Pith. sign in

REVIEW 5 major objections 5 minor 1 references

PET Tracer Separation Using Conditional Diffusion Transformer with Multi-latent Space Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A diffusion transformer conditioned on texture masks and per-tracer latent priors reports the best dual-tracer PET separation on brain and chest datasets.

desk verdict Separation claim is not supported as written: the model may be getting ground-truth single-tracer images as input, and the loss function as printed trains to copy the dual image. read the letter →

arxiv 2506.16934 v1 pith:JHXANHFO submitted 2025-06-20 eess.IV cs.CV

Bin Huang , Feihong Xu , Xinchong Shi , Shan Huang , Binxuan Li , Fei Li , Qiegen Liu This is my paper · ORCID
classification eess.IVcs.CV
keywords PETtracerseparationdual-tracerdiffusiontransformermulti-latentspacetexturemasklocalbinarypatternimage-domaingenerativemedicalimaging
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

PET scanners see all positron-emitting tracers as identical gamma energies, so a dual-tracer scan is a mixture the scanner cannot tell apart. This paper proposes MS-CDT, a diffusion-transformer model that separates the mixed image into per-tracer images by conditioning on texture masks and on a separate latent-space prior for each tracer. On synthetic dual-tracer brain and chest datasets built by fusing registered single-tracer scans, the method reports the highest PSNR/SSIM among five compared deep-learning baselines, especially for the harder tracer in each pair. The paper's claim, if it holds, is that conditioning a generative model on tracer-specific texture and latent priors is an effective route to dual-tracer PET separation in the image domain.

What carries the argument

The load-bearing mechanism is the combination of the Texture Mask Conversion Block (TMCB) and the Latent Prior Extraction Block (LPEB) inside a diffusion transformer (DTM). The TMCB applies the local binary pattern operator to each tracer PET, producing binary texture masks that are multiplied with the image to form texture conditions; this guides the model toward salient structural edges and, through a weighted fusion at the output, counteracts over-smoothing. The LPEB concatenates the dual-tracer image with each single-tracer reference, downsamples via pixel unshuffling, and compresses the result via residual blocks and linear layers into a latent vector per tracer, so the diffusion forward and reverse processes run on compact priors rather than full-resolution volumes. The estimated latent priors are injected as affine modulation parameters into the transformer's multi-head transposed attention and gated feed-forward layers, and the final separated image is a weighted combination of the transformer output and the texture-masked reconstruction. The low-dimensional latents are what make four diffusion steps sufficient and keep training tractable on a single 24 GB GPU.

What would settle it

Apply MS-CDT to a physically acquired simultaneous dual-tracer phantom scan whose per-tracer concentrations are known by construction and compare per-tracer PSNR/SSIM to the synthetic-fusion numbers; if the margin over DTM shrinks below the reported gains, the synthetic fusion assumption is the source of the improvement.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that dual-tracer PET images can be separated into their constituent tracer images by a diffusion transformer conditioned on two tracer-specific guides: a texture mask computed by local binary patterns and a low-dimensional latent-space prior extracted per tracer. Because all PET tracers emit annihilation photons of the same energy, the mixed signal carries no direct label of which tracer produced which count; the authors argue that these two conditions let the model disentangle tracer-specific structure without reconstructing the full mixed volume in the diffusion process. On the brain test set the method reports average PSNR/SSIM of 41.325/0.9098 for 18F-DOPA and 54.632/0.9625 for 18F-FDG, and on the chest test set 45.80/0.8709 for 18F and 39.18/0.6707 for 68Ga, each above the iVAN, DL-based, MPRnet, Pix2pix, and DTM baselines. The ablation study attributes the gain to adding the multi-latent prior and then the texture condition on top of the plain transformer.

Load-bearing premise

The dual-tracer training and test pairs are pixel-wise sums of flexibly registered single-tracer scans, so the paper assumes those fusions faithfully represent true simultaneous dual-tracer PET acquisitions with shared attenuation, scatter, and timing.

