REVIEW 5 major objections 5 minor 23 references
Plasma-CycleGAN: Plasma Biomarker-Guided MRI to PET Cross-modality Translation Using Conditional CycleGAN
T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read Conditioning MRI-to-PET synthesis on the plasma Aβ42/40 biomarker consistently improves generated PET image quality and amyloid-positivity classification; the best configuration, CycleGAN with latent concatenation, is named Plasma-CycleGAN.
desk verdict A clinically motivated but overclaimed GAN conditioning study; the concat variant may help, but the headline claim is contradicted by the paper's own Table 1 and lacks statistical support. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is latent-space concatenation of the biomarker: a scalar normalized plasma Aβ42/40 value is expanded into a single-channel $16 \times 16 \times 16$ tensor and concatenated with the generator's $128 \times 16 \times 16 \times 16$ bottleneck feature map, yielding a 129-channel map that a $1 \times 1$ convolution reduces back to 128 channels. This injects the biological information precisely where the encoder's structural features are about to be decoded into PET, and the paper reports that this placement—unlike adding the value to the input image or to the feature map—improves performance consistently across Pix2pix, CycleGAN, and ShareGAN.
What would settle it
Re-run the comparison with a validation-based selection procedure: pick the conditioning method on a held-out validation set, then compare Plasma-CycleGAN with plain CycleGAN on a test set used only once, and repeat the split to estimate uncertainty; if the SSIM or classification gap disappears, the paper's central claim is not supported.
Extended reading notes
Core claim
The central claim of the paper is that incorporating blood-based biomarker information, specifically the normalized plasma Aβ42/40 ratio, into deep generative models improves MRI-to-PET translation. The paper tests three conditioning placements for each of three baselines (Pix2pix, CycleGAN, ShareGAN) and finds that expanding the scalar biomarker to the size of the bottleneck feature map and concatenating it as an extra channel improves SSIM, PSNR, and MSE for every baseline, while the other two placements are unstable or harmful. For CycleGAN the reported gain is from SSIM 0.808 to 0.822, PSNR 24.65 to 25.07, and MSE 250.33 to 227.00, and amyloid-positivity classification accuracy jumps from 0.620 to 0.815. The paper therefore names CycleGAN plus latent concatenation Plasma-CycleGAN and presents it as the first cross-modality MRI-to-PET translation model conditioned on blood-based biomarkers.
Load-bearing premise
The whole result rests on the assumption that the reported improvement is a real effect of adding the blood-test value and not just a lucky pick among the twelve models the authors compared.
Editorial extensions
If this is right
- If the reported numbers are representative, synthetic PETs produced by Plasma-CycleGAN would carry enough amyloid signal to separate amyloid-positive from amyloid-negative brains with 0.815 accuracy, compared with 0.620 for the unconditioned CycleGAN.
- The fact that latent concatenation helped all three baselines suggests the gain comes from where the biomarker enters the network, not from a particular generator architecture.
- Improved SSIM, PSNR, and SUVR correlation together imply the generated images are closer to real PET both pixel-wise and in the uptake-value distribution used clinically.
- The paper's stated next steps are to apply the same conditioning idea to diffusion models and to incorporate p-tau217, a biomarker it notes has shown accuracy comparable to cerebrospinal fluid measures.
Reading between the lines
- A natural next experiment, which the paper does not run, is to test whether latent concatenation of plasma Aβ42/40 also improves diffusion-based MRI-to-PET synthesis; the consistency across three GAN baselines is suggestive but untested for diffusion models.
- If the central claim holds, a blood draw and a routine MRI could in principle serve as a PET surrogate for amyloid assessment, making amyloid screening more accessible; this clinical workflow is not evaluated in the paper.
- The injection mechanism is generic to scalar clinical variables, so the same design could be tried with p-tau217 or other biomarkers; the paper names p-tau217 as future work but does not test it.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies whether incorporating the plasma Aβ42/40 ratio into three MRI-to-PET synthesis baselines (Pix2pix, CycleGAN, ShareGAN) improves generated PET quality. Three conditioning mechanisms are compared: adding the biomarker to the input image, adding it to a latent feature map, and concatenating it as an extra channel in latent space. The paper reports that the concatenation strategy consistently improves all three baselines, names the CycleGAN variant Plasma-CycleGAN, and further evaluates SUVR correlation and amyloid-positivity classification on the synthesized PET images. The authors claim that this is the first integration of blood-based biomarkers into MRI-to-PET cross-modality translation.
