acceptodds
Under review as a conference paper at ICLR 2027

UniCTPET: Unifying 3D CT and PET Generation through Adaptive Structural Exchange

Abstract

Generative modeling offers a promising way to alleviate the scarcity of paired CT/PET data, yet existing methods typically address different generation tasks with separate models, limiting knowledge reuse across tasks. Unifying these tasks requires accommodating heterogeneous structural constraints and cross-modal dependencies that differ in direction and evolve throughout generation. We propose a unified 3D latent flow-matching framework that supports single-modality, structurally conditioned, CT-to-PET, and joint CT/PET generation within one model, without task-specific retraining. At its core, Time Task Structure Exchange (TTSE) adaptively regulates cross-modal structural interactions using modality feature states and flow time, accommodating one-way conditioning in CT-to-PET synthesis and bidirectional coordination in joint generation without explicit task representations. Complementarily, Skeletal Lesion Guidance (SLG) provides spatial control by injecting multi-scale skeletal and lesion conditions through lightweight encoders, without replicating the backbone encoder. Experiments on three whole-body PET/CT datasets comprising 1,599 paired scans demonstrate improved generation fidelity and downstream lesion segmentation over competing generative methods. Averaged across datasets, our framework improves PET PSNR in joint generation by 3.10 dB over the strongest competing baseline, while TTSE improves CT-to-PET PSNR by 2.15 dB over standard cross-attention. These findings highlight adaptive structural exchange as an effective mechanism for unifying generation tasks with heterogeneous conditions and modality dependencies.

Primary area:
generative models
Keywords:
Medical Image Generation, CT–PET, Multimodal Generation, Cross-Modal Synthesis, 3D Generative Model, Rectified Flow

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_B1JtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

UniCTPET is a shared 3D latent flow-matching framework for single-modality, structurally conditioned, CT-to-PET, and joint CT/PET generation. It introduces TTSE for time- and feature-dependent cross-modal exchange and SLG for skeletal and lesion guidance. Experiments on three datasets demonstrate improved generation quality and downstream lesion segmentation over competing methods.

Critical strengths
  1. Useful task unification. The framework accommodates different input-output configurations within one model, with evaluation on three large-scale datasets comprising 1,599 paired scans.

  2. Comprehensive experimental evaluation. The paper covers multiple generation tasks and downstream lesion segmentation, providing extensive empirical evidence.

Critical weaknesses
  1. Limited Novelty and Insight. The motivation for unifying different CT/PET generation tasks is reasonable. However, the proposed methods mainly build on existing techniques such as cross-attention, feature gating, and multi-scale conditioning. The paper lacks a sufficiently compelling new insight beyond combining these components, making the methodological novelty somewhat incremental for ICLR.

  2. Insufficient Validation of the Core Contribution. TTSE is the central contribution, but its adaptive structural exchange mechanism is not thoroughly validated or understood. The current ablations mainly compare TTSE with standard cross-attention, without isolating its key design choices. Although Fig. 1 provides some analysis of exchange strength and gradient alignment, the paper lacks deeper diagnostics explaining how the adaptive exchange works, why it improves performance, and whether it actually reduces interference between generation tasks.

Questions for authors
  1. What exactly is conditioned on during evaluation? Please specify the skeletal and lesion inputs available to every method in Tables 2–3 and explain how reference images are assigned for PSNR/SSIM evaluation.

  2. More detailed ablations are needed to validate TTSE, including fixed, time-only, and feature-only gating, as well as removing shared or differential structural features. The authors should also analyze the learned gating behavior across generation tasks and flow timesteps.

  3. As I understand it, the paired conditional generation tasks are closely related to medical image translation. The authors should include more representative medical image translation baselines, rather than focusing mainly on generative models, to provide a more comprehensive comparison.

  4. More details on the VQGAN tokenizer training are needed, including its architecture, loss functions, and training configurations. It would also be helpful to report reconstruction performance (e.g., PSNR and SSIM) for CT and PET.

  5. The motivation and novelty of SLG are less convincing than those of TTSE. Its differences from existing conditioning approaches should be better explained.

  6. The paper claims that SLG is lightweight, but provides no computational evidence. A comparison with ControlNet in terms of parameters, GPU memory, and computational cost is needed.

  7. The motivation for the two-stage training strategy is unclear. Why is TTSE trained separately rather than jointly with the backbone? The choice of 200k and 5k training steps also needs justification, ideally through comparisons with alternative training schedules.

