{"id":"07097899-fe6f-4303-b174-44cd44922237","arxiv_id":"2507.06581","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A rotation-based direction-aware convolution plus a tubular fusion module ranks first on the ATM22 airway segmentation leaderboard, with gains concentrated in branch detection at a small precision cost.","lead":"TfeNet is a new neural network for outlining airways in CT lung scans, built around a convolution operation that rotates its sampling pattern to follow the branching, tree-like shape of the airways. The authors report the top mean score (94.95%) on the ATM22 benchmark, the largest public airway dataset, which matters for bronchoscopic navigation and robot-assisted lung examinations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ATM22 margin (0.42 MeanScore) is well within reported per-case SD (~4); the leaderboard rank does not by itself establish the claimed superiority.","rationale":"The paper proposes a genuinely novel operator (DAConv) and provides extensive evaluation, a public repository, and a CUDA reimplementation of DSConv; these are real contributions. I read the BAS experiments as the most controlled evidence and the ATM22/AIIB23 leaderboard results as the headline. The single most load-bearing point is the statistical relation between the reported margins and the reported per-case standard deviations. The ATM22 margin over the runner-up is 0.42 MeanScore points with per-case SDs around 4; with n=150, the standard error of the difference is about 0.5, so the observed margin is not distinguishable from zero. This directly undercuts the abstract's 'highest overall score' claim as evidence of method superiority. The same pattern holds for the BAS comparisons and for the DAConv-vs-DSConv ablation, where the gains are about one point with SDs of 6–8 on roughly 20 cases. I do not consider this a refutation: the consistent direction of the effects and the controlled BAS setup suggest the method may help, but the paper's own numbers do not support the strength of the claim. The reader's weakest_assumption about leaderboard comparability (ensemble mixing, test-set tuning, q=pi/4 transfer) is valid and related, but I would put the statistical indistinguishability first because it applies even if comparability is perfect. Minor issues like the misprinted IOU in Eq. (21) and the two-stage ensemble reporting are addressable and further support a CONDITIONAL rather than REJECT. Since the reader already issued CONDITIONAL, my stress-test does not change the verdict.","tokens_in":20438,"tokens_out":7998,"duration_ms":88375,"concrete_test":"Using the ATM22 test-set statistics in Table 5 (TfeNet 94.95±3.90, Timi 94.53±4.68, n=150 per method), compute a two-sample Welch's t-test and a bootstrap 95% confidence interval for the MeanScore difference. If the CI includes zero, the 'highest overall score' claim should be reported as a leaderboard rank rather than a statistically supported performance gain, and the abstract and Section 5.2 wording should be tempered accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim that TfeNet 'achieves the highest overall score of 94.95%' on ATM22, and the sub-claim that rotation-parameterized DAConv improves tubular feature extraction over offset-based deformable convolution, rest on mean differences that fall well inside the reported per-case variability. On the ATM22 test set (Table 5), TfeNet's MeanScore is 94.95±3.90 (n=150) versus Timi's 94.53±4.68; the 0.42-point margin yields a Welch t ≈ 0.84 (p ≈ 0.40) and a 95% CI for the difference that includes 0. On BAS (Table 2), TfeNet's MeanScore advantage over nnUNet is 1.00 (87.37±8.12 vs 86.37±6.92, n=20), also not significant. Even the core ablation (Table 7) comparing DAConv45 to DSConv shows TD/BD gains of 0.97/1.41 with standard deviations of 6–8 on n≈20, again within noise. These are not marginal findings being overinterpreted; the reported numbers themselves do not distinguish the method from the runner-ups. The two-stage ensemble protocol on ATM22/AIIB23 (nnUNet Output1 plus TfeNet Output2) further complicates the comparison, but the statistical insufficiency is more fundamental: even under perfectly controlled comparability, the evidence as presented does not establish superiority. The consistent direction of the in-house BAS trends suggests a real effect may exist, so this is a 'strengthen the evidence' concern rather than a refutation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes TfeNet, a 3D airway