{"id":"8482a11f-5764-433f-b630-ea57eec7e358","arxiv_id":"2502.05320","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"FH-Seg, a U-Net variant with full-scale skip connections and hierarchical soft attention, reports higher Dice and F1 scores than prior methods on a new 16,212-image renal vasculature dataset.","lead":"A new deep learning model, FH-Seg, combines skip connections and attention gates to outline kidney blood vessels and lesions in microscope images more accurately than prior methods. The authors also built a large annotated renal vasculature dataset, though it is not yet publicly released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The LRV single-class non-overlapping labels make artery and its subcomponents mutually exclusive, so the reported class-wise Dice/F1 do not yet establish a well-defined multiclass SOTA.","rationale":"The most load-bearing condition for the central claim is the definition of the segmentation task. FH-Seg cannot be state-of-the-art for fine-grained renal vasculature segmentation if the labels used to train and evaluate it do not jointly respect the anatomy. The paper's own label description (§4) says the 'artery' class includes the lumen, tunica intima, tunica media, and external elastic regions, yet the labeling methodology forces a single, non-overlapping class per pixel and never combines artery with its subcomponents. Because each image in Table 1 is assigned to one category, a pixel that is part of tunica media is labeled 'tunica media' in that category's images, 'artery' in artery images, and implicitly background in lumen or wall images. A six-way softmax with cross-entropy cannot represent both 'tunica media is inside artery' and 'tunica media vs artery are exclusive classes'. If the loss instead uses per-class binary masks, the paper does not say how unlabeled structures are excluded from the loss; either way the reported class-specific Dice/F1 on separate test subsets do not demonstrate a coherent joint segmentation. The abstract's headline numbers are also the tunica-media-only row, so the strongest claim is not supported as a dataset-wide average. I agree with the reader's weakest assumption; a label-consistency audit and hierarchy-aware reevaluation would settle the concern. If the audit shows expert masks overlap while FH-Seg predicts disjoint classes, the central claim is not interpretable. The paper should release the dataset and clarify or mask the loss. This supports the reader's CONDITIONAL verdict; no change.","tokens_in":10001,"tokens_out":9471,"duration_ms":99799,"concrete_test":"On a held-out set of patches for which expert annotators provide both coarse 'artery' masks and fine subcomponent masks (tunica intima, tunica media, artery wall) for the same regions, run the released FH-Seg model and check whether the predicted 'artery' mask contains the predicted 'tunica media' and 'tunica intima' masks. If the model systematically predicts disjoint regions where the expert masks overlap, the single-class label scheme has suppressed the anatomical inclusion relation; then re-run the comparison with a hierarchy-aware loss or masked loss that ignores unannotated structures and compare the reported Dice/F1 gaps to Omni-Seg.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The dataset's labeling methodology (§4, 'Labeling Methodology') defines the task in a way that conflicts with the anatomy described in the same section. 'Artery' is said to encompass the lumen, tunica intima, tunica media, and external elastic regions, yet the paper assigns each pixel a single class and never co-labels artery with its subcomponents. Since Table 1 splits images by category, the same kind of pixel is labeled 'tunica media' in some images, 'artery' in others, and background in lumen-only or wall-only images. A standard six-class softmax/one-hot training treats these as exclusive alternatives, so the network cannot simultaneously learn that tunica media is part of the artery and that the two labels are mutually exclusive. If the loss is instead per-class binary, the paper does not state how unlabeled structures are excluded from the loss. The per-class Dice/F1 in Tables 2–3 are evaluated on category-specific test subsets, so they do not establish a coherent joint six-class segmentation; the abstract's 71.23/73.06 are the tunica-media row, not a global average. Thus the central SOTA claim rests on a task definition that is not demonstrably well-posed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FH-Seg, a Residual U-Net variant that augments full-scale skip connections with learnable hierarchical soft attention gates for fine-grained renal vasculature segmentation. The authors also introduce the LRV dataset, containing 16,212 annotated 2048x2048 patches derived from whole-slide images, with six anatomical/pathological classes: lumen, tunica intima, tunica media, artery, artery wall, and hyaline. Experiments on LRV compare FH-Seg against U-Net, DeepLabV3, Swin UNETR, and Omni-Seg in per-class Dice and F1 scores, and an ablation study examines the contribution of each proposed component. The central claim is that FH-Seg achieves state-of-the-art performance on this dataset, with the abstract citing 71.23% Dice and 73.06% F1.","tokens_in":10278,"tokens_out":4249,"duration_ms":46689,"significance":"If the central claim holds, the paper would provide a useful architecture for fine-grained vascular segmentation and a substantial new dataset for renal pathology. The code release and the