{"id":"df6c2e24-180a-4e89-a5b2-5e2063e3476a","arxiv_id":"2608.12658","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Orthogonal seeding with three perpendicular annotated slices improves slice-propagation CT organ segmentation by 21.9% Dice over a single axial seed, with the gain due to seed geometry rather than annotation count or retraining.","lead":"This paper shows that using three perpendicular seed slices (axial, coronal, sagittal) at inference time, instead of a single axial slice, substantially improves 3D organ segmentation in CT volumes when combined with a simple distance-weighted fusion. A generalist reader should care because this is a cheap, training-free change that could improve label-efficient medical segmentation tools.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The budget-matched 3-axial control is confounded by the fusion rule: with sigma=30, the exact ground-truth chain near its seed is diluted to ~60% weight, so the 3-axial degradation (Dice≈0.29 near seed) may reflect fusion, not coplanarity.","rationale":"The reader's conditional verdict is well placed because the only experiment that separates seed orthogonality from annotation count, the 3-axial control, is executed with a fusion rule that systematically disadvantages coplanar seeds. With sigma=30 and seeds roughly 25-50 slices apart, the chain whose seed coincides with the current slice receives only about 60% weight after normalization, so the exact ground-truth mask is diluted by errors from the other two axial chains. This is exactly what Fig. 2 shows: per-slice Dice collapses to ~0.29 at the nearest-seed bin. The paper interprets this as evidence against coplanar seeds, but a nearest-seed fusion rule would eliminate this artifact; therefore the causal claim 'orthogonality rather than count' is not secure. A second, related gap is that the method section promises an equal-weight averaging ablation (ωc≡1) for the correspondence backbone to isolate the distance-weighting contribution, yet no such result appears in the results section. The missing ablation matters because the orthogonal gain could come substantially from the weighting scheme rather than from seed geometry per se. I considered the other weaknesses flagged by the reader, such as ground-truth seed selection and the missing axial-only TransMorph baseline, but these are explicitly disclosed in the paper and affect the generality rather than the internal validity of the main comparison. The proposed check, re-running the 3-axial control with nearest-seed fusion or a sigma sweep, would directly settle whether the 3-axial degradation is an artifact. Until that check is performed, the conditional verdict is appropriate; my stress-test does not move the verdict.","tokens_in":8131,"tokens_out":8585,"duration_ms":88163,"concrete_test":"Recompute the 3-axial row of Table 1 with the same three axial seeds but with per-slice nearest-seed fusion (equivalently, take the chain with largest ωc, or set sigma→0), and sweep sigma in {1,5,10,30} for both 3-axial and 3-orthogonal. If the 3-axial per-slice Dice at the seed-adjacent bin rises above 0.8 and the cohort-mean Dice exceeds the 1-axial baseline under any sigma/rule, the reported 3-axial degradation is a fusion artifact and the 'orthogonality rather than count' claim loses its main support. If 3-axial remains below 3-orthogonal under all fusion rules, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that gains come from seed orthogonality rather than annotation count rests entirely on the 3-axial control. That control is executed with the same distance-weighted fusion of Eq. 5, sigma=30, with all three seeds on the axial axis. At a voxel on one seed plane, the chain carrying the exact ground-truth mask receives weight 1, but the neighboring axial chains, whose seeds are ~25-50% of the organ extent away, still receive exp(-25/30)≈0.43 and exp(-50/30)≈0.19, diluting the exact mask to ~60% of the fused probability. This quantitatively explains the per-slice Dice drop to ~0.29 near the seed in Fig. 2: the fusion rule, not coplanarity, handicaps the 3-axial control. A nearest-seed or sigma→0 fusion would restore the ground-truth chain's dominance near each seed and likely make 3-axial improve over 1-axial, undermining the 'orthogonality rather than count' conclusion. The paper also states in Sec. 2 that an equal-weight averaging ablation (ωc≡1) is reported for the correspondence backbone, but no such result appears in Sec. 3; without it, the contribution of