{"id":"8ca17012-b5be-4f41-9a82-c30d098a76d4","arxiv_id":"1908.04346","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SkrGAN synthesizes medical images by generating structural sketches from noise and rendering them into full images, reporting improved FID, SWD, MS-SSIM, and vessel segmentation over DCGAN, ACGAN, WGAN, and PGGAN.","lead":"SkrGAN is a two-stage GAN that first draws a structural sketch from random noise and then colors it in to synthesize medical images such as retinal fundus, X-ray, CT, and MRI. The authors report better image quality scores than four baseline GANs and show that pretraining a vessel segmentation network on synthetic images improves its accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eqs. (1)-(3) never train GP on GS's outputs: no term evaluates GP(GS(z)), so the claimed sketch-guided generator is not actually optimized end-to-end.","rationale":"The strongest claim is state-of-the-art synthesis in four modalities and improved segmentation after pretraining. For that claim, what must be true is that a single generator G = GP ∘ GS is trained so that sketch quality is constrained by final image quality, or at least that GP is competent on the sketch distribution produced by GS. The printed objective does not establish either: Eq. (1) depends only on DS, and Eq. (2)/L1 depend only on real y. The full objective in Eq. (3) is a sum of separate min-max games, not a composition. Even if one corrects the sign flips, there is no gradient path from GP losses into GS. This is a distinct and more basic issue than the reader's concern about sketch fidelity: a high-fidelity sketch prior would not help if the renderer never sees generated sketches during training. The concern is about internal consistency, not disagreement with community consensus. It is concrete and testable by inspecting the training graph. The appropriate disposition remains conditional: if the authors release code or an updated objective that includes GP(GS(z)), the concern is resolved; if they confirm the objective as written, the central claim is not supported. I therefore keep the reader's conditional verdict, hence UNCHANGED, and partially agree with the reader's weakest-assumption identification.","tokens_in":6188,"tokens_out":8413,"duration_ms":97753,"concrete_test":"Instrument the released (or re-implemented) training graph: insert a gradient hook on the GS parameters to record the contribution of the GP losses (Ladv and L1) to ∂L/∂theta_GS. If that contribution is exactly zero, GP is never trained on GS(z), confirming that the claimed two-module generator is not trained as a whole; then add a term evaluating GP(GS(z)) and re-run the four FID rows in Table 1. If the FID values are unchanged, the current results do not exercise the missing coupling; if they change, the paper's equations omit a necessary part of the method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eqs. (1)–(3), even after correcting the apparent real/fake sign flips, define two independent training problems. Eq. (1) trains GS to match the real sketch distribution via DS. Eq. (2) and L1 train GP on real sketch–image pairs (y, x): L1 = λ E_{(x,y)} ||GP(y) − x||_1 and the corrected Ladv = E_{(x,y)} log DP(x, y) + E_y log(1 − DP(GP(y), y)). Neither term contains GS(z). Thus the gradient of the total objective with respect to GS is independent of GP, and the color renderer never sees a generated sketch during training. At inference the paper feeds GS(z) into GP (Fig. 2), so GP is evaluated on an input distribution it was not trained on. The abstract's claim that SkrGAN uses a 'sketch prior constraint to guide medical image generation' is therefore not realized by the equations as written: there is no constraint coupling sketch generation to final image quality. The reported FID/SWD gains and the segmentation improvement could in principle come from the supervised sketch-to-image translator alone, not from the proposed two-module unconditional GAN. This is not merely a sign typo; it is a missing composition in the objective. If the authors' code contains an additional loss evaluating GP(GS(z)), that term must be reported, because without it the central mechanism is undefined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SkrGAN, an unconditional GAN for medical image synthesis that decomposes generation into a sketch guidance module (GS) and a color render mapping module (GP). Sketch drafts are extracted from real images using Sobel edge detection, Gaussian filtering, and morphological operations. GS is trained to generate sketches from noise under a sketch discriminator DS; GP is trained to translate sketches into color images under a color discriminator DP and an L1 loss. The paper evaluates the method on retinal fundus, chest X-ray, lung CT, and brain MRI datasets using MS-SSIM, SWD, and FID, and demonstrates a vessel-segmentation data-augmentation experiment on DRIVE. The abstract claims state-of-the-art synthesis and improved segmentation through SkrGAN-generated images.","tokens_in":6502,"tokens_out":2562,"duration_ms":30978,"significance":"If