size of the annotated dataset are concrete strengths. However, the claimed superiority is currently undercut by two issues: the headline metrics are not aggregate scores but the Tunica media row, and the label schema makes the joint six-class task internally inconsistent. The dataset may still be valuable, and the architectural components are simple enough to be reusable, but the SOTA claim as written is not yet well-defined.","major_comments":[{"comment":"The abstract's headline numbers, '71.23% Dice, 73.06% F1', are the Tunica media row of Table 3 (and the corresponding 'Total' row in Table 4), not an aggregate over all six classes. The statement that FH-Seg outperforms Omni-Seg by 2.67 and 2.13 percentage points is likewise specific to that class. The paper should either report a properly averaged Dice/F1 across the six classes or explicitly and consistently present the per-class nature of these claims; otherwise the central quantitative claim is overstated.","section":"Abstract; §5.1, Tables 3-4"},{"comment":"The single-label-per-pixel annotation strategy conflicts with the anatomical description in the same section. The text states that the 'artery' label encompasses the lumen, tunica intima, tunica media, and external elastic regions, yet the methodology forbids co-labeling artery with its subcomponents. Because each category in Table 1 corresponds to a separate image subset, the same kind of tissue (e.g., tunica media) is labeled as 'artery' in some images, as 'tunica media' in others, and as background in lumen-only or wall-only images. A standard six-class one-hot training cannot learn that tunica media is part of the artery while also treating these labels as mutually exclusive, and the per-class metrics in Tables 2-3, evaluated on category-specific subsets, do not demonstrate a coherent joint six-class segmentation. The authors must specify the exact training loss, state how unlabeled structures are excluded or treated, and report results on a consistent label space.","section":"§4, Labeling Methodology; Tables 2-3"},{"comment":"The architecture is not described consistently enough to reproduce. Equations (3)-(4) define full-scale skip connections that aggregate the same-level encoder feature, the decoder feature from the level above, and the decoder feature from the level below, while Algorithm 1 only concatenates the current decoder feature with the same-level encoder skip feature. Similarly, the text in §3.4 says the HSA is applied after the full-scale skip connections and before each upsampling step, but in Algorithm 1 the attention weights are computed after the entire decoder loop. The dense full-scale connections in Figure 2 also appear to show more connections than the 'immediate preceding and subsequent levels' described in the text. Please align the equations, pseudocode, and figure with the actual implementation.","section":"§3.3-§3.4; Algorithm 1; Figure 2"}],"minor_comments":[{"comment":"The dataset size is reported inconsistently: the abstract and Table 1 state 16,212 images, while §4 states 16,214 images. Please unify these numbers.","section":"Abstract; §4; Table 1"},{"comment":"There is a typo in the introduction: 'Sgementation' should be 'Segmentation'.","section":"§1"},{"comment":"Figure 2 depicts dense full-scale connections, but the text says the implementation uses only immediate neighboring levels. Please clarify which connections are actually used and ensure the figure and text do not contradict each other.","section":"§3.3; Figure 2"},{"comment":"The ablation claims that HSA improves average Dice by 1.1% and FS by 1.3%, but the numbers in Table 4 are per-class and the 'Total' row is not an additive combination of the individual components. Please specify how the average is computed and clarify whether the reported improvements are averaged over all six classes.","section":"§5.3, Table 4"},{"comment":"The F1 metric is never defined. Please state whether it is the per-class F1 score and how it is computed from precision and recall.","section":"§5, Tables 2-3"},{"comment":"Figure 6 is captioned 'Bbox plot' but no explanation of what the box plot shows is provided in the text. Please add a description or remove the figure.","section":"Figure 6"},{"comment":"Section 7, titled 'New or Breakthrough Work to be Presented', is an unusual section for a journal-style paper and mostly repeats the conclusion. It should be removed or integrated into the conclusion.","section":"§7"}],"recommendation":"major_revision","confidential_remarks":"The label-schema issue is the main obstacle: the paper's evaluation strategy does not convincingly support the claimed joint six-class SOTA. The dataset itself is a potentially valuable resource, and the architecture is simple enough to be a reasonable baseline contribution, but the authors need to either reformulate the task as per-class binary segmentation with clear handling of unlabeled structures or provide a consistent multiclass label space. The paper should also avoid presenting a single-class result as the headline metric. The self-citation of Omni-Seg as the main baseline is appropriate given the authors' prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the LRV dataset is the real contribution, and the architecture is an incremental but sensible combination of existing full-scale skip (U-Net3+) and attention gate (Attention U-Net) ideas. The paper is worth engaging, but the headline SOTA claim does not survive close reading.