Editorial extensions

If this is right

  • A single static dual-tracer scan could replace two separate single-tracer scans, with the separated images produced computationally afterward.
  • Because each tracer owns its own latent space, the architecture extends in principle to three or more tracers by adding additional latent channels.
  • Texture-mask conditioning gives a concrete way to counter the over-smoothing that generative PET reconstruction often produces, which is the reason the authors give for the SSIM gains.
  • The reported results on a second dataset (chest, 18F and 68Ga) suggest the mechanism is not limited to a single tracer pair or body region.
  • Estimating compact latent priors with only four diffusion steps points toward inference fast enough for clinical use, if the reported quality holds on real acquisitions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct stress test the authors did not run would use tracers whose uptake patterns overlap much more heavily than FDG and DOPA; if separation accuracy drops sharply, the method may be learning to separate by overall uptake location rather than by tracer-specific texture.
  • The same texture-latent conditioning recipe could be applied to other mixed-signal imaging problems, such as dual-isotope SPECT or multi-parametric MRI, where several biological signals share one acquisition.
  • The authors' threshold sweep finds 180-255 the best texture range, which suggests the LBP mask is acting as a high-activity edge prior; replacing it with a learned edge detector or anatomical mask would reveal whether the hand-crafted texture is essential.
  • Extending the latent-space conditioning from static images to dynamic frames could connect this work to kinetic-model-based dual-tracer reconstruction, an avenue the authors flag as future work in the sinogram or list-mode domain.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes MS-CDT, a conditional diffusion transformer for separating dual-tracer PET images. The method introduces a texture mask conversion block (TMCB) based on local binary patterns and a multi-latent space prior (MSP) extraction block (LPEB), and it trains a diffusion transformer to predict the MSP rather than full images. The authors evaluate on synthetic dual-tracer data created by registering and fusing separately acquired single-tracer PET scans for brain (18F-DOPA / 18F-FDG) and chest (18F / 68Ga) datasets, reporting PSNR/SSIM/NRMSE improvements over iVAN, a DL-based method, MPRnet, Pix2pix, and DTM. They also perform ablation studies and a threshold sensitivity analysis for the texture mask.

Significance. If the inference-time data flow is as claimed—namely, that only the dual-tracer composite is used as input at test time—the paper makes a useful contribution to PET tracer separation by combining diffusion and transformer architectures with latent-space conditioning. The manuscript's strengths include a publicly released code repository, a clear architectural motivation for using multi-latent priors, and an attempt at a generalization experiment on a second tracer pair and body region. However, the central quantitative claim is currently undermined by an unresolved ambiguity: the text and figures repeatedly state that the two single-tracer PETs are fed into the latent prior extractor, which would make the reported separation numbers measure reconstruction from ground-truth inputs rather than separation from the dual-tracer image alone. The training loss in Eq. (14) similarly appears to compare outputs to the dual-tracer image, not to the individual tracer ground truths. These issues must be resolved before the reported superiority over baselines can be accepted.

