{"id":"067871d2-6fcc-438a-ba98-b296c4b5f218","arxiv_id":"2506.21245","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"GANet-Seg couples a pretrained normal-brain GAN with a U-Net for brain tumor segmentation, but its own Table 2 contradicts the abstract's claim of better HD95 than baselines.","lead":"This paper pairs a pretrained GAN that models normal brain MRI with a U-Net segmenter, using the GAN's reconstruction and adversarial feedback to refine tumor boundaries. The reported gains are mixed and the evaluation borrows baseline numbers from another study, so the results are not reliable as stated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 contradicts the abstract: GANet-Seg's WT Dice (81.28) trails UMamba (85.69), and its WT/ET HD95 values are worse than every listed baseline, so the claimed 'both lesion-wise Dice and HD95' superiority is unsupported by the paper's own results.","rationale":"The reader's weakest assumption was that baseline numbers from Heras Rivera et al. [26] may not be comparable to the authors' self-selected 36-case test set, which would invalidate comparative claims. That is a valid and serious concern. My stress-test identifies a different, more direct failure: even taken entirely at face value, the paper's own Table 2 contradicts the abstract's blanket claim of superiority. On WT Dice, the model is worse than two baselines; on WT and ET HD95, it is worse than all baselines. Only the lesion-wise Dice columns and TC HD95 support the claim. This internal inconsistency does not depend on external benchmark provenance, so it is more load-bearing than the comparability issue alone: a corrected protocol cannot rescue the claim unless the reported numbers themselves change. The reader's verdict of REJECT remains appropriate, but the justification should include this direct contradiction, not only the cross-paper comparison concern. The concrete test is a simple row-by-row inspection of Table 2; no new experiments are required to see that the reported evidence fails the stated claim. My agreement is partial because the reader's concern is related but distinct; the internal contradiction is the stronger reason for rejection.","tokens_in":8521,"tokens_out":3257,"duration_ms":35319,"concrete_test":"Directly inspect Table 2 and check, for each of WT/ET/TC, whether GANnet-Seg exceeds every listed baseline on the stated metrics. The check yields: WT Dice 81.28 < UMamba 85.69 and < Optimized U-Net 81.44; WT HD95 27.06 > all baselines; ET HD95 50.13 > all baselines. If the claim is interpreted as global superiority on both Dice and HD95, it fails on the face of the table without any reimplementation. If the authors intend a narrower claim, they must specify it and revise the abstract accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the abstract is that the model achieves high sensitivity and accuracy in both lesion-wise Dice and HD95 metrics than the baseline. The paper's own Table 2 does not support this without qualification. On WT Dice, GANnet-Seg reports 81.28, which is lower than both UMamba (85.69) and Optimized U-Net (81.44). On HD95, GANnet-Seg is worse than every listed baseline for WT (27.06 vs 14.51, 22.30, 22.85) and ET (50.13 vs 37.48, 48.47, 48.83). Only on TC HD95 (13.95) and on the lesion-wise Dice columns is GANnet-Seg consistently better. The abstract is not merely imprecise; it asserts a comparative advantage that the reported numbers refute for major regions and metrics. Even if we set aside the independent concern that baseline numbers were copied from another paper with a different test split, the internal table is enough to invalidate the blanket claim of superiority. This is a load-bearing correctness issue because the paper's contribution is explicitly framed as beating the baseline on these metrics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GANet-Seg, a brain tumor segmentation framework that combines a U-Net segmentation module with a GAN pretrained on normal brain MRI. The GAN is used to reconstruct occluded tumor regions, and an edge-guided adversarial loss derived from a Laplacian-filtered mask and a PatchGAN discriminator is added to refine boundaries. The authors report results on a self-selected 36-case test set from BraTS 2020 and compare against Optimized U-Net, Swin-UNETR, nnU-Net, and UMamba, with baseline numbers taken from a separate paper. The abstract claims high accuracy in both lesion-wise Dice and HD95 relative to baselines, and the paper frames the method as reducing dependence on fully annotated data.","tokens_in":8866,"tokens_out":3924,"duration_ms":51492,"significance":"If the reported results were valid, the idea of using a pretrained normal-brain GAN as a source of anomaly-aware feedback for tumor segmentation would be a useful contribution to semi-supervised medical image segmentation, especially for datasets with scarce annotations. The paper also provides explicit loss formulations and a clear architectural diagram, which are helpful starting points for replication. However, the central comparative claim is contradicted by the paper's own table, and the evaluation protocol does not permit a fair comparison with the cited baselines. The current evidence does not support the stated contributions, so the significance as presented is low.","major_comments":[{"comment":"The abstract