{"id":"e2e6bfab-2181-48bc-ae32-7cb69290eb8a","arxiv_id":"2504.14737","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SuperCL pre-trains segmentation networks by using superpixel-based pseudo masks to create intra-image and inter-image contrastive pairs, improving Dice scores on eight medical datasets.","lead":"This paper introduces SuperCL, a contrastive learning pre-training method that uses superpixel maps to define positive pixel pairs and inter-image weak labels for medical image segmentation. A smart generalist might read it because it reports consistent accuracy gains on eight medical datasets with only 10% or 25% of labels, which could reduce the need for expert annotation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA gains may reflect test-set hyperparameter tuning and unreported seeded-subset variance rather than a robust advantage; a seeded rerun with a held-out validation split is needed.","rationale":"I agree with the reader's conditional verdict and with the importance of the superpixel semantic assumption, but I judge the evaluation-bias concern to be more load-bearing for the central claim. Even if superpixels are semantically imperfect, SuperCL could still improve segmentation by acting as a feature-smoothing regularizer; the empirical gains reported in Table I would remain evidence of effectiveness. By contrast, if the comparison is biased by tuning hyperparameters on the test datasets and by unreported variance, the claim of superiority over 12 methods collapses regardless of the mechanism. The reader's weakest_assumption focuses on Eq. 5 and the ILCP positives, but their rationale already lists the missing error bars, unspecified subset selection, and hyperparameter selection. My concern is therefore partially overlapping with the reader's, but I would elevate the evaluation-validity issue to the primary risk. The suggested test directly addresses that risk: a seeded rerun with a validation split and variance reporting would determine whether the reported gains are robust or artifacts. This does not require changing the reader's conditional verdict; it reinforces the conditions under which the central claim could be accepted.","tokens_in":23032,"tokens_out":4131,"duration_ms":38322,"concrete_test":"Rerun the 10% and 25% label fine-tuning experiments for SuperCL and the strongest baseline (e.g., PCL) on MMWHS, CHAOS, and Spleen with 5 independent seeds each, using a fixed labeled-subset selection procedure (e.g., random patient-split with a documented seed) and a held-out validation split for all hyperparameter choices (superpixel cluster number, stride, and loss weights). Report mean +/- std DSC for each method. If the validation-selected hyperparameters fail to reproduce the reported gains, or if the SuperCL advantage falls within one standard deviation of the baseline, the claim 'SuperCL outperforms existing 12 methods' is not empirically supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SuperCL outperforms 12 existing methods rests on the empirical comparison in Tables I and II, but the evaluation protocol contains a selection-bias risk that could invalidate the reported gains. In Section IV-B and Fig. 8, the superpixel cluster number is chosen as 100 based on downstream performance on ACDC and MMWHS with 25% labels; likewise Table IV selects stride = 1 on the same datasets. These are the same datasets and label budgets used in the main comparison tables, so the headline improvements (e.g., 3.15% DSC on MMWHS at 10% labels) are partially the product of tuning hyperparameters on the test sets. No held-out validation split is described for any hyperparameter, including the loss weights lambda_1, lambda_2, lambda_3 and temperature tau. Additionally, the labeled-subset selection for 10%/25% fine-tuning is unspecified: it is not stated whether subsets are stratified by patient, slice, or random with which seeds, and no error bars or multiple-seed results are reported. On small datasets such as MMWHS (20 patients) and HVSMR (10 patients), the reported gaps could easily fall within run-to-run variance, making the SOTA claim unsupported. The concern is not about the method's theoretical soundness but about whether the empirical evidence is sufficient to establish superiority, which is exactly what the abstract asserts.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SuperCL, a self-supervised contrastive pre-training method for medical image segmentation. SuperCL uses SLIC superpixels to define two contrastive pair generation strategies: ILCP, which treats pixels within the same superpixel as positive pairs in an intra-image pixel-level contrastive loss, and IGCP, which uses ASP (averaging features per superpixel region) and CCL (connected components on a 1-nearest-neighbor graph) to build a weak label for an inter-image instance-level contrastive loss. The method is evaluated by pre-training a U-Net encoder on CHD, BraTS2018, and KiTS2019, then fine-tuning on eight downstream datasets (ACDC, MMWHS, HVSMR, CHAOS, Spleen, ISIC, Heart, Hippocampus) with 10% and 25% of the training labels, comparing against 12 baselines. The authors report consistent DSC improvements and additionally show that SuperCL can initialize semi-supervised methods (BCP, CauSSL), with ablations isolating the contributions of ILCP and IGCP and sensitivity studies on stride and