{"id":"a1e27c4a-6559-4e40-b504-f236462b90fe","arxiv_id":"2507.17779","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A self-supervised contrastive masked autoencoder pre-training scheme for UNet improves coronary artery segmentation from X-ray angiography in low-data settings.","lead":"CM-UNet pre-trains a UNet with a contrastive masked autoencoder on unlabeled X-ray angiograms, then fine-tunes it on a small labeled set. On FAME2, it reaches a Dice score of 0.626 with only 18 annotated images, versus 0.375 for a supervised UNet, suggesting SSL can cut annotation needs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The capacity confound in Table I is not cleanly interpretable: CM-UNet's 121.49M parameter count includes SSL-only branches that §III.B says are discarded at fine-tuning, so the claimed superiority over other SSL methods is untested under a matched inference architecture.","rationale":"The reader's conditional verdict is reasonable. The paper's headline low-data improvement is large and directionally robust, but the comparison underpinning the method-specific claim is not controlled. My refinement of the reader's weakest assumption is that the parameter mismatch is more than a confound: it is internally inconsistent with the model description in §III.B. Because of this, the decisive check is not simply to make baselines bigger; it is first to determine what CM-UNet's fine-tuned inference graph actually contains. The absence of error bars strengthens the concern: with one seed, a 0.06-0.10 Dice gap could be within run-to-run variation. The private dataset limits external generalization but is not a flaw in the internal comparison. I therefore keep the existing CONDITIONAL verdict; no change in outcome, but the condition should explicitly require a capacity/architecture-matched re-evaluation and repeated-seed reporting.","tokens_in":9861,"tokens_out":8712,"duration_ms":91470,"concrete_test":"Use the released code to instantiate the fine-tuned CM-UNet exactly as in §III.B and print the parameter count after removing the pre-training-only branches (momentum encoder, feature decoder, projection heads). If that count is close to 31M, re-run all SSL baselines at the 79:1 ratio with identical backbone width, optimizer, batch size, learning-rate schedule, and at least 5 seeds, changing only the pre-training loss. Report Dice mean±std. If CM-UNet's margin over Models Genesis/MAE falls within seed noise (overlapping error bars), the claim that CM-UNet outperforms other SSL methods is not established; if a 121M from-scratch UNet matches CM-UNet, the gain is capacity rather than the SSL objective.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the parameter-count/capacity comparison behind the claim that CM-UNet outperforms other SSL methods. In Table I, CM-UNet is listed with 121.49M parameters versus 31.04-37.70M for the baselines, and the text interprets this as evidence that 'the increased complexity enhances segmentation accuracy.' However, §III.B says that after pre-training only the online reconstruction encoder E_r and pixel decoder D_r are retained and 'combined to form a UNet model with skip connections.' If that is what is fine-tuned, the deployed model is architecturally the same UNet used for all baselines, and the 121.49M count must include the momentum encoder, feature decoder, and projection heads used only during pre-training. Table I would then report pre-training parameter counts, not the parameters of the models that compute the segmentations. If instead a 121M network is really fine-tuned, then no capacity-matched baseline exists and the comparison is confounded by architecture size. Under either reading, the +0.063 Dice margin over the best other SSL method (0.626 vs 0.563) cannot be attributed specifically to the contrastive masked autoencoder objective. The paper itself lists the missing ablation of the reconstruction and contrastive branches as a limitation, and no error bars or repeated-seed results are reported for the Dice comparisons. The broader claim that SSL pre-training helps with 18 fine-tuning images is still supported by the large 0.375-to-0.626 gap to the no-pretraining baseline, but the method-specific superiority claim is not.