{"id":"5fe107ee-5a3a-4046-a172-1c5de4e05c0c","arxiv_id":"2608.12274","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"NA-UNETR, a neighborhood-attention transformer pretrained on CTA and fine-tuned with LoRA on 20 scans, reaches 45.64% Dice for LAD segmentation, a small and not statistically significant gain over baselines.","lead":"NA-UNETR, a new transformer network, segments the tiny left anterior descending coronary artery from non-contrast planning CT scans, where the vessel is barely visible. It beats previous models by a few percentage points on a 20-patient dataset, but the improvement is not statistically significant.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LAD-SEG superiority claim is fragile: the 3.10 pp Dice gain is within the 0.10-0.53 inter-observer range and is not statistically significant, so without multi-observer labels the ranking may reflect contouring style rather than anatomical accuracy.","rationale":"Good-faith reading: the architecture is coherent, the ablations are thorough, and the paper explicitly acknowledges that LAD-SEG results are not statistically significant and that the achieved DSC lies within expert variability. However, the abstract's wording ('improving Dice by 3.10 percentage points over nnU-Net') presents the gain as an improvement without the significance caveat, and the conclusion states that NA-UNETR 'enables accurate segmentation' of LAD on non-contrast CT. The decisive weakness is not the loss formulation or the transformer design; it is the evaluation target. On non-contrast CT, LAD boundaries are inherently ambiguous, and a single-annotator ground truth cannot distinguish 'better anatomy' from 'matches this annotator's style.' Combined with n=20 and p>0.05, the LAD-SEG ranking is not robust standalone evidence. The ImageCAS results (79.49% Dice, p<0.05) support the method's general segmentation ability on clearer CTA, but they do not validate the specific non-contrast LAD claim. The multi-observer check described above would settle whether the concern lands; until then, CONDITIONAL is the appropriate verdict, requiring either independent-label validation or a softened claim that tracks the statistical evidence.","tokens_in":18017,"tokens_out":6933,"duration_ms":65927,"concrete_test":"Obtain at least two additional independent physician contour sets for the 20 LAD-SEG scans. Train all compared models on the original labels as done here, then score each model against each new observer's labels and against a held-out subset of the original labels. If NA-UNETR's DSC/HD95 advantage over nnU-Net and Swin UNETR does not persist on at least one independent observer's labels, or if the advantage is not larger than the inter-observer Dice spread, the claim that NA-UNETR improves LAD delineation is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that NA-UNETR improves LAD delineation on non-contrast CT rests on a single set of physician-delineated labels on only 20 scans, in a regime where the paper itself reports inter-observer Dice of 0.10-0.53 (Section I). The reported 3.10 pp DSC advantage over nnU-Net (45.64 vs 42.54, Table 1) is smaller than this label noise floor; it falls entirely within the stated manual-contour variability. Section IV.B also reports Mann-Whitney p>0.05 for these differences, and the 5-fold validation was used both for hyperparameter selection (Section IV.C, Tables 3-5) and for final performance reporting, adding optimistic bias. Every model was trained and evaluated on the same single-annotator contours, so the comparison is internally consistent, but it cannot establish that NA-UNETR is anatomically better: an architecture whose inductive bias happens to match the annotator's contouring style would score higher without being more clinically accurate. The Discussion concedes that the achieved DSC lies within expert variability, so the distinction between 'better model' and 'different contouring style' is left unresolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes NA-UNETR, a 3D encoder-decoder transformer that replaces global attention with Neighborhood Attention and Dilated Neighborhood Attention, and combines Dice-Focal with a noise-perturbed Hausdorff loss weighted by learnable homoscedastic uncertainties. The model is pretrained on 1,000 ImageCAS CTA volumes and fine-tuned with LoRA on 20 non-contrast LAD-SEG scans. On LAD-SEG it reports the highest Dice (45.64%), clDice (44.39%), and lowest HD95 (38.16 mm) among eight baselines, with ASD 10.01 mm; on ImageCAS it reports 79.49% Dice and 1.02 mm ASD. Ablations address architecture depth, kernel sizes, residual blocks, LoRA rank, loss components, preprocessing, and postprocessing. The LAD-SEG improvements are reported as not statistically significant (p>0.05).","tokens_in":18294,"tokens_out":8648,"duration_ms":78867,"significance":"The