superpixel cluster number.","tokens_in":23322,"tokens_out":3463,"duration_ms":32999,"significance":"If the reported gains are robust, SuperCL would be a useful addition to the medical-image contrastive-learning toolbox: it replaces manually tuned thresholds for contrastive pair generation with a superpixel-derived grouping, and it is evaluated far more broadly than is typical in this area (8 downstream datasets, 12 baselines, ablations, backbone generalization, and semi-supervised integration). The proposed modules (ASP and CCL) are simple, plausible, and clearly described. The main unresolved issue is not the method's internal logic but the strength of the empirical evidence supporting the central 'outperforms 12 methods' claim, given the evaluation protocol described below.","major_comments":[{"comment":"The same concern applies to the loss weights (lambda_1, lambda_2, lambda_3 = 1.0, 1.0, 0.5) and temperature (tau = 0.1) in Section III-A and the Implementation Details: no sensitivity analysis or validation-based selection is described for these values, and if they were chosen with knowledge of the test-set results, the reported gains are again subject to selection bias.","section":"Table IV and Fig. 8"},{"comment":"This issue is load-bearing for the abstract's central claim, which is stated as a definitive comparison ('outperforms existing 12 methods'). At minimum, the manuscript should present error bars for the key comparisons in Tables I and II, or clearly state that the reported values are single runs and therefore only indicative.","section":"Section IV-D (Implementation Details)"},{"comment":"The claim that superpixels provide semantically valid positive pairs is the load-bearing assumption of ILCP, but the paper does not provide any direct evidence that the SLIC superpixel boundaries align with anatomical or tissue boundaries in the pre-training datasets, nor does it quantify the sensitivity of the result to this assumption. Fig. 6 and Fig. 8 show the effect of superpixel method and cluster count on downstream DSC, which is indirect evidence, but it would strengthen the paper to report, for a few representative slices, the fraction of superpixels that straddle label boundaries (using the available pre-training annotations) or otherwise to test the robustness of the method to the superpixel granularity. The current ablation varies K from 1 to 1024, which is useful, but the interpretation of the results as evidence for boundary alignment is not established.","section":"Section III-B, Eq. (5)"}],"minor_comments":[{"comment":"The manuscript contains numerous typos and grammatical errors that should be corrected before publication, including 'avaliable', 'Resonably', 'performace', 'simlar', 'avereged', 'untilizing', 'pesudo', 'datset', and 'Simens' (e.g., in Sections I, III-C, IV-A, IV-B, and the dataset descriptions).","section":"Throughout"},{"comment":"The text refers to 'Section 2.2' and 'Section 2.3' when discussing ILCP and IGCP, but the actual sections are III-B and III-C; please update the cross-references.","section":"Section III-B"},{"comment":"The definition of the averaged superpixel feature P is ambiguous: each P_c is a scalar (the mean of the channel-averaged feature over superpixel region c), yet later P is used as an image of size h x w. Please clarify that P is formed by filling each superpixel region with the scalar value P_c, and state explicitly which image (the original X or the augmented view I_1/I_2) is used to compute the SLIC superpixel map S, since non-spatial augmentations change pixel intensities and could affect the superpixel grouping.","section":"Section III-C, Eq. (6)"},{"comment":"The statement in the discussion of Fig. 8 that 'our SuperCL among all the cluster numbers except for 150 outperforms the PCL baseline' would be more informative if the figure showed error bars or repeated runs, since the differences among many cluster numbers appear small (within about 1% DSC); as written, the figure may not support the conclusion that K = 100 is uniquely or significantly better.","section":"Section IV-B, Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a well-motivated method and an unusually broad experimental study, but the evaluation protocol has a clear selection-bias risk: hyperparameters are chosen on the same datasets and label budgets that are later used for the headline comparisons, and no repeated runs or error bars are provided. This is fixable through a revised evaluation protocol (held-out validation splits, multiple seeds, and a clear description of labeled-subset sampling), and the core methodological contribution is sound enough to warrant a major revision rather than rejection. I would also urge the editor to require the authors to state explicitly that the reported numbers are single runs if that is the case, so readers can calibrate the strength of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zeng et al. propose SuperCL, which uses SLIC superpixels to generate pseudo masks for two contrastive objectives: intra-image pixel-level pairs (ILCP) and inter-image instance-level pairs (IGCP, via their ASP/CCL modules). The core idea is sensible—use local structure to define positives instead of hand-set thresholds—and the specific combination is new relative to WCL, PCL, and GCL. The empirical investment is real: 8 downstream datasets, 12 baselines, ablations for each component, superpixel-method comparison, backbone generalization, and a semi-supervised integration. The ablations consistently show ILCP contributes more than IGCP, which fits the intuition that pixel-level grouping matters for dense prediction.