\n\nWhat is genuinely good: 16,212 annotated patches from 5,600 arteries with six fine-grained structures—lumen, intima, media, artery, artery wall, hyaline—is a substantial annotation effort and the first dataset of this granularity for renal vasculature. The ablation in Table 4 shows each component adds roughly 1–1.3% Dice, which is honest evidence the combination helps. The paper compares against relevant baselines (U-Net, DeepLabV3, Swin UNETR, Omni-Seg) on the same data, and the code is released. Citations look appropriate; Omni-Seg being from the same group is fine since it is used as a baseline and ablation base.\n\nSoft spots, in rough order of seriousness. First, the label methodology in Section 4 assigns every pixel one class and never co-labels 'artery' with its subcomponents. That makes the six classes mutually exclusive in training, even though the text says artery includes intima, media, and external elastic regions. A lumen-only image implicitly treats the actual vessel wall as background, while a wall image treats wall as foreground. A standard six-class softmax gets conflicting supervision, and the per-category test sets sidestep this rather than resolving it. The stress-test note lands. Second, the abstract's 71.23 Dice / 73.06 F1 are the tunica media row, not a global average; no aggregated Dice is reported at all. Third, 'full-scale' is a misnomer: Eq. 3 and the surrounding text say only immediate neighboring decoder levels are connected, which is a local skip strategy, not full-scale. Fourth, no error bars or multiple runs are reported, and the dataset is not currently public, so the numbers cannot be reproduced.\n\nNone of this kills the paper, but it needs major revision and dataset release before the numbers can be trusted. For renal pathology and medical image segmentation researchers, the dataset is the takeaway; the architecture is a useful baseline. I'd send it to review rather than desk reject, with a clear request to fix the metric and task definition before acceptance.","headline":"A substantial new dataset and a modest architecture combo; the abstract overstates the result by quoting one class's Dice/F1, and the label scheme makes the joint six-class claim shaky.","tokens_in":10793,"tokens_out":4674,"would_cite":false,"duration_ms":50926,"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":"FH-Seg combines full-scale skip connections and hierarchical soft attention gates to reach state-of-the-art fine-grained renal vasculature segmentation on the new LRV dataset.","keywords":["renal pathology","image segmentation","skip connections","attention gates","fine-grained segmentation","renal vasculature","histopathology","U-Net"],"falsifier":"Compute the class-averaged Dice and F1 across all six categories for FH-Seg and Omni-Seg; if FH-Seg does not lead by roughly 2.67 and 2.13 points on that average, the abstract's headline comparison is not what it appears to be.","tokens_in":9831,"feed_emoji":"🔬","tokens_out":10013,"duration_ms":89873,"temperature":0.7,"pith_summary":"FH-Seg is a segmentation framework that tries to solve fine-grained renal vasculature segmentation by adding full-scale skip connections and hierarchical soft attention gates to a Residual U-Net. The paper also introduces the Large Renal Vasculature (LRV) dataset, 16,212 image-mask pairs covering six vascular classes, to support training and evaluation. If the claim is right, this is a concrete step toward automatic quantification of vascular changes in kidney pathology, where manual annotation is labor-intensive and existing models lose small or irregular structures. The paper reports that FH-Seg beats the Omni-Seg baseline by 2.67 points in Dice and 2.13 points in F1.","feed_headline":"FH-Seg beats Omni-Seg on fine-grained renal vessel segmentation","feed_subtitle":"New 16,212-image renal vasculature dataset and attention-gated U-Net lift Dice on all six vascular classes.","key_machinery":"The two mechanisms that carry the argument are the full-scale skip connections ($S_F$) and the learnable hierarchical soft attention gates (HSA). $S_F$ concatenates, at each decoder level, encoder features from all scales up to that level and decoder features from all scales above it, after resampling to a common resolution; this is what lets the network merge fine anatomical detail with coarse semantic context. HSA computes an attention coefficient $\\beta_i^m$ for each feature map from a gating signal and a transformed input, through a resampler, a linear transformation $\\varphi$, ReLU and sigmoid activations, and applies it by element-wise multiplication, suppressing noise before features are combined. The backbone is a Residual U-Net, and the dataset's labeling scheme, one class per pixel with the whole-artery label never co-occurring with its subcomponent labels, is what makes the six-class task well-defined for this architecture.","core_discovery":"The paper's central claim is that a Residual U-Net enhanced with full-scale skip connections and learnable hierarchical soft attention gates, called FH-Seg, delivers state-of-the-art fine-grained segmentation of renal vasculature on a new Large Renal Vasculature (LRV) dataset. The full-scale skip connections let each decoder level fuse encoder features from all earlier scales and decoder features from all later scales, while the hierarchical soft attention gates compute per-scale attention coefficients that suppress non-core information before feature fusion. On the LRV dataset, FH-Seg reports 71.23% Dice and 73.06% F1 on the tunica media