the claimed results hold, the paper would make a useful contribution by showing that an explicit structural sketch prior can improve unconditional medical image generation and provide a data-augmentation signal for segmentation. The application to four modalities and the DRIVE segmentation transfer experiment are appropriate tests, and the use of established metrics (FID, SWD, MS-SSIM) is commendable. However, the significance is currently undercut by formal problems in the stated objectives: the equations as printed do not implement the claimed sketch-guided generator, and the reported quantitative comparisons lack error bars or significance tests. The central mechanism must be correctly defined and verified before the empirical claims can be assessed.","major_comments":[{"comment":"The adversarial objective for the sketch guidance module is printed with inverted roles: Ls = Ez[log(DS(GS(z⊙l)))] + Ex[log(1−DS(y))]. Under the standard convention that DS outputs the probability of a real input, this trains DS to label generated sketches as real and real sketches as fake, and trains GS to move its output toward the fake label. If this is a typo, the correct form must be given, because the subsequent derivation and the interpretation of the training dynamics depend on it.","section":"Sec. 2.2, Eq. (1)"},{"comment":"The color render mapping objective has the same inversion: Ladv writes Ez,y[log(DP(GP(y),y))] for the generated pair and E(x,y)[log(1−DP(x,y))] for the real pair. As written, DP is rewarded for calling translated sketches real and original image-sketch pairs fake, which is the opposite of the standard conditional GAN loss. The authors should correct the sign convention or state explicitly if a nonstandard label convention is intended; as printed, the loss is not a valid adversarial objective.","section":"Sec. 2.3, Eq. (2)"},{"comment":"The total objective in Eq. (3) does not contain any term that evaluates GP on the output of GS: no loss depends on GP(GS(z)). Equation (2) trains GP on pairs of real sketches y and real images x, while Eq. (1) trains GS alone. Therefore the gradients with respect to GS are independent of GP, and the color renderer is never trained on generated sketches, even though Fig. 2 and the inference procedure feed GS(z) into GP. This is not a cosmetic issue: the claimed sketch-prior constraint guiding final image generation is not realized by the equations as written. The reported FID/SWD improvements and the segmentation gain could, in principle, be produced by the supervised sketch-to-image translator alone, with the unconditional sketch generator playing no role in final image quality. The authors must report the actual objective used in training; if GP(GS(z)) is indeed trained, that term must be added to Eq. (3), and if it is not, the central claim of the paper needs to be substantially revised.","section":"Secs. 2.2–2.3, Eqs. (1)–(3)"},{"comment":"The quantitative comparison reports mean SWD, MS-SSIM, and FID computed from 100 generated images per method, with no variance, confidence intervals, or significance tests across independent training runs. FID and SWD are notoriously noisy at this sample size, and for some entries the gap to the runner-up is small (e.g., brain MRI FID of 27.51 versus 33.76 for PGGAN). The authors should report standard deviations over repeated generation/training runs or provide a statistical test to support the claim that SkrGAN is consistently better than the baselines.","section":"Sec. 3, Table 1"},{"comment":"The paper provides no quantitative validation that the handcrafted sketch draft representation preserves the anatomical structures that matter (small vessels, disc/cup boundaries, bone edges). This is load-bearing because the segmentation pretraining uses the generated sketches as labels: if the sketch extractor drops small vessels, the reported +8.78% sensitivity improvement may reflect a denoised or simplified label signal rather than a genuinely better structural prior. The authors should compare their sketch extraction against the full images at the relevant resolutions and report, for example, vessel detection recall on the DRIVE ground truth, or an ablation that varies the sketch-generation parameters.","section":"Secs. 2.1 and 3, vessel segmentation experiment"}],"minor_comments":[{"comment":"There are several typographical errors that should be corrected: 'Freshet Inception Distance' should be 'Fréchet Inception Distance', 'comparsions' should be 'comparisons', and Table 2 contains 'whithout' for 'without'. The PGGAN reference also contains 'et. al.o'.","section":"Throughout"},{"comment":"The description of the evaluation protocol is incomplete: it is not stated how the real-image set for FID/SWD is defined, whether all real images are used, and whether the 100 generated images are produced after a fixed number of training iterations or after convergence. These details affect reproducibility.","section":"Sec. 3, datasets and protocol"},{"comment":"The notation 'arg[ min_GS max_DS Ls + min_GP max_DP (Ladv+L1) ]' is nonstandard