Research map / Image generation / Autoregressive and latent generation

Staged and Physics-Grounded Learning Framework with Hyperintensity Prior for Pre-Contrast MRI Synthesis

ICML 2025 · Image generation › Autoregressive and latent generation · approach: Retrieval and matching with generative features

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 SPHERE, a two-stage physics-grounded framework that segments contrast uptake with a hyperintensity prior and then inpaints Pre-Contrast MRI from Post-Contrast scans, with an artifact-correction variant SPHERE-Art.

Problem

Pre-Contrast MRI is often missing or corrupted, yet it is needed as a reference for quantitative and comparative contrast-enhanced MRI analysis; direct Post-to-Pre image-to-image synthesis produces blurry or distorted contrast details.

Approach

Derive from the SPGR signal model that Pre-Contrast signal can be expressed via a learnable contrast uptake mask, then decompose synthesis into a segmentation stage (NestedUNet plus a soft-threshold hyperintensity prior) and an "inpainting+" stage that reconstructs the Pre-Contrast image using Post-Contrast and the mask.

Key results

• SPHERE reaches PSNR 36.82 dB and SSIM 0.9026 on Site A, versus 35.25 dB and 0.8735 for the next best (BICEPS); Site B PSNR 35.36 dB vs 34.50 dB.
• SPHERE-Art reaches PSNR 37.12 dB on artifact removal vs 33.49 dB for UNet.
• In a two-reader Likert study on 15 cases, Syn-Pre scored higher than SOC-Pre on perceived image quality (3.96 vs 3.64) and anatomical alignment (3.96 vs 3.75), but slightly lower on tissue visualization and diagnostic value.

Measured effects reported in the paper

Surprising findings

• Synthesized Pre-Contrast images sometimes exceed standard-of-care Pre-Contrast images in alignment and artifact levels, attributed to inter-acquisition misalignment.
• SPHERE is about 2x slower at inference than baselines (0.45 vs ~1.1 images/s).

Limitations

• Inference is roughly 2x slower than baselines due to three NestedUNet components.
• Uptake mask assumes uniform contrast concentration per region, which simplifies gadolinium pharmacokinetics.
• Small private datasets from two sites with only 13 and 10 test cases; the framework is only shown for brain MRI at 3T with two contrast agents.

Experimental setup

Datasets: Site A private MPRAGE cohort (126 cases) Site B private MPRAGE cohort (159 cases) spine fine-tuning set (5 patients) breast fine-tuning set (3 cases)

Benchmarks: Site A test set (13 cases) Site B test set (10 cases) motion artifact removal test set low-dose simulation spine and breast transfer tasks

Metrics: PSNR SSIM CNR LPIPS CFS GMSD IRC EIS RMSE NPS

Baselines: UNet Att-UNet NestedUNet SwinIR UKAN MambaIR BICEPS R2-AttUNet

Builds on: NestedUNet BICEPS subtraction imaging pix2pix-style GAN loss

Closest papers by meaning

• Anatomical Consistency and Adaptive Prior-informed Transformation for Multi-contrast MR Image Synthesis via Diffusion Model (CVPR 2025 · similarity 0.86)
• Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction (AAAI 2024 · similarity 0.83)
• Fast MRI for All: Bridging Access Gaps by Training without Raw Data (NeurIPS 2025 · similarity 0.83)
• D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in MRI Reconstruction (NeurIPS 2025 · similarity 0.83)
• Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.83)
• LeFusion: Controllable Pathology Synthesis via Lesion-Focused Diffusion Models (ICLR 2025 · similarity 0.82)