segmentation network combining a novel Direction-Aware Convolution (DAConv) that rotates linear kernel segments via learned angles, and a Tubular Feature Fusion Module (TFFM) that fuses x/y/z DAConv features with a standard convolution via residual connections. The network is evaluated on the BAS, ATM22, and AIIB23 datasets. The paper claims the highest MeanScore (94.95%) on ATM22 test set, best MeanScore on BAS (87.37%), and second overall score on AIIB23 validation (89.26%). The authors also report ablations on BAS comparing DAConv against linear and DSConv constructions, and provide public code.","tokens_in":20667,"tokens_out":6130,"duration_ms":59991,"significance":"If the reported performance differences are real, the paper makes a modest but useful contribution to tubular-structure segmentation: DAConv provides a constrained, interpretable alternative to offset-based deformable convolutions, and the full architecture achieves competitive results on three benchmarks. The strengths are the public code, the disclosure of the loss-function switch after validation inspection, and the controlled in-house BAS comparison with several baselines. However, the headline claims of superiority are not yet supported by the reported statistics: all key comparisons have margins within the per-case standard deviations, and no significance tests are reported. The paper's central claim therefore requires additional evidence or reframing before it can be considered established.","major_comments":[{"comment":"The abstract and Section 5.2 claim that TfeNet 'achieves the highest overall score of 94.95%' on ATM22. The margin over the runner-up (Timi, 94.53) is 0.42 points, while the reported per-case standard deviation is 3.90 for TfeNet and 4.68 for Timi on n=150. This difference is far from significant (approximate Welch t ≈ 0.84, p ≈ 0.40, 95% CI containing 0). Similarly, the BAS results in Table 2 (87.37 vs 86.37, n=20) are within noise. The paper should report appropriate significance tests (paired where applicable, with multiplicity correction) or effect sizes, and the wording should be moderated if the differences are not statistically meaningful.","section":"§5.2, Table 5"},{"comment":"The rotation limit q is selected by comparing DAConv45, DAConv60, and DAConv90 on the BAS test set (Table 7), and this selected value q=π/4 is then used in all experiments, including the BAS results in Table 2. This constitutes test-set tuning and makes the reported BAS superiority non-independent; the same concern applies to the transfer of q to other datasets. The authors should select q on a validation split, or at minimum justify the choice from domain knowledge, and disclose this limitation.","section":"§6.1, Table 7"},{"comment":"The central sub-claim that rotation-based kernels improve over offset-based deformable convolutions is not directly tested. Table 7 compares linear, DSConv, and DAConv within the same TFFM, but DSConv is a continuity-constrained snake convolution, not a standard offset-based deformable convolution. The only offset-based baseline, DCUNet in Table 2, differs in the overall architecture (3D-UNet without TFFM), so the difference cannot be attributed to the kernel parameterization. An ablation arm using classic deformable convolution (with learned offsets) inside the same TFFM is needed to support the stated sub-claim.","section":"§6.1, Table 7 and §2.2"},{"comment":"The AIIB23 and ATM22 results are obtained with a two-stage pipeline in which nnUNet serves as Output1 and TfeNet as Output2 (Section 4.2; Section 5.3). The leaderboard comparators in Tables 5 and 6 are not necessarily two-stage or ensemble-based, so the comparison may conflate the method's contribution with the benefits of ensembling/refinement. Reporting the Output2-only result (or the single-stage variant of TfeNet) on these datasets would clarify the contribution of DAConv and TFFM.","section":"§5.3 and §4.2"}],"minor_comments":[{"comment":"The text says the BAS test set contains 'more than 20 scans' while Table 1 lists 20; the numbers should be reconciled.","section":"§4.1, Table 1"},{"comment":"The sentence 'Compared to Linear, TD and BD increase by 2.