and mixes the minimization and maximization arguments with the outer arg expression; this should be rewritten as separate min-max problems or as a single combined objective with clearly scoped optimization variables.","section":"Eq. (3)"},{"comment":"The definition DS = D_S^(n)···D_S^(0) and GS = G_S^(n)···G_S^(0) is not explained with enough precision; it would be clearer to state that the generator and discriminator are progressive-growing networks following PGGAN, rather than introducing notation that suggests a composition of separate networks.","section":"Sec. 2.2"},{"comment":"The statement that 'a small change of λ does not affect much the performance' is not supported by an ablation or sensitivity analysis; adding a short sensitivity experiment would strengthen the claim and the reproducibility of the method.","section":"Sec. 3, hyperparameters"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is not the sign flips in Eqs. (1) and (2), which are easy to fix as typos, but the complete absence of a coupling term between GS and GP in the stated objective. If the authors' code trains GP on GS(z) outputs, the manuscript is missing a core equation and needs only a reporting fix; if it does not, the paper's central claim that the sketch prior guides medical image generation is not supported by the presented method. I would ask the editor to have the authors clarify this point explicitly in the revision, ideally by providing the actual training objective and, if available, code. The comparison protocol also needs error bars before the SOTA claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece is the use of a hand-crafted edge-based sketch prior (Sobel plus Gaussian plus morphological opening/closing) inside an unconditional medical-image GAN. The split into a sketch generator and a sketch-to-color renderer has precedent in S2-GAN's style/structure decomposition, so this is an incremental architectural contribution, but a reasonable one. The experiments cover four modalities (fundus, X-ray, CT, MRI) with FID, SWD, MS-SSIM against DCGAN, ACGAN, WGAN, PGGAN, and a DRIVE vessel-segmentation augmentation test. The sensitivity gain from 0.778 to 0.846 with U-net pretraining is a practically interesting result, assuming it is real.\n\nNow the soft spots. The loss functions in Eqs (1) and (2) are printed inverted: they would train the discriminators to call fake images real and real images fake. That is likely a sign typo, but it must be corrected. More seriously, even after fixing that, Eq (3) is a sum of two independent training problems. The sketch generator GS is trained only against the sketch discriminator DS; the renderer GP is trained only on real sketch-image pairs (y,x). There is no term involving GP(GS(z)). So the objective never couples the two modules, and at inference GP receives input from a distribution (generated sketches) that it never saw during training. The described 'sketch prior constraint to guide medical image generation' is not realized by the equations as printed. If the actual code includes a loss on the full generator, that term must be reported; without it, this is a two-stage pipeline, not an end-to-end sketch-guided GAN. This is a load-bearing issue, not cosmetic.\n\nThe quantitative reporting has smaller problems: mean metrics over 100 images with no error bars or significance tests; MS-SSIM on unpaired generation is ambiguous; Table 1 appears to have copy-paste errors (brain MRI FID values for DCGAN and ACGAN are identical to lung CT). Table 2 gives one pretraining run with no variance. None of this falsifies the central result, but it makes the effect size hard to judge.\n\nCitation pattern is fine; self-citations are background.\n\nBottom line: this paper deserves referee time, but with major revision. The authors need to provide the real training objective, correct the sign flips, add error bars or significance tests, fix Table 1, clarify MS-SSIM, and ideally ablate the renderer trained on generated versus real sketches. The audience is medical-imaging GAN researchers and people using synthetic data for segmentation. I wouldn't cite it in its current form.","headline":"SkrGAN is a plausible two-stage sketch-then-render generator for medical images, with a useful data-augmentation experiment, but the equations as printed don't train the two stages together; that has to be fixed before the central claim is trustworthy.","tokens_in":7005,"tokens_out":6529,"would_cite":false,"duration_ms":71273,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Decomposing medical image generation into a sketching step and a color-rendering step preserves fine anatomical structures, and this paper reports that the resulting unconditional GAN outperforms four comparison GANs across four…","keywords":["medical image synthesis","generative adversarial networks","sketch prior","unconditional generation","data augmentation","retinal fundus","vessel segmentation","structure preservation"],"falsifier":"Generate one fixed noise vector through SkrGAN, then compare the generated sketch with the Sobel/morphological