{"id":"86cc5f44-7614-41d6-9cec-136b77353a71","arxiv_id":"2411.15719","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A benchmark of DDPM and LDM on colon pathology patches shows DDPM slightly outperforms LDM, and a pretrained model can generate images at unseen patch sizes, with mixed classification gains.","lead":"This paper tests how well diffusion-based image generators produce synthetic colon polyp tissue patches, comparing two model families and several sampling methods. It also explores whether a model trained at one resolution can generate new field-of-view sizes, which could help expand scarce pathology datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The patch-size/FOV claim is under-specified: §3.4.2 never defines how a 128x128 DDPM is conditioned on 'Patch Size', and Table 4's 'Patch reshaped' wording allows resizing as a confound; without code inspection or a resizing baseline, the central novelty is unsupported.","rationale":"The reader's identified weakness—missing counts, splits, and repeated runs—is real and partly overlaps with mine, but I see the more load-bearing problem at the level of the novel experiment itself. The paper's abstract highlights 'adjustment of image size during data generation can simulate varying fields of view' as a striking observation. Table 4 is the only evidence, yet Section 3.4.2 contains no description of how patch size is encoded, no altered network, and no training-time conditioning on patch size. A standard DDPM from [21] is not conditioned on output resolution; class and timestep embeddings are the only conditioning described in §3.2. Thus either the model is run at multiple resolutions via its convolutional architecture, which is a different and weaker claim, or the images are reshaped, which would invalidate the comparison. The 'Patch reshaped' column header is ambiguous and leans toward the latter. This ambiguity prevents a good-faith reader from evaluating the central novelty from the text alone. The classification-augmentation claim is also shaky—Table 5 shows Real+Generated accuracy decreasing on PKGH 336 (94.06 to 90.76)—but that weakness is secondary and would be addressed by the same request for dataset sizes, splits, and repeats. I therefore agree only partially with the reader's choice of weakest assumption, and I recommend no change to the CONDITIONAL verdict: the paper is plausible but must provide the generation mechanism and a resizing control before the patch-size claim can be accepted.","tokens_in":11793,"tokens_out":7414,"duration_ms":68588,"concrete_test":"Inspect the released repository's sampling code to determine whether the 'Patch Size' argument changes the spatial dimensions of the noise passed through the U-Net or resizes a 128x128 sample. Then run a control: from the trained 128x128 DDPM, generate 10,000 images and down/upsample them to 96x96 and 224x224; compute FID against the same real patches used in Table 4. If control FIDs match Table 4's 33.71 and 41.37 within sampling noise, the FOV-simulation claim is an artifact of resizing. Report all FID computations with the number of samples and bootstrap confidence intervals.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4.2 introduces 'Patch Size' as a novel hyperparameter used during image generation and Table 4 reports FID scores for outputs from 64x64 to 224x224 from a model trained only at 128x128 (§3.2). The manuscript never states how the U-Net, which is described with timestep and class embeddings only, is made to produce these other spatial sizes. If 'Patch Size' merely changes the shape of the noise tensor fed to a fully convolutional U-Net, the result is standard spatial extrapolation, not the 'prompting' claimed in the abstract; if the 128x128 output is resized to the target size, Table 4 measures interpolation artifacts rather than FOV simulation. The phrase 'Patch reshaped' in Table 4 makes the latter reading plausible. Since this experiment is one of the paper's two novel claims and is the basis for 'adjustment of image size ... simulate varying fields of view,' the absence of a mechanism description makes the central claim unreproducible from the manuscript. Additionally, FID/KID values in Tables 3 and 4 are reported without the number of real/generated patches or repeated runs, so the 19.08 versus 18.45 ordering is not yet statistically grounded.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a comparative empirical study of diffusion generative models (DDPM and LDM) for synthesizing histopathology image patches from a colon polyp dataset (KGH/PKGH). The authors report FID and KID scores across sampling methods (DDPM, DDIM, Epsilon Scaling) on two fields of view (224 and 336), propose a 'patch size prompting' experiment in which a 128x128-trained model generates images at other spatial sizes to simulate different FOVs, and evaluate ResNet-50 classification accuracy on real, generated, and augmented training sets. The main claims are that DDPM outperforms LDM on these data, that patch-size prompting can simulate varying FOVs, and that synthetic data can improve classifier accuracy when combined with real data.","tokens_in":12084,"tokens_out":3561,"duration_ms":32788,"significance":"If the claims were fully supported, the paper would provide a useful baseline comparison of diffusion models in computational pathology and an interesting observation about generating histopathology patches at multiple spatial sizes from a single trained model. The