major comments (5)
  1. [III-C, III-E, Fig. 2] The inference-time input graph of MS-CDT is not specified. Section III-C states that "the dual-tracer PET and the two single-tracer PETs are respectively combined into an MSP," and Section III-E says the LPEB concatenates the dual-tracer PET and the two single-tracer PETs before downsampling. If this describes the test-time procedure, then the reported PSNR/SSIM values in Tables I and V are not produced from the dual-tracer input alone; the ground-truth single-tracer images are provided to the latent-prior extractor, so the evaluation measures reconstruction of supplied targets rather than separation from the composite. The paper never explicitly states that single-tracer PETs are used only during training, and it does not describe how the MSP is obtained at inference from the dual-tracer image alone. Please state the inference-time data flow precisely and, if only the dual-tracer image is used, specify how each tracer's latent prior L_i is derived at test time.
  2. [Eq. (14)] The texture loss LOSSTM in Eq. (14) computes ||I_Dual - Ihat_DOPA||_1 and ||I_Dual - Ihat_FDG||_1, i.e., it compares both separated outputs to the dual-tracer composite image rather than to the corresponding single-tracer ground truths. As printed, this objective trains the transformer to reproduce the input rather than to separate the tracers, which contradicts the stated goal of the method. If this is a typographical error (the terms should presumably be ||I_DOPA - Ihat_DOPA|| and ||I_FDG - Ihat_FDG||), please correct it; otherwise the training objective itself is not a separation loss and the reported results are unexplained.
  3. [V, Fig. 13] The texture threshold (180, 255) used in the final reported results was selected by comparing four thresholds on the same test set. This is a form of test-set fitting and makes the reported PSNR/SSIM/NRMSE values optimistic. The threshold should be chosen on a validation split (or via cross-validation) and then evaluated on a held-out test set; alternatively, the paper should report the sensitivity of the main comparison to the threshold choice.
  4. [IV-A, IV-D] The test cohorts are very small (5 brain patients, 120 chest slices) and no error bars, standard deviations, or significance tests are reported for the quantitative comparisons. For example, the claimed brain-dataset advantage over DTM is about 1.3 dB for 18F-FDG and 2.4 dB for 18F-DOPA; without variability measures it is impossible to tell whether these differences are meaningful. Please report per-patient/per-slice statistics with a statistical test (e.g., paired t-test or Wilcoxon signed-rank test) or at least provide error bars.
  5. [IV-A, VI] The evaluation is entirely on synthetic dual-tracer data created by registering separately acquired single-tracer scans and fusing them with 3D-Slicer. As the authors note, there was no true simultaneous dual-tracer acquisition. The paper should explicitly acknowledge that such fusion does not reproduce the same attenuation, scatter, and cross-tracer contamination physics as a real simultaneous scan, and it should temper the clinical claim in the abstract accordingly. A concrete way to address this is to validate on a small set of truly simultaneous dual-tracer acquisitions or to simulate dual-tracer data through Monte Carlo from a single physical phantom.
minor comments (5)
  1. [Table I] The table header says "NMSE" while the text and Section IV-C consistently refer to NRMSE; please unify the abbreviation.
  2. [Eq. (21)-(22)] The definitions of CR and COV are inconsistent between the text and the equations: the text says CR compares the maximum pixel value in the lesion region to the mean in the liver region, but Eq. (21) reads CR = Max_gray / mu_white; similarly Eq. (22) uses sigma_white but the text describes sigma_gray / mu_gray. Please align these definitions.
  3. [Fig. 14-15] The captions for Figures 14 and 15 are identical, and both refer to "patient I" and "Tracer A/B"; one of them should refer to the appropriate patient and tracer labels, and the panel ordering (MS-CDT-180 vs MS-CDT-200) should be consistent.
  4. [Fig. 9-10] The captions describe the images as "cross-sectional" while the text refers to them as transverse sections; please use consistent anatomical plane terminology throughout.
  5. [IV-D] The sentence describing Fig. 11 and 12 in the chest generalization section refers to "18F-DOPA and 18F-FDG" tracers, but the chest dataset uses 18F and 68Ga; please correct the tracer names in the figure captions and text.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline numbers are partly selected on the test set, and the inference-time input graph for the latent-prior extractor is unspecified, so the separation claim is not fully supported as written.

  1. fitted input called prediction [Section V (Discussion), texture-threshold selection; final PSNR/SSIM in Tables I and IV]
    "Four texture feature thresholds were tested, and the results show that the texture features with a threshold of (180, 255) exhibit the best performance. ... The objective of this experiment is to identify the optimal texture feature threshold to optimize the model's performance."

    The threshold (180,255) is chosen as the best of four settings evaluated on the same test set that produces the headline PSNR/SSIM. The reported MS-CDT figures are thus the maximum over a small grid, so part of the claimed performance is a selection artifact rather than an a priori prediction of an unseen test set.

  2. other [Section III-C and III-E, Fig. 2 caption; inference-time use of single-tracer PETs]
    "Secondly, through a LPEB, the dual-tracer PET and the single-tracer PET are converted into latent priors respectively. ... The dual-tracer PET and the two single-tracer PETs are respectively combined into an MSP."

    The manuscript never states that single-tracer PETs are restricted to training; the method text and Fig. 2 caption describe the MSP as formed from the dual-tracer PET and the two single-tracer PETs. As written, the latent priors conditioning the transformer are extracted from the ground-truth single-tracer images, so the separated outputs would be decodes of the targets' own latent codes and the reported PSNR/SSIM would measure reconstruction of provided inputs, not separation from the dual-tracer composite alone.

full rationale

The strongest claim (Section IV-D) is not independently derived in the text. The only concrete reduction I can exhibit is the threshold selection: Section V chooses (180,255) as the best of four thresholds on the same test set used for Tables I and IV, so the headline PSNR/SSIM are fitted values, not predictions under a fixed model. A second text-level problem is that the inference-time input graph for LPEB is never specified; Section III-C/E and Fig. 2 describe the MSP as being formed from dual-tracer and single-tracer PETs, and if single-tracer PETs remain inputs at inference, the separation reduces to decoding the ground truth's latent space. I do not count this as definite circularity because the paper may intend those images for training supervision only, but the omission is load-bearing and the claim is not yet supported as written. Eq. (14), as printed, also regresses both outputs to I_Dual, which would train reproduction of the input rather than separation; this is likely a typo but it is another missing-support flag. The architecture comparison (DTM vs MS-CDT) retains some independent content, so the score is moderate, not maximal.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that synthetic fused single-tracer images represent true dual-tracer PET, and on the network's learned latent and texture representations. One test-set-fitted threshold and an unspecified fusion weight are free parameters. No new physical entities are introduced.