6 comments

Thank you for your detailed comments. We clarify the methodological and implementation details below and outline our planned additional analyses and experiments.

R1. Conditioning inputs and evaluation. For structurally conditioned generation in Tables 2–3, the generative baselines receive skeletal information and lesion masks through ControlNet, while our method uses SLG to incorporate the same conditions. The cross-modal methods, IL-CLDM and CoCoLIT, instead use VQGAN-encoded CT latents as conditioning. PSNR/SSIM are computed against the corresponding real CT or PET image associated with the conditioning inputs, over the full volume, including background and brain regions.

R2. TTSE ablations and gating behavior. The original submission compares TTSE with standard cross-attention. We are conducting additional ablations with fixed, time-only, and feature-only gating, as well as removing shared or differential structural features. We are also analyzing gating behavior across CT-to-PET and joint generation at different flow timesteps, and will report the results before the rebuttal/discussion period ends.

R3. Medical image translation baselines. The original submission includes IL-CLDM (IEEE JBHI, 2025; published online in 2024) and CoCoLIT (AAAI 2026) as CT-to-PET translation baselines. Table 3 reports PSNR/SSIM of 16.38 dB/0.173 for IL-CLDM and 17.33 dB/0.621 for CoCoLIT, compared with 36.99 dB/0.943 for our method. We also plan to add Uni-Brain (MICCAI 2026) as an additional baseline and report the results before the rebuttal/discussion period ends.

R4. VQGAN configuration and reconstruction quality. CT and PET use separate 3D VQGANs with residual blocks, four-head bottleneck self-attention, and 4× downsampling along each spatial axis. Their hidden widths are 96/48, respectively; both use 24 normalization groups, an 8,192-entry codebook, and four latent channels. Parameter counts excluding discriminators are 30.87M/7.78M. A 64³ patch maps to a 4×16×16×16 latent tensor. The flow model uses continuous pre-quantization latents, with quantization applied during decoding.

Each VQGAN was trained for 200 epochs on random 64³ crops with batch size 1, using Adam (initial learning rate 1e-4, betas 0.5/0.9). The learning rate decreased to 5e-5 and 1e-5 after epochs 20 and 40. The generator objective is:

LG=L1+0.3Lperc+0.25MSE⁡(ze,sg⁡(zq))+0.1Ladv.L_G=L_1+0.3L_{\mathrm{perc}}+0.25\operatorname{MSE}(z_e,\operatorname{sg}(z_q))+0.1L_{\mathrm{adv}}.

The perceptual term uses VGG-based LPIPS on slices sampled along three axes; the adversarial term uses least squares. The codebook is updated by EMA with decay 0.99.

Across 321 volumes per modality, reconstruction PSNR/3D SSIM are 37.06±1.28 dB / 0.975486±0.010472 for CT and 54.70±2.46 dB / 0.995977±0.000601 for PET (mean±SD). Metrics use full volumes normalized to [0,1], including background, with PSNR data range 1 and a 7³ SSIM window.

R5. SLG motivation. Our motivation is to unify conditional and unconditional generation within one model, allowing structural guidance to be introduced as needed while preserving a reusable generative backbone. ControlNet is commonly used for such guidance, but requires duplicating the full U-Net encoder and middle block. We designed SLG to provide sufficient skeletal and lesion guidance through independent, lightweight condition encoders that inject features directly into the backbone.

R6. SLG efficiency. We measured SLG and ControlNet as standalone conditioning modules, excluding the backbone and other model components.

MetricSLGControlNet
Parameters28.11M635.13M
Peak inference memory1.83 GiB4.33 GiB
Peak training memory14.27 GiB23.99 GiB

SLG reduces parameters by 95.6%, peak inference memory by 57.7%, and peak training memory by 40.5%.

R7. Training schedule. The backbone was trained for 200k steps, and TTSE and the baseline ControlNet modules were each trained for 50k steps. The “5k” reported in the original manuscript is a typographical error, which we will correct.

Training TTSE with the backbone fixed preserves the established generation capability while learning cross-modal interactions. We also completed joint training from the outset: CT-to-PET reaches 40.18 dB PSNR and 0.97496 SSIM, but unconditional joint generation collapses. Images illustrating this collapse are available in our anonymous GitHub repository.