paper's empirical framing is appropriate, and the FID/KID plus classification evaluation is a standard toolkit. However, the significance is currently limited by three unresolved issues: the patch-size generation mechanism is not specified in a way that makes the result reproducible or distinguishable from trivial resizing; all generative metrics are point estimates with no sample-size information or uncertainty quantification; and the classification results are internally inconsistent with the paper's stated conclusion that synthetic data increases accuracy. The code is promised in the GitHub repository, which is a positive step, but the manuscript itself does not provide enough detail to assess the central novelty.","major_comments":[{"comment":"The 'patch size prompting' experiment is the paper's principal novelty, but the manuscript never describes the mechanism by which a model trained at 128x128 generates images at 64x64, 96x96, 160x160, 192x192, or 224x224. Section 3.2 describes the U-Net as having timestep and class embeddings only, with no size embedding or text conditioning. Table 4's column label 'Patch reshaped' suggests that the generated images may simply be resized to the target patch size. If the model is fully convolutional and the noise tensor is initialized at different shapes, that is standard spatial extrapolation rather than 'prompting patch size'; if the 128x128 output is resized, then Table 4 measures interpolation artifacts rather than FOV simulation. The abstract and Section 1 claim that 'adjustment of image size during data generation can simulate varying fields of view,' but without a precise description of the generation procedure or a baseline of simply resizing real/generated 128x128 images, this claim is not reproducible from the manuscript.","section":"§4.1, Tables 3 and 4"},{"comment":"FID and KID scores are reported as single point estimates with no confidence intervals, no standard deviations, and no statement of the number of generated images used for evaluation. The paper also does not state the number of real patches per class or per dataset. Without this information, the observed differences (e.g., FID 19.08 for PKGH 224 vs. 18.45 for PKGH 336, or the ordering of DDPM vs. LDM) cannot be distinguished from sampling noise. The authors should report the number of real and generated samples used in each metric computation, and ideally provide bootstrapped confidence intervals or results over multiple generation runs. This is load-bearing because Table 3 is the main evidence for the DDPM-over-LDM comparison and Table 4 is the main evidence for the patch-size claim.","section":"§4.3, Table 5, Conclusion"},{"comment":"The classification results are internally inconsistent with the paper's conclusion that synthetic data 'increasing accuracy' and that 'larger FOV values yielded better FID scores and higher classification accuracy.' For PKGH 224, adding generated data to real data improves accuracy from 89.95% to 90.75%, but for PKGH 336, the same augmentation decreases accuracy from 94.06% to 90.76%. The abstract and contributions claim that synthetic data can 'boost' classifier performance, yet the PKGH 336 result directly contradicts this. Moreover, the accuracy differences (e.g., 0.8 points on PKGH 224) are not accompanied by confidence intervals, repeated runs, or a significance test, so the authors' conclusion that synthetic data improves classifiers is not statistically grounded. The paper should hedge its claims, analyze why augmentation hurts on PKGH 336, and report variance across training runs.","section":null}],"minor_comments":[{"comment":"The sentence 'In DDIM, setting η = 0 eliminates noise, making it equivalent to DDPM' is inaccurate: η = 0 makes DDIM fully deterministic, which is not equivalent to DDPM's stochastic reverse process. This should be corrected to state that η = 0 gives deterministic DDIM sampling.","section":"§3.1"},{"comment":"The patch extraction procedure is under-specified: the paper says patches were extracted at FOV 224 from 20X magnification and resized to 128x128, but it does not state the original pixel dimensions before resizing, the total number of patches per class, or the number of WSIs used. This information is essential for assessing dataset balance and the representativeness of the FID/KID scores.","section":"§4.2, Figure 7"},{"comment":"The caption of Figure 7 says the model generates 'histology slices at various resolutions, from 64x64 to 224x224,' but Table 4 also includes 96x96, 160x160, and 192x192; the figure caption should list all evaluated sizes or refer the reader to Table 4.","section":"§5"},{"comment":"The conclusion states 'DDPM and LDM showed comparable performance despite their architectural differences,' but Table 3 shows a consistent 4-6 point FID gap favoring DDPM; 'comparable' is an overstatement. Also, the conclusion's claim that 'larger FOV values yielded better FID scores' holds only between 224 and 336 in Table 3, not across the extended patch-size sweep in Table 4.","section":null}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a GitHub link, but the code was not provided for review. Given that the central novelty depends on the implementation of 'patch size prompting,' I strongly recommend that the editor request the code or a detailed algorithmic appendix before any further decision. The paper is also missing basic statistical reporting (sample sizes, confidence intervals, repeated runs) that is standard for generative model comparisons. If the patch-size experiment turns out to be a simple resize of a 128x128 output, the novelty claim would collapse; the authors should be asked to clarify this explicitly in a revision. The classification augmentation claim is also contradicted by the PKGH 336 result and should be reframed or supported with additional experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a solid incremental comparison of DDPM and LDM on a private colon polyp dataset, with one genuinely novel experiment (generating at patch sizes not seen in training) that is currently too under-specified to evaluate. The FID/KID benchmark is useful, but the classification claims outrun the evidence.