sketch of the rendered color image; if thin vessels or bone edges diverge, the sketch prior is not preserving the structures claimed. A second check is to train with random line drawings of similar density in place of the real sketch drafts: if FID and segmentation scores remain unchanged, the specific sketch representation is doing none of the work.","tokens_in":6001,"feed_emoji":"🩻","tokens_out":11128,"duration_ms":101532,"temperature":0.7,"pith_summary":"This paper proposes SkrGAN, an unconditional generative adversarial network that splits medical image synthesis into a sketching stage and a color-rendering stage, following the order in which a human painter works. The sketch guidance module turns random noise into a structural sketch, and the color render mapping embeds that sketch into a full color image, so fine foreground structures such as retinal vessels, ribs, and bone edges are not left to the image discriminator alone. The authors report lower Fréchet Inception Distance, lower Sliced Wasserstein Distance, and higher MS-SSIM than four comparison GANs on retinal color fundus, chest X-ray, lung CT, and brain MRI. They further report that pretraining a U-net on 2,000 synthetic retinal pairs, using generated sketches as labels, raises vessel-segmentation sensitivity on DRIVE from 0.778 to 0.846. If the claim is right, SkrGAN is a general-purpose, label-free data-augmentation method for medical imaging.","feed_headline":"Sketch-first GAN outperforms four rivals in medical image synthesis","feed_subtitle":"By generating structural sketches before color, SkrGAN preserves vessels and bones across fundus, X-ray, CT, and MRI.","key_machinery":"The load-bearing mechanism is a sketch prior enforced by two coupled modules. The sketch guidance module $G_S$ maps random noise to a structural sketch and answers to a sketch discriminator $D_S$ that compares against Sobel-and-morphology sketch drafts of real images; the color render mapping $G_P$, built on a U-net, embeds the sketch representation and produces the final color image under an image discriminator $D_P$ plus an $\\ell_1$ reconstruction term. The two objectives are optimized jointly, and the sketch stage grows progressively in resolution. This isolates the anatomical geometry in a low-dimensional sketch space before the model attempts to match texture and background appearance.","core_discovery":"SkrGAN's central claim is that an unconditional GAN can generate medical images with anatomically faithful foreground structures if the generator first commits to a structural sketch and only then renders color. The sketch guidance module $G_S$ is trained against a sketch discriminator $D_S$ on sketch drafts extracted from real images by Sobel edge detection, Gaussian low-pass filtering, and morphological opening/closing; the color render mapping $G_P$, a U-net, is trained against an image discriminator $D_P$ and an $\\ell_1$ reconstruction loss. The full objective joins the sketch-guidance loss with the color-rendering loss, with progressive growing used in the sketch stage. On the paper's experiments, this architecture performs better than DCGAN, ACGAN, WGAN, and PGGAN on FID, SWD, and MS-SSIM across four modalities, and its synthetic images raise U-net vessel-segmentation sensitivity on DRIVE from 0.778 to 0.846. The intended contribution is a general-purpose structure-preserving generator and data-augmentation tool that needs neither paired data nor manual labels.","pith_inferences":["Editorial inference: the sketch representation is probably interchangeable; if so, SkrGAN should transfer to ultrasound, OCT, or dermoscopy by swapping the sketch-preparation recipe, a test the paper does not run.","Editorial inference: the gains could come mainly from constraining the image discriminator's input space rather than from the specific Sobel representation; this could be tested by replacing real sketch drafts with random line drawings of matched density.","Editorial inference: because the generated sketches double as labels, SkrGAN is a source of weakly labeled training data; a natural extension is to measure how segmentation accuracy scales beyond 2,000 synthetic pairs.","Editorial inference: the unconditional setting leaves room for a conditional extension that controls anatomy or pathology class, which would help class-imbalanced medical datasets; the paper does not explore this."],"forward_implications":["On four modalities (retinal color fundus, chest X-ray, lung CT, brain MRI), SkrGAN reports lower FID and lower Sliced Wasserstein Distance, and higher MS-SSIM, than DCGAN, ACGAN, WGAN, and PGGAN.","The generator requires no paired data, labels, or depth maps, because the sketch prior is derived automatically from training images.","The synthetic images can be used as data augmentation: on DRIVE, pretraining a U-net with SkrGAN pairs raises sensitivity from 0.778 to 0.846, with accuracy and AUC also increasing.","The same architecture transfers across modalities without per-modality redesign, which makes it a candidate general-purpose augmentation tool when