free parameters (3)
  • Texture threshold (TMC) = 180 (from candidates 120, 150, 180, 200)
    Selected in Section V as the threshold giving the best PSNR/SSIM/NRMSE on the test set; the final reported results use this value, making it a test-set-fitted parameter.
  • Fusion weight alpha in Eq. (7) = not specified
    The final image is a weighted sum alpha*I_hat + (1-alpha)*U_hat, but the paper never states the value of alpha; it is a tunable parameter whose choice affects all reported results.
  • Diffusion steps T = 4
    The paper states 'T is set to 4' in Section IV-B; this is a hand-chosen design parameter for the diffusion process.
assumptions (4)
  • standard math Gaussian forward and reverse diffusion processes (Eqs. 1, 2, 11-13) are valid for the latent representations.
    The method assumes standard DDPM/DDIM sampling formulas apply to the MSP latents without modification.
  • domain assumption A dual-tracer PET image can be represented as a registered pixel-wise fusion of two single-tracer PET images.
    Section IV-A creates all dual-tracer data by registering and fusing single-tracer scans; the model is never trained or tested on a true simultaneous dual-tracer acquisition.
  • domain assumption LBP texture masks capture image structure relevant to tracer separation and improve detail preservation.
    TMCB uses local binary patterns to derive TMC (Eq. 5) and the authors assert this guides the model to salient structures; this is a modeling assumption without independent validation.
  • ad hoc to paper The multi-latent prior L_i (MSP) is a sufficient and useful representation for separating each tracer.
    LPEB compresses dual- and single-tracer PET into latent vectors that the diffusion model predicts and the transformer consumes; the paper asserts this improves disentanglement but provides no analysis of information sufficiency.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PET Tracer Separation Using Conditional Diffusion Transformer with Multi-latent Space Learning." pith.science (2026). https://pith.science/paper/JHXANHFO

@misc{pith2026250616934,
  author       = {Pith},
  title        = {Pith review of: PET Tracer Separation Using Conditional Diffusion Transformer with Multi-latent Space Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JHXANHFO}},
  note         = {Machine review of arXiv:2506.16934}
}
read the original abstract

In clinical practice, single-radiotracer positron emission tomography (PET) is commonly used for imaging. Although multi-tracer PET imaging can provide supplementary information of radiotracers that are sensitive to physiological function changes, enabling a more comprehensive characterization of physiological and pathological states, the gamma-photon pairs generated by positron annihilation reactions of different tracers in PET imaging have the same energy, making it difficult to distinguish the tracer signals. In this study, a multi-latent space guided texture conditional diffusion transformer model (MS-CDT) is proposed for PET tracer separation. To the best of our knowledge, this is the first attempt to use texture condition and multi-latent space for tracer separation in PET imaging. The proposed model integrates diffusion and transformer architectures into a unified optimization framework, with the novel addition of texture masks as conditional inputs to enhance image details. By leveraging multi-latent space prior derived from different tracers, the model captures multi-level feature representations, aiming to balance computational efficiency and detail preservation. The texture masks, serving as conditional guidance, help the model focus on salient structural patterns, thereby improving the extraction and utilization of fine-grained image textures. When combined with the diffusion transformer backbone, this conditioning mechanism contributes to more accurate and robust tracer separation. To evaluate its effectiveness, the proposed MS-CDT is compared with several advanced methods on two types of 3D PET datasets: brain and chest scans. Experimental results indicate that MS-CDT achieved competitive performance in terms of image quality and preservation of clinically relevant information. Code is available at: https://github.com/yqx7150/MS-CDT.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references · 1 canonical work pages

  1. [1]

    Diffusion Transformer Meets Random Masks: An Advanced PET Reconstruction Framework

     M. Wartski, and A. Sauvanet. "18F -FDG PET/CT in pancreatic adenocarcinoma: a role at initial imaging staging?" DII, vol. 100, no. 12, pp. 735-741, 2019.  L. Qiu, et al., "The role of 18F-FDG PET and PET/CT in the evaluation of primary cutaneous lymphoma." Nucl. Med. Commun, vol. 38, no. 2, pp. 106-116, 2017.  S. S. R. Mukku, et al., "Clinical u...

    work page Pith review arXiv 2019

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.