R3. Additional medical image translation baseline. Following our previous response, we have completed the Uni-Brain (MICCAI 2026) comparison. We evaluated 321 cases using 50-step DDIM sampling. Uni-Brain achieves an overall PSNR of 36.885±2.380 dB and SSIM of 0.93736±0.00707 (mean±SD), both slightly lower than our model’s 36.99 dB PSNR and 0.943 SSIM. Dataset-specific results are reported below.

DatasetCasesPSNR (dB)SSIM
DeepPSMA2034.442±2.4110.93820±0.00611
FDG18137.834±2.0370.93456±0.00672
PSMA12035.862±2.1460.94144±0.00559
Overall32136.885±2.3800.93736±0.00707

Thank you for your comprehensive rebuttal, which has addressed most of my concerns. However, the methodological novelty of the paper remains relatively modest. Therefore, I decide to maintain my original rating.

Reviewer_Atm8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UniCTPET, which is a unified 3D latent flow-matching framework that support single-modality, structurally conditioned, CT-to-PET, and joint CT/PET generation within single model, without task-specific retraining.

Its key techniques include: (1) Time Task Structure Exchange (TTSE), which adaptively adjusts the exchange of structural information between the CT and PET branches based on modal feature states and flow time. (2) Skeletal Lesion Guidance (SLG), which injects multi-scale skeletal and lesion information into the generative backbone via a lightweight conditional encoder.

The experiments covered three whole-body PET/CT datasets, comprising a total of 1,599 scan pairs, and were validated using image quality metrics, ablation experiments and lesion segmentation tasks trained on generated images. The results demonstrate that the proposed method outperforms the compared generative methods across multiple single-modality, joint and CT-to-PET generative metrics.

Critical strengths
  1. From the methodology perspective, unified task design is highly valuable. UniCTPET covers four input–output configurations within a single shared model, utilising the spatial and temporal states of different modalities to express unidirectional CT-to-PET conditional dependencies and bidirectional joint generation. TTSE and SLG address cross-modal exchange and external structural control respectively; the overall design aligns well with the problems posed in the paper. If this framework is fully validated, it could reduce the need to maintain independent models for each task, thereby having a positive impact on the final recommendations.

  2. The empirical evaluation is broad and systematic. The paper evaluates the methods on three PET/CT datasets with different tracers, disease compositions and data scales, reporting results for PSNR, SSIM, NMAE, NMSE, NCC, downstream Dice/HD95 and ablation results. Notably, TTSE achieves a 2.15 dB improvement in PSNR for CT-to-PET compared to standard cross-attention, the replacement experiments involving SLG and TTSE provide evidence that both modules contribute to performance. These results indicate that this method warrants further investigation.

Critical weaknesses
  1. The experiments demonstrate task unification, but do not isolate the benefits of unified training or adaptive structural exchange. In the results, it shows that UniCTPET can perform multiple generation tasks within a single model, and achieve strong task-level performance relative to the evaluated baselines. However, this paper has not yet validated the specific benefits of unified model through controlled comparisons. Particularly, this work has not compared the unified model with single-task models that use the same backbone network, model capacity, data and training budget but were trained separately.
    Similarly, the evaluation of TTSE was conducted solely by replacing it entirely with standard cross-attention. The respective contributions of flow time, model state, adaptive gating and the model of structural differences were not analysed separately. Furthermore, the paper does not report the number of parameters, memory consumption, training time or inference cost, meaning that the efficiency advantages claimed for its unified and lightweight design cannot be quantified.

  2. The current evaluation does not directly validate cross-modal consistency in joint CT/PET generation. The evaluation of joint generation primarily employs independent CT and PET metrics, such as PSNR and SSIM. These metrics measure the similarity of each modality to its corresponding reference image, but do not directly assess whether anatomical structures, lesion locations and tracer uptake patterns are consistent between the generated CT/PET image pairs. Consequently, the current evaluation provides only indirect evidence for the claim that TTSE achieves bidirectional cross-modal coordination. The downstream Dice/HD95 results provide indirect evidence that the generated images preserve lesion-related information. However, because separate segmentation models appear to be trained for CT and PET, these results assess modality-specific utility rather than consistency within a generated CT/PET pair. Although the t-SNE analysis in Figure 4 provides useful distribution-level visualisation, it cannot replace a direct assessment of pairwise cross-modal consistency.

Questions for authors
  1. Can the authors provide a controlled comparison between UniCTPET and separately trained single-task counterparts matched in backbone architecture, model capacity, data, and training budget? Such a comparison would clarify whether unified training provides knowledge sharing or positive transfer, beyond allowing multiple tasks to be performed by one model.

