{"id":"7d9a65b2-ee52-42fa-8d24-d5d91cc8e321","arxiv_id":"2501.10128","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FECT fuses cell, tissue, and edge features, achieving 65.8% weighted F1 on the seven-class BRACS breast cancer classification benchmark.","lead":"This paper describes FECT, a system that classifies breast cancer tissue images by combining features from cells, tissue, and the edges between them, and reports the best score so far on a standard seven-class benchmark. A generalist might read it to see how fusing pathologist-inspired cues can nudge automated cancer classification accuracy upward, though the gain over prior methods is modest and depends on extra annotation data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on edge features that require the authors' own BRACS segmentation masks, and no experiment separates mask availability from the fusion architecture.","rationale":"The reader correctly identified the load-bearing weakness: the edge branch depends on precise segmentation masks from the authors' prior work, and no experiment isolates the contribution of those masks from the contribution of the edge-feature architecture. My reading of Sections 3.4, 3.6, and Table 3 confirms that the ablation only compares feature sets, not mask sources. The paper's stated data availability does not include the masks, and no code is provided, so the method cannot be independently reproduced or re-run with alternative masks. This is a correctness and fairness risk, not merely a reproducibility nicety: the improved weighted F1 over ScoreNet is only 1.4 percentage points with overlapping standard deviations, so the extra supervision could plausibly account for the entire reported advantage. A conditional verdict is therefore appropriate: the central claim is plausible but not established until the mask-dependence is tested. I agree with the reader's assessment and recommend no change to the verdict.","tokens_in":16495,"tokens_out":3819,"duration_ms":43183,"concrete_test":"Re-run the full FECT pipeline, keeping cell and tissue branches and the SVM unchanged, but compute edge features as in Section 3.4 under three mask conditions: (a) the original [19] masks; (b) masks obtained from a foreground/nucleus segmentation model trained only on the BRACS training split without the manual masks; (c) a control mask with the same statistics but no tissue correspondence. If the Cell+Tissue+Edge weighted F1 under (b) stays within about 0.5 pt of the 65.75 in Table 3, the masks are not the source of the gain. If under (b) falls below the 64.4 ScoreNet result, the SOTA claim is not established without releasing the [19] masks and the extra supervision they provide.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 4.2, Table 1) is that FECT achieves a new state-of-the-art 65.8% weighted F1 on BRACS. Section 3.4 builds the edge branch on 'precise segmentation labels on breast cancer tissue images' from the authors' own [19], created by a semi-automated pipeline and not released with this paper. None of the compared methods use these masks, so the comparison is not apples-to-apples. The Table 3 ablation removes edge features as a whole, but it does not test whether the +3.3 pt gain from adding edge features to Cell+Tissue comes from the proposed edge/graph aggregation or simply from the extra semi-automated mask supervision entering through the edge patches and KNN graph. If the masks themselves provide the discriminative information, the 1.4 pt margin over ScoreNet and 4.3 pt margin over HACT-Net could vanish, and the claimed architectural advantage would not be established. Moreover, as specified the method is not reproducible without access to the unpublished [19] masks.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FECT proposes a multimodal architecture for seven-class breast tissue classification on BRACS ROIs. It extracts cell features via HoverNet segmentation and an attention-based aggregator; tissue features via ResMTUNet; and edge features via uniform sampling along contours of tissue masks, Nyström-attention aggregation of edge patches, and a KNN-based graph over cells at edges. The three feature vectors are weighted-concatenated and classified with an SVM. On BRACS, the paper reports a weighted F1 of 65.8% (Acc 66.37%), claiming a new state-of-the-art, and ablation experiments showing all three feature types contribute.","tokens_in":16758,"tokens_out":6958,"duration_ms":63278,"significance":"The interpretability motivation, matching pathologists' use of edge and stromal cues, is attractive, and a systematic ablation across feature subsets (Table 3) plus per-class error bars (Table 1) are useful assets. If the result holds, a 65.8% weighted F1 on the BRACS seven-class task would be a meaningful benchmark. However, the central SOTA claim is confounded by the use of extra mask supervision unavailable to baselines, and several methodological details are missing, so the current form does not yet establish the architectural advantage.","major_comments":[{"comment":"The edge branch is built on the 'seven-category segmentation masks' from the authors' own prior work [19], which are semi-automatically generated on BRACS. None of the baselines in Table 1 have access to these masks. Table 3 ablates edge features as a whole but does not separate the value of the mask-supervision from the value of the proposed edge-patch/KNN-graph aggregation. Since the Cell+Tissue baseline is 62.48% F1 and adding edge features gives 65.75%, the +3.27 pt could in principle be obtained by feeding the masks (or simple mask-derived features) to a baseline. Without a control that injects the same mask information into a baseline model, or a FECT variant that uses the masks without the edge graph/aggregator, the claim that the proposed architecture is superior is not supported.","section":"Section 3.4 and Section 3.6, Table 3"},{"comment":"The manuscript does not state whether the baseline numbers were obtained by re-running the methods under the same experimental protocol (same BRACS train/test split, same 512×512 resizing, same preprocessing and augmentation) or are copied from the original papers. If they are copied, differences in evaluation setup could change the margins; if they are re-run, details of the reimplementation and hyperparameters should be given. The authors should either re-run all baselines under identical conditions and report them, or explicitly cite the source for each column and justify comparability.","section":"Section 4.2, Table 1"},{"comment":"The fusion weights α, β, γ are selected by grid search, but the paper never states whether the search was performed on a held-out validation split or on the test set, and it does not report the selected values or the edge weight γ. If the grid search was done on the test set, the reported 65.75% F1 is optimistically biased. The exact validation protocol and the chosen weights must be disclosed.","section":"Section 3.5 and Section 4.3"},{"comment":"'Overall classification accuracy (Acc)' is defined as 'the unweighted average of recall for each category', i.e., balanced accuracy, not the standard fraction of correctly classified samples. This mismatch is carried through Tables 2–4 and makes the 'Acc' column non-comparable to accuracy values in prior work. The authors should either use standard overall accuracy or rename the metric, and should state which version is used in each table.","section":"Section 4.1, Tables 2–4"},{"comment":"The edge extraction procedure is not reproducible as described. The number and density of contour sampling points, the parameter K in the KNN graph, and the way the edge graph is pooled into a feature vector (no GNN layer or readout is specified) are all missing. Moreover, the segmentation masks from [19] are not included in the Data Availability statement, so a third party cannot build the edge branch without re-deriving the masks. At minimum, the missing hyperparameters should be provided and the masks (or a detailed protocol for regenerating them) should be released.","section":"Section 3.4"}],"minor_comments":[{"comment":"The abstract contains the grammatical error 'we proposes'; it should be 'we propose'.","section":"Abstract"},{"comment":"The phrase 'seven-category segmentation masks ... as proposed in [19]' is opaque; the cited paper's title concerns invasive carcinoma segmentation, so the reader cannot tell whether [19] actually provides multi-class masks or just the segmentation method. Clarify the provenance and contents of the masks.","section":"Section 3.6"},{"comment":"The statement that FECT shows 'exceptional performance' on ADH should be tempered because Table 1 shows ScoreNet's ADH F1 (46.7) is higher than FECT's (45.2).","section":"Section 4.2"},{"comment":"When describing the optimizer, consider specifying whether weight decay is used and whether any warm-up schedule accompanies the step-decay learning rate.","section":"Section 4.1"},{"comment":"The legend says the best results are bold and second-best underlined, but in the extracted text the underlining is not visible; please ensure the table rendering is unambiguous.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The authors should be asked to provide the BRACS segmentation masks or a concrete protocol for regenerating them, and to clarify whether the Table 1 baselines were re-run under identical conditions. The weight-grid-search validation procedure is a critical detail that must be disclosed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: FECT is a legitimate benchmark paper with a modest gain (65.8 vs 64.4 weighted F1) on BRACS, and the multi-feature design is clinically motivated. The soft spot is real: the edge branch depends on the authors' own unreleased segmentation masks, and no experiment separates mask supervision from the fusion architecture.\n\nWhat's new: the specific combination of cell/tissue/edge features with a KNN adjacency graph and weighted SVM fusion; the focus on edge features to separate IC from DCIS is a sensible, interpretable prior. The paper does several things right: results are averaged with std across runs, Table 1 includes many baselines from different families, ablations show each feature and pairwise combinations, there is a classifier comparison, and t-SNE/case visualizations. That is more evidence than most papers at this level.\n\nSoft spots: the main 'state-of-the-art' claim is not fully established. Section 3.4 and 3.6 say the edge features require precise seven-category segmentation masks from [19], produced by a semi-automated pipeline and not released. The baselines in Table 1 do not use these masks, so FECT gets extra supervision. The ablation in Table 3 removes the whole edge branch, but does not test whether the gain comes from the masks rather than from the edge sampling, aggregation, or graph design. So the 1.4 pt margin over ScoreNet and 4.3 pt over HACT-Net could partly be a supervision advantage. Also, hyperparameters (alpha, beta, gamma, K, edge patch size, sampling density) are not given, and there is no code. Table 2 and Table 3 report slightly different numbers for the same FECT (65.75 vs 65.8), which is minor but should be reconcilable. The 'significantly outperforms' language in the abstract is stronger than the evidence supports.