\n\nWhat's new: comparative FID/KID for two FOVs (224, 336) on a 5-class colon polyp dataset; DDPM gets FID 19.08/18.45, competitive with published numbers. The patch-size experiment is the interesting part: training one DDPM at 128x128 and generating at 64 to 224. If it truly generates at those resolutions, that's neat. But the paper never says how. Is the U-Net fully convolutional and able to take arbitrary spatial input? Or are you reshaping the 128x128 output, as Table 4's \"Patch reshaped\" wording suggests, then computing FID on resized images? Those are very different things. Without a mechanism description or a baseline of resizing real 128x128 patches to those sizes, the claim that patch size simulates FOV is unsupported. That is the main soft spot, and the stress-test note lands correctly here.\n\nOther issues: the classification experiment is internally inconsistent. Real+generated improves PKGH224 by 0.8 points but lowers PKGH336 by 3.3 points, yet the abstract and conclusion claim synthetic data increases accuracy. That overreach needs to be fixed. Missing details (number of patches per class, number of generated images used for FID/KID, train/test split, repeated runs) mean the small accuracy differences could easily be noise. The inline \"(author?)\" citation and the dangling \"Section\" reference in Section 3.4.3 show the manuscript is unfinished.\n\nWhat's good: the comparison is honest in reporting all numbers, and the FID values are plausible. The dataset is private but the method section is clear enough that others could replicate on their own slides. Code is promised, though I didn't verify the repo.\n\nVerdict: this deserves a serious referee, but only if the patch-size mechanism is clarified and the classification conclusion is reined in. I'd bring it to reading group as a case study in how a novel-sounding generation experiment can be under-specified. I wouldn't cite it yet.","headline":"Useful comparative FID/KID on a private colon polyp dataset, but the patch-size novelty is under-specified and the classification claims overreach; deserves review with major revision.","tokens_in":12585,"tokens_out":2792,"would_cite":false,"duration_ms":25428,"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":"This paper establishes that DDPM, a pixel-space diffusion model, can synthesize high-quality colon polyp histopathology patches (FID 19.08 for FOV 224), and that a single 128x128-trained model can generate other patch sizes to simulate…","keywords":["diffusion generative models","latent diffusion models","computational pathology","histopathology image synthesis","field of view","FID","KID","synthetic data augmentation"],"falsifier":"Run the reported comparison with fixed numbers of generated samples and per-class counts, and repeat the ResNet-50 training over multiple seeds; if the 90.75% versus 89.95% accuracy difference on PKGH 224 falls within the seed-to-seed standard deviation, or if DDPM's FID on a held-out patch set no longer beats LDM's, the central claims would be falsified.","tokens_in":11655,"feed_emoji":"🔬","tokens_out":7153,"duration_ms":58893,"temperature":0.7,"pith_summary":"This paper tries to establish that diffusion generative models, especially the pixel-space DDPM, can produce high-quality synthetic histopathology patches of colonic polyps, reporting FID scores of 19.08 and 18.45 for two fields of view. It further claims that a model trained at a single 128x128 patch size can generate plausible images at other sizes, effectively simulating different fields of view without retraining. The authors argue that synthetic data can support classifiers, showing that adding generated patches to real PKGH 224 data raises a ResNet-50's accuracy from 89.95% to 90.75%. If true, synthetic generation would help address privacy and data-scarcity problems in computational pathology.","feed_headline":"Synthetic colon polyp images reach FID 19.08","feed_subtitle":"A single DDPM can generate multiple patch sizes, simulating different fields of view without retraining.","key_machinery":"The central objects are Denoising Diffusion Probabilistic Models (DDPM), which add and remove Gaussian noise directly in pixel space, and Latent Diffusion Models (LDM), which run diffusion in a lower-dimensional latent space learned by a VQ-autoencoder. Classifier-free guidance mixes conditional and unconditional noise predictions to steer generation toward one of the five tissue classes. The paper's novel knob is prompting patch size at sampling time: with training fixed at 128x128, requesting other output sizes changes the field of view of the generated images, which is the mechanism claimed to simulate multiple FOVs from one model.","core_discovery":"On its own terms, the paper reports that