training sets are small."],"supporting_citations":[{"why":"supplies the adversarial training framework that underlies both the sketch discriminator and the color discriminator.","marker":"[4]"},{"why":"provides the progressive-growing backbone for the sketch guidance module and is one of the comparison baselines.","marker":"[7]"},{"why":"motivates the sketching-then-rendering decomposition that defines the architecture.","marker":"[12]"},{"why":"provides the U-net backbone for the color rendering module and for the segmentation network in the augmentation experiment.","marker":"[15]"},{"why":"supplies the DRIVE dataset and ground-truth vessel annotations used to evaluate the segmentation benefit.","marker":"[16]"},{"why":"defines the Sliced Wasserstein Distance used to compare real and synthetic image distributions.","marker":"[13]"},{"why":"one of the comparison baselines whose structural distortions motivate the sketch prior.","marker":"[14]"},{"why":"one of the comparison baselines the paper must beat on all four modalities.","marker":"[1]"},{"why":"one of the comparison baselines, included because it conditions synthesis on auxiliary class information.","marker":"[11]"},{"why":"source of the chest X-ray dataset used in the modality experiments.","marker":"[8]"}],"fun_headline_variants":["Sketch-first GAN preserves vessels and bones in synthetic medical images","Unconditional GAN sketches structures before rendering color for medical images","Sketch-then-render GAN boosts medical image synthesis without paired data","SkrGAN: sketch-guided generation improves medical image realism"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's results depend on the hand-made sketch representation (Sobel edges, smoothing, and morphological cleaning) capturing exactly the anatomical structures that matter; if that sketch drops thin vessels or blurs bone boundaries, both the image-quality gains and the segmentation improvement would shrink.","fun_headline_variants_meta":{"raw":{"variants":["Sketch-first GAN preserves vessels and bones in synthetic medical images","Unconditional GAN sketches structures before rendering color for medical images","Sketch-then-render GAN boosts medical image synthesis without paired data","SkrGAN: sketch-guided generation improves medical image realism"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000548,"raw_usage":{"total_tokens":2639,"prompt_tokens":985,"completion_tokens":1654,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":1580}},"tokens_in":601,"tokens_out":1654,"duration_ms":12399,"temperature":1.0,"reasoning_tokens":1580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:53:52.260556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate one fixed noise vector through SkrGAN, then compare the generated sketch with the Sobel/morphological sketch of the rendered color image; if thin vessels or bone edges diverge, the sketch prior is not preserving the structures claimed. A second check is to train with random line drawings of similar density in place of the real sketch drafts: if FID and segmentation scores remain unchanged, the specific sketch representation is doing none of the work.","supporting_citations":[{"cited_title":"al.: Generative adversarial networks","cited_arxiv_id":null,"evidence_quote":"supplies the adversarial training framework that underlies both the sketch discriminator and the color discriminator."},{"cited_title":"al.o: Progressive growing of gans for improved quality, stability, and variation","cited_arxiv_id":null,"evidence_quote":"provides the progressive-growing backbone for the sketch guidance module and is one of the comparison baselines."},{"cited_title":"Psychology of Aesthetics, Creativity, and the Arts 6(2), 124–136 (2012)","cited_arxiv_id":null,"evidence_quote":"motivates the sketching-then-rendering decomposition that defines the architecture."},{"cited_title":"al.: Ridge-based vessel segmentation in color images of the retina","cited_arxiv_id":null,"evidence_quote":"supplies the DRIVE dataset and ground-truth vessel annotations used to evaluate the segmentation benefit."},{"cited_title":"Foundations and Trends R© in Machine Learning 11(5-6), 355–607 (2019)","cited_arxiv_id":null,"evidence_quote":"defines the Sliced Wasserstein Distance used to compare real and synthetic image distributions."},{"cited_title":"international conference on learning representations (2016)","cited_arxiv_id":null,"evidence_quote":"one of the comparison baselines whose structural distortions motivate the sketch prior."},{"cited_title":"arXiv (2017)","cited_arxiv_id":null,"evidence_quote":"one of the comparison baselines the paper must beat on all four modalities."},{"cited_title":"In: ICML","cited_arxiv_id":null,"evidence_quote":"one of the comparison baselines, included because it conditions synthesis on auxiliary class information."},{"cited_title":"al.: Identifying medical diagnoses and treatable diseases by image-based deep learning","cited_arxiv_id":null,"evidence_quote":"source of the chest X-ray dataset used in the modality experiments."}],"review_version":1}