{"id":"e37a7678-510e-4ab1-b685-96ee24bb50e2","arxiv_id":"2411.13147","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GraphCL adds graph-based structure modeling and a correlation clustering loss to the BCP framework, reporting higher Dice scores on ACDC, LA, and Pancreas-NIH.","lead":"This paper introduces GraphCL, a semi-supervised medical image segmentation method that augments the bidirectional copy-paste training framework with graph neural network alignment and a graph clustering loss. The method reports higher Dice scores than published baselines on three medical imaging benchmarks, though the gains are modest and the evaluation has several reporting inconsistencies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"On ACDC 10%, the full GraphCL reports Dice 89.31, below SA-only (89.52) and LCC-only (89.53), contradicting the claim that combining components yields the best results; this inconsistency undermines the central empirical claim.","rationale":"The reader's weakest assumption focused on the underspecified DSA and the potential for the clustering loss to collapse to trivial assignments. My concern is complementary and more directly tied to the central empirical claim: the paper's own ablation data contain a concrete contradiction where the full method is not the best configuration on a primary metric. This indicates that the reported experimental evidence is internally inconsistent, which is a load-bearing problem for the claim of universal superiority. The reader's CONDITIONAL verdict already accounts for the need for revisions, so I do not change the verdict; however, this specific inconsistency should be explicitly addressed before the results can be considered reliable. The proposed concrete test would settle whether the contradiction is a typo or a real phenomenon.","tokens_in":14783,"tokens_out":14555,"duration_ms":134795,"concrete_test":"Run the ACDC 10% labeled ablation (Baseline, SA-only, LCC-only, and full GraphCL) under the paper's exact protocol with three random seeds and report mean plus/minus standard deviation for Dice. If the full model does not outperform both single-component variants in at least two of three seeds, the component-complementarity claim is false; if it does, the reported Table 4 is erroneous and the remaining results still need independent verification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim that GraphCL outperforms state-of-the-art semi-supervised medical image segmentation methods is not reliably supported by the paper's own reported numbers. In the ACDC 10% ablation (Table 4), the full model (SA+LCC) achieves Dice 89.31, which is lower than both SA-only (89.52) and LCC-only (89.53). This directly contradicts the text in Section 4.4 stating that 'incorporating both SA and LCC achieves the best results.' If the full model is not the best configuration on one of the three benchmarks, then the apparent improvements over SOTA in Tables 1-3 may be due to noise, seed, or cherry-picking rather than the proposed components. Additionally, Table 6 is internally garbled, with a 1-layer GCN entry showing Dice 0.030 and 95HD 80.51, which are not plausible segmentation metrics for the ACDC dataset; no error bars or code are provided, so the reproducibility of the headline numbers cannot be independently assessed. These internal contradictions and missing evidence make the abstract's unqualified claim unsubstantiated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GraphCL, a semi-supervised medical image segmentation method built on the BCP teacher-student framework. It adds two graph-based components: a structure-aware alignment module that constructs an instance graph from DSA-derived structure scores and propagates features with a GCN, and a correlation clustering loss L_CC computed from a patch-affinity matrix. The method is evaluated on LA, ACDC, and Pancreas-NIH with Dice, Jaccard, 95HD, and ASD, reporting improvements over BCP and other baselines in most settings.","tokens_in":15090,"tokens_out":6741,"duration_ms":62113,"significance":"The direction is plausible: using graph structure and clustering as a regularizer for SSMIS is a reasonable extension of the BCP line, and the reported Dice/Jaccard gains are consistently positive across three datasets. However, the paper's central claim is empirical, and the current evidence is weakened by internal inconsistencies in the ablations, an implausible table entry, incorrect text-vs-table reporting, and the absence of variance information and code. The contribution is potentially useful if those issues are resolved.","major_comments":[{"comment":"In Table 4, on ACDC with 10% labeled data, the full model (SA+LCC) reaches Dice 89.31, while both SA-only (89.52) and LCC-only (89.53) are higher. This directly contradicts the Section 4.4 statement that 'incorporating both SA and LCC achieves the best results,' and it undermines the component-combination claim that supports the method. Additionally, on LA with 10% labeled data, LCC-only (88.79) is below the baseline (89.39), which contradicts the text's claim that adding SA or LCC individually improves performance over the baseline. The authors should report results with error bars or multiple seeds and explain these discrepancies.","section":"4.4, Table 