Evolve.Pub
Sign inScan your paper
arXiv paper· medical image analysis (cross-modal image synthesis)· method paper

3D CT-to-PET Translation via Latent Brownian Bridge Diffusion

Sarita Mourya, Francesco Di Feola, Pierangelo Veltri, Paolo Soda

View original paper ↗arXiv:2609.12860

cs.CV · cs.AI

Quick verdict

3D latent Brownian-bridge diffusion + contrastive VAE for CT-to-PET; lesion-level gains in lung credible, but volume-level win over Pix2Pix unproven.

Summary

The paper proposes X-Bridge, a 3D CT-to-PET translation framework built on a latent Brownian Bridge Diffusion Model (BBDM). A VAE with a shared encoder is first trained on paired CT-PET patches using KL, L1, perceptual, adversarial, and InfoNCE contrastive losses to align anatomical and metabolic latents; a BBDM then learns the bridge between CT and PET latent embeddings, with 200-step DDIM sampling and patch stitching to reconstruct whole volumes. Evaluation uses the FDG-PET-CT-Lesions dataset (864 training/150 test scans, lung region via MOOSE masks) and an external ENHANCE.PET cohort (579 scans, Tuebingen overlap excluded). Compared with CPDM, a 3D LDM, and Pix2Pix, X-Bridge reports higher lesion-level PSNR, SSIM, Dice1.5/Dice2.5, SUVmax, and SUVmean (Wilcoxon signed-rank with Bonferroni, p<0.05), better MTV/TLG preservation, and an advantage for small lesions; external volume-level PSNR is highest (29.91) with the smallest internal-external variation. However, Pix2Pix attains better internal PSNR (33.30 vs 29.99) and SSIM (0.849 vs 0.779), and CPDM's SSIM edges X-Bridge on both cohorts, so volume-level superiority is not established. Main limitations: lung-only evaluation, lesion-level metrics impossible externally, uniform patch-averaging artifacts, and no uncertainty quantification for this inherently under-constrained task.

Review complete

Assessment at a glance

medical image analysis (cross-modal image synthesis)

5

risk flags identified

● high 0

● medium 3

● low 2

Reproducibility readinessmedium
Code & data availableyes
Review confidencehigh
AI text signal1/20 passages flagged
Deep research viableCode-reproducibleNeeds GPU

4 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

  • X-Bridge (contrastive VAE + latent Brownian bridge diffusion) improves lesion-level PSNR, SSIM, Dice1.5/Dice2.5, SUVmax and SUVmean over CPDM, a 3D LDM, and Pix2Pix (Fig 2, Wilcoxon signed-rank with Bonferroni, p<0.05).
  • Small-lesion metabolic activation is better preserved than by competing methods, with the largest advantage in the smallest lesion-size quartile (Fig 4).
  • X-Bridge shows the best cross-dataset robustness, with internal-external PSNR variation of 0.27% and the highest external PSNR (29.91, Table I).
  • Median lesion-level MTV and TLG are better preserved than by competing methods at SUV thresholds 1.5 and 2.5 (Fig 3).

Strengths

  • Volumetric latent formulation with cross-modal alignment

    Unlike the 2D slice-wise CPDM [5], X-Bridge operates on 32³ patches in a shared VAE latent space (Sec II-A), adding an InfoNCE contrastive term (Eq 5) to close the CT/PET modality gap before bridge diffusion — a sensible fit of BBDM [7] to 3D data.

  • Clinically grounded, lesion-centric evaluation

    Beyond PSNR/SSIM, Sec IV-B defines SUVmax/SUVmean, MTVα, TLGα, and Dice at SUV thresholds 1.5 and 2.5 (Eqs 13–16), and Fig 4 stratifies performance by lesion-size quartiles — directly probing the small-lesion failure mode that volume-level averages hide.

  • Genuine external validation with leakage control

  • Reproducibility scaffolding

Weaknesses

  • Volume-level narrative contradicts Table I

  • Unexplained baseline substitution

  • MTV/TLG treated as higher-is-better

  • Thin statistical reporting

  • Core novelty never ablated; key details missing

Full Review Report

$1.99 · one-time

Unlock the complete preliminary analysis Evolve produced for this paper.

Report preview

Scope & Claims

X-Bridge is a two-stage, patch-based 3D framework translating CT into FDG-PET: (1) a shared-encoder VAE trained on paired CT/PET lung patches with L1, KL, perceptual, patch-adversarial, and InfoNCE contrastive objectives (Eqs 4-10); (2) a latent Brownian Bridge Diffusion Model (BBDM [7]) learning the stochastic bridge from CT to PET latents (Eqs 11-12), decoded via DDIM and stitched into whole volumes. The paper claims improved lesion-level fidelity over CPDM, a 3D LDM, and Pix2Pix; better small-lesion preservation; superior MTV/TLG preservation; and cross-dataset robustness (Δ PSNR 0.27%). The Introduction is commendably careful: synthetic PET is framed as "a probabilistic estimate of plausible metabolic activity", not a replacement for acquired PET.

The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.

  • All 4 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 X-Bridge and how does it synthesize PET from CT?

It is a two-stage 3D deep-learning model. A variational autoencoder with a shared encoder first compresses paired CT/PET lung patches into a latent space, using contrastive learning to align the two modalities. A Brownian bridge diffusion model then learns the stochastic path from a CT latent to a PET latent; decoded patches are stitched into a whole 3D PET volume.

Can synthetic PET from X-Bridge replace a real PET scan?

No. The authors explicitly position it as a probabilistic estimate of plausible metabolic activity inferred from CT, intended to support lesion analysis, triage, and computational workflows — not as a clinical replacement for acquired FDG-PET.

What data was the model trained and tested on?

Internally, the FDG-PET-CT-Lesions dataset (1,014 whole-body FDG-PET/CT scans from University Hospital Tuebingen; 864 training/150 test, lung region only via MOOSE masks). Externally, 579 scans from the multi-center ENHANCE.PET dataset (Leipzig and Careggi), after excluding Tuebingen-overlapping cases to avoid data leakage.

Did X-Bridge beat every competing method on every metric?

No. It leads the lesion-level metrics (PSNR, SSIM, Dice at SUV 1.5/2.5, SUVmax/mean) and the external volume-level PSNR (29.91), but Pix2Pix achieves higher internal PSNR (33.30 vs 29.99) and SSIM (0.849 vs 0.779), and CPDM's SSIM is slightly higher on both datasets.

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.

  • Quantitative lesion-level results (Figs 2-4) exist only as plots in the extracted text: axis ranges, test names, and qualitative claims are present, but medians, means, and per-group sample sizes cannot be read from the extraction.
  • Fig 5's qualitative image quality could not be assessed from the text extraction; the assessment of visual claims relies on the authors' descriptions.

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.

Scan your paper
3D CT-to-PET Translation via Latent Brownian Bridge Diffusion | Evolve.Pub