Research map / Image generation / Fast diffusion sampling
Anatomical Consistency and Adaptive Prior-informed Transformation for Multi-contrast MR Image Synthesis via Diffusion Model
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 APT, a conditional diffusion model for multi-contrast MR image synthesis that conditions on BiomedCLIP-based anatomical features fused across contrasts, initializes from a Bézier-transformation prior codebook, and refines with iterative unrolling.
Problem
Synthesizing missing multi-contrast MR images with diffusion models while maintaining anatomical consistency and effectively integrating the complementary information across contrasts.
Approach
Combines a mutual information fusion module with an anatomical consistency contrastive loss on frozen BiomedCLIP features to condition a diffusion model via cross-attention, and uses a two-stage inference: a prior codebook of Bézier-transformed, similarity-matched images provides a structured starting point at an intermediate diffusion step, followed by weighted-average iterative unrolling near th
Key results
- On BraTS 2021, APT reaches 25.52 dB PSNR / 0.920 SSIM for FLAIR synthesis and 27.44 dB / 0.951 for T2-w, outperforming six SOTA baselines in most contrasts.
- Ablation on ADNI shows adaptive prior selection improves T1-w PSNR over random prior selection (e.g., 24.81 vs 20.89 dB on BraTS T2-w context), and the full U+A+I variant gives the best ADNI T1-w result of 20.82 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the APT anatomical alignment and adaptive prior-informed two-stage synthesis framework | APT · BraTS 2021 multi-contrast MR synthesis | PSNR: 24.36 -> 25.52; 25.13 -> 25.88; 24.49 -> 24.54; 26.10 -> 27.44 | Helped |
| add the APT anatomical alignment and adaptive prior-informed two-stage synthesis framework | APT · ADNI multi-contrast MR synthesis | PSNR: 20.40 -> 20.82; 23.97 -> 24.56; 24.75 -> 24.48 | Mixed |
| add adaptive prior information to the unified synthesis model (ablation) | Conditional diffusion model (U) · BraTS 2021 T1Gd | PSNR: 22.02 -> 24.81 | Helped |
| add adaptive prior information to the unified synthesis model (ablation) | Conditional diffusion model (U) · ADNI T1-w | PSNR: 19.42 -> 20.45 | Helped |
| add Stage 2 iterative unrolling with weighted averaging (ablation) | U+A synthesis · ADNI T2-w | PSNR: 23.71 -> 24.56 | Helped |
| replace zero prior with adaptive prior selection (ablation) | APT prior-guided synthesis · BraTS 2021 T1Gd | PSNR: 20.74 -> 24.81 | Helped |
| replace Gaussian noise prior with adaptive prior selection (ablation) | APT prior-guided synthesis · BraTS 2021 T1Gd | PSNR: 21.06 -> 24.81 | Helped |
| replace random prior selection with adaptive prior selection (ablation) | APT prior-guided synthesis · BraTS 2021 T1Gd | PSNR: 20.89 -> 24.81 | Helped |
| increase Stage 2 unrolling iterations from 10 to 50 (ablation) | APT · BraTS 2021 T1Gd | PSNR: 24.83 -> 24.54 | Hurt |
| increase Stage 2 unrolling iterations from 25 to 50 (ablation) | APT · BraTS 2021 T1Gd | PSNR: 24.71 -> 24.54 | Hurt |
| increase Stage 2 unrolling iterations from 50 to 100 (ablation) | APT · BraTS 2021 T1Gd | PSNR: 24.54 -> 23.75 | Hurt |
| increase Stage 2 unrolling iterations from 50 to 200 (ablation) | APT · BraTS 2021 T1Gd | PSNR: 24.54 -> 23.60 | Hurt |
Surprising findings
- Starting the diffusion from noise (M=1000) is worse than starting from intermediate steps with the prior, and very early start steps (M=20-40) degrade performance sharply.
- Performance drops for ADNI when only PD is available as input, attributed to low gray/white matter contrast in PD images.
Limitations
- Evaluated only on 2D slices from two public brain MRI datasets
- Performance degrades when only PD is available in ADNI
- Iterative unrolling beyond N=100 yields diminishing or negative returns
Experimental setup
Datasets: BraTS 2021 ADNI
Benchmarks: BraTS 2021 ADNI
Metrics: PSNR SSIM
Baselines: MM-GAN HiNET ResViT ADM SynDiff M2DN
Builds on: Latent diffusion models (conditional diffusion model) BiomedCLIP DDPM Bézier curve transformation
Closest papers by meaning
- Staged and Physics-Grounded Learning Framework with Hyperintensity Prior for Pre-Contrast MRI Synthesis (ICML 2025 · similarity 0.86)
- Rethinking Diffusion Model for Multi-Contrast MRI Super-Resolution (CVPR 2024 · similarity 0.86)
- Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction (AAAI 2024 · similarity 0.86)
- Self-Supervised Diffusion MRI Denoising via Iterative and Stable Refinement (ICLR 2025 · similarity 0.85)
- AnaDiffusion: Anatomically Compositional Latent Diffusion for Controllable 3D Brain MRI Generation (NeurIPS 2026 · similarity 0.85)
- Autoregressive Image Diffusion: Generation of Image Sequence and Application in MRI (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.