AI is rapidly reshaping breast cancer screening, with recent large-scale trials -- including the MASAI trial and ScreenTrustCAD -- showing that AI-assisted reading performs comparably to traditional double reading by radiologists. However, progress is constrained by three fundamental data challenges: scarcity of annotated medical images, patient privacy regulations that limit data sharing, and uneven representation of certain patient groups in available datasets.
Breast imaging has specific gaps that make these challenges acute. There are very few mammograms from young women who undergo non-routine imaging, almost no longitudinal imaging sequences that track the same patient over time, and limited collections of paired multimodal data (matched mammography, ultrasound, and MRI from the same individual). These gaps make it difficult to train AI systems that generalize well across the full range of patients seen in clinical practice.
This review, covering literature from 2020 through April 2025, examines how synthetic data and generative AI -- principally GANs and diffusion models -- can address these barriers by generating realistic artificial mammograms. Three types of generated data are distinguished: synthetic data (not from any real patient), augmented data (expanded datasets combining real and transformed images), and generative data (images produced by explicit AI models such as GANs or diffusion models). Most recent breast imaging research involves generative data used as part of augmented training sets.
Generative Adversarial Networks (GANs) are the most established generative approach in breast imaging. A generator network creates synthetic images while a discriminator network tries to distinguish them from real ones; through this competition, the generator learns to produce highly realistic outputs. GANs have been applied to density transformation, lesion synthesis, dataset augmentation, and contralateral breast modeling. Their main limitations are a risk of bias amplification, mode collapse (where the generator produces limited variety), and high computational cost.
Diffusion models work differently: they iteratively transform random noise into a structured image, guided by conditioning information such as text descriptions or target density levels. These models offer high fidelity and precise control over generated image attributes -- for example, producing a mammogram with a specified density category and lesion morphology at a defined location. They are still in early adoption for medical imaging and require large training datasets and significant computing resources.
Transformer-based generative models use self-attention mechanisms to capture long-range relationships across an image, making them suited for high-resolution synthesis where global structure matters. They are the newest of the three approaches in breast imaging and currently have limited published studies, but their scalability suggests growing relevance as the field matures. All three approaches have distinct validation needs: reader studies, detection benchmarking, and equity assessments across breast density and demographics.
The most direct application of synthetic data is dataset augmentation: adding AI-generated mammograms to training sets to expose models to more variation and reduce overfitting. One study used a U-Net architecture with self-attention to insert or remove lesions in mammogram patches, and showed that training with these synthetic images raised classifier AUC from 0.829 to 0.846. Another used CycleGAN to progressively convert extremely dense mammograms to lower density equivalents, revealing cancers hidden by overlapping tissue. StyleGAN3 generated 10,000 synthetic ultrasound images for the LightweightUNet model, raising classification accuracy from 87% to 96%.
A second application is class balancing: using synthetic images to correct imbalances between rare and common cases, or between underrepresented and well-represented patient groups. GANs conditioned on the contralateral breast have been used to generate synthetic normal mammograms that help models identify occult cancers, raising AUC from 0.70 to 0.77. Conditional diffusion models, such as the MAM-E framework, can produce mammograms with specific density levels and lesion characteristics for targeted augmentation of rare phenotypes.
Synthetic data also supports radiology education and simulation-based training. A randomized study showed that residents who practiced with AI-generated mammogram cases achieved measurably improved cancer detection sensitivity. One system allowed residents to insert simulated lesions into synthetic mammograms in a game-like environment with immediate feedback. High-resolution CycleGAN-generated images have been used to simulate detecting lesions in both fatty and dense breast tissue, allowing personalized training that targets each trainee's specific weaknesses in a risk-free setting.
A major concern in AI-assisted mammography is representation bias: if training data is dominated by images from Caucasian women in high-income countries, the resulting AI may perform significantly worse for patients from other ethnic backgrounds, younger age groups, or underserved communities. This creates a compounding disadvantage where populations already underserved by healthcare infrastructure are also underserved by the AI tools meant to help them. Researchers analyzing global mammography AI datasets have confirmed geographic and demographic imbalances that raise serious concerns about equitable deployment.
Generative AI offers a concrete strategy for closing these gaps. By training conditional GANs to produce synthetic images representing minority subgroups, researchers can augment training data with examples that would otherwise be extremely scarce. One study showed that augmenting training data with synthetic health records for minority patients improved model fairness and predictive performance for Black patients and female patients. In the imaging domain, the same approach could be applied to generate synthetic mammograms for women under 40, patients with rare tumor subtypes, or underrepresented imaging equipment types.
