Bridging Modalities on the Cortex: Surface-based MRI to PET Translation with a Diffusion Bridge
Yitong Li, Alexandra Samoylova, Fabian Bongratz, Timo Grimmer, Dennis M. Hedderich, Igor Yakushev, Christian Wachinger
cs.CV · cs.AI
Quick verdict
First surface-based MRI→PET diffusion bridge; strong results, but gains over bridge baselines are narrow and MRI vs syn-PET used different readers.
Summary
The paper presents DB-SUiT, a framework that translates structural MRI into FDG-PET directly on the cortical surface instead of in voxel space. Cortical thickness from FreeSurfer serves as the source signal; PET values sampled along pial-white-matter trajectories on an ico5 spherical mesh form the target. A diffusion bridge (VP formulation, EDM preconditioning, deterministic probability-flow ODE with Heun solver) learns the mapping; the denoiser, SUiT, couples spherical ResNets with bottleneck Transformers and is conditioned on age, sex, and 36 subcortical volumes via adaLN and cross-attention, trained with an EDM-weighted denoising loss plus a surface-gradient regularizer. On ADNI and a single-site TUM cohort, DB-SUiT reports the best MAE (0.0430/0.0438), PSNR (26.17/25.69 dB), and PCC (0.897/0.868), reducing MAE by 5.3-8.6% versus the strongest bridge baselines and about 22% versus the surface-projected volumetric method SiM2P. In three-way dementia classification on the in-house cohort, synthetic PET surfaces reach 77.3% balanced accuracy, a 14.2% relative gain over the best MRI configuration, exceeding real PET volumes (69.5%) but staying below real PET surfaces (79.3%). In a blinded reader study on 62 external subjects using the ADNI-trained model without fine-tuning, synthetic PET scored 85.5% versus 75.8% for MRI and 95.2% for real PET. Key limitations: narrow, significance-untested margins over the strongest baselines; ablation deltas within likely noise; and MRI and PET read by different clinicians.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- DB-SUiT is the first MRI-to-PET translation framework that generates metabolic signals directly on the cortical manifold via a diffusion bridge with a spherical U-shaped Transformer backbone.
- It achieves the best reconstruction fidelity on both datasets (MAE 0.0430 ADNI / 0.0438 in-house), beating diffusion-bridge baselines and the surface-projected volumetric SiM2P method.
- Synthetic PET surfaces raise three-way dementia classification to 77.3% BACC, a 14.2% relative gain over the best MRI configuration and 11.3% over real PET volumes, approaching real PET surfaces (79.3%).
- An ADNI-trained model applied without fine-tuning to an external cohort containing unseen bvFTD achieves 85.5% blinded reader accuracy, versus 75.8% for MRI and 95.2% for real PET.
- Inference through the probability-flow ODE is fully deterministic, eliminating sampling-related variability in deployment.
Strengths
Native surface formulation is well-motivated and consistently validated
Generating PET on the ico5 mesh with vertex-wise MRI-PET correspondence avoids voxel partial-volume artifacts; the lobe-wise analysis (Table II) shows the lowest MAE in all six lobes on both ADNI and in-house data, so the gain is spatially robust rather than region-specific.
Fair volumetric baseline handling
SiM2P volumetric outputs are projected onto the surface with the identical pipeline (Table I footnote), cleanly isolating the representation benefit: 21.4% (ADNI) and 22.9% (in-house) MAE reduction over 0.0547/0.0568.
Multi-level clinical validation chain with zero-shot transfer
Deterministic deployment-friendly inference
Code and protocol details released
Weaknesses
Thin margins over strongest baselines, untested for significance
Ablation conclusions rest on deltas within noise
Reader-study design confounds modality with reader
Unexplained negative result for extra MRI features
Missing PASTA baseline from the same group
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DB-SUiT (arXiv 2609.20147) is a method paper proposing the first MRI-to-FDG-PET translation framework that generates metabolic maps directly on the cortical surface rather than in voxel space. Its three headline claims: (1) a diffusion bridge with a Spherical U-shaped Transformer (SUiT) yields state-of-the-art surface PET synthesis on ADNI and a single-site TUM cohort; (2) synthetic PET surfaces materially improve automated dementia classification (+14.2% relative BACC over the best MRI configuration, 11.3% over real PET volumes); (3) in a blinded reader study on 62 external subjects, an ADNI-trained model applied without fine-tuning reaches 85.5% diagnostic accuracy versus 75.8% (MRI) and 95.2% (real PET), demonstrating cross-cohort and cross-pathology (bvFTD) generalization.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DB-SUiT?
DB-SUiT is a deep-learning framework that synthesizes FDG-PET signals directly on the brain's cortical surface from structural MRI (cortical thickness) using a diffusion bridge model with a spherical Transformer backbone, so the synthetic metabolism map stays vertex-aligned with the cortex rather than being generated in voxel space.
How accurate is MRI-derived synthetic PET for dementia diagnosis?
In the paper's blinded reader study on 62 external subjects, synthetic PET surfaces reached 85.5% diagnostic accuracy (three-class CN/AD/bvFTD), versus 75.8% for structural MRI and 95.2% for real PET, read by board-certified physicians.
Does it work for dementia types not seen during training?
Yes, partially: the model trained only on ADNI (which contains no FTD cases) was applied without fine-tuning to an external hospital cohort including behavioral-variant FTD, and the reader study suggests diagnostically useful output, though this rests on a single external site and readers.
Is the code and data available?
The authors state code is available at github.com/ai-med/DB-SUiT. The ADNI data are publicly available under a data-use agreement; the in-house clinical cohort and reader-study data are not shared.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Equations (4)-(6) and one row of Table IV are garbled by PDF text extraction; the EDM preconditioning derivation could not be fully verified.
- Figures 1-4 are not analyzable from extracted text; qualitative fidelity and z-score claims rest on the authors' descriptions.
- In-house TUM acquisition details (scanner, PET protocol) are absent from the extracted text.
- arXiv v1 preprint; no journal peer review record was available to this review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.