{"id":"ea748914-bc40-4e72-9bc5-6474b1413c00","arxiv_id":"2412.00787","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TSUBF-Net, a U-Net style 3D network with cross-layer attention and a Sobel smoothness loss, reports state-of-the-art segmentation metrics on adenoid hypertrophy CT scans.","lead":"This paper presents TSUBF-Net, a 3D deep learning network for segmenting enlarged adenoids in children's head CT scans, reporting better boundary accuracy than earlier models on their private dataset. A generalist reader might care because accurate adenoid segmentation could support safer surgical planning for pediatric sleep apnea.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Superiority over SOTA on AHSD rests on one private split with no error bars and a λ chosen on the test set; the 1.13 DSC edge over nnUNet is unverified.","rationale":"After reading the paper in full, the central claim is exactly the abstract's 'superior to SOTA' on the private AHSD. The paper is honest about AHSD being private and describes the clinical motivation well; the proposed modules are plausible and the public ACDC/MSD-Lung results are competitive. However, the AHSD evidence has three compounding weaknesses: (i) a single random split with 38 test volumes and no variance estimate; (ii) no annotation reliability evidence for the very boundary that the method is designed to handle; and (iii) λ=0.1 selected on the same test set used for the headline comparison. These are not mere presentation issues: they directly affect whether the 1.13 DSC gap over nnUNet is real or within noise. The internal contradiction about λ=1.0 vs 0.1 and the swapped dataset descriptors in the conclusion (ACDC(CT), MSD-Lung(MRI)) lower confidence but are not the primary load-bearing issue. The proposed concrete test—releasing data/code and running repeated splits/seeds with paired statistics—would settle whether the superiority claim is reproducible. This supports the reader's CONDITIONAL verdict; I recommend no change to that verdict.","tokens_in":19898,"tokens_out":6015,"duration_ms":51211,"concrete_test":"Release AHSD under a data-use agreement and the training code, then rerun the Table 2 comparison with 5-fold cross-validation (or at least three different random 189/38 splits) and three seeds per method, reporting mean±SD per metric. Compute the paired difference in DSC and HD95 between TSUBF-Net and nnUNet and between TSUBF-Net and UNETR++, and test whether the advantage survives (e.g., Wilcoxon signed-rank or bootstrap CI). If the DSC difference over nnUNet shrinks below about 1 point or the 95% CI includes 0, the 'superior to SOTA' claim should be softened to 'competitive on our dataset.' Additionally, have a second clinician re-contour 10 random test cases to verify that model-vs-annotator agreement is not better than annotator-vs-annotator agreement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2 is the entire basis for the central claim. It reports single-run scores on one private 189/38 split of AHSD. TSUBF-Net's advantage over the strongest baseline nnUNet is only 1.13 DSC (92.26 vs 91.13) and 1.13 mm HD95 (7.03 vs 8.16), while its advantage over UNETR++ is larger. With 38 test volumes, HD95 is dominated by the worst boundary outlier; one ambiguous or mislabeled posterior boundary can change HD95 by several mm. The annotation protocol is not described: no number of raters, no inter-observer agreement, and no explicit rule for the clinically ambiguous posterior border, even though the paper emphasizes that this border is indistinct. In addition, the best λ=0.1 in Table 5 was selected by evaluating λ∈{1.0,0.5,0.1} on the same test set, so the reported 4.57% enhancement includes selection bias; the public-dataset results are more credible but do not rescue the AHSD SOTA claim because AHSD is the only adenoid benchmark. The text also contradicts itself on which λ is best: Section 4.4 says λ=1.0 gives the best indicators, while Table 5 shows λ=0.1 is best, further undermining the reliability of the ablations. Code and AHSD are not released, so none of the critical numbers can be independently recomputed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TSUBF-Net, a 3D U-shaped segmentation network for adenoid hypertrophy in CT, with three contributions: the Trans-Spatial Perception (TSP) module, the Bi-directional Sampling Collaborated Fusion (BSCF) module, and a Sobel-gradient smoothness term