4"},{"comment":"Section 4.3 states that on ACDC with 10% labeled data GraphCL outperforms the second-best approach by an average of 26.01% across all four metrics, and by 29.65% with 5% labeled data. These numbers do not match Table 2, where absolute Dice/Jaccard differences over BCP are below 3 percentage points and 95HD/ASD are measured in distance units. The average-improvement figures need to be recalculated or removed.","section":"4.3, ACDC paragraph"},{"comment":"Table 6 reports Dice of 0.030 and 0.008 with Jaccard of 0.016 and 0.004 for the '1 layer' configuration on the ACDC dataset, along with 95HD of 80.51 and 40.94. These values are implausible for the evaluated segmentation task and are not discussed in the text. Because this table is used to support the claim that deeper GCN placement is optimal, it cannot serve as evidence in its current form.","section":"4.4, Table 6"},{"comment":"The Data Structure Analyzer is only characterized by its output X_sa used in Eq. (13); its architecture, input, and output dimension h are never specified. Similarly, the 'k-less' correlation clustering in Eq. (17) uses a softmax MLP whose output dimension implicitly defines the number of clusters, but that dimension and the handling of the softmax assignments are not given. Without these details, the method is not reproducible, and the risk that L_CC in Eq. (18) degenerates to trivial cluster assignments cannot be assessed.","section":"3.3.1-3.3.2"},{"comment":"The abstract states that GraphCL outperforms state-of-the-art semi-supervised medical image segmentation methods, but on Pancreas-NIH Table 3 shows ASD of 2.12, which is worse than CoraNet's 1.89. The claim should be qualified per metric and per dataset. In addition, no error bars or multiple-seed statistics are provided anywhere, and no code is released, so the robustness of the headline improvements cannot be assessed.","section":"Abstract; Table 3"}],"minor_comments":[{"comment":"The expression W = F·F^T - Max(F·F^T) / τ is ambiguous; parentheses should clarify whether the maximum is subtracted before or after division by τ, and the role of τ as a temperature should be stated consistently with Section 4.5.","section":"Eq. (15)"},{"comment":"The text says κ controls the weight of the structure-aware alignment, but Eq. (19) multiplies κ by the clustering loss L_CC. This inconsistency should be corrected.","section":"Section 4.5"},{"comment":"The visualization is described as 'kernel dense estimations' but the axes, color scale, and the feature layer used are not identified, making it hard to interpret.","section":"Figure 5"},{"comment":"The claim of being the first to model data structure information for SSMIS is not substantiated; given the graph-based SSL surveys and GNN segmentation work cited in Section 2, the novelty claim should be softened or placed in a more precise context.","section":"Introduction"},{"comment":"Some references contain typos (e.g., 'NeruIPS' should be 'NeurIPS'); please proofread the reference list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from a careful revision of the reported numbers and an honest statement of variability. If the ACDC 10% ablation finding is robust, the claims in the abstract and Section 4.4 need to be scaled back. I would not recommend acceptance in current form, but the method is a reasonable extension and the issues are addressable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the one thing to know: GraphCL is a plausible incremental extension of BCP for semi-supervised medical image segmentation. The main tables report 1-2 point Dice gains over BCP on LA, ACDC, and Pancreas-NIH. That is the empirical claim. What is new is the combination of BCP with a GCN alignment module and a correlation clustering loss derived from pairwise feature affinities. The package as a whole is not in the cited literature, and the writing is reasonably clear. The ablations show that SA and LCC each help in most rows, which is a point in the authors' favor.\n\nThe soft spots are serious but not fatal. The ACDC 10% ablation (Table 4) reports the full model at Dice 89.31, below both SA-only (89.52) and LCC-only (89.53), while the text says combining both achieves the best. That contradiction is load-bearing because the abstract rests on the full model beating SOTA. If the full model is not the best configuration on one of the three benchmarks, the headline gains look like seed luck rather than a property of the method. Table 6 also has a 1-layer GCN row with Dice 0.030 and 95HD 80.51, which is not a plausible segmentation result for ACDC; that looks like a typo, but in a submission it is unacceptable. There are no error bars, no code, and the text's reported average improvements (e.g., 29.65% on ACDC 5%) do not match the tables. The DSA in Eq. 13 is underspecified, and the clustering loss L_CC is computed from the same features being optimized, with no analysis of collapse or of the implied number of clusters in the softmax output. All of this is fixable, but it currently prevents me from trusting the central claim.