However, synthetic data is not a reliable fix on its own. As one analysis noted, generative models are only as good as the data they are trained on -- biases present in real training data will be reflected and can be amplified in generated outputs. The field is developing fairness metrics for synthetic data and debiasing pipelines that verify generated datasets truly enhance representation rather than replicating prejudicial patterns. Clinical deployment of AI tools should include subgroup performance analyses across age, breast density, and demographics to identify bias before any system is used at scale.
Generative AI in breast imaging raises unresolved medico-legal questions about liability. When an AI tool misses a cancer or triggers an unnecessary biopsy, the current regulatory position in the United States (as stated by the Federation of State Medical Boards in 2024) holds the interpreting radiologist -- not the software developer -- accountable. The reasoning is that AI tools are assistive, with the radiologist retaining final interpretive responsibility. However, as AI systems move toward greater autonomy, this attribution of liability is increasingly contested.
An emerging alternative is a shared accountability model, where responsibility is distributed based on the level of AI autonomy, the transparency of its recommendations, and how closely the clinician adhered to standard of care. Some legal scholars have proposed requiring AI algorithms to carry malpractice insurance or legal personhood in highly autonomous contexts, treating a malfunctioning AI like a faulty medical device rather than a tool misused by a clinician. The field does not yet have established case law or regulatory frameworks to resolve this debate.
A separate and serious concern is deepfake medical imaging: the use of GANs to inject or remove lesions in clinical mammograms, creating convincing falsified images that could mislead diagnosis. Researchers have demonstrated this is technically feasible, raising scenarios such as fabricated cancers for insurance fraud or removed tumors to conceal negligence. Proposed countermeasures include cryptographic image signatures and blockchain-based provenance tracking to verify that images have not been tampered with. A related privacy risk is model memorization, where a generative model reproduces outputs too similar to specific training patients, enabling re-identification. Mitigation requires explicit de-duplication checks and, in sensitive settings, differential-privacy training techniques.
Despite their demonstrated benefits, synthetic images face a fundamental validation problem: there are no standardized protocols for verifying that synthetic mammograms are clinically equivalent to real ones. The field currently relies on three informal criteria -- reader studies where radiologists cannot distinguish synthetic from real images, statistical analyses comparing pixel and texture distributions, and evidence that synthetic augmentation improves task performance without introducing visible artifacts. None of these alone is sufficient, and the combination is not formally required by any regulatory body.
A structural weakness in many studies is the use of enriched datasets, which contain a much higher proportion of cancer-positive cases than would appear in a real screening population. This design improves statistical power but produces accuracy estimates that cannot be extrapolated to routine clinical use. Current generative models also perform unevenly across lesion types: subtle microcalcification clusters, ductal carcinoma in situ, architectural distortions, and rare histological subtypes are underrepresented in most training corpora, meaning synthetic images for these categories may be unrealistic or systematically biased.
From a regulatory standpoint, neither the FDA nor the EMA has yet established specific guidelines for the validation, documentation, or use of synthetic data in medical AI. The FDA's 2021 action plan for AI in medical devices does not address synthetic data. The 2024 ESGAR consensus statement and AUR tutorial have begun to fill this gap by recommending task-specific validation that includes radiologist reader studies and algorithmic benchmarking across representative populations, but formal regulatory requirements remain absent. This creates a significant barrier to clinical adoption, as institutions lack a clear compliance pathway.
For synthetic data to be adopted responsibly in clinical breast imaging, several preconditions must be met simultaneously. Technically, generative models need standardized task-specific validation protocols, explicit safeguards against model memorization of training patients, and performance evaluation stratified by lesion type and patient subgroup. The technology's current gap -- unrealistic synthesis of subtle microcalcifications, rare subtypes, and non-mass enhancement -- must be directly addressed before these lesion types are used in training data.
Institutionally, a structured deployment approach is recommended. Governance and privacy committees should review how generative models were trained and how source data were handled. Statistical comparisons should verify that synthetic images reflect the local patient population. Models should be tested on independent real-world datasets with and without synthetic augmentation, with particular scrutiny of performance changes in vulnerable subgroups. This internal validation process should precede any synthetic data being introduced into machine-learning pipelines for clinical use.
The broader pathway forward requires multidisciplinary collaboration among radiologists, AI developers, legal experts, policymakers, and patient representatives. The review frames this as a question not just of technical performance but of ethics, equity, and trust: generative AI can be a force multiplier for health equity if used thoughtfully, but misapplication risks baking in existing disparities or eroding trust through deepfake threats. The authors advocate for retaining human judgment within AI-assisted workflows -- ensuring that AI serves as a support for clinical responsibility rather than a substitute for it.