{"id":"25494839-0d77-4d7f-b393-6514710c7087","arxiv_id":"2501.03891","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A superpixel flood-fill correction on multi-depth CAM pseudo-masks gives 71.08% mIoU on the BCSS breast cancer segmentation benchmark, a small gain over PistoSeg's 70.55%.","lead":"This paper refines weakly supervised tissue segmentation in breast cancer slides by filling each superpixel with its dominant class label after combining activation maps from several network depths. It reports 71.08% mIoU on BCSS, slightly above the prior 70.55%, which could make pixel-level labelling less necessary in computational pathology.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed boundary improvement is contradicted by the paper's own per-class results: LYM IoU drops from 0.6098 (WSSS-Tissue) to 0.6048 after superpixel floodfill, indicating majority-class smoothing rather than true boundary correction.","rationale":"The reader identified the SLIC superpixel alignment assumption as the weakest point. The paper's own Table 1 provides direct evidence that this assumption fails for the LYM class, the most heterogeneous and fragmented tissue type in BCSS. The drop in LYM IoU relative to WSSS-Tissue is small but in the wrong direction for a method advertised as improving boundary delineation. The proposed floodfill may nonetheless raise aggregate mIoU by smoothing larger classes, but this does not validate the mechanism as a boundary corrector. A controlled ablation and boundary-aware evaluation would settle the issue. Since the reader's conditional verdict already requires exactly this evidence, my concern does not change the recommended decision; it sharpens the condition and points to a concrete test.","tokens_in":5365,"tokens_out":10267,"duration_ms":103370,"concrete_test":"On the BCSS validation/test split, reproduce the base WSSS-Tissue model and then apply Algorithm 1 over a grid of SLIC parameters (S in {20,40,80}, m in {10,30,50}) and tau in {0.5,0.7,0.9}. For each configuration, report per-class mIoU and boundary F-score (or recall of LYM objects) before and after floodfill. If LYM IoU decreases systematically while TUM/STR/NEC increase, the refinement is confirming majority smoothing rather than boundary correction; if the best configuration fails to reproduce the reported 0.7108 mIoU, the claimed SOTA is not robust to the undisclosed parameters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that superpixel floodfill 'significantly improves tumor microenvironment boundary delineation' is undermined by the paper's own Table 1. Relative to WSSS-Tissue [11], the method incorporating the floodfill module improves TUM, STR, and NEC IoU but decreases LYM IoU from 0.6098 to 0.6048 (Sec. 4.3). This is exactly the failure mode predicted by the weakest-assumption analysis: lymphocytic infiltrate is the most fragmented class, so majority-vote floodfill (Algorithm 1) erases minority pixels within mixed-tissue superpixels. The paper reports no ablation isolating Algorithm 1, no SLIC parameters S, m, or threshold tau (Sec. 3.2), and no error bars. The reported +1.38 mIoU over WSSS-Tissue is therefore not shown to arise from genuine boundary correction; it could be a majority-class smoothing bias. For a method whose stated purpose is boundary delineation, regressing on the smallest tissue class is directly contrary to that claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a weakly supervised semantic segmentation (WSSS) method for histopathology images that combines multi-layer CAM pseudo-mask fusion with a superpixel-based floodfill refinement. Using only image-level labels, the method trains a classification network (ResNet38d) and a segmentation network, then refines the final segmentation by assigning each SLIC superpixel the dominant predicted class when the majority fraction exceeds a threshold tau. On the BCSS breast cancer dataset, the method reports an mIoU of 71.08%, which is higher than the previously reported PistoSeg (70.55%). The paper claims state-of-the-art results and improved tumor microenvironment boundary delineation.","tokens_in":1294,"tokens_out":2824,"duration_ms":58693,"significance":"If the result is robust, the contribution is a simple post-processing addition to existing CAM-based WSSS pipelines for histopathology, with potential practical value because it requires no additional annotation. The authors provide a clear algorithm description, per-class IoU numbers, and qualitative visualizations. However, the manuscript currently does not substantiate that the improvement stems from genuine boundary correction: the per-class LYM IoU decreases relative to WSSS-Tissue, the key hyperparameters are unreported, no ablation isolates the floodfill module, and the comparison to baselines is not shown to be statistically grounded. The central claim is therefore