  2. How do the authors directly evaluate the anatomical and functional consistency within jointly generated CT/PET pairs? The current per-modality PSNR and SSIM results assess CT and PET separately, but do not establish whether lesion locations, anatomical structures, and tracer-uptake patterns are mutually consistent across modalities.

  3. Were all training, validation, and test sets split at the patient level? If any patient had multiple scans, how was overlap between the splits prevented?

6 comments

Thank you for your constructive feedback. We respond to each comment below and outline the additional experiments planned to address your concerns.

R1. Unified versus single-task training. The original submission already includes DDPM and DDIM single-modality generation baselines using the same backbone architecture and configuration as our method. We plan to extend this comparison to separately trained single-task flow-matching models with the same backbone and configuration, and report the results in an updated response before the rebuttal/discussion period ends.

R2. Cross-modal consistency. We are arranging a physician assessment of jointly generated CT/PET pairs, with consistency scores covering anatomical correspondence, lesion localization, and tracer uptake relative to CT anatomy. We will update the evaluation protocol and results before the rebuttal/discussion period ends.

R3. Dataset partitioning. FDG and PSMA were split at the patient level, with all scans from each patient assigned to the same partition. FDG contains 811 unique patients, with 649/162 allocated to training/testing; PSMA contains 378 unique patients, with 302/76 allocated to training/testing, following an approximately 4:1 ratio. DeepPSMA does not provide patient identifiers, so we used a random sample-level split of 78/20 for training/testing.

R1. Unified versus single-task training (additional results). Following up on our previous response, we have now completed the additional experiments with separately trained single-task flow-matching models for conditional single-modality generation. These models use the same backbone architecture and experimental settings as our method, complementing the DDPM and DDIM baselines already included in the original submission. Results across the three datasets are shown below, with brackets indicating the 95% confidence intervals of the means.

\hline \text{Dataset} & \text{Modality} & \text{PSNR (dB)} & \text{SSIM} \\\\ \hline \text{DeepPSMA} & \text{CT} & 24.62\\;[23.63,\\,25.61] & 0.8441\\;[0.8103,\\,0.8778] \\\\ \text{DeepPSMA} & \text{PET} & 37.22\\;[36.09,\\,38.35] & 0.9754\\;[0.9700,\\,0.9809] \\\\ \hline \text{FDG} & \text{CT} & 22.35\\;[21.16,\\,23.54] & 0.7357\\;[0.6966,\\,0.7749] \\\\ \text{FDG} & \text{PET} & 41.30\\;[39.46,\\,43.15] & 0.9650\\;[0.9582,\\,0.9719] \\\\ \hline \text{PSMA} & \text{CT} & 26.53\\;[24.91,\\,28.14] & 0.9029\\;[0.8661,\\,0.9396] \\\\ \text{PSMA} & \text{PET} & 40.39\\;[38.62,\\,42.17] & 0.9804\\;[0.9762,\\,0.9846] \\\\ \hline \end{array}$$ **R2. Cross-modal consistency (additional results).** Following up on our previous response, we have now completed a radiologist assessment of jointly generated CT/PET pairs. A radiologist scored lesion consistency in the generated CT and PET images, as well as cross-modal consistency between the paired volumes. The detailed results are presented in Figure 7 of the latest revised manuscript PDF.

Thank you for your rebuttal and additional experiments. The new single-task baselines and radiologist assessment of cross-modal consistency improve the experimental validation, which address my concerns. The explanation of R3 (data partitioning) looks reasonable for me. However, I still concern for the specific benefits of unified training and individual contributions of TTSE. For now, i will maintain my original score, but i remain open to further discussion and will consider the other reviewers' perspectives before making my final decisions.

Reviewer_zeuVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UniCTPET, a unified 3D flow-matching framework for CT/PET generation that aims to support four tasks within a single shared backbone: single-modality CT generation, single-modality PET generation, CT→PET conditional generation, and bi-modal joint CT/PET generation. The method maps CT and PET volumes into latent space with pretrained VQGAN encoders, models them with a shared 3D U-Net flow-matching backbone distinguished by modality embeddings, and introduces two modules: Time Task Structure Exchange (TTSE), which adaptively regulates cross-modal information exchange using structural features and flow time without an explicit task representation, and Skeletal Lesion Guidance (SLG), a lightweight conditional guidance module that injects skeletal and lesion-mask features without replicating the full U-Net encoder. The claimed contributions are the unified multi-task framework, TTSE for alleviating exchange-pattern conflicts, and SLG for efficient structural conditional guidance. Evaluation is reported on AutoPET-FDG, AutoPET-PSMA, and DeepPSMA for generation quality and downstream lesion segmentation.