added to the loss. The method is evaluated on a private adenoid dataset (AHSD), where it is claimed to outperform prior state-of-the-art methods (HD95 7.03, IoU 85.63, DSC 92.26), and on two public benchmarks, ACDC and MSD-Lung, where it is competitive or favorable. The paper also reports ablation studies showing the contribution of each module and the sensitivity of the Sobel loss weight.","tokens_in":20217,"tokens_out":4313,"duration_ms":39616,"significance":"If the empirical claims are accepted, the paper offers a useful architectural recipe for 3D segmentation of structures with weak or ambiguous boundaries, and it addresses a genuinely under-studied clinical task. The inclusion of public-dataset results (ACDC, MSD-Lung) is a real strength and provides some evidence that the proposed modules generalize beyond the private AHSD data. The ablations, despite the issues noted below, also give a clear decomposition of the contributions of TSP and BSCF. However, the central SOTA claim is an empirical benchmark result on a private dataset, and the current experimental design does not yet support that claim as stated: there are no error bars, the test set is used for hyperparameter selection, the ground-truth annotation protocol is unspecified, and one key metric definition appears mis-specified.","major_comments":[{"comment":"The headline superiority claim rests on a single random split of a private dataset (189 training / 38 test volumes) with no error bars, no multiple seeds, and no cross-validation. The reported advantage over nnUNet is small in absolute terms (DSC 92.26 vs 91.13, HD95 7.03 vs 8.16), and HD95 is an outlier-sensitive boundary metric on a test set of only 38 volumes. The authors should report per-volume distributions, confidence intervals, and ideally results across multiple splits or seeds, so the reader can assess whether the advantage is stable.","section":"Section 4.1 and Table 2"},{"comment":"The Sobel loss weight lambda is selected by evaluating lambda in {1.0, 0.5, 0.1} on the AHSD test set, since no validation split is described. This means the reported 4.57% enhancement over the baseline includes selection bias, and the comparison with fixed-hyperparameter baselines is not fully fair. The authors should either introduce a separate validation split for hyperparameter selection or use nested cross-validation, and should describe the selection procedure explicitly.","section":"Section 4.3.4 and Table 5"},{"comment":"There is a direct internal contradiction about which lambda value is best. Section 4.4 states that 'when the lambda is set to 1.0, the three evaluation indicators all get the best results,' but Table 5 shows lambda = 0.1 gives the best HD95 (7.03), IoU (85.63), and DSC (92.26), while lambda = 1.0 gives HD95 7.78, IoU 84.63, and DSC 91.68. This contradiction must be resolved; as written, it undermines the reliability of the ablation conclusions.","section":"Section 4.4 vs Table 5"},{"comment":"The HD95 definition appears mis-specified. The standard definition is HD95(Y,P) = max( h95(Y,P), h95(P,Y) ), where h95 is the 95th percentile of the directed distances. Equation (10) writes HD95(Y,P) = max( dYP + dPY ), and the surrounding text describes 'maximum 95th percentile distance' in a way that conflates the two directed distances. The authors should correct the equation and state precisely which implementation was used, since this metric is central to the claimed improvement.","section":"Equation (10)"},{"comment":"No annotation protocol is described for AHSD: the paper does not state the number of annotators, whether there was any inter-observer agreement assessment, or how the clinically ambiguous posterior boundary was defined and resolved. Given that the paper repeatedly emphasizes that this boundary is indistinct and clinically important, the quality and consistency of the ground-truth labels are load-bearing for the reported benchmark. This information should be added, or the corresponding limitation should be explicitly acknowledged.","section":"Section 4.1"}],"minor_comments":[{"comment":"The conclusion mislabels the datasets: it refers to 'ACDC(CT)' and 'MSD-Lung(MRI)', whereas Section 4 correctly identifies ACDC as MRI and MSD-Lung as CT.","section":"Section 5"},{"comment":"There is a typo: 'Swim-Transformer' should be 'Swin-Transformer'.","section":"Section 2.4"},{"comment":"The loss expression in Equation (5) lacks clear parentheses; as