\n\nThe soft spot is the evaluation protocol, and it is load-bearing. The cluster number K and the stride in ILCP are chosen in Section IV-B by looking at downstream DSC on ACDC and MMWHS with 25% labels—the same datasets and label budget used in the main comparison tables. There is no described validation split for hyperparameters, no error bars or repeated runs, and the selection of the 10%/25% labeled subsets is unspecified. The loss weights and temperature are given as empirical settings with no sensitivity study. On MMWHS (20 patients) and HVSMR (10 patients), a different slice sample can plausibly swing several DSC points; the reported gains of 3.15%, 5.44%, 7.89% could partly be a product of that variance plus test-set tuning. This is not a theoretical flaw in the method—it is a fixable protocol gap. Re-run with held-out validation for hyperparameters, report multiple seeds, and spell out how labeled subsets are drawn. The fact that improvements appear across all 8 datasets, not just the tuned ones, suggests something real is there, but the magnitude is not yet supported. I am less worried about the superpixel-boundary concern; the ablations show ILCP helps, so the positives are apparently informative enough, even if failure cases are not discussed.\n\nCitation pattern is fine; the self-cited LNSNet is relevant prior work. The paper is aimed at researchers doing self-supervised pre-training for segmentation with scarce labels. It deserves a serious referee: the method is interesting and the experimental breadth is worth engaging. I would send it to review with a request for a validated re-run. I'd cite the method idea, but not the empirical claim, until that rerun happens.","headline":"Superpixel-guided contrastive pre-training that looks genuinely useful, but the reported SOTA gains rest on test-set hyperparameter tuning and unstated seed variance; worth a careful peer review, not yet a claim to trust.","tokens_in":23862,"tokens_out":5464,"would_cite":true,"duration_ms":43664,"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":"SuperCL pre-trains a U-Net with SLIC superpixel pseudo masks to define positive contrastive pairs, and it outperforms 12 methods on 8 medical segmentation datasets under limited annotations.","keywords":["contrastive learning","medical image segmentation","superpixel","self-supervised pre-training","limited annotations","SLIC","U-Net","intra-image contrastive pairs"],"falsifier":"Corrupt the superpixel pseudo mask in a controlled way, for example by randomly reassigning 10–50% of superpixel labels or by shifting boundaries away from organ edges, and re-run pre-training with identical settings; if downstream DSC on MMWHS and CHAOS does not drop as corruption increases, the reported gains are not caused by the superpixel grouping itself.","tokens_in":22839,"feed_emoji":"🩻","tokens_out":9755,"duration_ms":79123,"temperature":0.7,"pith_summary":"The paper proposes SuperCL, a contrastive pre-training method for medical image segmentation when annotated data are scarce. Its central idea is to let superpixel maps act as pseudo labels: pixels that fall in the same SLIC superpixel are treated as positive pairs for an intra-image contrastive loss, and averaged superpixel features feeding a connected-components step on a nearest-neighbor graph define positive pairs across images. This design removes the manually set thresholds that earlier medical contrastive methods rely on. The paper reports that SuperCL outperforms 12 baselines on 8 CT and MRI datasets, with DSC gains of 3.15, 5.44, and 7.89 points over the previous best on MMWHS, CHAOS, and Spleen at 10% annotations.","feed_headline":"Superpixel pre-training lifts medical segmentation by 7.89 DSC","feed_subtitle":"Pixels in one superpixel become positive pairs, so pre-training needs no thresholds and beats 12 methods on 8 datasets.","key_machinery":"The load-bearing object is the superpixel pseudo mask $S \\in \\mathbb{R}^{B \\times h \\times w}$, generated by SLIC and downsampled to feature-map resolution. Pixel-wise, it defines the positive set $\\Omega_{ILCP}$ for the intra-image loss $L_{intra}$ in Eq. (5). Instance-wise, its per-cluster averaged features $P_c$ feed the ASP module, and the connected-components weak label $W$ from a top-1 nearest-neighbor graph defines the positive set for $L_{inter}$ in Eq. (9). The machinery converts raw pixel intensities into a grouping prior that substitutes for manually tuned thresholds.","core_discovery":"The central claim is that superpixel structure is a reliable, annotation-free grouping signal for contrastive learning in medical images. SuperCL computes a superpixel map per image, flattens it into a pseudo mask aligned with the encoder feature map, and applies a supervised contrastive loss that pulls together features of pixels sharing a superpixel cluster (ILCP, Eq. (5)). For inter-image pairs, it averages the feature map inside each superpixel (ASP) and runs connected-components labeling on a top-1 nearest-neighbor graph over those averaged features (CCL), producing weak labels that define