{"id":"962c40d8-ecba-4c06-9c53-c5097043d87a","arxiv_id":"2507.19874","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DiffCode uses a task-adaptive vector-quantized codebook bank plus latent diffusion to improve all-in-one medical image restoration, showing gains on MRI super-resolution, CT denoising, and PET synthesis.","lead":"This paper introduces DiffCode, a three-stage framework that combines vector-quantized codebooks with latent diffusion to restore medical images across MRI, CT, and PET in one unified model. If it works as claimed, it could simplify clinical image processing pipelines by replacing separate task-specific restoration tools with a single adaptable system.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed latent diffusion enhancement of codebook retrieval is supported only by a ~0.1 dB end-task gain, with no direct retrieval-accuracy or distribution-alignment evidence; the weak 5-layer/T=8 denoiser may not provide the asserted mapping power.","rationale":"The reader's weakest assumption identified the latent diffusion component as unverified, and my stress-test converges on the same point, sharpened: the only evidence for the mechanism is a small end-task metric gain, while the network capacity and step count (5 linear layers, T=8) make the asserted 'powerful mapping capability' questionable. This is load-bearing because the paper's contribution is explicitly framed as 'latent diffusion-enhanced prior retrieval'; if retrieval accuracy is not improved, the central claim is not established. The proposed concrete test directly measures retrieval success and distribution alignment, which would settle the causal role of the diffusion stage. I also note the stochastic inference and lack of variance reporting, which compounds the small margins. However, the weaknesses are addressable with additional analysis, and the architecture and ablations are otherwise coherent, so the reader's CONDITIONAL verdict remains appropriate; I do not recommend changing it to accept or reject based on this review pass.","tokens_in":8974,"tokens_out":6856,"duration_ms":74645,"concrete_test":"Reproduce Stage II and, on the MRI test set (or a public subset), compute the codebook retrieval accuracy before and after the latent diffusion: encode each HQ test image with E_vq to obtain ground-truth code indices (via RQ), then, from the corresponding LQ image, retrieve code indices using (a) raw LQ features without diffusion and (b) the diffusion-refined features after the full T=8 reverse process. Report top-1 index match rate, and also compare the feature distribution distances (e.g., MMD or FID) between the estimated features and the ground-truth z_n features, with and without diffusion. In addition, run the full DiffCode inference with at least 5 different random seeds for the diffusion initialization and report mean ± std of PSNR; if the 95% confidence interval of the V3-vs-DiffCode gap overlaps zero, the claimed enhancement is not significant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is the claim that a latent diffusion model 'enhances prior retrieval' from the task-adaptive codebook bank (Secs. 1 and 2.2). The only quantitative support is the final PSNR/SSIM/RMSE difference in Table 2(a): adding the latent diffusion stage changes average PSNR from 34.5273 (V3) to 34.6186 (DiffCode), a gain of 0.0913 dB, with no error bars or significance testing. No direct measurement of codebook retrieval accuracy (e.g., whether retrieved code indices match ground-truth HQ codes) or of the estimated feature distribution before vs. after diffusion is reported. This matters because the diffusion denoiser is specified as '5 linear layers' with T=8 total steps and β linearly increasing to 0.99 (Sec. 3.1). Such a small, nearly linear model may be unable to 'iteratively refine the latent feature distribution' as claimed; the observed PSNR gain could instead come from the added stochasticity or extra parameters acting as a regularizer. If the latent diffusion does not actually improve retrieval, the paper's stated mechanism is unsupported, and the architecture reduces to a VQ codebook bank plus TARS, which is a smaller increment over prior work (e.g., AMIR). The stochastic initialization from random Gaussian noise at inference also introduces run-to-run variability, and since no seeds or variance are reported, the small margins in Table 1 (best average PSNR 34.6186 vs. AMIR 34.3585) may not be reproducible. The concern is not that the method is wrong, but that the evidence for the central causal claim is missing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DiffCode, a three-stage framework for all-in-one medical image restoration. Stage I builds a task-adaptive vector-quantized codebook bank with residual quantization on high-quality images for each task. Stage II trains a small latent diffusion model, conditioned on low-quality features, to estimate features that are then quantized against the codebook bank; the stated goal is to improve prior retrieval by refining the latent feature distribution. Stage III decodes the retrieved codebook features into reference images and concatenates them with the low-quality input to guide a task-aware routing restoration backbone. The authors report state-of-the-art results on MRI super-resolution, CT denoising, and PET synthesis, with average PSNR 34.6186, SSIM 0.9336, and RMSE 12.3728, and they provide component ablations in Table 