Research map / Image generation / Autoregressive and latent generation

AnaDiffusion: Anatomically Compositional Latent Diffusion for Controllable 3D Brain MRI Generation

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AnaDiffusion, a part-to-whole latent diffusion framework that generates left/right hemisphere and cerebellar-brainstem parts, assembles them into a scaffold, and injects it into a whole-brain denoiser for controllable 3D brain MRI generation.

Problem

Monolithic 3D brain MRI diffusion models often fail at local anatomical fidelity and region-level controllability, and mask-conditioned methods need dense subject-specific segmentation inputs.

Approach

Factorize the brain into part-specific latent diffusion models (shared canonicalized hemisphere model plus a cerebellar-brainstem model), assemble their outputs in image space, re-encode into whole-brain latent space, and inject the scaffold at an intermediate noise level so a whole-brain denoiser refines it into a coherent volume.

Key results

• Achieves lowest whole-brain MedicalNet FID of 36.16 [19.16, 59.61] (x10^4) versus 40.92 for standard LDM and 134.40 for VAE-GAN.
• Lowest cerebellum SynthSeg Cohen's |d| of 0.169 [0.007, 0.482] among compared methods.
• Part editing target transfer MS-SSIM of 0.9427 vs 0.7952 for Segm. cLDM on left hemisphere replacement.

Surprising findings

• Separate hemisphere models give lower left-hemisphere, CB, and seam FID than the shared canonicalized hemisphere model, though the whole-brain and right-hemisphere FID favor sharing; CIs overlap.
• Off-target preservation is slightly lower for AnaDiffusion than for the segmentation-conditioned LDM (0.9363-0.9496 vs 0.9447-0.9512).

Limitations

• Multi-stage design adds training and inference complexity compared with a monolithic LDM.
• Fixed anatomy-driven factorization does not solve tissue-level calibration and is not validated for severe mass effect or displaced anatomy.
• Assumes approximate MNI152 correspondence; fixed decomposition may miss other meaningful organizations.

Experimental setup

Datasets: ADNI

Benchmarks: ADNI subject-disjoint test split

Metrics: MedicalNet FID seam FID SynthSeg Cohen's |d| MS-SSIM

Baselines: VAE-GAN HA-GAN LDM Segmentation cLDM MorphLDM ControlNet LDM Grid-based LDM

Builds on: latent diffusion models DDPM DDIM Med-DDPM MorphLDM

Closest papers by meaning

• Anatomical Consistency and Adaptive Prior-informed Transformation for Multi-contrast MR Image Synthesis via Diffusion Model (CVPR 2025 · similarity 0.85)
• MicroDiffusion: Implicit Representation-Guided Diffusion for 3D Reconstruction from Limited 2D Microscopy Projections (CVPR 2024 · similarity 0.83)
• PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.83)
• AMD: Anatomical Motion Diffusion with Interpretable Motion Decomposition and Fusion (AAAI 2024 · similarity 0.83)
• Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis (CVPR 2025 · similarity 0.83)
• Denoising Diffusion via Image-Based Rendering (ICLR 2024 · similarity 0.83)