% and 2.57%, respectively' contains an incomplete percentage; should read '2.32%' or the correct value.","section":"§6.1"},{"comment":"The definition of the sampling locations in Eq. (1) is confusing: the set lists (x−1,1,1), (x−2,1,1), ..., (x+1,1,1), (x+2,1,1), which is not the expected ordered set for a k×1×1 kernel. Please define the coordinates consistently and use proper indexing in Eqs. (4)–(6).","section":"§3.1, Eq. (1)"},{"comment":"The phrase 'profits from its innovative encoder-decoder structure' should be 'benefits from its innovative encoder-decoder structure'.","section":"§2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound in its implementation but the central benchmark claims are statistically weak. If the authors cannot provide significance tests or robust effect-size evidence, the claims should be substantially softened. The test-set-based selection of q is a concern for the editorial desk. The availability of code and the in-house BAS comparison are strengths that could allow a revision to succeed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2507.06581 (TfeNet for airway segmentation). The genuinely new piece is DAConv: a linear convolution kernel whose two halves rotate by learned angles, parameterized as four angles per axis. That is distinct from deformable convolution's free offsets and DSConv's chain offsets, and the rotation algebra in Eq. (5) is internally consistent. The TFFM module is simple—axis-wise concatenation of DAConv features with a 3x3 conv and a residual path—but the ablation (Table 7) shows it helps, and the BAS comparisons against DSConv and deformable baselines are fair and reasonably controlled. Code is public. That's a solid engineering contribution.\n\nThe soft spots are about the strength of the claims, not the existence of the method. On the ATM22 test leaderboard, TfeNet's MeanScore is 94.95 vs Timi's 94.53, a 0.42-point margin with per-case SDs around 3.9–4.7 on n=150. By my calculation that's a Welch t ≈ 0.84, p ≈ 0.4. The paper calls this \"the highest overall score,\" which is true as a ranking but not established as a meaningful difference. The BAS margin over nnUNet (1.00 point, SDs 6–8, n=20) and the DAConv-vs-DSConv ablation gains (TD +0.97, BD +1.41, SDs 6–8) sit in the same zone: consistent direction, but within noise. The AIIB23 result is second place, but that pipeline uses nnUNet as Output1 plus TfeNet as Output2—an ensemble compared to single-model leaderboard entries. Also, Eq. (21) as printed defines IOU as 1 - TP/(TP+FP+FN), which is actually 1 - IoU. That's a typo, but it should be fixed.\n\nOne thing I'd push back on in the stress-test note: the \"peeking\" concern. They disclose choosing q=π/4 on the BAS test set and switching to Tversky after seeing ATM22 validation misses; that's overfitting to val/test to some degree, but it's not hidden, and the in-house comparisons are still informative. The bigger issue is statistical: the headline rank is fragile.\n\nWho is this for? People working on tubular structure segmentation or airway annotation pipelines. It's worth a serious referee, because the operator is novel and the evaluation is broad—but the referee should ask for significance tests or paired comparisons, single-stage vs ensemble numbers for the challenge submissions, and a corrected IOU equation. If those come back solid, TfeNet is a useful addition.\n\nRecommendation: send to peer review. Not a desk reject.","headline":"Novel rotation-parameterized kernel with broad benchmark work, but the SOTA headline rests on margins that are within per-case noise.","tokens_in":21358,"tokens_out":2759,"would_cite":false,"duration_ms":26736,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","92C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"TfeNet claims that rotating convolution kernels, not offset-based deformation, recovers fine airway branches and reports the top ATM22 score of 94.95%.","keywords":["airway segmentation","direction-aware convolution","tubular structure extraction","deformable convolution","feature fusion","computed tomography","class imbalance","topology preservation"],"falsifier":"A controlled replication would settle it: train the top ATM22 entries and TfeNet as single models under an identical protocol, with the same patches, loss, and GPU budget, and compare MeanScore on the same 150-case test split; if the 94.95% figure does not stay on top, or if an offset-based deformable convolution with equal parameter count and a continuity regularizer matches DAConv's TD and BD, then rotation is not the mechanism carrying the result.","tokens_in":2188,"feed_emoji":"🫁","tokens_out":2725,"duration_ms":132825,"temperature":0.7,"pith_summary":"TfeNet is an attempt to fix a specific, practical failure in airway segmentation: standard and deformable convolutions lose