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

Deep learning models for glioma classification often struggle with domain shifts and the lack of large, annotated datasets. The authors address this by developing a framework for 3D volumetric MRI synthesis that can generate realistic, anatomically consistent images using only a few target-domain examples (few-shot learning).

The proposed Anatomically-conditioned Latent Diffusion Model (ALDM) operates in two stages. First, a 3D variational autoencoder (VAE) is trained on a data-rich source domain to learn a compact, anatomically meaningful latent representation of brain MRI volumes. Second, a conditional latent diffusion model (LDM) is trained to generate target-domain images. This diffusion process is guided by tumor segmentation masks via a ControlNet and FiLM-style modulation, ensuring that the synthesized volumes maintain structural integrity and spatial alignment with the provided anatomical priors. The model uses classifier-free guidance and spatially weighted loss functions to balance anatomical accuracy with image diversity.

ALDM significantly outperformed existing GAN and hybrid baselines in extreme few-shot scenarios (using only 16 target images). It achieved a superior Fréchet Inception Distance (FID) of 85.40 and a downstream classification AUC of 0.987. Qualitative analysis demonstrates that the model successfully preserves sharp pathology boundaries and cross-modal consistency, effectively transferring structural knowledge from the data-rich source domain to the data-scarce target domain.

This research provides a robust tool for clinical data augmentation in low-resource settings. By synthesizing high-quality, anatomically plausible 3D MRI volumes, ALDM helps mitigate the impact of dataset scarcity and domain heterogeneity, potentially improving the performance of downstream diagnostic models without requiring extensive additional manual labeling or data collection.