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CM-UNet, a self-supervised pretraining framework for coronary artery segmentation from X-ray angiography. It combines a masked reconstruction branch with a contrastive branch during pretraining on unlabeled images, then retains the reconstruction encoder and decoder as a UNet that is fine-tuned on small labeled subsets. On the FAME2 dataset, the authors report that CM-UNet reaches a Dice of 0.626 with only 18 fine-tuning images versus 0.375 for a supervised UNet, and that its Dice drop when fine-tuning data shrinks from 500 to 18 images is 15.2% compared with 46.5% for the supervised baseline. The paper also benchmarks four other SSL methods, reports clinical metrics such as artery diameter difference and Bland-Altman agreement, and releases source code.","tokens_in":10183,"tokens_out":9070,"duration_ms":113995,"significance":"If the central comparisons were clean, the paper would make a useful practical contribution: it addresses a real annotation bottleneck in coronary angiography, evaluates on a clinical dataset with held-out test images, and provides clinical metrics beyond Dice. The release of source code and the use of a fixed seed for reproducibility are strengths, as is the breadth of SSL baselines considered. However, the head-to-head claim that CM-UNet outperforms other SSL methods currently rests on a capacity confound and on unreplicated single-run numbers, so the magnitude and attribution of the observed gains are not yet established.","major_comments":[{"comment":"The capacity comparison that underlies the claim that CM-UNet outperforms other SSL methods is confounded. §III.B states that after pre-training only the reconstruction encoder E_r and pixel decoder D_r are retained and combined into a UNet; this is the same four-level UNet architecture described in §IV.C for all baselines. Table I nevertheless reports 121.49M parameters for CM-UNet versus 31.04-37.70M for the baselines, and the text interprets this as \"increased complexity enhances segmentation accuracy.\" Since 121.49M appears to include the momentum encoder, feature decoder, and projection/prediction heads used only during pre-training, the deployed segmentation model is not 121.49M parameters. If the 121.49M model is actually the fine-tuned model, then no capacity-matched baseline exists. Under either reading, the +0.063 Dice margin over the best baseline in Table I cannot be attributed specifically to the contrastive masked autoencoder objective. Please report the parameter count of the fine-tuned CM-UNet and add a same-capacity baseline or an SSL baseline with a matched inference architecture.","section":"§III.B, Table I, §IV.C"},{"comment":"All segmentation results are reported as single numbers with no variance. Table I reports one Dice value per method at the 79:1 ratio, Fig. 3 plots curves without error bars or confidence bands, and §V calls the 4.3% improvement over the non-pre-trained model \"statistically significant (p<0.05)\" without naming a test or describing the distribution. With 18 fine-tuning images, random initialization, augmentation, and pretraining stochasticity can materially change results. Please run multiple seeds (or repeated fine-tuning) and report means with standard deviations or confidence intervals, and state the hypothesis test used for any significance claim.","section":"Fig. 3, Table I, §V"},{"comment":"The paper itself lists the absence of an ablation of the reconstruction and contrastive components as a limitation. This omission is load-bearing because §III.A motivates the combined objective as balancing \"global\" and \"fine-grained\" features, and the main numerical advantage of CM-UNet is presented as the product of this combination. A controlled ablation (reconstruction-only, contrastive-only, and combined within the same framework and budget) is needed to establish that the fusion, rather than one of the branches, drives the gains; the baselines MAE and MoCo do not substitute for this because they use different objectives, implementations, and pre-training budgets.","section":"§III.A, §V"},{"comment":"The claim that CM-UNet \"outperforms competing SSL approaches\" is not supported on all metrics in Table I. For ADD, CM-UNet scores 0.965, which is worse than Model Genesis's 0.893; the sentence reporting a -13.33% reduction in ADD therefore needs a clearly stated reference value. If the comparison is to the mean of the SSL baselines, the table implies about a 16.8% reduction, not 13.33%; if the comparison is to the best baseline, the claim is false on ADD. Please recompute and report all percentage improvements against an explicitly named reference, and qualify the outperformance claim per