typeset, it is not clear whether the cross-entropy term is inside or outside the sum over classes. Please reformat.","section":"Equation (5)"},{"comment":"The text defines FLOPs as 'Floating Point Operations Per Second', but the table reports FLOPs in G (giga floating-point operations), which is a count, not a rate. Please correct the terminology.","section":"Section 4.3.4"},{"comment":"The attention formula appears to have a typo: it should be Softmax( Q K^T / sqrt(d) ) * V rather than Softmax( Qs, K^T / sqrt(d) ) * V.","section":"Equation (2)"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim depends entirely on a private dataset and the authors' own implementation, with no code or data release. If the journal's policy strongly values reproducibility for benchmark claims, this is a scope concern in addition to the technical issues. The public-dataset results are more credible and should be emphasized in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nTwo things to know about this paper. First, it is a serious first attempt at 3D segmentation of adenoid hypertrophy in CT, with a new private dataset (AHSD, 227 hypertrophy cases, 189/38 train/test split) and a U-shaped architecture with two added modules: a trans-spatial attention module (TSP) and a bi-directional fusion block (BSCF), plus a Sobel gradient smoothness term in the loss. Second, the headline \"SOTA on AHSD\" should be taken with a large grain of salt: it rests on one private split, no error bars, and the best lambda=0.1 was selected by evaluating on the test set. The 1.13 DSC edge over nnUNet (92.26 vs 91.13) is within the noise you would expect from a single 38-volume test split.\n\nWhat is actually new: the dataset and task, and the combination of modules. The public dataset results are the strongest part. On MSD-Lung, TSUBF-Net beats UNETR++ by about 3.7 DSC, and on ACDC it is essentially tied (92.68 vs 92.83). That shows the architecture is not junk. The TSP module, with shared Q/K across height/width/depth attention, is a reasonable way to get inter-layer spatial context with modest parameter growth, and the BSCF fusion idea is sensible for combining down-sampled and up-sampled features.\n\nSoft spots, in order of severity. (1) The AHSD benchmark is not independently verifiable: no code, no data release, no annotation protocol, no inter-observer agreement. The posterior border is clinically ambiguous and the paper says so, but never describes how the annotators handled it. (2) The ablation is compromised by test-set tuning of lambda; reporting a 4.57% enhancement from the best of three values is not a fair estimate of the loss's benefit. (3) The text contradicts itself: Section 4.4 says lambda=1.0 gives the best indicators, while Table 5 shows lambda=0.1 is best. (4) Eq. 10 misstates HD95; it is a percentile of distances, not the max of a sum. (5) The conclusion swaps dataset labels, calling ACDC \"CT\" and MSD-Lung \"MRI.\" These are presentation errors, but they do not help trust.\n\nThe central architectural idea is plausible and the public evidence is competitive. This paper deserves a serious referee, but the AHSD claim needs substantive revision. I would send it to review with a request for code/data, multiple seeds, and correction of the internal inconsistencies. The likely readership is the pediatric airway imaging community and anyone working on loss functions for fuzzy boundaries.","headline":"Plausible first 3D adenoid segmentation paper whose headline AHSD SOTA claim rests on one private split, while the public-dataset results are the more credible evidence.","tokens_in":20721,"tokens_out":2955,"would_cite":false,"duration_ms":28183,"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":"TSUBF-Net, a 3D U-shaped segmentation network with trans-spatial attention, bidirectional fusion, and a Sobel-gradient smoothness loss, outperforms prior methods for segmenting adenoid hypertrophy in CT scans, reporting DSC 92.26, IoU…","keywords":["adenoid hypertrophy","3D medical image segmentation","CT segmentation","trans-spatial attention","feature fusion","Sobel loss","U-shaped network","pediatric airway"],"falsifier":"Take the 38 held-out AHSD CT volumes, have a second clinician independently redraw the adenoid boundaries, and recompute HD95 and DSC for TSUBF-Net and UNETR++ against that new ground truth; if the margin over UNETR++ shrinks below the reported 3.99 DSC points or flips