Anatomically-conditioned Latent Diffusion Model for Data-Efficient Few-Shot Cross-Domain 3D Glioma MRI Synthesis

Researchers propose ALDM, an anatomically-conditioned latent diffusion model that synthesizes 3D brain MRI scans from limited data to improve glioma classification across medical imaging centers. The framework achieves superior synthetic image quality and clinical classification performance with only 16 target images, addressing a critical challenge in medical AI where domain shifts and data scarcity limit model generalization.

Medical imaging AI faces a fundamental constraint: high-quality annotated datasets remain scarce, especially for rare conditions like gliomas, and models trained on one hospital's data often fail when deployed elsewhere due to domain shifts in imaging protocols and equipment. ALDM tackles this challenge through a sophisticated two-stage architecture that separates anatomical learning from synthesis. The 3D variational autoencoder captures structural priors from abundant source-domain data, while the conditional latent diffusion model generates new volumes guided by tumor masks, enabling the system to learn from just 16 target examples—an extreme few-shot scenario that mirrors real clinical constraints.

The technical innovation lies in leveraging ControlNet conditioning to maintain anatomically plausible structures while generating diversity in imaging characteristics. This approach addresses a core pain point in medical AI deployment: clinicians cannot wait for thousands of annotated scans before deploying diagnostic systems. By achieving 0.987 AUC on downstream classification tasks while producing visually realistic synthetic data, ALDM demonstrates that generative models can meaningfully augment clinical datasets without requiring massive labeled collections.

The broader implications extend beyond glioma detection. Healthcare institutions operating in low-resource settings or dealing with rare diseases can now adopt similar frameworks to bridge domain gaps and reduce annotation burden. This work validates that diffusion-based synthesis outperforms GANs in medical contexts, likely because diffusion models better preserve fine diagnostic details critical for pathology detection. Success here could accelerate adoption of synthetic data augmentation across oncology, radiology, and other imaging-dependent specialties where domain generalization remains unsolved.

• →ALDM achieves 0.987 AUC classification performance using only 16 target MRI images, demonstrating extreme few-shot learning viability in medical imaging.
• →Two-stage architecture combining VAE anatomical priors with ControlNet-guided diffusion outperforms GAN and hybrid baselines on synthetic data quality metrics.
• →Framework addresses critical clinical need for domain-shift adaptation across imaging centers without requiring expensive re-annotation of datasets.
• →Synthetic data generation preserves pathology boundaries and cross-modal consistency, enabling practical clinical deployment in data-scarce settings.
• →Open-source implementation enables broader adoption across healthcare institutions seeking to augment limited imaging datasets.