class, and its per-class results on lumen, artery wall, and hyaline exceed the Omni-Seg baseline by 1.57, 2.03, and 1.23 Dice points respectively; the paper states an overall advantage of 2.67 Dice and 2.13 F1 over Omni-Seg. Ablations attribute about 1.1 points of average Dice gain to the attention gates and about 1.3 points to the full-scale skip connections.","pith_inferences":["The headline 71.23% Dice and 73.06% F1 figures in the abstract are the tunica media class scores, not a six-class average; a class-averaged comparison could rank the methods differently.","The single-class labeling rule means the model never sees an image where the whole artery and its subcomponents are labeled together; a hierarchical multi-label training scheme could capture containment relations and may improve consistency on complex vessels.","All experiments use 512x512 crops from 2048x2048 patches; using the model on full whole-slide images would require a tiling strategy, and the reported scores do not directly measure whole-slide performance.","The comparison is carried out on a single dataset created by the authors; external validation on other renal pathology cohorts would show whether the gains generalize beyond the LRV distribution."],"forward_implications":["The LRV dataset gives the renal pathology community a shared, fine-grained benchmark with six classes, so future methods can be compared on the same 16,212-image split.","Full-scale skip connections plus hierarchical attention can be dropped into a standard U-Net-style backbone, so the design is a practical modification rather than a new architecture family.","Per-class gains on lumen, artery wall, and hyaline suggest the method reduces both over-segmentation from noise and missed small lesions, which is where earlier models fail.","The ablation numbers indicate that the attention gates and the skip connections are complementary: each adds roughly one Dice point, and together they add about 2.4 points."],"supporting_citations":[{"why":"U-Net backbone and the standard baseline that FH-Seg extends and beats.","marker":"[3]"},{"why":"Introduced the full-scale skip-connection idea that FH-Seg adapts to renal vasculature.","marker":"[5]"},{"why":"Provided the attention-gate mechanism that the hierarchical soft attention gates build on.","marker":"[4]"},{"why":"Omni-Seg is the main baseline; the paper's claimed 2.67 Dice and 2.13 F1 advantages are measured against it.","marker":"[17]"},{"why":"Swin Unetr transformer baseline used in the comparison.","marker":"[8]"},{"why":"U-Net++ nested skip connections are the relevant multi-scale skip-connection prior.","marker":"[33]"},{"why":"DeepLabV3 baseline used in the comparison.","marker":"[7]"}],"fun_headline_variants":["FH-Seg: SOTA renal vessel segmentation via full-scale learning","New renal vasculature dataset powers attention-gated FH-Seg","Full-scale skip connections boost renal vessel Dice to 71.23","FH-Seg outpaces Omni-Seg on 16K renal artery images","Hierarchical attention gates sharpen renal vasculature segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes that each pixel should be assigned exactly one of the six vascular classes, so the 'artery' label and its subcomponent labels (tunica intima, tunica media, artery wall) never appear together in the same image.","fun_headline_variants_meta":{"raw":{"variants":["FH-Seg: SOTA renal vessel segmentation via full-scale learning","New renal vasculature dataset powers attention-gated FH-Seg","Full-scale skip connections boost renal vessel Dice to 71.23","FH-Seg outpaces Omni-Seg on 16K renal artery images","Hierarchical attention gates sharpen renal vasculature segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1324,"prompt_tokens":1004,"completion_tokens":320,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":231}},"tokens_in":620,"tokens_out":320,"duration_ms":3422,"temperature":1.0,"reasoning_tokens":231,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:46:43.715718+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the class-averaged Dice and F1 across all six categories for FH-Seg and Omni-Seg; if FH-Seg does not lead by roughly 2.67 and 2.13 points on that average, the abstract's headline comparison is not what it appears to be.","supporting_citations":[{"cited_title":"U-net: Convolutional networks for biomedical image segmen- tation,","cited_arxiv_id":null,"evidence_quote":"U-Net backbone and the standard baseline that FH-Seg extends and beats."},{"cited_title":"Unet 3+: A full-scale connected unet for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Introduced the full-scale skip-connection idea that FH-Seg adapts to renal vasculature."},{"cited_title":"Omni-seg: A scale-aware dynamic network for renal pathological image segmentation,","cited_arxiv_id":null,"evidence_quote":"Omni-Seg is the main baseline; the paper's claimed 2.67 Dice and 2.13 F1 advantages are measured against it."},{"cited_title":"Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,","cited_arxiv_id":null,"evidence_quote":"Swin Unetr transformer baseline used in the comparison."},{"cited_title":"Unet++: A nested u-net architecture for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"U-Net++ nested skip connections are the relevant multi-scale skip-connection prior."},{"cited_title":"An integrated iterative annotation technique for easing neural network training in medical image analysis,","cited_arxiv_id":null,"evidence_quote":"DeepLabV3 baseline used in the comparison."}],"review_version":1}