states that the model achieves \"high sensitivity and accuracy in both lesion-wise Dice and HD95 metrics than the baseline,\" but Table 2 does not support this. For the whole tumor (WT), GANet-Seg reports Dice 81.28 versus UMamba's 85.69, and for WT HD95 it reports 27.06 versus UMamba's 14.51. For enhancing tumor (ET) HD95, GANet-Seg reports 50.13 versus UMamba's 37.48, and it is also worse than the other listed baselines on these metrics. The only region where GANet-Seg is consistently better is the tumor core (TC), and the lesion-wise Dice columns generally improve. This is a load-bearing discrepancy because the paper's contribution is explicitly framed as outperforming baselines on both Dice and HD95.","section":"Abstract and Table 2"},{"comment":"The baseline results for Optimized U-Net, nnU-Net, Swin-UNETR, and UMamba are not re-run on the authors' test set; they are copied from reference [26]. The authors use a self-selected set of 36 BraTS 2020 samples, and there is no evidence that the preprocessing, the exact test split, the evaluation code, or the metric definitions match those used in [26]. Without a common evaluation protocol, every comparative statement in Table 2 is invalid. The paper should either reproduce the baselines under the identical protocol or restrict claims to descriptive results without comparative language.","section":"Section 5.5.2 and Section 5.1.2"},{"comment":"The claim that the method \"minimizes the dependency on fully annotated data\" is not supported by the experimental design. The U-Net is trained with full ground-truth masks using cross-entropy and Dice losses, and the size consistency loss in Eq. (6) uses S_label, the labeled pixel count from ground truth, for every training slice. The GAN pretraining uses unlabeled normal-brain data, but the segmentation training still requires dense pixel-level annotations for all training subjects. The paper should either quantify the reduction in annotation demand or soften this claim substantially.","section":"Section 5.4.1, Appendix A.3.2, and Abstract"},{"comment":"No variance, confidence interval, or statistical significance is reported for any metric, and the test set consists of only 36 cases. Given that the reported differences between GANet-Seg and UMamba are small for several metrics (e.g., ET Dice 79.81 vs. 77.41), the absence of any uncertainty measure makes it impossible to determine whether the observed differences are meaningful. At minimum, the authors should report per-case standard deviations or confidence intervals and state the statistical test used.","section":"Section 5.5 and Table 2"}],"minor_comments":[{"comment":"The model name is inconsistently written as \"GANet-Seg\" in the title and abstract and \"GANnet-Seg\" in Section 5.5.2 and Table 2; please unify the terminology.","section":"Throughout"},{"comment":"\"Unbuntu 24.02\" should be \"Ubuntu 24.04\" or the correct version; this appears to be a typo.","section":"Section 5 (Experimental Results)"},{"comment":"The sentence \"The GAN pretraining module to learn the is designed to learn the distribution of normal brain\" contains a duplicated fragment and should be rewritten.","section":"Section 4.1.1"},{"comment":"Figures 4, 5, and 6 duplicate the captions and content of Figures 1 and 2 from the main text; the appendix should be removed or the figures should be cross-referenced rather than repeated.","section":"Appendix E"},{"comment":"References [3] and [19] appear to cite the same BRATS paper; please consolidate the duplicated reference.","section":"References"},{"comment":"The notation for the lesion-wise Dice coefficient is unclear: the denominator is written as (|P_i|+|T_i|)/2 without showing the factor 2 in the numerator, which can confuse readers about the equivalence to the standard Dice definition.","section":"Appendix D, Eq. (18)"},{"comment":"The description of the edge map operation \"a max-pooling with stride size of 5×5\" is ambiguous; it should specify the pooling window and stride separately.","section":"Section 4.1.3"}],"recommendation":"reject","confidential_remarks":"The central comparative claim is contradicted by the paper's own Table 2, and the baseline comparison is not methodologically valid. The paper would require a major redesign of the evaluation and a re-framing of the contribution before it could be considered for publication. I do not see a path to acceptance within the scope of a minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a plausible architectural idea wrapped in an evaluation that doesn't support the claims. The paper combines a pretrained PatchGAN discriminator on normal-brain MRI, a U-Net segmenter, occlusion-based reconstruction, and an edge-guided adversarial loss. That combination is new as far as I know, and the underlying intuition—using the GAN's sensitivity to abnormal structure to guide boundary refinement—is worth exploring. The authors also acknowledge in the discussion that their current model does not delineate tumor sub-regions, which is an honest limitation statement.\n\nThe problems are in the results and their presentation. The abstract claims 'high sensitivity and accuracy in both lesion-wise Dice and HD95 metrics than the baseline.' Table 2 does not support that. GANet-Seg's WT Dice (81.28) is below UMamba (85.69), and its WT and ET HD95 (27.06, 50.13) are worse than every listed baseline. Only on TC HD95 and lesion-wise Dice does it lead. So the blanket superiority claim is false on its face.