metric.","section":"Table I, §IV.D1"}],"minor_comments":[{"comment":"The sentence \"only 1% of the training dataset (18 images)\" is inconsistent with a 1,390-image training set; 18 images is approximately 1% of all 1,738 images. Clarify the denominator for all PT-FT ratios.","section":"§IV.D1"},{"comment":"The relationship among the contrastive encoder E_c, the momentum encoder, and the projection heads is unclear. The caption says the momentum encoder outputs are compared with the feature decoder outputs, while the text says E_c processes X_k^B. Specify which encoder is the momentum encoder, whether gradients flow through it, and how negative examples are collected.","section":"§III.A3, Fig. 2"},{"comment":"The text says Fig. 6 visualizes reconstruction results, but the caption describes segmentation outputs; align the text, the caption, and the figure panels.","section":"§IV.D3, Fig. 6"},{"comment":"There are several typos that should be corrected: \"employes\" in §III.A2, \"pPT-FT ratios\" in the Fig. 5 caption, and \"Pytorch\" in §IV.C.","section":"§III.A2, Fig. 5, §IV.C"},{"comment":"The Bland-Altman and PCC analyses are reported only for CM-UNet at two PT-FT ratios; reporting the same clinical metrics for the best SSL baseline would make the clinical-validity comparison more informative.","section":"§IV.D2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central finding is real: self-supervised pre-training on unlabeled angiograms helps a lot when you only have 18 labeled images for fine-tuning. The gap between no pre-training (Dice 0.375) and CM-UNet (0.626) is large, and the trend across data ratios in Fig. 3 is consistent. That part is solid, and it is the main thing most readers will take away.\n\nWhat's new: a systematic comparison of SSL methods (Model Genesis, MAE, SparK, MoCoV2, CMAE-style) for coronary artery segmentation on the FAME2 dataset, and a new application of contrastive masked autoencoders to X-ray angiography. The components are borrowed, but the benchmark and the low-data evaluation are not in the prior literature. The code is public, which is a plus.\n\nThe soft spots are in the method-specific claims. First, Table I lists CM-UNet at 121.49M parameters versus 31-37M for baselines, but Section III.B says fine-tuning retains only the reconstruction encoder and pixel decoder, which form the same UNet used everywhere. So that parameter count is for the pre-training machinery, not the deployed model. The paper then says the higher parameter count 'demonstrates that the increased complexity enhances segmentation accuracy,' which is either misleading or indicates a real capacity confound. Note the deployed model is not capacity-mismatched at inference; the confound is at pre-training, where CM-UNet uses substantially more capacity and compute. You cannot attribute the +0.063 Dice margin specifically to the objective. Second, there are no error bars or repeated-seed runs for the Dice numbers, so a 0.626 vs 0.563 difference could be within noise. Third, the paper itself acknowledges the missing ablation of the reconstruction and contrastive branches. That is not fatal, but it means the design choices are not isolated.\n\nAlso, the FAME2 data is not public; the code is, but the results are not independently reproducible without access to the dataset.\n\nOverall: this deserves serious referee time. The broad claim about SSL for low-data segmentation is well supported and useful; the specific claim about CMAE superiority needs tightening. I would ask for error bars, an ablation, and a clear statement of what parameters are being reported before acceptance.","headline":"The broad claim that SSL pre-training helps in low-data coronary segmentation holds up; the specific CMAE superiority claim is muddied by a parameter-count ambiguity and missing error bars.","tokens_in":10752,"tokens_out":4864,"would_cite":true,"duration_ms":50005,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that CM-UNet, a UNet pre-trained with a contrastive masked autoencoder on unlabeled coronary X-ray angiography, keeps segmentation accurate when fine-tuned on as few as 18 annotated images, where a conventionally trained…","keywords":["coronary artery segmentation","X-ray angiography","self-supervised learning","contrastive masked autoencoder","low-data fine-tuning","UNet","Dice score","transfer learning"],"falsifier":"Train a plain UNet with the same 121.49-million-parameter architecture and the same 18-image fine-tuning