thin distal branches, so predicted airway trees come out fragmented. The paper proposes deforming convolution kernels by rotation instead of by learned offsets, splitting each linear kernel at its center and rotating the two arms by learnable angles so the sampling pattern always stays a connected line segment or polyline lying along the airway. Combined with a residual fusion module that mixes x-, y-, and z-direction responses, the network claims the highest overall score (94.95%) on ATM22, the largest airway dataset, and the best topology scores on the BAS benchmark. That matters because distal-airway reachability is the metric that determines whether bronchoscopic navigation and robotic systems can be deployed clinically.","feed_headline":"Rotating kernels top airway segmentation at 94.95%","feed_subtitle":"A rotation-based kernel swap beats offset deformable convolutions at tracing fine distal airway branches.","key_machinery":"The load-bearing object is DAConv, a direction-aware convolution: a linear kernel of length $k$ is split at its center and the two arms are each rotated twice (around the y-axis then z-axis for the x-direction kernel) by angles learned per location, with the chain $3\\times3\\times3$ conv, instance norm, tanh, and scale $q\\in(0,\\pi]$ turning features into angles bounded by $q=\\pi/4$. The rotation matrices $R_x(\\theta)$, $R_y(\\theta)$, $R_z(\\theta)$ define the new sampling positions, so the kernel becomes a bendable line segment or polyline in 3D instead of a fixed straight bar, and trilinear interpolation reads the image at fractional coordinates. Around it sits the TFFM, which fuses DAConv responses from the x, y, and z axes with a standard $3\\times3\\times3$ branch and preserves the input via residual addition. Training uses the General Union Loss with local-imbalance weights (Tversky loss in ATM22's first stage), patch-based 128$^3$ sampling, and a two-stage Output1 plus Output2 protocol, with nnU-Net serving as Output1 on AIIB23; post-processing keeps the largest connected component and fills holes, which directly supports the reported topology scores.","core_discovery":"The paper claims that rotation-based kernel deformation is better than offset-based deformation for tree-like tubular structures, and that this mechanism drives top-ranked airway segmentation. DAConv takes the linear kernels $k\\times1\\times1$, $1\\times k\\times1$, and $1\\times1\\times k$ along the three axes, splits each kernel at its center $(k-1)/2$, and rotates the two halves around the travel axis by angles $\\theta_1,\\dots,\\theta_4$ produced on the fly from the input feature map via a $3\\times3\\times3$ convolution, instance normalization, tanh, and a scale factor $q=\\pi/4$; the resulting fractional sampling positions are read with trilinear interpolation. Because the two arms remain connected through the fixed rotation center, the kernel can bend but never disintegrate into scattered points, which is the failure the paper attributes to deformable convolutions. The Tubular Feature Fusion Module concatenates the three directional DAConv responses with a $3\\times3\\times3$ context branch and adds the input back through a $1\\times1\\times1$ residual, preserving faint airway signal. On this machinery the paper reports the best MeanScore (87.37%), tree-length detection (84.27%), and branch detection (77.51%) on BAS; the top DSC (95.49%) and MeanScore (94.95%) on the ATM22 test set; and second place (89.26%) on the fibrosis-focused AIIB23 validation set.","pith_inferences":["My read: DAConv's four learned angles are far fewer free parameters than deformable convolution's per-location offsets, so part of the reported gain may be a low-degree-of-freedom regularization effect; a parameter-matched comparison between rotation and offset designs would separate the geometry from the regularization.","My read: the fixed $q=\\pi/4$ was tuned on the BAS test set and carried to ATM22 and AIIB23; fibrosis-distorted airways on AIIB23 may want a different rotation range, so a per-dataset or learned $q$ is a natural extension that could widen or shrink the reported margins.","My read: the ATM22 margin over the runner-up (94.95 vs 94.53) is smaller than the reported per-case standard deviation of 3.90, so the exact leaderboard position is fragile even if the in-house BAS comparison stands; the mechanism claim and the precise ranking should be weighed differently.","My read: the two-stage pipeline, with nnU-Net as Output1 on AIIB23, means the