\n\nNone of this kills the paper. The architecture is reasonable, the evaluation is on a public dataset, and the clinical motivation is sound. It deserves a serious referee, but the authors should be asked to release the masks or run baselines with the same masks, and to provide a control where edge features are derived from a generic segmentation model (e.g., HoverNet masks alone) rather than the seven-category masks.\n\nWho it's for: people working on BRACS or feature fusion in histopathology. I'd read it if I were in that area, but I wouldn't cite it until the mask question is resolved. Send to review, with a request for the mask-ablation and code.","headline":"A plausible, modest benchmark gain on BRACS, but the edge-feature branch relies on the authors' unreleased segmentation masks and the SOTA claim is not yet isolated from that extra supervision.","tokens_in":17252,"tokens_out":2269,"would_cite":false,"duration_ms":20780,"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":"FECT, a model that fuses cell, tissue, and edge features, reports a new best weighted F1 of 65.8% on the seven-class BRACS breast tissue classification benchmark.","keywords":["Breast cancer classification","Feature fusion","Transformer","Attention-based aggregator","Multimodal","BRACS dataset","Pathological image classification","Edge features"],"falsifier":"Replace the semi-automated seven-category masks with automatically predicted masks and rerun the edge branch; if the weighted F1 advantage over ScoreNet disappears, the claimed benefit of the fusion architecture is actually the benefit of the extra mask supervision.","tokens_in":16339,"feed_emoji":"🔬","tokens_out":11673,"duration_ms":102674,"temperature":0.7,"pith_summary":"Breast tissue diagnosis requires looking at different scales: individual cells, overall tissue architecture, and the boundary where a lesion meets surrounding tissue. This paper argues that existing classifiers rely on only one of these scales, and that a model which extracts and fuses all three can do better. The authors build FECT, which segments cells with HoverNet, encodes tissue structure with ResMTUNet, samples patches along tissue edges and builds a cell-adjacency graph at those edges, then weights and concatenates the three feature sets into an SVM classifier. On the seven-category BRACS benchmark, FECT reports a weighted F1 score of 65.8%—an average of per-category F1 scores weighted by class size—and an accuracy of 66.37%, ahead of the compared methods, with the clearest gains in normal tissue, ductal carcinoma in situ, and invasive carcinoma. If the result holds, the takeaway is that boundary-level morphology—the feature pathologists use to separate invasive from in situ disease—deserves a place in automated classifiers, not just cell and tissue features.","feed_headline":"Fusion of cell, tissue, and edge features tops BRACS benchmark","feed_subtitle":"Three-level feature fusion reaches 65.8% weighted F1 on the seven-class BRACS task, besting prior published methods.","key_machinery":"The load-bearing mechanism is the three-branch feature extraction and fusion pipeline. Each branch is designed to capture one scale a pathologist attends to: cells (nucleus segmentation plus attention-based aggregation), tissue (ResMTUNet global architecture), and edges (contour-sampled patches encoded by a vision transformer, aggregated with Nyström attention, and linked by a KNN adjacency graph of cells near the boundary). The fused representation is a weighted concatenation $X_f = [\\alpha X_C, \\beta X_T, \\gamma X_E]$ fed to an SVM. The ablation design is what carries the argument: comparing single branches against their combinations shows that the full fusion outperforms every subset, and that the edge branch contributes most when the other two are present.","core_discovery":"The paper's central claim is that FECT—Fused features of Edges, Cells, and Tissues—sets a new best result on the BRACS Tissue Regions of Interest task by combining three complementary views of each image. The cell branch uses HoverNet to segment nuclei and an attention aggregator to combine per-cell embeddings. The tissue branch uses ResMTUNet to encode overall tissue architecture. The edge branch samples 64×64 patches along the boundaries of the tissue mask, encodes them with a vision transformer pretrained on ImageNet, aggregates them with Nyström attention, and constructs a KNN graph of cells at those edges to capture the tumor microenvironment. The three feature vectors are weighted, concatenated, and classified by an SVM. On the BRACS test set, FECT achieves a weighted F1 of 65.8% and accuracy of 66.37%, surpassing the compared methods, including HACT-Net (61.5% weighted F1) and ScoreNet (64.4% weighted F1), and the ablation study shows that edge features, although weak on their own, improve classification when added to cell and tissue features, especially for ductal carcinoma in situ and invasive carcinoma.","pith_inferences":["A direct test of the edge branch's portability is to replace the semi-automated seven-category masks with automatically predicted or coarser masks and measure how much of the FECT gain remains.","The margin over the closest compared method