plausible but not yet demonstrated.","major_comments":[{"comment":"The per-class results contradict the claim that the superpixel floodfill improves boundary delineation: compared with WSSS-Tissue [11], LYM IoU drops from 0.6098 to 0.6048, while the fragmented nature of lymphocytic infiltrate makes it exactly the class most vulnerable to majority-vote floodfill (Algorithm 1). Since no ablation isolates Algorithm 1 from the multi-layer fusion and training changes, the +1.38 mIoU over WSSS-Tissue cannot be attributed to boundary correction. Please add an ablation that toggles the refinement module on the same backbone and reports per-class changes.","section":"Section 4.3, Table 1"},{"comment":"The method's key hyperparameters—lambda_1, lambda_2, lambda_3, SLIC average cluster size S, compactness m, and threshold tau—are never reported, and no code is provided. Without these values or a sensitivity analysis, the experiments are not reproducible and the reported 71.08% mIoU cannot be independently verified or compared against variants.","section":"Section 3.2, Algorithm 1, and Section 4.2"},{"comment":"The paper does not state whether the baseline numbers (HistoSegNet, SEAM, C-CAM, WSSS-Tissue, PistoSeg) were re-run under the same protocol or quoted from the original publications, and it provides no error bars or significance tests. Given that the improvement over PistoSeg is only 0.53 mIoU points, the state-of-the-art claim requires either re-run baselines with matched evaluation or a statistical analysis over multiple seeds or splits.","section":"Section 4.3, Table 1"},{"comment":"The dataset description incorrectly states that pathologists 'classify the images without providing pixel-level mask information,' even though the evaluation uses pixel-level ground truth on the BCSS test set. This inconsistency should be corrected, and details of the patch generation and train/validation/test split should be clarified.","section":"Section 4.1"}],"minor_comments":[{"comment":"The description of SEAM [9] does not match the cited reference; reference [9] is 'Weakly supervised semantic segmentation by pixel-to-prototype contrast' (CVPR 2022), not the SEAM paper. Please correct the citation.","section":"Section 2.2"},{"comment":"The phrase 'With siperpixels' should read 'With superpixels', and 'multi-layer segmenation backbone' has a typo.","section":"Section 1"},{"comment":"The pseudocode uses the misspelled variable names 'classprefined' and 'prefined'; these should be 'classrefined' and 'refined'.","section":"Section 3.2, Algorithm 1"},{"comment":"The text says '10 epoches' and 'training rate'; these should be '10 epochs' and 'learning rate'.","section":"Section 4.2"},{"comment":"The phrase 'great performance' is vague; consider reporting concrete numbers or stating the relative improvement explicitly.","section":"Abstract and Section 4.3"},{"comment":"The figure caption does not explain how the 'SP-boundary' panel is produced; please clarify what is shown in that panel.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope, and the evaluation uses a held-out test set with pixel-level ground truth, so there is no circularity concern. The main issue is that the central boundary-correction claim is not supported by the reported per-class numbers or by an ablation; this is fixable within the manuscript's scope by adding a controlled ablation, reporting hyperparameters, and clarifying the baseline comparison protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a simple, honest paper: take WSSS-Tissue's multi-layer CAM pseudo-masks, add a SLIC superpixel floodfill step, and report a new best mIoU on BCSS. I believe the numbers are what they say, but the boundary-correction story does not survive their own per-class table.\n\nWhat's new is the specific combination and the benchmark result. The components are known—WSSS-Tissue supplies the multi-layer pseudo-supervision, and superpixel refinement of CAMs exists in the wider WSSS literature. Applying this to breast cancer histopathology and showing a small improvement is a legitimate extension, not a conceptual advance. To the paper's credit, the writing is clear, the setup is standard, the dataset is public, and the total mIoU is internally consistent with the per-class numbers.\n\nThe big soft spot is the LYM class. WSSS-Tissue got 0.6098 LYM IoU; after the floodfill it drops to 0.6048. That is exactly the signature of majority-vote smoothing: lymphocytic infiltrate is the most fragmented class, so within mixed-tissue superpixels the minority pixels get overwritten. The paper never isolates Algorithm 1 in an ablation, so we do not know whether the +1.38 mIoU over WSSS-Tissue comes from genuine boundary alignment or from a smoothing bias that helps the larger classes. The SLIC parameters (S, m, tau) are not reported. There are no error bars or significance tests, and the gain over PistoSeg is 0.53 points, which could easily be noise. The baselines are quoted from their papers rather than re-run.