{"id":"d31926d0-1c78-47b9-b847-d874a4e4362c","arxiv_id":"2501.03466","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DGSSA trains retinal vessel segmenters on procedurally generated vascular structures rendered into pseudo-fundus images by an improved Pix2Pix, plus PixMix style mixing, gaining about 0.38 average Dice points over the prior best method AADG.","lead":"This paper combines two data-augmentation tricks to help retinal vessel segmentation software cope with images from scanners it has never seen: procedurally generated fake vessel patterns are rendered into realistic fake fundus images with a GAN, while photometric mixing adds style diversity. Tested on four public datasets, the method edges out prior domain-generalization approaches by a small margin in average Dice score.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Target-domain Pix2Pix generators may leak into source training pool; the leave-one-out exclusion rule is unstated and determines validity of the 77.98% SOTA claim.","rationale":"The reader's weakest assumption is the same concern I would flag, and it is genuinely load-bearing. Unlike secondary issues such as the absence of error bars or the small 0.38-point margin over AADG, a leak of target-domain information through the per-dataset Pix2Pix generator changes what the experiment measures: it turns a domain-generalization evaluation into a protocol that may include unlabeled target-domain appearance through synthetic images. The manuscript gives real support for the method's direction: the ablations are internally consistent, the thin-vessel analysis (13.1% relative improvement) is a credible mechanistic signal, and the computational overhead claim is modest. Those pieces do not, however, rescue the headline SOTA claim if the evaluation protocol is contaminated. The concrete test is straightforward and would settle the issue: rerun the leave-one-out experiments with explicit exclusion of target-trained generators, or confirm from code/logs that such exclusion already happened. If the protocol is already clean, the paper should state it clearly; if not, the 77.98% average DSC cannot be read as a domain-generalization result. Because the reader already conditioned acceptance on this point, my stress-test does not move the verdict.","tokens_in":21240,"tokens_out":3210,"duration_ms":33511,"concrete_test":"Ask the authors for the exact per-fold data-generation recipe, and rerun the leave-one-out protocol under a strict exclusion rule: for each target D, train the Pix2Pix generator only on the remaining three source datasets (or use only source-trained generators) to produce structural pseudo-images, then compare fold DSC with Table 1. If the held-out-target DSC drops by more than roughly 0.5 points or the AADG margin vanishes, the reported SOTA claim is an artifact of target leakage. Also require a one-line protocol statement and the training configuration so the check is reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (77.98% average DSC, SOTA) depends on a clean leave-one-out protocol: when dataset D is the held-out target, no information about D may enter training. Section 4.2 states that the improved Pix2Pix model is trained \"independently on each dataset\" to synthesize pseudo-retinal images from 100 space-colonization masks per dataset, and Section 3.1 says the generated images are then combined with the original dataset for segmentation training. The paper never states the exclusion rule for the target fold. If the target-trained generator contributes pseudo-images to the source-only training pool, the segmentation model is trained on synthetic images whose appearance and vascular statistics are conditioned on the target dataset's unlabeled images, even though the target ground-truth masks are not used. That would make the reported 77.98% DSC a measure of adaptation to target appearance, not domain generalization to a fully unseen domain, and would invalidate the comparison with AADG and DoFE. The wording in Section 4.1 that \"only the training images from the multi-source domains were input into the network\" is ambiguous: it does not explicitly prohibit pseudo-images produced by a target-trained generator from being included, and Section 4.2's per-dataset training description strongly suggests all four generators are built regardless of the fold.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DGSSA, a domain-generalization framework for retinal vessel segmentation that combines structural augmentation (space-colonization-generated vascular structures rendered into pseudo-retinal images by an improved Pix2Pix model) with stylistic augmentation (PixMix with photometric augmentations and uncertainty perturbations). The segmentation network is DeepLabv3+ with a MobileNetV2 backbone. The method is evaluated under a leave-one-out protocol on STARE, HRF, DRIVE, and CHASEDB1, reporting an average DSC of 77.98% and claiming state-of-the-art performance over prior DG methods such as AADG, DoFE, and ELCFS. The paper also includes ablations, a thin/thick-vessel decomposition, hyperparameter sensitivity analysis, and statistical comparisons.","tokens_in":21508,"tokens_out":4091,"duration_ms":37521,"significance":"If the reported results are valid, the central idea is interesting: adding procedurally generated structural diversity to style augmentation is a cheap and potentially effective DG mechanism for tubular structures like retinal vessels. The paper has several strengths: the structural generation pipeline is described algorithmically, the ablation study isolates structural and stylistic contributions, the thin-vessel analysis directly addresses the clinically