Critical strengths

Strength 1: The problem framing—unifying single-modality, conditional cross-modal, and joint bi-modal generation under one backbone—is practically relevant and conceptually clean. CT/PET workflows often require different input–output configurations, and training a separate model for each is wasteful. A shared generative backbone with task-adaptive cross-modal exchange is a sensible direction, and the paper correctly identifies that CT→PET generation and joint generation impose different information-transfer requirements.

Strength 2: TTSE is a plausible and non-trivial mechanism for time- and structure-dependent cross-modal exchange. Conditioning the exchange gate on both modality-specific structural features and flow-time embeddings, rather than using a fixed cross-attention or an explicit task ID, is a reasonable way to handle one-way versus bidirectional dependencies. The ablation against standard cross-attention, if the numbers are verified, would support the claim that adaptive exchange strength matters.

Critical weaknesses

Weakness 1: The provided manuscript is materially corrupted, and the central quantitative table is missing, which prevents evaluation of the paper’s main empirical claims. Page 1 consists entirely of line numbers, Pages 4 and 5 are largely repeated indices, and Table 2—the main quantitative comparison for single-modality CT and PET generation—contains only the header row in the provided file. Table 3 is described in the text but the table itself is not readable. Several parts of Sections 3.2–3.3 and the appendix are truncated or corrupted. This is not a cosmetic issue: the paper’s headline claims include large quantitative improvements (e.g., CT PSNR gains of 1.60–1.89 dB, PET PSNR gains of 1.21–1.64 dB, and a CT→PET PSNR gain of 19.66 dB over CoCoLIT), and without the actual tables these claims cannot be verified. For a methods paper whose contribution is primarily empirical, this is a serious evaluability failure and supports rejection in the current form.

Weakness 2: The empirical evidence does not sufficiently establish that the unified framework and TTSE are responsible for the reported gains, and several results are implausibly large or inconsistent with the downstream evaluation. The CT→PET improvement over CoCoLIT—19.66 dB PSNR and 0.322 SSIM—is far outside the range typically observed for a fair comparison and suggests a possible mismatch in task setup, conditioning, evaluation protocol, or baseline configuration. At the same time, downstream CT lesion segmentation Dice is extremely low (0.0754–0.1680), which weakens the practical value of CT synthesis for lesion-level downstream use. The paper also does not compare against recent unified or multi-task generation methods such as MedSyn or TotalGen, despite discussing them in related work. Moreover, the shared backbone still uses modality-specific VQGANs and modality embeddings, and TTSE is trained only in a second stage after the backbone is frozen; the contribution of TTSE relative to the shared backbone and the two-stage training schedule is not cleanly isolated. For a paper whose central claim is that one unified framework can handle all four tasks without task-specific retraining, these issues make the empirical case less convincing than the paper claims.

Questions for authors
  1. The provided manuscript has corrupted pages and missing table bodies, including Table 2 and Table 3. Can you provide a complete, readable version? Without the actual quantitative results, the paper’s central claims cannot be assessed.

  2. The reported CT→PET gain over CoCoLIT is 19.66 dB PSNR and 0.322 SSIM. Can you clarify the exact task setup, conditioning inputs, training data, and evaluation protocol used for CoCoLIT and for your method? A gain of this magnitude requires a careful explanation to rule out an unfair comparison.

  3. Downstream CT lesion segmentation Dice is very low (0.0754–0.1680). What does this imply about the practical utility of the generated CT images for lesion-level tasks, and how do these numbers compare with training on real CT or on images from the strongest baseline?

  4. Why are recent unified or multi-task generation methods such as MedSyn and TotalGen not included in the comparison? Since the paper’s main claim is unification across tasks, comparisons against other unified methods would materially affect the assessment.

  5. TTSE is trained only in Stage 2 after the backbone is frozen. How much of the improvement comes from TTSE itself versus from the two-stage training schedule and the shared backbone? Can you report a variant where TTSE is trained jointly with the backbone from the start?

  6. The framework still uses separate CT and PET VQGANs and modality embeddings. In what sense is the latent space “shared,” and how does this affect the claim of a unified generative model?