Significance. If the claims were substantiated, the paper would introduce a clinically meaningful, low-cost conditioning variable into cross-modality image translation, with potential diagnostic utility. The use of a moderately large ADNI dataset, multiple baselines, and a subject-level train/validation/test split is a reasonable experimental setup. The clearest strength is that the idea is concrete and the reported numbers are falsifiable. However, the paper's central claim as written is contradicted by its own Table 1, and the remaining positive results are presented without significance testing or proper model selection, so the practical significance of the proposed method is not currently established.
major comments (5)
- [Abstract and Sec. 3.2, Table 1] The claim that "BBBMs integration consistently enhances the generative quality across all models" is contradicted by the paper's own results. In Table 1, Pix2pix+image has SSIM 0.683 vs. baseline 0.766, Pix2pix+add has SSIM 0.465 vs. 0.766, and ShareGAN+image has SSIM 0.556 vs. 0.704; these are substantial degradations, not enhancements. The Table 1 caption stating "All models achieved improved performance after incorporating BBBMs in all metrics" is factually incorrect for these rows. The text in Sec. 3.2 later admits that "the performance was unstable or decreased in Pix2pix and ShareGAN." The abstract, introduction, and conclusion should be corrected to state the narrower, accurate finding that latent-space concatenation produced small directional improvements.
- [Sec. 3.2, Table 1] The surviving claim that latent-space concatenation improves all three baselines is not statistically supported. The SSIM differences are +0.014 (CycleGAN), +0.016 (Pix2pix), and +0.006 (ShareGAN), while the reported standard deviations are on the order of 0.07–0.13. The PSNR differences are +0.42, +0.34, and +0.05 dB, and the MSE differences are also small relative to the standard deviations. No paired significance tests, confidence intervals, or effect-size statistics are provided, even though paired tests on the 186 test images should be straightforward. Without such tests, the observed differences are within the noise level and do not support the claim of consistent improvement.
- [Sec. 3.1 and Sec. 3.2] The model selection procedure is problematic. Table 1 reports test-set performance for all 12 configurations, and Sec. 3.2 states "we only use concatenation in later experiments" immediately after presenting those test-set results. This constitutes selecting the best variant based on the same test set used for the final evaluation, which inflates the reported performance and makes the subsequent Plasma-CycleGAN comparisons (Tables 2 and 3) difficult to interpret. No validation-based selection, hold-out procedure, or correction for multiple comparisons is described. A proper approach would be to select the integration method on the 242-image validation set and to report test-set results only for the selected configuration.
- [Sec. 3.3, Table 3] The classification accuracy improvement (CycleGAN 0.620 to Plasma-CycleGAN 0.815) is reported without confidence intervals, significance tests, or a description of the number of unique subjects in the test set. Because the data consist of 1338 images from 456 individuals and the split is by subject, the 186 test images are not independent; multiple images can come from the same person. The effective sample size for classification is therefore smaller than 186, and treating images as independent observations inflates the apparent statistical strength. A per-subject bootstrap, a cluster-robust test, or reporting subject-level aggregated predictions is needed to support the claim that Plasma-CycleGAN achieves the best accuracy.
- [Sec. 3.3, Table 2] The SUVR correlation results are likewise reported without accounting for the clustered structure of the data. Pearson correlation coefficients and p-values computed on 186 images from a smaller number of subjects do not satisfy the independence assumption. The claim that "incorporating BBBM information enhanced the correlation in all models" should be reassessed with a cluster-robust correlation or a per-subject analysis. The differences between PCC values (e.g., 0.777 to 0.807 for CycleGAN) are also presented without confidence intervals, so it is unclear whether they are meaningful.
minor comments (5)
- [Table 1] The caption "All models achieved improved performance after incorporating BBBMs in all metrics" is inaccurate and should be revised to reflect the actual rows, several of which show degraded SSIM, PSNR, or MSE.