\n\nWho this is for: people working on semi-supervised medical image segmentation, especially those building on BCP. The method is worth a serious look, but the evidence needs to be cleaned up. I would send this to peer review rather than desk reject: the idea is not silly, the gains are large if real, and referees can pressure the authors into releasing code and fixing the tables. I would not cite it in my own work until the ACDC 10% contradiction is resolved and the code is out. It's a good reading group case study in how to read ablation tables critically.","headline":"A plausible incremental method for semi-supervised medical segmentation whose central empirical claim is undermined by an internal contradiction in the ACDC ablation.","tokens_in":15555,"tokens_out":3449,"would_cite":false,"duration_ms":33772,"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 modeling medical images as graphs and adding a correlation-clustering loss improves semi-supervised segmentation, reporting top Dice scores on LA, ACDC, and Pancreas-NIH.","keywords":["semi-supervised learning","medical image segmentation","graph neural networks","correlation clustering","graph convolutional networks","teacher-student framework","volumetric segmentation","structure-aware alignment"],"falsifier":"Fix everything in the pipeline but replace the Data Structure Analyzer's structure scores with random noise; if the reported gains persist, the structure-aware graph is not doing the claimed work. In the same spirit, inspect the learned cluster assignments $S$ on a validation fold: if voxels from different organs fall mostly into one or two clusters, the clustering objective has collapsed and the reported gains must come from the copy-paste baseline.","tokens_in":14589,"feed_emoji":"🩺","tokens_out":8368,"duration_ms":74390,"temperature":0.7,"pith_summary":"The paper tries to establish that semi-supervised medical image segmentation improves when the model explicitly represents graph structure and clusters features with a graph neural network. It argues that existing methods lean on complex training strategies while ignoring structural relationships between samples and between voxels. GraphCL builds a dense instance graph from CNN features, propagates information through a graph convolutional network, and adds a correlation-clustering loss that groups similar voxels without needing a preset number of clusters. On three benchmarks (LA, ACDC, Pancreas-NIH) with 5–20% labeled volumes, it reports higher overlap-based Dice and Jaccard scores and generally lower boundary distances than the compared semi-supervised methods. If the claim holds, adding graph-structure modeling to teacher-student pipelines is a simple and effective way to use unlabeled data.","feed_headline":"GraphCL reports top Dice scores on three medical benchmarks","feed_subtitle":"Graph structure plus a clustering loss beats strong baselines using only 5–20% labeled scans.","key_machinery":"The load-bearing object is a graph built from the network's own features. Each sample in a mini-batch becomes a node, and a Data Structure Analyzer produces structure scores $X_{sa}$; the adjacency matrix is $\\hat{A} = X_{sa}X_{sa}^\\top$, on which a graph convolutional network propagates features as $Z = \\hat{D}^{-1/2}\\hat{A}\\hat{D}^{-1/2}XW$. A second, voxel-level graph uses the correlation of deep features, $W = F F^\\top - \\max(F F^\\top)/\\tau$, and a GNN plus an MLP with softmax produces the cluster assignment matrix $S$. The correlation-clustering loss $L_{CC} = -\\operatorname{Tr}(W S S^\\top)$, added to the segmentation loss with weight $\\kappa$, is the mechanism intended to make similar voxels cluster together without a predefined number of clusters.","core_discovery":"GraphCL's central claim is that jointly modeling data structure—at the image level through a structure-aware graph and at the voxel level through graph clustering—improves semi-supervised segmentation in one unified network. The paper says this is the first work to model data-structure information for semi-supervised medical image segmentation. On the LA and ACDC datasets, it claims to outperform the compared state-of-the-art methods on all four reported metrics at both 5% and 10% labeled ratios; on Pancreas-NIH, it reports higher Dice, Jaccard, and 95HD than the best compared method. The reported best Dice scores are 88.80% and 90.24% on LA, 88.68% and 89.31% on ACDC, and 83.15% on Pancreas-NIH. The driving objective is the correlation-clustering loss $L_{CC} = -\\operatorname{Tr}(W S S^\\top)$, which rewards giving similar voxels the same cluster assignment while separating dissimilar voxels.","pith_inferences":["The paper leaves implicit that the clustering loss could be tested as a standalone regularizer on other semi-supervised baselines; if it helps them too, the benefit is generic rather than tied to the specific graph construction.","Because the adjacency matrix and cluster assignments are computed from the same features the loss optimizes, a direct check of cluster purity against ground-truth anatomy would show whether $S$ actually corresponds to organs or is merely a feature regularizer.","The ASD result on Pancreas-NIH, where GraphCL does not beat the best baseline on average surface distance, suggests that boundary-distance behavior is not uniformly improved; a natural follow-up is to weight the clustering loss per class or combine it with an explicit boundary term."],"forward_implications":["If GraphCL's empirical