relevant challenge of capillary segmentation, and the authors attempt statistical comparisons and hyperparameter sensitivity checks. However, the headline SOTA claim rests on a small 0.38-percentage-point average DSC margin over AADG, the paper does not provide code or error bars for the main table, and the leave-one-out protocol as written leaves open a serious data-leakage ambiguity. The significance of the contribution is therefore conditional on resolving these reporting and protocol issues.","major_comments":[{"comment":"The leave-one-out protocol does not state whether the Pix2Pix generator trained on the target dataset and its generated pseudo-images are excluded from the source training pool when that dataset is the held-out target. Section 4.2 says the improved Pix2Pix is trained \"independently on each dataset\" to synthesize 100 images per dataset, and Section 3.1 says the generated images are combined with the original dataset for segmentation training. The current text therefore permits a target-trained generator to contribute pseudo-images with target-domain appearance to the source-only training pool, which would violate the DG setting and invalidate the comparison with AADG and DoFE. Please state the exclusion rule explicitly for each fold and confirm that only generators trained on the three source datasets are used in that fold's training set.","section":"Section 4.1 and Section 4.2"},{"comment":"The prose claims in Section 4.3 are numerically inconsistent with Table 1. The text says AADG achieves a 3.29% average DSC improvement over baseline, but Table 1 gives AADG average 77.60 vs baseline 75.13 (a 2.47-point gap). It then says DGSSA improves another 3.78% over baseline, but Table 1 gives DGSSA average 77.98 (2.85 points over baseline), and the stated \"0.60% increase on the HRF dataset\" is also inconsistent with the table's HRF values (DGSSA 72.66 vs baseline 72.23, a 0.43-point gap). Please correct these deltas and rephrase the SOTA claim to reflect the actual 0.38-point average margin over AADG.","section":"Section 4.3, Table 1"},{"comment":"The statistical support for the SOTA claim is under-specified. Table 5 reports paired t-test p-values, but the paper does not state whether the unit of analysis is each fold (n=4), per-image DSC, or per-pixel DSC; no standard deviations or confidence intervals accompany the central Table 1 results. Given the small average margin over AADG, the claim that DGSSA \"surpasses existing methods\" is not yet quantitatively supported. Please report per-fold results with standard deviations across repeated runs and clarify the t-test unit, sample size, and whether the test is paired across the same test images.","section":"Section 4.3 and Section 5.4, Tables 1 and 5"}],"minor_comments":[{"comment":"There are two algorithms labeled \"Algorithm 1\" (the space colonization procedure and the PixMix procedure); renumber the second as Algorithm 2 and update the in-text reference in Section 3.3.","section":"Section 3.2 and Section 3.3"},{"comment":"The dataset name appears as \"CHASEDB\" in the abstract but \"CHASEDB1\" elsewhere; please standardize the spelling.","section":"Abstract and throughout"},{"comment":"The opening sentence says \"We conducted a comprehensive evaluation of the impact of AADG\" but the paper evaluates DGSSA; this appears to be a typo.","section":"Section 4.1"},{"comment":"These equations restate the loss functions but do not include the unpaired-data terms defined in Eqs. (5)-(7) and (11), and the hyperparameter values (λ_L1=100, λ_adv=0.2, λ_1=0.3, λ_GP=10) are only given in the text; please align the equations with the full loss definitions.","section":"Section 3.4, Eqs. (17)-(18)"},{"comment":"The tested ranges for λ_1, λ_adv, λ_GP, and λ_L1 are not stated in the caption or text; please specify them so the sensitivity analysis is reproducible.","section":"Figure 5"},{"comment":"The caption refers to \"domains a, b, c, and d\" while the columns are labeled A, B, C, D, and AVG; please clarify the notation and define what AVG averages over, and note explicitly which rows are borrowed from AADG rather than re-evaluated.","section":"Table 1 caption"},{"comment":"The paper does not mention code or data release; providing the implementation and pretrained generators would substantially strengthen reproducibility.","section":"General reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent incremental contribution with a plausible underlying idea: combining structural synthesis with style augmentation for DG in tubular structures. The main obstacle is not the novelty but the rigor of the evaluation: the leave-one-out protocol is ambiguous about whether target-trained generators leak into source training, and the prose overstates the numerical margins. I believe these issues are fixable within the scope of a revision, but they must be addressed before the SOTA claim can be accepted. The self-citations (refs. 13-15) appear in the UDA background and do not affect the central claim, so I do not see a citation-pattern concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible and cheap augmentation idea for retinal vessel domain generalization, and the thin-vessel analysis is the most credible part, but the paper never closes a protocol gap that could invalidate its headline DG result, and the SOTA edge is 0.38 DSC points against borrowed numbers.