- [Table 1] There is a typographical error in the Pix2pix+add cell: "259.09.01±271.76" should likely be "259.09±271.76" or a corrected value.
- [Sec. 2.3] The loss function for CycleGAN is written with inconsistent notation: after defining L(G1,G2,D1,D2), the identity loss term is written as "λidtLidt" without defining Lidt; please clarify whether this is the same as Lide used for ShareGAN.
- [Sec. 3.2] The statement that "CycleGAN showed the highest robustness against disturbance" is not supported by any experiment or quantitative criterion; if robustness is intended as a claim, it should be formalized and measured.
- [Sec. 3.1] The sentence "The difference of validation and testing sets was to avoid data leakage" is confusing; splitting by subject avoids leakage, whereas validation and test sets serve the standard model-selection and final-evaluation purposes. Please clarify the intended roles of the validation and test subsets.
Circularity Check
No circularity: plasma biomarkers are input covariates, and the evaluated metrics compare generated images to ground-truth PET, not to the conditioning inputs.
full rationale
The paper's chain is empirical rather than derivational. The plasma Aβ42/40 values are supplied as conditioning covariates (expanded and added or concatenated at input or latent feature maps), never computed from the generated PET or from the quality metrics. The headline measures SSIM, PSNR, and MSE compare synthesized PET to real PET, so an improvement cannot be true by construction from the scalar biomarker input. The only self-citation is the MCSUVR>1.19 threshold from the authors' prior study [14]; this threshold is an external amyloid-positivity cutoff applied to the generated images, not a parameter fitted in this paper, and the SUVR classification is a downstream diagnostic-use check rather than the definition of image quality. The abstract's "consistently enhances across all models" contradicts Table 1 for several non-concat variants, and the model-selection procedure (evaluating all 12 configurations on the test set and then reporting the winner) is a soundness and statistical concern, but those are not circularity. No equation in the paper defines an output in terms of the claimed result or renames a fitted value as a prediction.
Assumptions & free parameters
free parameters (5)
- CycleGAN loss weights (lambda1, lambda2, lambda_idt) =
10.0, 10.0, 0.3
- ShareGAN loss weights (lambda1, lambda2, lambda_idt) =
10.0, 10.0, 0.5
- Learning rate =
0.0002
- Training epochs =
100
- Data split and augmentation parameters =
not reported
assumptions (4)
- domain assumption Plasma Aβ42/40 ratio provides information about PET amyloid uptake not fully captured by structural MRI.
- domain assumption Co-registered paired MRI and PET from ADNI are correctly aligned and preprocessed.
- standard math Standard GAN training assumptions: adversarial objectives converge and cycle-consistency preserves content.
- domain assumption The random split keeps all images of one individual in the same subset.
Cite this review
Pith. "Pith review of Plasma-CycleGAN: Plasma Biomarker-Guided MRI to PET Cross-modality Translation Using Conditional CycleGAN." pith.science (2026). https://pith.science/paper/7C4YVE3R
@misc{pith2026250102146,
author = {Pith},
title = {Pith review of: Plasma-CycleGAN: Plasma Biomarker-Guided MRI to PET Cross-modality Translation Using Conditional CycleGAN},
year = {2026},
howpublished = {\url{https://pith.science/paper/7C4YVE3R}},
note = {Machine review of arXiv:2501.02146}
}
read the original abstract
Cross-modality translation between MRI and PET imaging is challenging due to the distinct mechanisms underlying these modalities. Blood-based biomarkers (BBBMs) are revolutionizing Alzheimer's disease (AD) detection by identifying patients and quantifying brain amyloid levels. However, the potential of BBBMs to enhance PET image synthesis remains unexplored. In this paper, we performed a thorough study on the effect of incorporating BBBM into deep generative models. By evaluating three widely used cross-modality translation models, we found that BBBMs integration consistently enhances the generative quality across all models. By visual inspection of the generated results, we observed that PET images generated by CycleGAN exhibit the best visual fidelity. Based on these findings, we propose Plasma-CycleGAN, a novel generative model based on CycleGAN, to synthesize PET images from MRI using BBBMs as conditions. This is the first approach to integrate BBBMs in conditional cross-modality translation between MRI and PET.