clinical problem is well motivated and the proposed system is computationally efficient, with 19.6M trainable parameters, 314.1B FLOPs, and a public code release. The baseline comparison is comprehensive, the ImageCAS gains are reported as statistically significant, and the paper is appropriately cautious about clinical deployment. If the LAD-SEG result held under independent validation, the work would be a useful contribution to cardiac substructure segmentation in radiotherapy planning. However, the primary evidence on the in-house dataset is weak: n=20, p>0.05 for all model differences, and the reported gains fall within the inter-observer variability range cited by the authors. The contribution therefore currently rests mainly on the ImageCAS results and the architectural/ablation analysis rather than on a demonstrated LAD-SEG superiority.","major_comments":[{"comment":"The central claim that NA-UNETR improves LAD delineation on LAD-SEG is not supported by the reported statistics. The 3.10 percentage point Dice advantage over nnU-Net (45.64 vs 42.54) and the 2.96 mm HD95 advantage over Swin UNETR are accompanied by Mann-Whitney p>0.05 for the differences, and the sample is only n=20 with per-case standard deviations of 4-8% Dice. The paper itself cites inter-observer Dice of 0.10-0.53 on non-contrast CT, so the observed differences lie entirely within the label-variability floor. The authors should report effect sizes and confidence intervals, pre-specify the primary comparison, and ideally evaluate on an independent test set or with multi-observer labels; otherwise the LAD-SEG superiority claim should be reframed as a preliminary observation.","section":"Section IV.B, Table 1; Section V"},{"comment":"The same 5-fold split is used both to select hyperparameters (NAT depth, kernel sizes, LoRA rank, and loss variants in Tables 3-5) and to report the final LAD-SEG performance in Table 1. No separate held-out test set or nested cross-validation is described, so the reported numbers are optimistically biased by model selection on the validation folds. The authors should fix all hyperparameters before evaluating on an untouched test split, or use nested cross-validation, or explicitly present the reported LAD-SEG numbers as internal-validation results rather than as final test performance.","section":"Section IV.C, Tables 3-5; Section III.B.1"},{"comment":"The evaluation uses a single physician-delineated contour set per scan, despite the paper's own citation of inter-observer Dice ranging from 0.10 to 0.53 on non-contrast CT. Under such label variability, a model's ranking may reflect which annotator's contouring style it happens to approximate rather than anatomical accuracy. The comparison is internally consistent, but it cannot establish that NA-UNETR is anatomically superior to the baselines. The authors should provide multi-observer labels or a consensus ground truth for at least a subset of scans, or explicitly restrict all LAD-SEG superiority claims to 'matches the reference contours in this dataset'.","section":"Section II.C.1; Section V"},{"comment":"Equation (11) does not define the standard HD95 metric. The formula HD95 = quantile95%(max_{x in boundary of prediction} min_{y in boundary of ground truth} ||x-y||) takes a 95th percentile of a single maximum, which is not meaningful, and it is one-sided rather than symmetric. Standard HD95 is the 95th percentile of the set of all directed nearest-neighbor distances from both boundaries to the other. The authors should correct the definition and confirm that the reported HD95 values were computed with the standard symmetric percentile distance; if the implementation matches the printed formula, the HD95 results should be recomputed.","section":"Section III.D, Eq. (11)"}],"minor_comments":[{"comment":"Table 5 contains two rows with very similar labels for standard preprocessing, with values 43.12 and 39.98 for DSC; the text states that standard preprocessing reduces NA-UNETR from 45.64 to 43.12 and nnU-Net from 42.54 to 39.98, but the table does not clearly identify which model the second 'Standard Preprocessing Only' row refers to. The row labels should be corrected.","section":"Table 5"},{"comment":"The Gaussian noise term epsilon with variance sigma_n^2 added to the Hausdorff loss is never specified; the paper should report the exact variance used and ideally ablate this noise term separately from the homoscedastic weighting, since the loss design is presented as one of the contributions.","section":"Section II.E.3, Eq. (8)"},{"comment":"Equation (2) introduces a learnable relative positional bias b(i,j) that is absent from Equation (1); the authors should state whether standard NA also uses this bias or whether its introduction in DiNA is a deliberate difference.","section":"Section