extend across the full drug discovery pipeline, from virtual screening and lead identification to ADMET (absorption, distribution, metabolism, excretion, and toxicity) analysis.
Managing drug resistance: Tumor heterogeneity and genetic mutations frequently lead to treatment resistance, complicating outcomes for patients. AI-driven models anticipate these changes by analyzing longitudinal genetic and proteomic data, allowing oncologists to modify treatment regimens proactively rather than reactively. Reinforcement learning algorithms can simulate various therapeutic scenarios, helping clinicians choose strategies that maximize efficacy while minimizing toxicity. Circulating tumor DNA (ctDNA) kinetics has emerged as a promising biomarker for monitoring treatment efficacy and guiding therapy adjustments in real time.
Dynamic, evolving treatment strategies: AI algorithms can analyze complex transcriptomic profiles to predict drug responses and identify optimal treatment combinations for individual patients. These AI-driven approaches enable the development of dynamic treatment strategies that adapt to the evolving nature of cancer progression. In lung cancer, for example, AI models integrate molecular information, radiomics, and patient characteristics to optimize both diagnosis and treatment. The same principles are being applied to breast cancer, where disease biology is heterogeneous and treatment must evolve as the tumor changes over time.
Challenges in equitable deployment: Despite these advances, the review underscores persistent challenges in ensuring data quality, mitigating algorithmic biases, and addressing disparities in access to AI-driven care. AI models trained on data from well-resourced institutions may not generalize to populations with limited healthcare infrastructure. Collaborative efforts among policymakers, healthcare providers, and technology developers are critical for overcoming these hurdles and ensuring that the benefits of precision oncology reach all patient populations equitably.
Algorithmic bias and transparency: The use of AI in breast cancer care raises significant ethical, legal, and social implications. The "black box" nature of AI systems creates concerns about transparency and accountability, particularly in clinical decision-making where a wrong recommendation can endanger lives. Algorithmic bias is a major concern: AI systems trained on data that underrepresents certain demographic groups may generate biased recommendations that fail to adequately address the needs of those populations. This creates ethical dilemmas around maintaining equitable and personalized care for diverse patient populations. Addressing these challenges requires broad stakeholder engagement, robust oversight systems, and proactive roles for regulators.
Regulatory gaps: The rapid advancement of AI technology has outpaced legal and regulatory development, creating a significant governance gap. Key issues include data privacy, security, accountability, and liability. There is a pressing need for tailored regulatory frameworks that address AI-specific challenges while still fostering innovation. The lack of clear regulations raises unresolved questions about responsibility in cases of AI-related errors or harm. Data protection, informed consent, and algorithmic transparency must be addressed through comprehensive governance frameworks that balance innovation with patient protection.
Patient autonomy and informed consent: Studies show that patients, physicians, and other healthcare professionals have varying levels of trust in AI-driven healthcare. The use of unexplainable machine learning algorithms may undermine the trusted doctor-patient relationship and patient autonomy. Patients may struggle to understand AI-driven decision-making, potentially compromising their ability to provide truly informed consent. AI tools may also generate recommendations that patients feel pressured to follow because they perceive them as "scientifically" grounded. Ensuring that consent remains informed and voluntary in an AI-driven environment is a key challenge that must be addressed.
Frameworks for responsible AI: Explainable AI (XAI) aims to bridge the gap between complex models and patient understanding by providing clear explanations of how decisions are made. The informed consent process must evolve to accommodate AI technology, potentially incorporating simplified explanations, visual aids, and interactive tools. Collaborative decision-making is emphasized: AI should not replace human judgment or patient involvement but rather act as a supportive tool. Healthcare professionals should engage patients in open discussions about AI use in their care, ensuring that values, preferences, and goals remain central to treatment decisions.
Current state of evidence: The review concludes that AI, particularly deep learning, has meaningfully advanced breast cancer diagnostics across mammography/DBT, ultrasound, MRI, and digital pathology, especially when embedded in clinician-centered workflows. Evidence is most mature in screening mammography, where AI has demonstrated clear improvements in sensitivity, specificity, reading time reduction, and recall rate reduction. Applications in other modalities and in multimodal fusion (combining imaging with clinicopathologic and genomic data) are promising but remain heterogeneous in quality and validation rigor.
Persistent performance risks: Real-world AI performance remains sensitive to dataset quality, population and vendor differences, acquisition protocols, probability calibration, and workflow design. Diminished performance on external datasets and miscalibration are recurrent risks that require explicit mitigation during both development and deployment. The authors emphasize that domain shifts across institutions and variability in reporting standards continue to undermine the generalizability of many published models. These risks are not theoretical; they represent practical barriers to safe clinical deployment.
Near-term priorities: Responsible clinical deployment requires multi-site external validation with representative datasets, bias mitigation strategies, interpretable outputs that support rather than replace clinical decision-making, and post-deployment monitoring for drift, safety, and equity. The authors identify second-reader and triage use cases as the most immediately actionable near-term applications, alongside the integration of routinely available clinical and pathological data to refine individualized risk assessment. Robust training protocols, cross-site harmonization, and predefined escalation pathways for uncertain cases are also priorities.
The augmentation principle: The review's central conclusion is that AI should augment, not replace, clinicians. When deployed within well-validated and transparently governed systems, AI can enable earlier detection, more consistent diagnosis, and personalized care. However, the authors caution that achieving this vision requires sustained collaborative effort among AI researchers, clinicians, regulators, ethicists, and patients. Compliance with privacy regulations, data governance standards, and evolving regulatory requirements must be built into every stage of the AI lifecycle, from initial model development through long-term clinical monitoring.