MRecover

Compare synthesized T2w-TSE directly against the as-acquired reference — generated from T1w alone.

Case 01

T1w MPRAGE -> TSE

Autoregressive synthesis from T1w image to T2w-TSE contrast, validated against the as-acquired reference for downstream ASHS subfield segmentation

MPRAGE source image

Scientific context

Designed to recover motion corrupted T2w TSE data

High-resolution T2w-TSE is the gold standard for hippocampal subfield analysis, but motion artifacts hit hardest in elderly and cognitively impaired patients. We synthesize it from T1w, which everyone already has.

Built for Scale

Full volume synthesized in under 30 seconds

Single-step ODE sampling on an NVIDIA A100, no iterative denoising, no prohibitive compute cost. Fast enough for large-scale cohort workflows, practical enough for retrospective datasets where T2w-TSE was never acquired.

Request access

Model access is managed on Hugging Face

MRecover weights are hosted as a gated Hugging Face model. Request access there with your name, email, affiliation, and intended research use.

  1. Open the model page
  2. Submit the Hugging Face access request
  3. Run huggingface-cli login
Open gated model

After approval, authenticate with Hugging Face before running the package.