II.A, Eqs. (1)-(2)"},{"comment":"The sentence reporting Levene's test says a single statistic of 0.1619 with p=0.8511 was computed 'across three key attributes', but it is unclear how one test combines voxel intensity, artery size, and boundary complexity. This should be clarified or the sentence removed.","section":"Section II.C.1"},{"comment":"The Mann-Whitney results on LAD-SEG are reported only as p>0.05; actual p-values and a statement about multiple comparisons would allow readers to assess the evidence, and if no correction is used this should be acknowledged.","section":"Section IV.B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope and the ImageCAS portion is strong, but the LAD-SEG claim is currently supported only by non-significant differences on 20 single-annotator cases. If the authors cannot supply independent validation or multi-observer labels, I would expect the LAD-SEG superiority claim to be substantially softened and the paper to stand primarily on the ImageCAS results, the efficiency analysis, and the ablation study. The HD95 formula issue should be verified during revision because it directly affects one of the headline metrics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a competent engineering study, not a breakthrough. The genuinely new bit is the evaluation of a transformer-based LAD segmentation pipeline on non-contrast CT, which nobody has done before. The architecture is a 3D adaptation of Neighborhood Attention Transformer with a UNETR-style decoder, plus LoRA fine-tuning and an uncertainty-weighted Dice-Focal-Hausdorff loss. Each piece exists; the combination is new, and the paper gives it a fair test against seven baselines.\n\nThe ImageCAS results are solid. NA-UNETR gets 79.49% Dice, beating the next best by about 1.2 points, and the Mann-Whitney U test says the differences are significant. The ablations are careful and mostly confirm the design choices. Computational cost is reasonable. The paper is honestly written in the body, at least. The Discussion explicitly says the achieved Dice on LAD-SEG lies within expert variability.\n\nNow the soft spots. The central clinical claim, that NA-UNETR improves LAD delineation on non-contrast CT, is statistically weak. It rests on 20 scans with a single annotator. The 3.10 point Dice gain over nnU-Net (45.64 vs 42.54) is within the 0.10–0.53 inter-observer Dice range that the paper itself cites. Section IV.B reports p>0.05 for all model differences on LAD-SEG. So the ranking may just reflect which model happens to match the annotator's contouring style. The abstract does not mention this, which is a problem: it states the improvements without the caveat.\n\nSecond, the 5-fold validation was used both for hyperparameter selection (LoRA rank, NAT depth, kernel sizes) and for final performance reporting. That is optimistic bias. There is no independent test set.\n\nThird, the Gaussian noise term added to the Hausdorff loss is ad hoc. It is not ablated separately, just lumped into the boundary loss. Either justify it or drop it.\n\nMy overall take: this is a useful contribution to the medical image segmentation literature, especially the ImageCAS result and the careful comparisons. But the LAD-SEG claim needs to be reframed as provisional, with confidence intervals and a clear statement that the differences are not statistically significant. If the authors can get multi-observer labels or at least a separate validation set, that would help. As it stands, I would send it to peer review with major revisions requested. The paper is coherent, the experiments are described in enough detail to reproduce, and the authors are not overselling in the body. The abstract needs work.\n\nRecommendation: engage with it seriously, but do not let the LAD-SEG claim stand as-is.","headline":"A competent transformer segmentation study whose ImageCAS gains hold up, but the LAD-SEG claim is within label noise and not statistically significant.","tokens_in":18894,"tokens_out":2929,"would_cite":false,"duration_ms":24720,"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":"NA-UNETR, a 3D transformer built on Neighborhood Attention, is claimed to segment the left anterior descending artery from non-contrast CT more accurately than the compared CNN and transformer baselines.","keywords":["left anterior descending artery","neighborhood attention","coronary artery segmentation","non-contrast CT","radiotherapy cardiac sparing","transformer segmentation","uncertainty-weighted loss","LoRA fine-tuning"],"falsifier":"Ask two additional expert readers to re-contour the same 20 LAD-SEG scans, form a consensus mask, and score NA-UNETR versus nnU-Net against the second reader and against the consensus; the claim of superiority is falsified if the margin disappears or flips sign on either the