\n\nThe new thing is real: instead of style-only augmentation, they procedurally generate vascular structures with a space colonization algorithm, render them into pseudo-fundus images with a modest Pix2Pix extension (multi-scale discriminator, unpaired training), and mix those into the source pool alongside PixMix style augmentation. The ablations are coherent and support the direction of the claim: structural augmentation alone lifts average DSC from 75.13 to 76.83, and the thin-vessel partition shows a 13.1% relative improvement over baseline, which is credible evidence that structural diversity is what helps the hard cases. They also report a U-Net variant and a hyperparameter sensitivity check. The paper is honest about the uncertainty perturbation giving no significant DSC gain (Table 2 and Section 5.4).\n\nThe soft spots are load-bearing. First, the leave-one-out protocol is underspecified. Section 4.2 says the Pix2Pix model is trained independently on each of the four datasets, and the paper never states that the generator for the held-out target is excluded from producing pseudo-images for the source training pool. If that generator is allowed in, the held-out numbers are contaminated by target appearance, and the comparison with AADG/DoFE is unfair. This is not a confirmed flaw, but it determines the validity of the 77.98% headline, and a referee must ask for an explicit statement. Second, the SOTA claim rests on a 0.38-point average DSC gain over AADG numbers that are copied, not reproduced, and no variance is reported despite 'repeated operations' being mentioned. That margin is within run-to-run noise for this kind of segmentation. Third, the prose in Section 4.3 has percentage claims (3.29%, 0.60%, 3.78%) that do not match the table deltas; sloppy, fixable, but it undercuts trust in the numbers.\n\nWho is this for: people working on domain generalization for fundus or tubular-structure segmentation, and method developers looking for cheap structural augmentation. It deserves a serious referee, but the SOTA sentence should not survive until the protocol is clarified and the baseline comparison is reproduced or given error bars.","headline":"Interesting structural-augmentation idea for retinal vessel DG, but a protocol gap could invalidate the headline result and the SOTA edge is within noise.","tokens_in":22119,"tokens_out":3167,"would_cite":false,"duration_ms":27390,"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":"Training on procedurally grown vessel structures plus style mixing raises cross-dataset retinal vessel segmentation Dice to 77.98%.","keywords":["domain generalization","retinal vessel segmentation","structural augmentation","space colonization algorithm","style augmentation","PixMix","medical image segmentation","fundus images"],"falsifier":"Look at the training files for each leave-one-out fold and see whether the pseudo-retinal images produced by the Pix2Pix model trained on the target dataset are present in the source pool; if they are, the 77.98% average Dice is not a measurement of generalization to an unseen domain.","tokens_in":20952,"feed_emoji":"👁️","tokens_out":6618,"duration_ms":57013,"temperature":0.7,"pith_summary":"This paper argues that a retinal vessel segmentation model generalizes to datasets it has never seen better when its training data are made more diverse in both image style and vessel structure, not just style. To that end it introduces DGSSA, which procedurally generates artificial vascular trees with a space-colonization algorithm, renders them into pseudo-retinal images with an improved Pix2Pix generator, and then further diversifies all training images with PixMix-style photometric mixing and uncertainty perturbation. In leave-one-out tests across four fundus datasets, the paper reports an average Dice similarity coefficient of 77.98%, about 0.38 percentage points above the style-only method AADG. If the result holds, procedural structure generation offers a cheap way to make segmentation models robust to new scanners and patient populations without collecting target-domain data.","feed_headline":"Synthetic vessels plus style mix lift DG Dice to 77.98%","feed_subtitle":"Structural plus stylistic augmentation beats style-only methods across four retinal datasets in leave-one-out tests.","key_machinery":"The load-bearing mechanism is the pairing of a procedural structure generator with a conditional image synthesizer. The space-colonization algorithm grows branching vascular networks by iteratively placing nodes toward attractor points, then applies Murray's-law-style thickening and post-processing to produce realistic vessel masks; an improved Pix2Pix model with a multi-scale discriminator converts these masks, together with real paired masks, into pseudo-retinal images. Style augmentation via PixMix with uncertainty perturbation then mixes each image with augmented variants and externally sourced images under random photometric operations. Together they provide the segmentation network with both structural and stylistic diversity during training, so that the evaluation protocol is the only part of the pipeline that sees the target domain.","core_discovery":"The central claim is that combining structural augmentation with stylistic augmentation improves domain generalization for retinal vessel segmentation more than either alone. The paper's full method, DGSSA, reports average DSC 77.98% across STARE (82.17), HRF (72.66), DRIVE (78.62), and CHASEDB1 (78.45), surpassing the AADG baseline (77.60) and other comparison methods. The authors attribute the gain to structure augmentation: space-colonization-generated vessel masks rendered by a multi-scale Pix2Pix model into pseudo-retinal images, which forces the segmentor to learn a wider range of vessel morphologies; style augmentation (PixMix plus uncertainty perturbation) then adds photometric variation on top of both original and synthetic images. Their ablation shows structure augmentation