protocol; if it reaches or exceeds Dice 0.626 without any pre-training, the reported gain is explained by capacity rather than by the contrastive masked autoencoder.","tokens_in":9663,"feed_emoji":"🩻","tokens_out":6767,"duration_ms":66536,"temperature":0.7,"pith_summary":"The paper tries to show that self-supervised pre-training can replace most manual annotation in coronary artery segmentation from X-ray angiography. It introduces CM-UNet, a UNet pre-trained with a contrastive masked autoencoder on unlabeled angiograms, then fine-tuned on a small labeled set. The central result is that with only 18 labeled images for fine-tuning, CM-UNet reaches a Dice score of 0.626, while a supervised UNet without pre-training drops to 0.375. The authors argue this matters because annotation of coronary vessels is costly and scarce, and a model that learns from unlabeled data could make automated stenosis assessment practical in more clinics.","feed_headline":"Self-supervised pretraining keeps artery segmentation alive on 18 images","feed_subtitle":"Combining contrastive and masked reconstruction pre-training trims the annotated-data cost for coronary X-ray analysis.","key_machinery":"The central object is the Contrastive Masked Autoencoder (CMAE) objective, adapted to a UNet backbone. One branch masks 65% of $16 \\times 16$ patches and reconstructs them with an encoder-decoder; a second branch applies heavier augmentations and feeds a momentum encoder whose outputs are pulled together with the reconstruction branch's features through an InfoNCE loss. The two losses are combined as $\\mathcal{L} = \\mathcal{L}_r + \\lambda_c \\mathcal{L}_c$, and after pre-training the online encoder and pixel decoder become the initialization for the segmentation UNet. This joint objective is what carries the argument: reconstruction preserves local vessel continuity, while contrastive alignment keeps representations discriminative across images.","core_discovery":"CM-UNet combines two self-supervised objectives into one pre-training stage: masked image reconstruction, which forces the network to recover fine vessel details, and contrastive learning, which aligns global representations of different augmented views of the same angiogram. After pre-training on unlabeled data from the FAME2 dataset, the reconstruction encoder and decoder are reused as the segmentation UNet and fine-tuned with a Dice plus cross-entropy loss. The paper reports that this pipeline outperforms both a supervised UNet and other self-supervised baselines in low-data regimes; at a 79:1 pre-training-to-fine-tuning ratio, CM-UNet scores 0.626 Dice versus 0.563 for the best alternative SSL method and 0.375 without pre-training. The authors conclude that self-supervised pre-training, especially the combined contrastive-masked objective, is what lets segmentation survive extreme label scarcity.","pith_inferences":["A capacity-matched test the paper leaves open: if a 121M-parameter UNet without SSL also reaches roughly 0.6 Dice on 18 images, the contribution of the contrastive-masked objective would need to be re-examined.","If the effect is real, the same two-stage recipe should transfer to other thin-structure segmentation tasks, such as cerebral or peripheral angiography, where labeled examples are even rarer.","The paper lists ablation of the reconstruction and contrastive components as future work; one predictable outcome is that reconstruction matters most for vessel continuity and contrastive learning for discriminating arteries from background clutter."],"forward_implications":["Fine-tuning CM-UNet on 18 labeled images instead of 500 costs 15.2% in Dice, whereas non-pre-trained models lose 46.5%, so the amount of annotation needed for usable coronary segmentation can shrink by more than an order of magnitude.","At the 79:1 data split, CM-UNet's Dice of 0.626 beats the best other self-supervised baseline (0.563) and the supervised UNet (0.375), suggesting the combined objective generalizes better than contrastive-only or reconstruction-only pre-training in this regime.","Across all tested pre-training-to-fine-tuning ratios, pre-trained models outperform the non-pre-trained model, and the gap widens as fine-tuning data shrinks, implying SSL becomes more valuable exactly when annotations are rarest.","CM-UNet's predicted vessel diameters correlate with manual ground-truth diameters (PCC 0.76 at 50:30 and 0.57 at 79:1), indicating the model preserves clinically meaningful measurements