Reference graph
Works this paper leans on
-
[1]
INTRODUCTION Alzheimer’s disease (AD) is a major neurodegenerative con- dition affecting millions worldwide, with the number of pa- tients and associated societal costs continually escalating. In the A/T/N classification system for AD diagnosis, brain amy- loid, tau pathology, and neurodegeneration are identified via PET scans using various tracers. Among...
work page Pith review arXiv 2025
-
[2]
METHODS 2.1. Data Acquisition All data used in this study were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and were freely available online at the LONI Image and Data Archive (IDA) data repository (https://ida.loni.usc.edu/). We downloaded and processed 1338 image instances of 456 individuals, in- cluding 31 AD, 231 mild cognitive...
work page 2017
-
[3]
EXPERIMENTS AND RESULTS 3.1. Model Training With 1338 paired MRI and FBP PET images, We randomly split the dataset into 910 training, 242 validation and 186 test- ing. The difference of validation and testing sets was to avoid data leakage. Data leakage was avoided by keeping all images of one individual in the same subset. All models were trained using a...
-
[4]
DISCUSSION AND CONCLUSION In this study, we investigated the impact of incorporating BBBM data in generative models for MRI to PET cross- modality translation. This is the first approach to combine BBBMs with generative models for cross-modality transla- tion from MRI to PET. We trained three models with and without including BBBMs on a subset of the ADNI...
-
[5]
As confirmed by the li- cense attached to the open-access data, ethical approval was not required
COMPLIANCE WITH ETHICAL STANDARDS This research study was conducted retrospectively using hu- man subject data available in ADNI repository and can be ac- cessed through the LONI Image and Data Archive (IDA) data repository (https://ida.loni.usc.edu/). As confirmed by the li- cense attached to the open-access data, ethical approval was not required
-
[6]
ACKNOWLEDGMENTS This work was supported by grants from the National In- stitutes of Health (RF1AG073424) and the State of Arizona via the Arizona Alzheimer Consortium. Algorithm develop- ment and image analysis for this study were partially sup- ported by the National Institute on Aging (R21AG065942, R01AG069453, RF1AG073424, and P30AG072980), and the Sta...
-
[7]
A/T/N: An unbiased descriptive classification scheme for Alzheimer disease biomarkers,
Clifford R Jack Jr, David A Bennett, Blennow, et al., “A/T/N: An unbiased descriptive classification scheme for Alzheimer disease biomarkers,” Neurology, vol. 87, no. 5, pp. 539–547, 2016
work page 2016
-
[8]
Yongsheng Pan, Mingxia Liu, Yong Xia, and Dinggang Shen, “Disease-image-specific learning for diagnosis- oriented neuroimage synthesis with incomplete multi- modality data,” IEEE transactions on pattern analy- sis and machine intelligence , vol. 44, no. 10, pp. 6839– 6853, 2021
work page 2021
Show all 23 references
-
[9]
BPGAN: Brain PET syn- thesis from MRI using generative adversarial network for multi-modal Alzheimer’s disease diagnosis,
Jin Zhang, Xiaohai He, et al., “BPGAN: Brain PET syn- thesis from MRI using generative adversarial network for multi-modal Alzheimer’s disease diagnosis,” Com- puter Methods and Programs in Biomedicine , vol. 217, pp. 106676, 2022
2022
-
[10]
Spatial adap- tive and transformer fusion network (STFNet) for low- count PET blind denoising with MRI,
Lipei Zhang, Zizheng Xiao, Zhou, et al., “Spatial adap- tive and transformer fusion network (STFNet) for low- count PET blind denoising with MRI,”Medical Physics, vol. 49, no. 1, pp. 343–356, 2022
2022
-
[11]
Bidirectional map- ping generative adversarial networks for brain MR to PET synthesis,
Shengye Hu, Baiying Lei, et al., “Bidirectional map- ping generative adversarial networks for brain MR to PET synthesis,” IEEE Transactions on Medical Imag- ing, vol. 41, no. 1, pp. 145–157, 2021