claim is right, graph-structure modeling is a viable source of improvement for semi-supervised medical segmentation, not just for supervised or weakly supervised tasks.","The k-less correlation-clustering loss can act as a regularizer that sharpens boundary delineation, according to the paper's ablation studies, without requiring the user to specify the number of clusters.","The reported gains are largest at the 5% labeled ratio on LA and ACDC, suggesting graph structure is most helpful when labeled data are scarcest.","The ablation over GCN placement implies the position of the graph module inside the encoder matters, with deeper insertion giving the best results."],"supporting_citations":[{"why":"Supplies the bidirectional copy-paste teacher-student framework and the mixed-sample construction that GraphCL builds on.","marker":"[2]"},{"why":"Defines the graph convolutional layer used to propagate features on the instance graph.","marker":"[15]"},{"why":"Provides the SS-Net baseline and the benchmark protocol for LA and ACDC.","marker":"[47]"},{"why":"Provides the CoraNet baseline and the Pancreas-NIH experimental setup.","marker":"[35]"},{"why":"Releases the ACDC dataset used in the two-label-ratio experiments.","marker":"[4]"},{"why":"Releases the LA dataset used in the 5% and 10% labeled experiments.","marker":"[48]"},{"why":"Releases the Pancreas-NIH dataset used in the 20% labeled experiment.","marker":"[32]"},{"why":"Supplies the mean-teacher weight update rule for the teacher network.","marker":"[40]"},{"why":"Introduces the V-Net backbone used for the 3D experiments.","marker":"[27]"}],"fun_headline_variants":["GraphCL uses graph clustering to beat SSL baselines on 3 medical sets","First graph-based clustering for semi-supervised medical segmentation","GraphCL: unified graph model cuts labeled data to 5% for top Dice","Clustering loss in GraphCL outperforms SOTA on LA, ACDC, Pancreas","GraphCL achieves top Dice on LA, ACDC, and Pancreas-NIH"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the graph built from the network's own features capturing true anatomical structure, and on the clustering loss pushing features into useful groups instead of collapsing to trivial assignments.","fun_headline_variants_meta":{"raw":{"variants":["GraphCL uses graph clustering to beat SSL baselines on 3 medical sets","First graph-based clustering for semi-supervised medical segmentation","GraphCL: unified graph model cuts labeled data to 5% for top Dice","Clustering loss in GraphCL outperforms SOTA on LA, ACDC, Pancreas","GraphCL achieves top Dice on LA, ACDC, and Pancreas-NIH"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000321,"raw_usage":{"total_tokens":1790,"prompt_tokens":908,"completion_tokens":882,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":781}},"tokens_in":524,"tokens_out":882,"duration_ms":8661,"temperature":1.0,"reasoning_tokens":781,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:47:25.179784+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix everything in the pipeline but replace the Data Structure Analyzer's structure scores with random noise; if the reported gains persist, the structure-aware graph is not doing the claimed work. In the same spirit, inspect the learned cluster assignments $S$ on a validation fold: if voxels from different organs fall mostly into one or two clusters, the clustering objective has collapsed and the reported gains must come from the copy-paste baseline.","supporting_citations":[{"cited_title":"Deeporgan: Multi-level deep convolutional networks for automated pan- creas segmentation","cited_arxiv_id":null,"evidence_quote":"Releases the Pancreas-NIH dataset used in the 20% labeled experiment."},{"cited_title":"Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-teacher weight update rule for the teacher network."},{"cited_title":"V-net: Fully convolutional neural networks for volumetric medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Introduces the V-Net backbone used for the 3D experiments."},{"cited_title":"Bidirectional copy-paste for semi-supervised medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the bidirectional copy-paste teacher-student framework and the mixed-sample construction that GraphCL builds on."},{"cited_title":"Semi-supervised classifi- cation with graph convolutional networks","cited_arxiv_id":null,"evidence_quote":"Defines the graph convolutional layer used to propagate features on the instance graph."},{"cited_title":"Exploring smoothness and class-separation for semi-supervised medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the SS-Net baseline and the benchmark protocol for LA and ACDC."},{"cited_title":"Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the CoraNet baseline and the Pancreas-NIH experimental setup."},{"cited_title":"Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? TMI, 37(11):2514–2525, 2018","cited_arxiv_id":null,"evidence_quote":"Releases the ACDC dataset used in the two-label-ratio experiments."},{"cited_title":"A global benchmark of algo- rithms for segmenting the left atrium from late gadolinium- enhanced cardiac magnetic resonance imaging","cited_arxiv_id":null,"evidence_quote":"Releases the LA dataset used in the 5% and 10% labeled experiments."}],"review_version":1}