2.","tokens_in":9378,"tokens_out":5043,"duration_ms":51400,"significance":"If the results are reproducible and the diffusion mechanism is genuinely improving prior retrieval, the task-adaptive codebook bank is a sensible and potentially useful contribution to all-in-one medical image restoration, where task heterogeneity is a recognized challenge. The paper is clearly written, the architecture is specified in detail, and the experimental comparison against eleven methods across three tasks is comprehensive. The component ablation shows that each module adds some quantitative benefit. However, the central novelty of the paper is the latent diffusion stage, and that stage is currently supported only by a small end-task metric gain with no direct evidence about retrieval behavior or distribution alignment; this weakens the significance of the claimed mechanism.","major_comments":[{"comment":"The central claim that the latent diffusion model 'enhances prior retrieval' is supported only by a 0.0913 dB average PSNR improvement (V3: 34.5273 vs. DiffCode: 34.6186) and a 0.0013 SSIM improvement, reported without error bars or significance tests. No direct evidence is provided that the diffusion process changes which codebook entries are retrieved or that the estimated features align with ground-truth HQ features. This is load-bearing because the diffusion stage is the paper's main novelty; I request direct retrieval analysis (e.g., retrieved code index agreement with the HQ code indices, feature-space distance or MMD before vs. after diffusion) and repeated-seed statistics to rule out that the gain comes from added stochasticity or parameter count rather than from improved retrieval.","section":"Sec. 2.2 and Table 2(a)"},{"comment":"At inference the reverse diffusion is initialized from randomly sampled Gaussian noise, but the experimental tables report a single run with no seed or variance information. Given that the largest reported advantage over the strongest baseline AMIR is 0.2601 dB average PSNR (Table 1), run-to-run variability could plausibly change the ranking. The authors should report mean and standard deviation over at least three to five seeds for all main comparisons and ablations, or fix a seed and state it explicitly.","section":"Sec. 2.2, Eq. (5)"},{"comment":"The training objective L_stage2 = ||z^n - \\hat z^n||_1 trains the denoiser to map Gaussian noise to HQ features conditioned on LQ features, and at inference the reverse process starts from pure noise. This is not a 'refinement' of the LQ-degraded distribution in the sense of gradually correcting an input feature; it is a conditional generation from noise. The paper does not explain why this procedure should land near the codebook manifold, and no codebook reconstruction loss or nearest-neighbor constraint is used in Stage II. I ask the authors to show that \\hat z^n lies close to the codebook manifold (e.g., nearest-neighbor distances before and after diffusion) and to ablate the initialization choice (Gaussian noise vs. LQ feature plus noise).","section":"Sec. 2.2, Eq. (6)"},{"comment":"The description 'subsampling full scans with a 12x dose reduction, and reconstructed using the standard OSEM method' omits essential details: how the sinogram subsampling is performed, the OSEM iteration and subset counts, whether any post-filtering is applied, and how the high-quality reference is exactly defined. Without these details the PET synthesis result cannot be reproduced or compared fairly against the cited prior work [9], and the quantitative gain on that task is not fully verifiable.","section":"Sec. 3.1, PET Synthesis dataset"}],"minor_comments":[{"comment":"The claim of being 'the first to exploit the VQ codebook prior for all-in-one MedIR' should be supported by a brief discussion of prior VQ-based restoration works and why none of them addresses the all-in-one setting; otherwise the novelty claim is too strong.","section":"Sec. 1, Contribution 1"},{"comment":"Several labels in the overview figure appear as unreadable '???' placeholders, particularly around the equations and block annotations; please provide a clean vector version so that the architecture can be verified.","section":"Fig. 1"},{"comment":"The implementation description gives NAF block counts and channels but omits the exact structure of the condition encoder E_lq, the number of learnable parameters, and the total compute or storage overhead of the codebook bank and diffusion stage; adding a parameter/FLOPs table would strengthen the efficiency claims in the Introduction. Also, reference [18] is missing its publication venue (ICLR 2017).","section":"Sec. 3.1 and References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in the style of a strong MICCAI-style paper and follows the authors' prior AMIR work. The main concern is that the headline novelty, the latent diffusion stage, currently has only a small end-task margin and no direct retrieval-level evidence; this is fixable with additional experiments rather than being an inherent flaw. I would reject only if the authors cannot produce such evidence in revision. No citation-pattern concerns beyond the 'first' claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhiwei,\n\nQuick take: this is a real, if modest, step forward in all-in-one medical image restoration. The novel part is the task-adaptive codebook bank — per-task VQ codebooks with residual quantization — plus a latent diffusion module that is supposed to make retrieval from that bank more faithful. The authors report steady gains over eleven baselines across MRI super-resolution, CT denoising, and PET synthesis, and their ablations show each of the three components contributes something. The writing is clear, the experimental setup is standard for the field, and the baselines are fair. If you work on MedIR, this gets your attention.