challenge numbers include an ensemble effect; the BAS results, where both TfeNet1 and TfeNet2 come from TfeNet alone, are the cleaner evidence for the DAConv mechanism."],"forward_implications":["If the mechanism claim holds, any tubular-structure segmentation task, including vessels, coronary arteries, and nerves, can swap offset-based deformable blocks for DAConv blocks and expect more connected predictions, since the paper frames DAConv as generic tubular feature extraction rather than an airway-specific trick.","The reported balance on ATM22, a spread of only 2.2% between the best and worst of the four metrics, means a single TfeNet model needs no accuracy-versus-continuity tuning knob, unlike leaderboard entries that score about 97% on tree length by sacrificing roughly 8 points of precision.","Because TFFM is plug-and-play, existing 3D U-Net-family models can gain branch continuity by replacing convolution blocks rather than redesigning losses or adding a second network.","The ablation's monotone pattern, with $q=\\pi/4$ beating $q=\\pi/3$ and $q=\\pi/2$, implies the benefit comes from restricting the rotation range, so practitioners should keep $q$ small rather than letting angles roam.","On BAS, combining Output1 with Output2 adds 1.99 to 2.66 points of tree-length and branch-detection score over Output1 alone, so the two-stage protocol is itself a large part of the topological gain, a corollary the paper states explicitly."],"supporting_citations":[{"why":"Supplies the BAS dataset, the General Union Loss, and the local-imbalance weighting used in every TfeNet training run.","marker":"Zheng et al., 2020"},{"why":"Introduces the refined local-imbalance weight (LIB) that sets voxel weights inside the GUL loss.","marker":"Zheng et al., 2021"},{"why":"Defines the ATM22 challenge, its 150-case test split, and the MeanScore protocol that produces the 94.95% headline.","marker":"Zhang et al., 2023"},{"why":"Defines the AIIB23 fibrosis dataset and the OverallScore formula behind the second-place 89.26% result.","marker":"Nan et al., 2023"},{"why":"Introduces deformable convolution, the offset-based approach DAConv is designed against, and the basis of the DCUNet control.","marker":"Dai et al., 2017"},{"why":"Dynamic Snake Convolution, the closest continuity-constrained prior work; the paper reimplements its kernel and outscored it in the ablation by 0.97% TD and 1.41% BD.","marker":"Qi et al., 2023"},{"why":"nnU-Net, the principal baseline and the Output1 stage on AIIB23, whose preprocessing choices are used throughout.","marker":"Isensee et al., 2020"},{"why":"3D U-Net, the encoder-decoder backbone into which the TFFM modules are inserted.","marker":"Çiçek et al., 2016"}],"fun_headline_variants":["Rotating kernels outdo deformable nets on airways","TfeNet's rotated kernels score 94.95% on ATM22","Bent kernels trace finer airway branches than offsets","Direction-aware rotation improves airway branch detection"],"cache_read_input_tokens":23168,"weakest_assumption_plain":"The headline ranking assumes that the published scores of rival methods are fair comparisons, with equal compute and no multi-model ensembling on their side, while TfeNet's numbers come from combining two model outputs and its 0.42-point lead over second place sits well below the 3.90-point case-to-case variation.","fun_headline_variants_meta":{"raw":{"variants":["Rotating kernels outdo deformable nets on airways","TfeNet's rotated kernels score 94.95% on ATM22","Bent kernels trace finer airway branches than offsets","Direction-aware rotation improves airway branch detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1650,"prompt_tokens":1097,"completion_tokens":553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":713,"completion_tokens_details":{"reasoning_tokens":489}},"tokens_in":713,"tokens_out":553,"duration_ms":6654,"temperature":1.0,"reasoning_tokens":489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:01:42.577967+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled replication would settle it: train the top ATM22 entries and TfeNet as single models under an identical protocol, with the same patches, loss, and GPU budget, and compare MeanScore on the same 150-case test split; if the 94.95% figure does not stay on top, or if an offset-based deformable convolution with equal parameter count and a continuity regularizer matches DAConv's TD and BD, then rotation is not the mechanism carrying the result.","supporting_citations":[],"review_version":1}