2021
-
[12]
MRI to PET Cross-Modality Translation using Globally and Lo- cally Aware GAN (GLA-GAN) for Multi-Modal Di- agnosis of Alzheimer’s Disease,
Apoorva Sikka, Virk, et al., “MRI to PET Cross-Modality Translation using Globally and Lo- cally Aware GAN (GLA-GAN) for Multi-Modal Di- agnosis of Alzheimer’s Disease,” arXiv preprint arXiv:2108.02160, 2021
2021 arXiv
-
[13]
Unpaired image-to-image translation using cycle-consistent adversarial networks,
Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE international conference on computer vi- sion, 2017, pp. 2223–2232
2017
-
[14]
PASTA: Pathology-Aware MRI to PET Cross-Modal Translation with Diffusion Mod- els,
Yitong Li, Igor Yakushev, Dennis M Hedderich, and Christian Wachinger, “PASTA: Pathology-Aware MRI to PET Cross-Modal Translation with Diffusion Mod- els,” arXiv preprint arXiv:2405.16942, 2024
2024 arXiv
-
[15]
FREA- UNet: frequency-aware U-Net for modality transfer,
Hajar Emami, Qiong Liu, and Ming Dong, “FREA- UNet: frequency-aware U-Net for modality transfer,” arXiv preprint arXiv:2012.15397, 2020
2012 arXiv
-
[16]
Ganbert: Gen- erative adversarial networks with bidirectional encoder representations from transformers for mri to pet synthe- sis,
Hoo-Chang Shin, Alvin Ihsani, et al., “Ganbert: Gen- erative adversarial networks with bidirectional encoder representations from transformers for mri to pet synthe- sis,” arXiv preprint arXiv:2008.04393, 2020
2008 arXiv
-
[17]
Plasma phosphorylated tau 217 and Aβ42/40 to predict early brain A β accumulation in people without cogni- tive impairment,
Shorena Janelidze, Nicolas R Barth ´elemy, et al., “Plasma phosphorylated tau 217 and Aβ42/40 to predict early brain A β accumulation in people without cogni- tive impairment,” JAMA neurology, vol. 81, no. 9, pp. 947–957, 2024
2024
-
[18]
Blood-based biomarkers for Alzheimer’s disease,
Antoine Leuzy, Niklas Mattsson-Carlgren, et al., “Blood-based biomarkers for Alzheimer’s disease,” EMBO molecular medicine , vol. 14, no. 1, pp. e14408, 2022
2022
-
[19]
Blood-based biomarkers for Alzheimer disease: map- ping the road to the clinic,
Harald Hampel, Sid E O’Bryant, Molinuevo, et al., “Blood-based biomarkers for Alzheimer disease: map- ping the road to the clinic,” Nature Reviews Neurology, vol. 14, no. 11, pp. 639–652, 2018
2018
-
[20]
Combin- ing blood-based biomarkers and structural mri measure- ments to distinguish persons with and without signifi- cant amyloid plaques,
Yanxi Chen, Yi Su, Jianfeng Wu, et al., “Combin- ing blood-based biomarkers and structural mri measure- ments to distinguish persons with and without signifi- cant amyloid plaques,” Journal of Alzheimer’s Disease, vol. 98, no. 4, pp. 1415–1426, 2024
2024
-
[21]
Image-to-image translation with conditional ad- versarial networks,
Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros, “Image-to-image translation with conditional ad- versarial networks,” in Proceedings of the IEEE confer- ence on computer vision and pattern recognition , 2017, pp. 1125–1134
2017
-
[22]
Joint learn- ing framework of cross-modal synthesis and diagnosis for Alzheimer’s disease by mining underlying shared modality information,
Chenhui Wang, Sirong Piao, Huang, et al., “Joint learn- ing framework of cross-modal synthesis and diagnosis for Alzheimer’s disease by mining underlying shared modality information,” Medical Image Analysis , vol. 91, pp. 103032, 2024
2024
-
[23]
Diagnostic accuracy of a plasma phosphorylated tau 217 immunoassay for Alzheimer disease pathology,
Nicholas J Ashton, Wagner S Brum, Di Molfetta, et al., “Diagnostic accuracy of a plasma phosphorylated tau 217 immunoassay for Alzheimer disease pathology,” JAMA neurology, vol. 81, no. 3, pp. 255–263, 2024
2024
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.