is about 1.4 weighted-F1 points, so a multi-seed replication with confidence intervals would be a natural next step before drawing strong clinical conclusions.","Because the invasive-versus-in-situ distinction is the clinical question the edge branch targets, the same edge-graph idea could transfer to other tumor types where invasion status matters.","A pathologist study could check whether the edge regions FECT attends to correspond to the areas a pathologist would confirm with myoepithelial markers such as P63."],"forward_implications":["If FECT's reported numbers replicate, the 65.8% weighted F1 is the new reference point for seven-class BRACS tissue classification, ahead of the 64.4% of ScoreNet and 61.5% of HACT-Net.","The ablations show the fused model beats every single-feature and every two-feature combination, so the paper's design argument is that the three scales are genuinely complementary.","Edge features, while the weakest branch on their own, are what push ductal carcinoma in situ and invasive carcinoma apart in the fused model; any future method targeting that pair should include boundary cues.","The SVM-with-weighted-concatenation configuration outperforms the other classifiers tested on the same fused features, making the fusion-and-classifier combination part of the claimed result."],"supporting_citations":[{"why":"Supplies both the tissue-feature backbone (ResMTUNet) and the seven-category epithelial masks that define where edge patches are sampled.","marker":"[19]"},{"why":"Is the dataset and benchmark on which FECT is trained and evaluated; the paper's claims are about BRACS.","marker":"[20]"},{"why":"Provides the nucleus segmentation used to locate cells for the cell-feature branch and for the edge-cell graph.","marker":"[45]"},{"why":"Supplies the attention mechanism used to aggregate single-cell features into an image-level cell representation.","marker":"[18]"},{"why":"Provides the pretrained weights for the vision transformer that encodes edge patches.","marker":"[47]"},{"why":"Supplies the Nyström attention aggregator used to combine high-dimensional edge patch features.","marker":"[48]"},{"why":"Is the strongest graph-based baseline (HACT-Net) that FECT compares against and surpasses on the BRACS test set.","marker":"[11]"},{"why":"Is the transformer-based baseline (ScoreNet) whose 64.4% weighted F1 is the closest published result FECT claims to beat.","marker":"[15]"},{"why":"Is the feature-fusion baseline (two-branch TGNet) used for the fusion-strategy comparison in Table 2.","marker":"[52]"}],"fun_headline_variants":["Edge, cell, tissue fusion sets new BRACS record","Triple-feature fusion beats prior best on breast cancer images","Fusing edges, cells, and tissues tops BRACS classification","New breast cancer classifier fuses three views for top F1","FECT fusion model achieves 65.8% F1 on BRACS benchmark"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The edge-feature branch is built on precise seven-category segmentation masks that were generated semi-automatically in prior work, and the paper never tests whether the performance gain comes from having those masks rather than from the fusion design itself.","fun_headline_variants_meta":{"raw":{"variants":["Edge, cell, tissue fusion sets new BRACS record","Triple-feature fusion beats prior best on breast cancer images","Fusing edges, cells, and tissues tops BRACS classification","New breast cancer classifier fuses three views for top F1","FECT fusion model achieves 65.8% F1 on BRACS benchmark"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1540,"prompt_tokens":954,"completion_tokens":586,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":497}},"tokens_in":570,"tokens_out":586,"duration_ms":5178,"temperature":1.0,"reasoning_tokens":497,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:22:50.771395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the semi-automated seven-category masks with automatically predicted masks and rerun the edge branch; if the weighted F1 advantage over ScoreNet disappears, the claimed benefit of the fusion architecture is actually the benefit of the extra mask supervision.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies both the tissue-feature backbone (ResMTUNet) and the seven-category epithelial masks that define where edge patches are sampled."},{"cited_title":"Brancati, A","cited_arxiv_id":null,"evidence_quote":"Is the dataset and benchmark on which FECT is trained and evaluated; the paper's claims are about BRACS."},{"cited_title":"Graham, Q","cited_arxiv_id":null,"evidence_quote":"Provides the nucleus segmentation used to locate cells for the cell-feature branch and for the edge-cell graph."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pretrained weights for the vision transformer that encodes edge patches."},{"cited_title":"14138–14148","cited_arxiv_id":null,"evidence_quote":"Supplies the Nyström attention aggregator used to combine high-dimensional edge patch features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the strongest graph-based baseline (HACT-Net) that FECT compares against and surpasses on the BRACS test set."},{"cited_title":"Stegmüller, B","cited_arxiv_id":null,"evidence_quote":"Is the transformer-based baseline (ScoreNet) whose 64.4% weighted F1 is the closest published result FECT claims to beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the feature-fusion baseline (two-branch TGNet) used for the fusion-strategy comparison in Table 2."}],"review_version":1}