on HD95, the boundary-superiority claim is not robust. Alternatively, retrain on several random 189/38 splits and check whether TSUBF-Net wins every time.","tokens_in":19724,"feed_emoji":"","tokens_out":7928,"duration_ms":65130,"temperature":0.7,"pith_summary":"TSUBF-Net is a 3D U-shaped segmentation network built for a task where previous work has been sparse: delineating adenoid hypertrophy in pediatric head CT. The paper argues that adenoid boundaries are unusually blurred, so the network adds a Trans-Spatial Perception (TSP) module that attends across CT slices, a Bi-directional Sampling Collaborated Fusion (BSCF) module that merges down-sampled and up-sampled features, and a Sobel-based loss that penalizes rough surfaces. On the authors' private 240-patient AHSD dataset, TSUBF-Net reports the best scores among the compared methods, with HD95 7.03, IoU 85.63, and DSC 92.26, beating nnUNet and UNETR++. If the result holds, it would give surgeons a quantitative, volumetric view of the adenoid before ablation surgery, where today they rely largely on endoscopic views and experience.","feed_headline":"New 3D network tops adenoid CT segmentation at 92.26","feed_subtitle":"Spatial attention, two-way fusion, and a Sobel smoothness term beat UNETR++ on a 240-patient pediatric CT set.","key_machinery":"The load-bearing pieces are three. The Trans-Spatial Perception (TSP) module is a four-head attention mechanism: one channel head plus three inter-layer spatial heads that attend along height, width, and depth, with shared query and key weights, so a voxel can gather context from neighboring slices. The Bi-directional Sampling Collaborated Fusion (BSCF) module replaces plain skip connections: up-sampled and down-sampled features are first aligned by a shared 3x3x3 convolution, then up-sampled features act as queries against down-sampled keys and values so contour information corrects segmentation features. The Sobel loss convolves the prediction with 3D Sobel kernels along x, y, and z and adds the gradient magnitude to the soft Dice and cross-entropy loss, explicitly favoring smooth surfaces.","core_discovery":"On its own terms, the paper's central claim is that the specific combination of TSP, BSCF, and Sobel loss, rather than any single component, is what solves fuzzy-boundary adenoid segmentation. The ablation table shows that TSP alone slightly lowers DSC relative to the UNETR++ baseline (87.78 versus 88.22) while cutting FLOPs; adding BSCF raises DSC to 90.91; adding both reaches 91.49; and the Sobel loss with lambda 0.1 brings the model to 92.26, a 4.57-point improvement over baseline. The paper reports TSUBF-Net as superior to all compared methods on AHSD, with HD95 of 7.03 versus 10.13 for UNETR++ and 8.16 for nnUNet, and also reports the best DSC on MSD-Lung (83.69) while being close to the best on ACDC (92.68, versus UNETR++ at 92.83). All reported AHSD numbers are single-model, without pre-training or ensembling.","pith_inferences":["The superiority claim rests on one private dataset with a single random 189/38 split and no inter-observer variability estimate; re-splitting or independent re-annotation could narrow or reverse the reported margins, so the headline numbers should be read as provisional until the dataset is released or validated externally.","The pattern across datasets suggests a trade-off: explicit boundary-smoothness regularization helps when the target has genuinely indistinct borders, such as adenoids and lung tumors, but can slightly hurt on clearly bounded organs like cardiac MRI; a learned or per-dataset lambda, which the paper itself floats, would be the natural extension.","The BSCF mechanism of letting up-sampled features query down-sampled features is a general recipe for any U-shaped segmenter where contour and semantics live at different scales; it could be grafted onto other backbones and tested on other ill-defined structures such as tumor margins or ground-glass opacities."],"forward_implications":["On AHSD, the full model's DSC advantage over UNETR++ is 3.99 points and over nnUNet is 1.13 points, with the largest reported gap in HD95 (7.03 versus 10.13 and 8.16), so the claim is specifically a boundary-accuracy improvement.","The same architecture transfers to a different small-target CT task: on MSD-Lung, TSUBF-Net reports DSC 83.69 versus 80.68 for UNETR++ and 80.14 for MedNeXt.","On ACDC, an MRI dataset with smooth organ boundaries, TSUBF-Net's mean DSC of 92.68 is essentially tied with UNETR++ at 92.83, suggesting the smoothness prior is neutral or slightly detrimental when boundaries are already well defined.","The ablation values imply that the BSCF fusion module, rather than the TSP attention module, accounts for most of the accuracy gain, while TSP chiefly buys lower computational cost.","The paper's stated future direction is to extend segmentation to tonsils, turbinates, and epiglottis to build a complete sleep-apnea database, which the architecture is positioned to support."],"supporting_citations":[{"why":"U-Net supplies the standard CNN baseline and the U-shaped architecture that TSUBF-Net builds on.","marker":"[21]"},{"why":"nnUNet is the strongest CNN competitor on AHSD and the source of the preprocessing pipeline.","marker":"[26]"},{"why":"UNETR++ is the main transformer-CNN baseline whose EPA block is the direct comparison for TSP and the ablation baseline.","marker":"[37]"},{"why":"Merged U-Net supplies the merge-gate idea that BSCF adapts for selective feature fusion.","marker":"[45]"},{"why":"UNETR is a transformer-encoder baseline compared on AHSD, ACDC, and MSD-Lung.","marker":"[36]"},{"why":"nnFormer is an interleaved transformer-CNN baseline compared on all three datasets.","marker":"[41]"},{"why":"Sobel edge detection supplies the operator used to build the gradient smoothness loss.","marker":"[43]"},{"why":"V-Net provides the soft Dice plus cross-entropy base loss that the Sobel term extends.","marker":"[42]"}],"fun_headline_variants":["TSUBF-Net: 92.26 DSC on adenoid CT segmentation","Adenoid CT segmentation: new 3D net hits 92.26 DSC","Bi-direction fusion boosts adenoid CT segmentation to 92.26","Sobel loss and fusion push adenoid CT segmentation to 92.26 DSC","TSUBF-Net outdoes UNETR++ on pediatric adenoid CT segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the manual ground-truth boundaries in the AHSD dataset are correct and that one random split into 189 training and 38 test patients represents the full variety of pediatric adenoid CT anatomy.","fun_headline_variants_meta":{"raw":{"variants":["TSUBF-Net: 92.26 DSC on adenoid CT segmentation","Adenoid CT segmentation: new 3D net hits 92.26 DSC","Bi-direction fusion boosts adenoid CT segmentation to 92.26","Sobel loss and fusion push adenoid CT segmentation to 92.26 DSC","TSUBF-Net outdoes UNETR++ on pediatric adenoid CT segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000563,"raw_usage":{"total_tokens":2768,"prompt_tokens":1139,"completion_tokens":1629,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":755,"completion_tokens_details":{"reasoning_tokens":1524}},"tokens_in":755,"tokens_out":1629,"duration_ms":9996,"temperature":1.0,"reasoning_tokens":1524,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:59:50.779820+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the 38 held-out AHSD CT volumes, have a second clinician independently redraw the adenoid boundaries, and recompute HD95 and DSC for TSUBF-Net and UNETR++ against that new ground truth; if the margin over UNETR++ shrinks below the reported 3.99 DSC points or flips on HD95, the boundary-superiority claim is not robust. Alternatively, retrain on several random 189/38 splits and check whether TSUBF-Net wins every time.","supporting_citations":[{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"U-Net supplies the standard CNN baseline and the U-shaped architecture that TSUBF-Net builds on."},{"cited_title":"Unetr++: Delving into efficient and accurate 3d medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"UNETR++ is the main transformer-CNN baseline whose EPA block is the direct comparison for TSP and the ablation baseline."},{"cited_title":"Merged u-net for bone tumors x-ray images segmentation,","cited_arxiv_id":null,"evidence_quote":"Merged U-Net supplies the merge-gate idea that BSCF adapts for selective feature fusion."},{"cited_title":"Unetr: Transformers for 3d medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"UNETR is a transformer-encoder baseline compared on AHSD, ACDC, and MSD-Lung."},{"cited_title":"Sobel edge detection algorithm,","cited_arxiv_id":null,"evidence_quote":"Sobel edge detection supplies the operator used to build the gradient smoothness loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"V-Net provides the soft Dice plus cross-entropy base loss that the Sobel term extends."}],"review_version":1}