\n\nThe soft spot is the evidence for the central mechanism. The claim that latent diffusion 'enhances prior retrieval' is supported only by a 0.09 dB difference in Table 2(a) (V3 vs DiffCode: 34.5273 vs 34.6186 PSNR), with no error bars, no significance test, and no direct measurement of retrieval accuracy or feature alignment. The diffusion model is small — five linear layers, eight steps, beta up to 0.99 — so it is not obvious it has the 'powerful mapping capability' the paper asserts. The gain could be regularization or just extra parameters. Worse, inference starts from randomly sampled Gaussian noise, so the reproducibility of that margin is an open question. This is a load-bearing missing analysis, not a nitpick: the diffusion module is the differentiator from AMIR.\n\nMinor issues: no code, thin details on the PET synthesis dataset, single-scanner CT data. All addressable.\n\nMy recommendation: send it to serious peer review, but require the authors to directly test the retrieval mechanism — e.g., compare retrieved code indices against ground-truth HQ codes, or show feature distributions before/after diffusion — and to report variance across seeds. If they cannot provide that evidence, the conclusion about 'enhanced retrieval' should be downgraded to 'a diffusion-based refinement step that yields small gains.' The central architecture is not broken; the paper just over-claims its mechanism.\n\nYou could bring this to a reading group as a case study in ablation interpretation. I wouldn't cite it yet, though, until the mechanism is substantiated.","headline":"A plausible codebook-prior architecture for all-in-one MedIR, undermined by weak evidence for its signature latent-diffusion component.","tokens_in":9946,"tokens_out":3982,"would_cite":false,"duration_ms":37018,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DiffCode proposes a task-adaptive codebook bank plus latent diffusion that restores MRI, CT, and PET images with one model and reports the best average scores among twelve methods.","keywords":["medical image restoration","all-in-one restoration","vector quantization","codebook prior","latent diffusion","MRI super-resolution","CT denoising","PET synthesis"],"falsifier":"On a held-out split, extract the true high-quality features for each test image, quantize them against the codebook bank, and compare the retrieved codes or feature distances from the diffusion path versus a one-step direct prediction from the degraded features. If the diffusion path does not improve retrieval accuracy or feature similarity, the paper's explanation for its gains is unsupported.","tokens_in":8758,"feed_emoji":"🩻","tokens_out":4697,"duration_ms":43350,"temperature":0.7,"pith_summary":"DiffCode is a unified framework that restores degraded medical images from three different modalities—MRI, CT, and PET—with one set of weights. The paper's central claim is that a task-adaptive bank of vector-quantized codebooks, one codebook per task, supplies high-quality prior features that compensate for the different kinds of information each degradation destroys, and that a short latent diffusion process makes retrieval from that bank accurate enough to matter. On the IXI MRI, LDCT, and clinical PET datasets, DiffCode reports the best average scores among twelve methods, with PSNR 34.6186, SSIM 0.9336, and RMSE 12.3728. The authors argue that this is the first effective use of a vector-quantized codebook prior for all-in-one medical image restoration.","feed_headline":"DiffCode beats 11 rivals across MRI, CT, and PET restoration","feed_subtitle":"A task-adaptive codebook plus eight-step latent diffusion lifts average PSNR to 34.62.","key_machinery":"The central object is the task-adaptive codebook bank: a set of vector-quantized codebooks, one per task, each containing 8,192 code items of dimension 256 and updated by residual vector quantization across eight iterations. The latent diffusion stage, built from a five-layer denoising network run for eight time steps, maps degraded-image-conditioned Gaussian noise back toward high-quality latent features before those features are quantized against the bank. A task-aware global routing module with four experts, one activated per image, routes each task's features through specialized subnetworks to reduce task interference.","core_discovery":"The paper proposes that the obstacle to all-in-one medical image restoration is not the restoration backbone but the retrieval of high-quality prior information: each task destroys different features, so a single shared prior cannot compensate for all losses. DiffCode therefore builds a codebook bank in which each task has its own residual-quantized codebook of high-quality latent features, and then trains a latent diffusion model that, conditioned on the degraded image's features, iteratively refines a noisy latent toward the correct high-quality feature before quantization. The retrieved codes are decoded into a reference image that is concatenated with the degraded input and fed to a restoration network with task-aware routing. The paper reports that this pipeline outperforms eleven task-specific, general, and all-in-one baselines on MRI super-resolution, CT denoising, and PET synthesis, and that each component—codebook bank, latent diffusion, and routing—contributes positively in the ablations.","pith_inferences":["A testable implication the paper leaves implicit is whether the diffusion retriever's benefit grows with degradation severity; if retrieval errors increase as blur or noise worsens, the fixed eight-step schedule may need task-dependent length.","The codebook bank could in principle be extended to unseen tasks by adding codebooks trained on small amounts of high-quality data, turning restoration into a plug-and-play prior library; the paper does not demonstrate this.","Comparing the retrieved code indices against ground-truth indices would isolate where the diffusion stage helps, since final PSNR conflates retrieval accuracy with the backbone's ability to ignore bad codes.","The same latent-diffusion-plus-codebook pattern may transfer to non-medical multi-degradation restoration, where task heterogeneity is equally pronounced; this is an extrapolation, not a claim in the paper."],"forward_implications":["One trained model can serve MRI super-resolution, CT denoising, and PET synthesis without task-specific fine-tuning, simplifying clinical deployment.","Adding a new task may only require appending a new codebook and retraining the retrieval stage, rather than redesigning the whole restoration network.","The codebook prior supplies high-quality reference images that could be used for downstream quality control or as inputs to other analysis models.","Because the codebook bank and diffusion retriever are built once and then frozen, they can guide any restoration backbone, not only the NAF-block architecture used here.","The reported gains over AMIR, the closest all-in-one baseline, suggest that the bottleneck in all-in-one restoration is prior retrieval rather than routing alone."],"supporting_citations":[{"why":"Supplies the residual vector quantization mechanism used to build the codebook bank.","marker":"[10]"},{"why":"Provides the vector-quantized codebook prior and self-reconstruction training strategy.","marker":"[19]"},{"why":"Provides the latent diffusion formulation, condition encoder, and training loss used in Stage II.","marker":"[21]"},{"why":"Supplies the task-adaptive routing strategy that TARS adapts and serves as the closest all-in-one baseline.","marker":"[22]"},{"why":"Motivates the use of latent diffusion models for refining feature distributions.","marker":"[17]"},{"why":"Provides the NAF blocks used in the encoders, decoders, and restoration backbone.","marker":"[2]"},{"why":"Provides the low-dose CT dataset used for CT denoising experiments.","marker":"[15]"},{"why":"Provides the PET degradation and reconstruction protocol used to generate low-quality PET images.","marker":"[9]"}],"fun_headline_variants":["Codebook prior plus latent diffusion unifies medical image restoration","DiffCode restores MRI, CT, PET with task-adaptive codebooks and diffusion","Task-adaptive codebook bank with diffusion refines image priors for MedIR","All-in-one medical image restoration: adaptive codebooks plus latent diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a five-layer diffusion network can recover the correct high-quality latent feature from Gaussian noise in only eight denoising steps, conditioned on the degraded image; the paper validates this only through final restoration metrics, not by checking the retrieved features themselves.","fun_headline_variants_meta":{"raw":{"variants":["Codebook prior plus latent diffusion unifies medical image restoration","DiffCode restores MRI, CT, PET with task-adaptive codebooks and diffusion","Task-adaptive codebook bank with diffusion refines image priors for MedIR","All-in-one medical image restoration: adaptive codebooks plus latent diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001428,"raw_usage":{"total_tokens":5779,"prompt_tokens":979,"completion_tokens":4800,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":4720}},"tokens_in":595,"tokens_out":4800,"duration_ms":32698,"temperature":1.0,"reasoning_tokens":4720,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:50:07.499515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out split, extract the true high-quality features for each test image, quantize them against the codebook bank, and compare the retrieved codes or feature distances from the diffusion path versus a one-step direct prediction from the degraded features. If the diffusion path does not improve retrieval accuracy or feature similarity, the paper's explanation for its gains is unsupported.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Provides the latent diffusion formulation, condition encoder, and training loss used in Stage II."},{"cited_title":"In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Supplies the task-adaptive routing strategy that TARS adapts and serves as the closest all-in-one baseline."},{"cited_title":"https://doi.org/10.7937/9NPB-2637","cited_arxiv_id":null,"evidence_quote":"Provides the low-dose CT dataset used for CT denoising experiments."},{"cited_title":"IEEE transactions on medical imaging37(6), 1478–1487 (2018)","cited_arxiv_id":null,"evidence_quote":"Provides the PET degradation and reconstruction protocol used to generate low-quality PET images."}],"review_version":1}