DDPM with class-conditioning and classifier-free guidance generates colon polyp patches that score FID 19.08 on the 224 field-of-view dataset and 18.45 on the 336 dataset, outperforming the LDM baseline on both datasets and under both DDPM and DDIM sampling. The same model trained at 128x128 produces patches at other sizes, with FID worsening away from the training size (161.01 at 64x64, 41.37 at 224x224) while preserving recognizable structure at intermediate sizes. In the downstream evaluation, training on real plus generated patches improves accuracy over real-only training on PKGH 224 (90.75% vs 89.95%), though the same augmentation slightly reduces accuracy on PKGH 336 (90.76% vs 94.06%).","pith_inferences":["Implicit in the patch-size results is that the trained model encodes multi-scale tissue structure; this could be tested by comparing generated patches across sizes against real patches of the corresponding magnifications.","The paper's accuracy differences are small and lack variance reporting, so the augmentation benefit on PKGH 224 may not survive a significance test; a repeated-seed experiment would settle it.","If synthetic patches can fool FID but not preserve subtype-specific diagnostic cues, downstream tasks like polyp classification may need their own quality metrics beyond FID and KID."],"forward_implications":["DDPM in pixel space is the stronger baseline for this histopathology generation task, so future method comparisons can use it as the reference point.","A single diffusion model can synthesize patches at untrained resolutions, giving a cheap way to create multi-field-of-view datasets without retraining.","Because the 336 FOV dataset produced both better FID and higher real-data classifier accuracy, FOV choice should be a reported and controlled variable in pathology generation studies.","Standard DDPM sampling is preferable to DDIM or epsilon scaling in this setting, since both alternatives produced higher FID and KID scores."],"supporting_citations":[{"why":"Defines the DDPM training and sampling procedure that produces the paper's main synthetic images.","marker":"[18]"},{"why":"Provides the LDM architecture used as the comparison baseline.","marker":"[32]"},{"why":"Supplies classifier-free guidance, used to condition generation on tissue class.","marker":"[33]"},{"why":"Provides the UNet architecture and guidance techniques the experiments build on.","marker":"[21]"},{"why":"Defines the FID metric used for all quality comparisons.","marker":"[37]"},{"why":"Defines the KID metric reported alongside FID.","marker":"[39]"},{"why":"The ResNet-50 classifier used to evaluate whether synthetic data helps classification.","marker":"[40]"},{"why":"Reported FID baseline for diffusion-based histopathology generation that the paper compares against.","marker":"[41]"},{"why":"Reported FID from a morphology-focused diffusion model used as the other external comparison.","marker":"[28]"}],"fun_headline_variants":["DDPM beats LDM on colon polyp FID 19.08","One diffusion model simulates multiple pathology fields of view","Class-conditioned DDPM tops pathology FID at 19.08","Synthetic colon polyps boost classifier to 90.75% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported experimental setup can support the conclusions, meaning the number of patches per class, the number of generated images used for FID and KID, the classifier's train/test split, and repeated runs with variance are all sufficient to make the small accuracy differences meaningful.","fun_headline_variants_meta":{"raw":{"variants":["DDPM beats LDM on colon polyp FID 19.08","One diffusion model simulates multiple pathology fields of view","Class-conditioned DDPM tops pathology FID at 19.08","Synthetic colon polyps boost classifier to 90.75% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000701,"raw_usage":{"total_tokens":3160,"prompt_tokens":936,"completion_tokens":2224,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2149}},"tokens_in":552,"tokens_out":2224,"duration_ms":16494,"temperature":1.0,"reasoning_tokens":2149,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:58:12.683651+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the reported comparison with fixed numbers of generated samples and per-class counts, and repeat the ResNet-50 training over multiple seeds; if the 90.75% versus 89.95% accuracy difference on PKGH 224 falls within the seed-to-seed standard deviation, or if DDPM's FID on a held-out patch set no longer beats LDM's, the central claims would be falsified.","supporting_citations":[{"cited_title":"Diffusion models beat gans on image synthesis","cited_arxiv_id":null,"evidence_quote":"Provides the UNet architecture and guidance techniques the experiments build on."},{"cited_title":"Gans trained by a two time-scale update rule converge to a local nash equilibrium","cited_arxiv_id":null,"evidence_quote":"Defines the FID metric used for all quality comparisons."},{"cited_title":"Generating synthetic data in digital pathology through diffusion models: a multi- faceted approach to evaluation","cited_arxiv_id":null,"evidence_quote":"Reported FID baseline for diffusion-based histopathology generation that the paper compares against."},{"cited_title":"A morphology focused diffusion probabilistic model for synthesis of histopathology images","cited_arxiv_id":null,"evidence_quote":"Reported FID from a morphology-focused diffusion model used as the other external comparison."}],"review_version":1}