\n\nNone of this is fatal. The method does improve overall mIoU, and the per-class gains on TUM, STR, and NEC are consistent. The LYM drop is small. But it undercuts the 'significantly improving boundary delineation' claim, at least for the class where boundary precision matters most.\n\nThe paper is for researchers working on CAM post-processing or weakly supervised segmentation in pathology. It is a useful data point on a public benchmark, and the failure mode (erosion of small objects) is worth knowing about. I would not build a project on the SOTA claim, but I would cite it as related work.\n\nA serious editor should send this to peer review. The result is specific and checkable, and the issues are addressable with a proper ablation, hyperparameter disclosure, and significance testing. With those, the claim might hold; without them, the contribution is a modest empirical trick.","headline":"Modest, plausible extension of WSSS-Tissue with superpixel floodfill; the mIoU gain is real but the boundary-correction claim is undercut by a regression on the LYM class.","tokens_in":6123,"tokens_out":3022,"would_cite":false,"duration_ms":25518,"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":"Superpixel floodfill refinement raises weakly supervised breast tissue segmentation to 71.08% mIoU.","keywords":["weakly supervised semantic segmentation","histopathology","breast cancer","BCSS dataset","class activation map","superpixel","floodfill","tumor microenvironment"],"falsifier":"Remove Algorithm 1 from the pipeline and evaluate on the same BCSS test split; the central claim collapses if the mIoU drop is not reproducible or if sweeping S, m, and tau produces results below the reported 71.08%.","tokens_in":5145,"feed_emoji":"🔬","tokens_out":6223,"duration_ms":57447,"temperature":0.7,"pith_summary":"This paper tries to show that a cheap post-processing step can solve the boundary problem of weakly supervised semantic segmentation on histopathology images. The authors take CAM pseudo-masks from a classifier trained only on image-level labels, merge CAMs from three network depths, and then cluster the raw image into SLIC superpixels and flood-fill each superpixel with its dominant predicted class. On the BCSS breast cancer dataset this produces a mean IoU of 71.08%, above the previous best of 70.55%, with the top per-class score on necrosis and no pretrained model or data augmentation. The authors' conclusion is that superpixel boundaries, because they follow natural color and texture edges in tissue, are a reliable correction signal for CAM masks in pathology.","feed_headline":"Superpixels lift weakly supervised tissue segmentation to 71.08% mIoU","feed_subtitle":"Using only image-level labels, superpixel floodfill edges past the prior best on breast cancer slides.","key_machinery":"The central object is Algorithm 1, the superpixel floodfill refinement. SLIC (Simple Linear Iterative Clustering) partitions the H&E patch into superpixels using spatial distance and color intensity; then for each superpixel, the algorithm counts pseudo-mask labels within it and, if the dominant class covers more than a threshold tau of the superpixel's pixels, reassigns the whole superpixel to that class. The mask being corrected comes from multi-layer pseudo-mask fusion, where CAMs from three network depths are combined through weighted cross-entropy losses. The floodfill's work is to suppress unnatural outlier blocks in CAM-derived masks and push boundaries onto the color and texture edges that superpixels follow.","core_discovery":"The central claim, stated on the paper's own terms, is that the residual error in weakly supervised tissue segmentation is largely a boundary-scale problem, and that a multi-level superpixel correction is enough to fix it. By extracting CAMs at three different CNN depths and fusing them in the loss function, the method captures tissue boundaries at different granularities; then Algorithm 1 imposes the superpixel partition on the fused mask, wiping out outlier blocks and aligning segmentation edges with the visual tissue distribution. Experimental results report 71.08% mIoU on BCSS, which the paper presents as state of the art for weak supervision on that benchmark, ranking first on necrosis and second on tumor and stroma. The authors also report that refining the segmentation-stage output improves performance, while refining the earlier CAM-stage pseudo-mask does not, which they interpret as evidence that premature CAM modification discards useful image-level information.","pith_inferences":["Beyond the paper, the dependence on SLIC parameters S, m, and tau means transfer to other organs (lung, colon, prostate) is not guaranteed until those parameters