alone accounts for most of the thin-vessel improvement, while the full combination reaches the highest average.","pith_inferences":["A testable extension is to vary the number of source domains and the number of generated structures per domain; the marginal value of structural augmentation may shrink once real source diversity is large.","The same space-colonization-plus-synthesis recipe could be applied to other tubular structures such as coronary arteries or bronchi, where a branching prior matches the anatomy.","The unresolved exclusion rule could be checked by rerunning one fold with target-trained generated images deliberately included and excluded; the difference would bound the contamination effect.","Because the paper reports that structure augmentation helps thin vessels most, a clinical deployment might measure generalization gains on microvessel density rather than overall Dice."],"forward_implications":["If correct, adding procedurally generated vascular structures to the training set improves cross-dataset retinal vessel segmentation beyond style-only augmentation.","The method requires no change to the segmentation architecture and no extra computation at inference, so it can be attached to existing segmentors.","The reported thin-vessel analysis implies the benefit concentrates on fine vessel morphology, where structure augmentation gives a 13.1% relative Dice improvement over baseline.","The combination is architecture-agnostic: the paper reports gains with both DeepLabv3+/MobileNetV2 and U-Net/ResNet backbones.","Using generated structures plus style mixing on the source domains could reduce the need for collecting labeled data from every new clinical site."],"supporting_citations":[{"why":"Provides the style-only domain generalization baseline (AADG) that DGSSA must beat, with the evaluation protocol and reported metrics used for direct comparison.","marker":"[21]"},{"why":"Supplies the space-colonization algorithm used to generate vascular-like branching structures.","marker":"[58]"},{"why":"Introduces the Pix2Pix conditional GAN that the paper improves with a multi-scale discriminator to render pseudo-retinal images.","marker":"[48]"},{"why":"Introduces PixMix, the style-mixing augmentation the paper builds on with uncertainty perturbation.","marker":"[41]"},{"why":"Sets the leave-one-out dataset partitioning and provides the DoFE comparison method in the domain generalization evaluation.","marker":"[19]"},{"why":"Provides the ELCFS comparison method and its continuous-frequency-space interpolation approach for cross-domain generalization.","marker":"[20]"}],"fun_headline_variants":["Synthetic structure + style mix lifts retinal vessel Dice to 77.98%","Structural and style augmentation hit 77.98% Dice across four retinal sets","DGSSA: 77.98% Dice via structural plus stylistic augmentation","Retinal vessel segmentation: combine structure and style for 77.98% Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported generalization scores are valid only if no image or generator trained on the held-out target dataset contributes to the source training pool, and the paper never states that exclusion rule.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic structure + style mix lifts retinal vessel Dice to 77.98%","Structural and style augmentation hit 77.98% Dice across four retinal sets","DGSSA: 77.98% Dice via structural plus stylistic augmentation","Retinal vessel segmentation: combine structure and style for 77.98% Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00091,"raw_usage":{"total_tokens":3906,"prompt_tokens":934,"completion_tokens":2972,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":2887}},"tokens_in":550,"tokens_out":2972,"duration_ms":20543,"temperature":1.0,"reasoning_tokens":2887,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:52:59.657381+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Look at the training files for each leave-one-out fold and see whether the pseudo-retinal images produced by the Pix2Pix model trained on the target dataset are present in the source pool; if they are, the 77.98% average Dice is not a measurement of generalization to an unseen domain.","supporting_citations":[{"cited_title":"AADG: Automatic augmentation for domain generalization on retinal image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the style-only domain generalization baseline (AADG) that DGSSA must beat, with the evaluation protocol and reported metrics used for direct comparison."},{"cited_title":"Modeling trees with a space colonization algorithm,","cited_arxiv_id":null,"evidence_quote":"Supplies the space-colonization algorithm used to generate vascular-like branching structures."},{"cited_title":"Image-to-image translation with conditional adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Introduces the Pix2Pix conditional GAN that the paper improves with a multi-scale discriminator to render pseudo-retinal images."},{"cited_title":"Pixmix: Dreamlike pictures comprehensively improve safety measures,","cited_arxiv_id":null,"evidence_quote":"Introduces PixMix, the style-mixing augmentation the paper builds on with uncertainty perturbation."},{"cited_title":"Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets,","cited_arxiv_id":null,"evidence_quote":"Sets the leave-one-out dataset partitioning and provides the DoFE comparison method in the domain generalization evaluation."},{"cited_title":"Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space,","cited_arxiv_id":null,"evidence_quote":"Provides the ELCFS comparison method and its continuous-frequency-space interpolation approach for cross-domain generalization."}],"review_version":1}