alternative observer or the consensus mask.","tokens_in":17798,"feed_emoji":"🫀","tokens_out":8671,"duration_ms":74685,"temperature":0.7,"pith_summary":"This paper tries to establish that a 3D transformer segmentation model built on Neighborhood Attention can outline the left anterior descending artery (LAD) on non-contrast, free-breathing CT better than established CNN and transformer baselines. The proposed NA-UNETR is pretrained on 1,000 contrast-enhanced coronary CT angiography volumes and then fine-tuned on only 20 planning CT scans using low-rank adapters, with a Dice-Focal plus Hausdorff loss balanced by learned uncertainty. On the institutional LAD-SEG set the paper reports 45.64% Dice, 38.16 mm HD95, and 10.01 mm ASD, the best among the compared models, with gains of 3.10 Dice points over nnU-Net and a 2.96 mm HD95 reduction over Swin UNETR. If the claim holds, it offers a computationally modest path toward cardiac substructure contours for thoracic radiotherapy planning, where the LAD is rarely segmented because manual delineation is so uncertain. The paper itself notes that expert manual contours overlap by only 10 to 53 percent on this task, so the reported margin sits within label variability.","feed_headline":"Attention network beats U-Net variants at LAD artery segmentation","feed_subtitle":"On non-contrast CT it scores 45.6% Dice and beats 3D U-Net baselines on boundary distance.","key_machinery":"The load-bearing mechanism is the Neighborhood Attention (NA) block: attention in which each query attends only to keys and values in a local $k \\times k \\times k$ window, plus Dilated NA (DiNA), which samples that window at increasing dilation. Alternating NA and DiNA in NAT blocks inside a UNETR-style encoder gives a local spatial inductive bias and an expanding receptive field without the cost of global attention, and the paper's ablations tie the best result to this pairing together with residual convolutions, variable kernel sizes, and an uncertainty-weighted Dice-Focal plus Hausdorff loss that turns overlap and boundary objectives into one adaptive training signal.","core_discovery":"On its own terms, the discovery is that alternating Neighborhood Attention (NA) with Dilated Neighborhood Attention (DiNA) inside a UNETR-style transformer lets a segmentation model track a vessel that occupies a tiny fraction of the volume: local windows preserve the tubular contour, dilation expands the receptive field to follow the vessel course, and the combination with an uncertainty-weighted Dice-Focal plus Hausdorff loss produces the best overlap, centerline, and boundary scores among all compared models. The paper states this as an architecture-and-training result, not a claim that LAD segmentation is solved: it reports 45.64% Dice on non-contrast CT, notes that manual contours themselves range from 0.10 to 0.53 Dice, and explicitly says the approach is not yet suitable for clinical deployment.","pith_inferences":["Because manual LAD contours on non-contrast CT overlap by only 10 to 53 percent, the reported 3.10-point Dice advantage over nnU-Net could measure which observer's contour style the model learned rather than true anatomical accuracy; a multi-observer consensus test is the direct way to separate style matching from accuracy.","The same NA/DiNA alternating pattern should transfer to other thin, low-contrast tubular targets such as the esophagus, coronary veins, or small airways, and a controlled test across several such structures would show whether the mechanism or the preprocessing pipeline carries the gain.","Boundary errors on LAD-SEG remain large even for the best model, with HD95 above 38 mm, so a centerline-aware loss or topology-preserving postprocessing may be a higher-leverage next step than further architectural changes.","Because the model is pretrained on CTA and fine-tuned on non-contrast CT, explicit modality alignment such as intensity normalization or domain adaptation might close the remaining boundary gap without requiring more annotated LAD scans."],"forward_implications":["A radiotherapy planning workflow could obtain LAD contours from non-contrast CT in a single forward pass, since NA-UNETR runs at 1.33 seconds per volume with 4.17 GB peak GPU memory.","Pretraining on contrast CTA followed by LoRA fine-tuning becomes a transfer recipe for small-data cardiac substructure tasks: training on the 20 LAD scans alone drops Dice from 45.64% to 36.39%.","The dilated variant's higher centerline Dice (44.39% versus 43.45%) suggests the vessel trajectory is preserved better, which matters more than volumetric overlap alone for estimating dose to the artery.","The same architecture reaches 79.49% Dice on a high-contrast 1,000-case coronary benchmark with statistically significant gains, indicating that the local-global attention benefit