even under scarce labels."],"supporting_citations":[{"why":"Supplies the Contrastive Masked Autoencoder pre-training objective that CM-UNet adapts to a UNet backbone.","marker":"[5]"},{"why":"Provides the FAME2 X-ray angiography dataset, including the 1,738 labeled images used for pre-training, fine-tuning, and testing.","marker":"[6]"},{"why":"Supplies the UNet architecture used as the backbone for both pre-training and downstream segmentation.","marker":"[8]"},{"why":"Serves as a generative self-supervised baseline whose 31.04M-parameter UNet is compared against CM-UNet.","marker":"[16]"},{"why":"Provides the masked autoencoder baseline that CM-UNet extends with a contrastive branch.","marker":"[21]"},{"why":"Supplies the MoCo contrastive baseline used to benchmark CM-UNet in low-data fine-tuning.","marker":"[18]"},{"why":"Serves as the sparse convolutional masked-modeling baseline (SparK) in the SSL comparison.","marker":"[22]"},{"why":"Defines the InfoNCE loss used to align the reconstruction and contrastive branches during pre-training.","marker":"[25]"}],"fun_headline_variants":["Self-supervised pre-training enables artery segmentation with 18 images","Combined SSL pre-training cuts annotation need for coronary segmentation","CM-UNet: masked + contrastive pre-training for low-label vessel segmentation","Self-supervised pretraining reduces label need for X-ray artery segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes CM-UNet's advantage comes from its self-supervised objective rather than from its much larger network, because no equally sized UNet or capacity-matched SSL baseline is trained.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised pre-training enables artery segmentation with 18 images","Combined SSL pre-training cuts annotation need for coronary segmentation","CM-UNet: masked + contrastive pre-training for low-label vessel segmentation","Self-supervised pretraining reduces label need for X-ray artery segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00079,"raw_usage":{"total_tokens":3493,"prompt_tokens":968,"completion_tokens":2525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":2450}},"tokens_in":584,"tokens_out":2525,"duration_ms":21091,"temperature":1.0,"reasoning_tokens":2450,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:57:47.912149+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a plain UNet with the same 121.49-million-parameter architecture and the same 18-image fine-tuning protocol; if it reaches or exceeds Dice 0.626 without any pre-training, the reported gain is explained by capacity rather than by the contrastive masked autoencoder.","supporting_citations":[{"cited_title":"Contrastive Masked Autoencoders are Stronger Vision Learners,","cited_arxiv_id":null,"evidence_quote":"Supplies the Contrastive Masked Autoencoder pre-training objective that CM-UNet adapts to a UNet backbone."},{"cited_title":"Fractional flow reserve-guided PCI versus medical therapy in stable coronary disease,","cited_arxiv_id":null,"evidence_quote":"Provides the FAME2 X-ray angiography dataset, including the 1,738 labeled images used for pre-training, fine-tuning, and testing."},{"cited_title":"U-Net: Convolutional Networks for Biomedical Image Segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the UNet architecture used as the backbone for both pre-training and downstream segmentation."},{"cited_title":"Models Gen- esis: Generic Autodidactic Models for 3D Medical Image Analysis,","cited_arxiv_id":null,"evidence_quote":"Serves as a generative self-supervised baseline whose 31.04M-parameter UNet is compared against CM-UNet."},{"cited_title":"Masked autoencoders are scalable vision learners,","cited_arxiv_id":null,"evidence_quote":"Provides the masked autoencoder baseline that CM-UNet extends with a contrastive branch."},{"cited_title":"Momentum con- trast for unsupervised visual representation learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the MoCo contrastive baseline used to benchmark CM-UNet in low-data fine-tuning."},{"cited_title":"Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling,","cited_arxiv_id":null,"evidence_quote":"Serves as the sparse convolutional masked-modeling baseline (SparK) in the SSL comparison."},{"cited_title":"Representation Learning with Contrastive Predictive Coding,","cited_arxiv_id":null,"evidence_quote":"Defines the InfoNCE loss used to align the reconstruction and contrastive branches during pre-training."}],"review_version":1}