\n\nWorse, the discussion says the framework lacks the ability to delineate the specific sub-regions (ET, TC), yet Table 2 reports ET and TC metrics. Either the model produces sub-region labels somewhere, or the numbers are computed by some other protocol. The paper doesn't explain this. This internal contradiction makes the results hard to trust.\n\nThe baseline numbers are copied from another paper [26] rather than re-run on the same 36-case test split. There are no error bars, and the reference list contains a placeholder 'John Doe' entry [15]. These are not minor. A paper that asks us to accept a new architecture's superiority needs a controlled comparison, and this doesn't have one.\n\nWhat's genuinely useful: the idea of pretraining a generator/discriminator on unlabeled normal brain data and using occlusion-based reconstruction as an anomaly prior is a reasonable direction for semi-supervised segmentation. The size-consistency loss and edge-guided adversarial feedback are sensible ingredients. But as presented, the evidence is insufficient to conclude these components help.\n\nWho is this for? Someone working on semi-supervised brain tumor segmentation might find the architecture section worth a skim, but they shouldn't rely on the numbers. I would not bring it to reading group.\n\nVerdict: desk reject. The idea is not dead on arrival, but the paper needs a major rewrite, a real benchmark protocol, and a resolution of the sub-region contradiction before it deserves referee time.","headline":"A plausible adversarial-GAN architecture for tumor segmentation undermined by a results table that contradicts its own abstract and discussion.","tokens_in":9356,"tokens_out":2409,"would_cite":false,"duration_ms":26024,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GANet-Seg claims that a U-Net trained with edge-weighted adversarial feedback from a pre-trained normal-brain GAN achieves more accurate brain tumor segmentation than four published supervised baselines, with tumor-core Dice of 88.84%.","keywords":["brain tumor segmentation","GAN","U-Net","adversarial learning","MRI","anomaly detection","BraTS dataset","lesion-wise Dice"],"falsifier":"Re-run Optimized U-Net, nnU-Net, Swin-UNETR, and UMamba on the exact 36-case split with the paper's preprocessing and then compare TC Dice and HD95 with GANet-Seg; if any baseline equals or exceeds the reported 88.84 TC Dice under identical conditions, the central comparative claim fails. A cheaper check is recomputing both volume-weighted and lesion-wise Dice on the same predictions, since the paper reports both in Table 2 and any large inconsistency would indicate metric or protocol drift.","tokens_in":8346,"feed_emoji":"🧠","tokens_out":8112,"duration_ms":81266,"temperature":0.7,"pith_summary":"This paper tries to establish that brain-tumor segmentation accuracy can be improved by coupling a U-Net with a GAN that has learned the appearance of normal brain MRI. The U-Net proposes a tumor mask; the GAN's generator fills the masked region with plausible normal tissue, and the discriminator's per-pixel abnormality scores are weighted by a Laplacian edge map and used as an adversarial training loss. On a 36-case test set drawn from BraTS 2020, the combined system reports higher tumor-core and lesion-wise Dice than four published baselines (tumor-core Dice 88.84 percent versus 82.07 for the strongest baseline). The wider claim is that abundant unlabeled normal-brain scans can substitute for a large share of pixel-level tumor annotations.","feed_headline":"Brain-tumor core Dice reaches 88.84 with GAN-guided U-Net","feed_subtitle":"A normal-brain GAN supplies boundary feedback that lifts tumor-core Dice past four published baselines.","key_machinery":"The central object is the pre-trained normal-brain GAN's discriminator, repurposed as a per-pixel abnormality scorer. A U-Net mask is expanded by a 5×5 max-pooling and passed through a 3×3 Laplacian filter to form an edge map; that map is dot-multiplied with the discriminator's output to build the edge-guided adversarial loss. Training proceeds in phases—segmentation loss alone for the first ten epochs, then sparsity, size-consistency, and adversarial losses with weights set inversely to gradient magnitudes—which stabilizes the joint U-Net/GAN optimization.","core_discovery":"GANet-Seg claims that the discriminator of a GAN pretrained only on normal brain tissue can act as a pixel-level tumor boundary supervisor, not just a whole-image anomaly detector. During training, the U-Net's mask occludes the suspected lesion, the generator reconstructs the occluded area as normal anatomy, and the discriminator's response, multiplied by the mask's downsampled edge map, produces an adversarial loss that concentrates on border disagreement. The reported consequence is segmentation that exceeds the cited baselines on the tumor-core region (WT 81.28, ET 79.81, TC 88.84 Dice) and on lesion-wise Dice (TC 86.28 versus 78.05), while the authors note the discriminator does not itself distinguish tumor sub-regions.","pith_inferences":["The same discriminator-feedback mechanism could transfer to other lesion types—multiple sclerosis lesions, stroke, or lung nodules—where