is not limited to low-contrast CT.","The reported compute stays near Swin UNETR in parameters and FLOPs and below UNETR, so the accuracy gain is not bought with a substantially larger model."],"supporting_citations":[{"why":"Supplies the 1,000-volume coronary CTA dataset used for pretraining and for the direct public benchmark comparison.","marker":"40"},{"why":"Supplies the Neighborhood Attention and Dilated Neighborhood Attention blocks, plus the depth and kernel configuration pattern that NA-UNETR adapts.","marker":"38"},{"why":"Supplies the UNETR transformer encoder-decoder backbone on which NA-UNETR is built and one of the compared baselines.","marker":"39"},{"why":"Defines the nnU-Net baseline whose Dice and boundary results the central comparison claims to beat.","marker":"23"},{"why":"Defines the Swin UNETR baseline that NA-UNETR compares against on HD95 and supplies the uniform-depth ablation configuration.","marker":"48"},{"why":"Supplies the LoRA parameter-efficient fine-tuning used to adapt pretrained encoders to the 20-scan LAD dataset.","marker":"45"},{"why":"Supplies the homoscedastic uncertainty weighting that balances the Dice-Focal and Hausdorff loss terms.","marker":"41"},{"why":"Supplies the differentiable Hausdorff loss that drives the boundary accuracy claimed for NA-UNETR.","marker":"44"},{"why":"Supplies the cardiac contouring atlas evidence that manual LAD contours overlap by only 0.10 to 0.53 Dice, the paper's stated caveat about the target.","marker":"18"}],"fun_headline_variants":["Neighborhood attention transformer refines LAD artery segmentation","Dilated attention blocks lift LAD Dice over U-Net baselines","New transformer cuts boundary error in tiny cardiac artery CT","Uncertainty-weighted loss helps transformer segment LAD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison rests on treating one physician's LAD contour as ground truth, yet the paper cites inter-observer Dice of 0.10 to 0.53 on this imaging type, so the model's 3.10-point improvement over the runner-up is smaller than the disagreement between experts.","fun_headline_variants_meta":{"raw":{"variants":["Neighborhood attention transformer refines LAD artery segmentation","Dilated attention blocks lift LAD Dice over U-Net baselines","New transformer cuts boundary error in tiny cardiac artery CT","Uncertainty-weighted loss helps transformer segment LAD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1389,"prompt_tokens":1077,"completion_tokens":312,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":245}},"tokens_in":693,"tokens_out":312,"duration_ms":3272,"temperature":1.0,"reasoning_tokens":245,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:09:59.938458+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ask two additional expert readers to re-contour the same 20 LAD-SEG scans, form a consensus mask, and score NA-UNETR versus nnU-Net against the second reader and against the consensus; the claim of superiority is falsified if the margin disappears or flips sign on either the alternative observer or the consensus mask.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 1,000-volume coronary CTA dataset used for pretraining and for the direct public benchmark comparison."},{"cited_title":"Hassani, S","cited_arxiv_id":null,"evidence_quote":"Supplies the Neighborhood Attention and Dilated Neighborhood Attention blocks, plus the depth and kernel configuration pattern that NA-UNETR adapts."},{"cited_title":"Hatamizadeh, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the UNETR transformer encoder-decoder backbone on which NA-UNETR is built and one of the compared baselines."},{"cited_title":"Isensee, P","cited_arxiv_id":null,"evidence_quote":"Defines the nnU-Net baseline whose Dice and boundary results the central comparison claims to beat."},{"cited_title":"Hatamizadeh, V","cited_arxiv_id":null,"evidence_quote":"Defines the Swin UNETR baseline that NA-UNETR compares against on HD95 and supplies the uniform-depth ablation configuration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LoRA parameter-efficient fine-tuning used to adapt pretrained encoders to the 20-scan LAD dataset."},{"cited_title":"Kendall, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the homoscedastic uncertainty weighting that balances the Dice-Focal and Hausdorff loss terms."},{"cited_title":"Karimi and S","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable Hausdorff loss that drives the boundary accuracy claimed for NA-UNETR."},{"cited_title":"Duane et al., A cardiac contouring atlas for radiotherapy, Radiotherapy and Oncology 122 , 416--422 (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the cardiac contouring atlas evidence that manual LAD contours overlap by only 0.10 to 0.53 Dice, the paper's stated caveat about the target."}],"review_version":1}