are reported and re-tuned for different tissue boundary statistics.","The threshold tau acts as an eraser for thin structures: with a high tau, isolated lymphocytes or narrow stromal strands could be swallowed by a neighboring dominant class, which may explain why LYM IoU (60.48%) trails the other classes.","A testable extension is to make the threshold class-aware or boundary-aware, so that superpixels intersected by known high-contrast edges are exempted from floodfill.","Because the method is pure post-processing, it could be composed with stronger pseudo-mask generators to push the boundary further."],"forward_implications":["Weakly supervised segmentation can reach competitive accuracy on breast cancer tissue without pixel-level annotations, lowering the cost of tumor microenvironment analysis.","Because the reported gain is additive and requires no pretrained model or augmentation, the same floodfill correction can be dropped onto other CAM-based WSSS pipelines.","Refinement should be applied to the segmentation stage rather than to the raw CAM, since the paper finds first-stage refinement has no effect.","The per-class balance (first on necrosis, second on tumor and stroma) implies the method does not improve average IoU by sacrificing difficult classes."],"supporting_citations":[{"why":"Defines the class activation map procedure that generates the initial pseudo-masks the method corrects.","marker":"[5]"},{"why":"HistoSegNet serves as a baseline WSSS method on histopathology, compared in Table 1.","marker":"[8]"},{"why":"SEAM is a baseline CAM-regularization method compared in Table 1.","marker":"[9]"},{"why":"C-CAM is a baseline causal CAM method compared in Table 1.","marker":"[10]"},{"why":"Supplies the multi-layer pseudo-supervision scheme and three-depth CAM fusion that this method extends.","marker":"[11]"},{"why":"PistoSeg is the prior state of the art on BCSS with 70.55% mIoU, the baseline the paper must beat.","marker":"[12]"}],"fun_headline_variants":["Superpixel floodfill sharpens CAM boundaries to 71.08% mIoU on breast slides","Weakly supervised tissue segmentation gets a superpixel edge — 71.08% mIoU","Superpixel correction lifts weakly supervised histopathology segmentation to SOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the small clusters drawn by the SLIC algorithm on H&E patches follow true tissue boundaries, so assigning each cluster the dominant CAM class improves the mask; if many clusters straddle two tissue types, or if the undisclosed parameters S, m, and tau are unsuitable, the floodfill would erase thin structures such as isolated lymphocytes.","fun_headline_variants_meta":{"raw":{"variants":["Superpixel floodfill sharpens CAM boundaries to 71.08% mIoU on breast slides","Weakly supervised tissue segmentation gets a superpixel edge — 71.08% mIoU","Superpixel correction lifts weakly supervised histopathology segmentation to SOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000906,"raw_usage":{"total_tokens":3846,"prompt_tokens":845,"completion_tokens":3001,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":2927}},"tokens_in":461,"tokens_out":3001,"duration_ms":18370,"temperature":1.0,"reasoning_tokens":2927,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:44:35.915725+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Remove Algorithm 1 from the pipeline and evaluate on the same BCSS test split; the central claim collapses if the mIoU drop is not reproducible or if sweeping S, m, and tau produces results below the reported 71.08%.","supporting_citations":[{"cited_title":"Weakly supervised semantic segmentation by pixel-to- prototype contrast","cited_arxiv_id":null,"evidence_quote":"SEAM is a baseline CAM-regularization method compared in Table 1."},{"cited_title":"C-cam: Causal cam for weakly supervised semantic segmentation on medical image","cited_arxiv_id":null,"evidence_quote":"C-CAM is a baseline causal CAM method compared in Table 1."},{"cited_title":"Histoseg- net: Semantic segmentation of histological tissue type in whole slide images","cited_arxiv_id":null,"evidence_quote":"HistoSegNet serves as a baseline WSSS method on histopathology, compared in Table 1."},{"cited_title":"Learning deep features for discriminative localization","cited_arxiv_id":null,"evidence_quote":"Defines the class activation map procedure that generates the initial pseudo-masks the method corrects."},{"cited_title":"Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-layer pseudo-supervision scheme and three-depth CAM fusion that this method extends."},{"cited_title":"Weakly-supervised semantic segmentation for histopathology images based on dataset synthesis and feature consistency constraint","cited_arxiv_id":null,"evidence_quote":"PistoSeg is the prior state of the art on BCSS with 70.55% mIoU, the baseline the paper must beat."}],"review_version":1}