normal anatomy is abundant and annotated lesions are scarce; the paper only demonstrates it for brain tumors.","A direct testable extension is to condition the generator or discriminator on sub-region labels so the adversarial signal can refine boundaries of edema and enhancing tumor separately, which the paper leaves as future work.","The reported TC Dice gap (88.84 versus 82.07) is large enough that re-running the baselines on the same 36-case protocol would be a strong check on whether the advantage survives a like-for-like comparison.","The edge-guided adversarial loss behaves like a boundary-aware regularizer, so combining it with contour-refinement post-processing is a plausible cheap add-on; the paper does not test this."],"forward_implications":["If the central claim holds, tumor-core segmentation can be improved without any new labeled sub-region data, since the adversarial signal comes from normal-brain images.","The framework's reliance on unlabeled normal MRIs implies that the labeled-data requirement for a usable segmentation model can be lowered, which matters where expert annotations are scarce.","Because the adversarial loss is concentrated on mask edges, the approach predicts that boundary refinement, rather than global appearance matching, is what drives the reported Dice gains.","The reported lesion-wise Dice improvements (e.g., TC 86.28 versus 78.05) imply better handling of small or fragmented tumor components, since lesion-wise metrics weight every lesion equally regardless of size.","The authors' stated 2D-slice limitation implies the current gains are captured from slice-level cues; extending to full 3D volumes is the paper's own next step and would test whether volumetric context adds further accuracy."],"supporting_citations":[{"why":"Establishes the pre-trained GAN approach to tumor detection with high sensitivity that GANet-Seg uses as its starting point.","marker":"[1]"},{"why":"Provides the U-Net encoder-decoder with skip connections that serves as the segmentation backbone.","marker":"[10]"},{"why":"Defines the multi-modal MRI benchmark and the tumor sub-region labels used for training and evaluation.","marker":"[3]"},{"why":"Supplies the unlabeled skull-stripped healthy-brain T1 scans used to pre-train the GAN on normal anatomy.","marker":"[18]"},{"why":"Provides the training cases, expert sub-region labels, and preprocessing conventions the evaluation relies on.","marker":"[19]"},{"why":"Supplies the contrast and brightness enhancement pipeline applied before segmentation to sharpen tumor boundaries.","marker":"[25]"},{"why":"Source of all baseline numbers (Optimized U-Net, nnU-Net, Swin-UNETR, UMamba) that GANet-Seg compares against.","marker":"[26]"}],"fun_headline_variants":["GAN pretrained on healthy tissue sharpens tumor-core Dice","Healthy-brain GAN supervises tumor boundaries, reduces annotation need","Adversarial loss from normal-brain GAN refines tumor mask edges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the published baseline scores in [26] were obtained under the same test cases, preprocessing, cropping, and metric protocol as the authors' self-selected 36-case BraTS 2020 test set, since none of the baselines was re-run here.","fun_headline_variants_meta":{"raw":{"variants":["GAN pretrained on healthy tissue sharpens tumor-core Dice","Healthy-brain GAN supervises tumor boundaries, reduces annotation need","Adversarial loss from normal-brain GAN refines tumor mask edges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000771,"raw_usage":{"total_tokens":3345,"prompt_tokens":810,"completion_tokens":2535,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":426,"completion_tokens_details":{"reasoning_tokens":2478}},"tokens_in":426,"tokens_out":2535,"duration_ms":25153,"temperature":1.0,"reasoning_tokens":2478,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:28:50.604760+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Optimized U-Net, nnU-Net, Swin-UNETR, and UMamba on the exact 36-case split with the paper's preprocessing and then compare TC Dice and HD95 with GANet-Seg; if any baseline equals or exceeds the reported 88.84 TC Dice under identical conditions, the central comparative claim fails. A cheaper check is recomputing both volume-weighted and lesion-wise Dice on the same predictions, since the paper reports both in Table 2 and any large inconsistency would indicate metric or protocol drift.","supporting_citations":[{"cited_title":"Deep neural network with generative adversarial networks pre-training for brain tumor classification based on MR images","cited_arxiv_id":null,"evidence_quote":"Establishes the pre-trained GAN approach to tumor detection with high sensitivity that GANet-Seg uses as its starting point."},{"cited_title":"The preprocessed connectomes project repository of manually corrected skull-stripped T1- weighted anatomical MRI data","cited_arxiv_id":null,"evidence_quote":"Supplies the unlabeled skull-stripped healthy-brain T1 scans used to pre-train the GAN on normal anatomy."},{"cited_title":"A Symmetrical Approach to Brain Tumor Segmen- tation in MRI Using Deep Learning and Threefold Attention Mechanism","cited_arxiv_id":null,"evidence_quote":"Supplies the contrast and brightness enhancement pipeline applied before segmentation to sharpen tumor boundaries."}],"review_version":1}