  7. Can you report model size, training time, inference latency, and GPU memory for the unified framework relative to training separate task-specific models? This is important for evaluating whether the unification actually reduces overall cost.

3 comments

Thank you for reviewing our manuscript. We address your comments below and outline our planned additional experiments.

R1. Manuscript integrity. We have verified that the originally submitted manuscript is intact, and Tables 2 and 3 are complete and readable. Please refer to these tables in the original PDF for the full quantitative results.

R2. Comparison with CoCoLIT. PSNR/SSIM were computed over the full image, including the background. CoCoLIT and IL-CLDM exhibit substantial background errors on our datasets, which enlarge the measured performance gap. For CoCoLIT, we trained the image-generation stage for 200k iterations and the ControlNet stage for 50k iterations, using VQGAN-encoded CT latents as conditioning. Both stages used batch size 1 and learning rate 0.00002, with the learning rate matching the official implementation. We also tried longer training and larger batch sizes without obtaining better results.

R3. CT lesion-segmentation performance. The original manuscript reports segmentation results from training on real data in the first row of Table 3. Real CT and most generated-image baselines also yield low Dice scores; the real-data scores are 0.287 on FDG and only 0.0823 and 0.1032 on PSMA and DeepPSMA, respectively. This is because, in tumor detection, CT provides anatomical information, while PET provides metabolic information that is more informative for lesion detection.

R4. Comparison with unified generation methods. MedSyn and TotalGen were both included in the original submission. Table 2 reports generation quality for each modality, while Table 3 reports segmentation performance on real datasets using the generated images and the overall score for joint generation.

R5. TTSE and two-stage training. We use a shared backbone to support unified generation across tasks, together with a two-stage schedule designed to limit gradient interference among heterogeneous tasks and preserve the backbone’s general generation capability, avoiding specialization toward individual tasks. TTSE addresses the distinct information-exchange requirements of CT-to-PET and joint generation by adapting cross-modal interactions to feature states and flow time, accommodating both one-way conditioning and bidirectional coordination.

We completed joint training from the outset. CT-to-PET improves from 36.99 to 40.18 dB PSNR and from 0.943 to 0.97496 SSIM, while conditional FDG PET generation improves from 37.78 to 41.56 dB PSNR. However, conditional CT generation decreases from 28.85/27.10/30.91 to 25.53/23.15/27.08 dB on DeepPSMA/FDG/PSMA, respectively. Conditional PET generation on DeepPSMA/PSMA also decreases from 42.67/41.27 to 38.89/39.97 dB. Meanwhile, unconditional joint generation completely collapses. Images illustrating this collapse are available in our anonymous GitHub repository.

R6. Latent spaces and unified generation. Sharing occurs across tasks within each modality: the same CT and PET tokenizers and their respective latent spaces are reused for single-modality, structurally conditioned, CT-to-PET, and joint generation. Across modalities, CT and PET retain separate VQGAN latent spaces, with modality embeddings identifying each modality. Our unified-generation claim refers to jointly modeling these modality-specific representations and supporting multiple generation tasks within one trained model, without task-specific retraining.

R7. Computational cost and task reuse. Our generative model has 744.73M parameters and uses 15 sampling steps. Full sampling takes 5.45 s for single-modality PET generation, 9.98 s for CT-to-PET, and 11.38 s for joint CT/PET generation. Even joint generation, our most time-consuming setting, samples faster than every compared baseline. Training and inference memory range from 34.06–61.07 GiB and 4.68–5.36 GiB, respectively. Complete parameter, timing, and memory comparisons are reported in Table 5 at the end of the supplementary material, included in the updated PDF.

The existing baselines are task-specific models that cannot be reused across tasks with different generation configurations. Our method avoids repeated training across these configurations by supporting multiple tasks within one trained model.

Dear Reviewer zeuV,

We would appreciate your feedback on our response above, particularly regarding the manuscript readability concern. Our response points to the relevant pages and tables in the original submission, including the MedSyn and TotalGen comparisons. Could you please let us know whether you are now able to access and read these sections? We would be happy to clarify any remaining issues.

Thank you for your time and consideration.

Dear Reviewer zeuV,

We would like to briefly follow up on our earlier response regarding manuscript readability and the baseline comparisons. Could you please let us know whether the relevant pages and tables in the original submission are now readable, and whether any concerns remain after considering our clarifications? We would be happy to address any specific outstanding issues.

Thank you again for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.