global positives (IGCP, Eq. (9)). The paper argues this exploits both intra-image pixel correlation and inter-image structural similarity while avoiding threshold tuning, and that it yields state-of-the-art segmentation performance when a U-Net is fine-tuned on 10% or 25% of downstream labels.","pith_inferences":["The method's ceiling is set by superpixel boundary quality: on lesions or organs with low intensity contrast, SLIC clusters will cross anatomical borders, and ILCP will then pull together features that should stay apart; comparing SLIC with boundary-aware superpixels would show whether segmentation gains track edge recall.","SuperCL is threshold-free only for the pairing step; cluster number, stride, and loss weights still need tuning, and the paper reports sensitivity only for cluster number and stride, not for all eight datasets.","ASP performs a fixed non-learnable average pooling inside each superpixel; making that pooling learnable or multi-scale is a natural extension that could strengthen the inter-image positives.","The semi-supervised experiments cover only ACDC; testing SuperCL as an initialization for semi-supervised methods on CT datasets would tell whether the benefit generalizes across modalities."],"forward_implications":["If the reported gains hold, pre-training on unlabeled CT/MRI volumes with SuperCL should improve U-Net segmentation at 10% and 25% annotation rates across multi-organ and ROI tasks, with the largest margins at the lowest label ratios.","The same pre-trained encoder transfers to U-Net variants (AttUNet, UCTransNet, BCDUNet, ResUNet, RollingUNet, UKAN), so the benefit is not tied to one decoder design.","SuperCL also serves as an initialization for semi-supervised pipelines: the paper reports that it raises BCP and CauSSL performance on ACDC at 10% and 20% labels.","Because the pseudo masks come from a generic algorithm, the strategy can be dropped into other dense-prediction encoders as long as a feature map can be aligned with the superpixel map."],"supporting_citations":[{"why":"SLIC generates the superpixel pseudo masks that define positive pairs for both ILCP and IGCP.","marker":"[12]"},{"why":"WCL supplies the weak-label nearest-neighbor graph idea that CCL adapts for inter-image positives.","marker":"[9]"},{"why":"GCL is the partition-based contrasting baseline that motivates SuperCL's threshold-free pair generation and is used in comparisons.","marker":"[3]"},{"why":"PCL is the position-based contrasting baseline that SuperCL builds on for Lins and compares against in ablations.","marker":"[4]"},{"why":"U-Net is the architecture whose encoder SuperCL pre-trains and which is fine-tuned for all downstream segmentation tasks.","marker":"[13]"}],"fun_headline_variants":["Superpixel positives beat threshold tuning in medical segmentation pretraining","No thresholds: superpixel contrastive learning wins on 8 medical datasets","Superpixel clusters as free labels: contrastive pretraining gains 7.89 DSC","Superpixel-guided contrastive learning: no thresholds, SOTA on 8 datasets","SuperCL: superpixel positives replace threshold tuning for medical pretraining"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that SLIC superpixels, computed from raw pixel intensities, group pixels into the same anatomical region, so that pixels sharing a superpixel are genuinely positive pairs rather than false ones.","fun_headline_variants_meta":{"raw":{"variants":["Superpixel positives beat threshold tuning in medical segmentation pretraining","No thresholds: superpixel contrastive learning wins on 8 medical datasets","Superpixel clusters as free labels: contrastive pretraining gains 7.89 DSC","Superpixel-guided contrastive learning: no thresholds, SOTA on 8 datasets","SuperCL: superpixel positives replace threshold tuning for medical pretraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000586,"raw_usage":{"total_tokens":2805,"prompt_tokens":1048,"completion_tokens":1757,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":1655}},"tokens_in":664,"tokens_out":1757,"duration_ms":10591,"temperature":1.0,"reasoning_tokens":1655,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:40:55.111928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Corrupt the superpixel pseudo mask in a controlled way, for example by randomly reassigning 10–50% of superpixel labels or by shifting boundaries away from organ edges, and re-run pre-training with identical settings; if downstream DSC on MMWHS and CHAOS does not drop as corruption increases, the reported gains are not caused by the superpixel grouping itself.","supporting_citations":[{"cited_title":"Weakly supervised contrastive learning,","cited_arxiv_id":null,"evidence_quote":"WCL supplies the weak-label nearest-neighbor graph idea that CCL adapts for inter-image positives."},{"cited_title":"Contrastive learning of global and local features for medical image segmentation with limited annotations,","cited_arxiv_id":null,"evidence_quote":"GCL is the partition-based contrasting baseline that motivates SuperCL's threshold-free pair generation and is used in comparisons."},{"cited_title":"Positional contrastive learning for volumetric medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"PCL is the position-based contrasting baseline that SuperCL builds on for Lins and compares against in ablations."}],"review_version":1}