Generative adversarial networks for medical and dental education

Generative adversarial networks (GANs) were first described by Ian Goodfellow and colleagues in 2014, they consist of two competing neural networks; a generator, which produces synthetic data, and a discriminator, which attempts to distinguish synthetic from real data. The Two networks are trained together in a zero-sum, game-theoretic setup, with the generator progressively improving its output as it tries to "fool" the discriminator.

Background

Generative models, including generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion models play an important role in the field of medical imaging. GANs generate new, synthetic samples that mimic the statistical properties of a real dataset without directly copying it. In fields where access to real data is limited by privacy law, cost, or scarcity, this generative capability has made GANs attractive for producing training and teaching material that behaves like authentic data without exposing identifiable or sensitive records.

Case- and scenario-based teaching in the health professions

Developing robust clinical judgement and decision-making through using case- scenario- or problem-based teaching/learning is GeneRally considered The Primary aim of medical/dental training. A distinct and growing application is the use of GANs to generate synthetic clinical images and case material for teaching, particularly where patient consent and privacy rules restrict the use of real cases. An example is the usage in plastic surgery where GANs have been used for facial recognition, burn estimation, scar prediction, and post-breast cancer reconstruction anomaly scoring.. GANs can be used for creation of synthetic ECGs to augment trainee ECG interpretation.

A 2026 scoping review examined GANs application in postgraduate orthodontic education. The review grouped GANs usage into five categories of GAN-generated orthodontic imagery:

• Facial de-identification of frontal or profile photographs
• Panoramic radiograph (OPG) synthesis
• Lateral cephalogram synthesis
• Cone-beam computed tomography (CBCT) synthesis
• Intraoral photograph synthesis

Across these studies, GAN-generated images were frequently indistinguishable from real images even to trained clinicians, particularly at moderate resolutions; fine anatomical detail — individual tooth morphology, cortical bone texture, and soft tissue — remained the most reliable cue for expert detection at very high resolution. Synthetic images also improved the accuracy of downstream diagnostic AI models when used to balance imbalanced training datasets, and AI-enhanced low-dose CBCT preserved clinical decision-making while reducing radiation exposure.

The review concluded that GANs hold promise for advancing case- and scenario-based orthodontic education by offering a way to build teaching libraries of realistic imagery without patient-consent constraints, but that current evidence derives largely from single-centre, single-population pilot studies with limited external validation, and that larger, multi-centre datasets and condition-specific models are needed before wider adoption.

Similar motivations — using GANs to create synthetic medical images or records that mimic real data's statistical properties while omitting identifiable information — appear more broadly in medical education and health data research, where GANs support cross-institutional data sharing, dataset augmentation, and model training under privacy regulations such as the HIPAA and the General Data Protection Regulation (GDPR).

Benefits

• Privacy preservation — synthetic cases can be used for teaching and assessment without exposing real patient or student data.
• Expanded case libraries — GANs can help address the shortage of rare or atypical cases available for case-based teaching by generating varied synthetic examples.
• Dataset augmentation for AI training — synthetic images generated for teaching purposes can also be reused to balance imbalanced datasets used to train diagnostic AI models.

Limitations and challenges

• Lack of standardisation across GAN studies, in terms of architectures, datasets, and evaluation metrics, complicates comparison between approaches.
• Limited external validation — much of the evidence to date comes from single-centre, single-population pilot studies.
• Detectability at high resolution — very high-resolution outputs and fine anatomical or structural detail remain the most reliable cues by which trained experts can distinguish synthetic from real images.
• Residual privacy risk — synthetic data can reduce but does not always eliminate the risk of re-identification, and may omit complex, disease-relevant detail present in real data.

Future directions

There is a need for larger, more diverse, multi-centre datasets; conditional GAN architectures tailored to specific conditions or presentations (for example, particular malocclusion types or facial forms in orthodontics); and standardised evaluation frameworks before GAN-generated material is adopted at scale in curricula and assessment.

See also

• Generative adversarial network
• Synthetic data
• Medical education
• Educational technology
• Artificial intelligence in education
• Orthodontics