the distance-weighting to the orthogonal gain is untested. The concern is not that the method fails, but that the causal interpretation is not yet established by the presented comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an inference-time change to slice-propagation segmentation: instead of propagating from a single axial seed, propagate from three orthogonal seeds (axial, coronal, sagittal) and fuse the resulting soft maps with a label-free distance-weighted rule (Eq. 5). The authors evaluate this on three backbones (axial Sli2Vol, triplanar Sli2Vol, and TransMorph) across multi-organ CT cohorts, reporting large gains in Dice, NSD, and AHD over the single-axis baseline. They argue that the gains stem from seed orthogonality rather than annotation count, citing a budget-matched 3-axial control that provides no benefit and can degrade performance. Statistical significance is assessed with paired Wilcoxon tests on per-case Dice.","tokens_in":8417,"tokens_out":5123,"duration_ms":53254,"significance":"If the causal claim holds, the contribution is significant: a training-free, architecture-free inference modification yields large accuracy gains on surface-sensitive metrics, with public code and a controlled comparison across training regimes and inference seed configurations. The use of multiple backbones, an unseen organ (spleen), and paired nonparametric tests are strengths. However, the central interpretation—that orthogonality, not annotation count, drives the improvement—rests on a 3-axial control that is potentially confounded by the distance-weighted fusion rule. The paper also promises an equal-weight averaging ablation that is not reported in the results. These issues require additional experiments or a more cautious claim.","major_comments":[{"comment":"The text states that \"for the correspondence backbone we also report equal-weight averaging (ωc≡1) as an ablation isolating the distance weighting,\" but no such result appears in Section 3. This missing experiment is exactly what would isolate the contribution of distance weighting in Eq. (5) and is needed to interpret the 3-axial control. The authors should either add the equal-weight ablation results for all configurations or explicitly remove the promise from the Method section.","section":"Section 3, last paragraph of Method (Sec. 2)"}],"minor_comments":[{"comment":"The phrase \"analytical ∼!3× increase in forward passes\" appears to contain a typographical artifact (\"!3\"); please correct to \"3×\".","section":"Section 3, Implementation Details"},{"comment":"The paper mentions the largest Dice gain on pancreas (+28%) but does not provide per-organ results in a table or figure; please include per-organ breakdowns for all metrics.","section":"Section 3, Quantitative Results"},{"comment":"Paired Wilcoxon results are reported only for Dice; since the headline claims include NSD and AHD, report significance tests for those metrics as well.","section":"Section 3, Statistical tests"},{"comment":"Figure 2's caption identifies the backbone as triplanar Sli2Vol, but the text around it sometimes refers to the axial Sli2Vol backbone; please ensure consistent labeling of which backbone is shown.","section":"Fig. 2 and surrounding text"}],"recommendation":"major_revision","confidential_remarks":"I recommend requesting the additional 3-axial fusion control (nearest-seed or sigma→0) and the promised equal-weight averaging ablation before acceptance. If the corrected 3-axial control also improves over the single-axis baseline, the paper's central attribution to orthogonality rather than annotation count will need to be substantially revised."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is better than the average MICCAI-style empirical piece, and worse than its own abstract claims. The headline number is real: on abdominal CT, orthogonal seed propagation with the Sli2Vol correspondence backbone improves Dice by 21.9%, NSD by 25.5%, and cuts AHD by 53.5% over the standard single-seed axial baseline. The finding that an axially trained network applied to off-axis seeds captures nearly all the gain (explicit triplanar training adds little) is a genuinely new and useful decomposition. The study is well built: three backbones, public training/test splits with no patient overlap, spleen as a held-out structure, paired Wilcoxon tests on 342 cases, and code released. The limitations section is candid about the GT-based seed selection and the absent axial-only TransMorph baseline.\n\nThe soft spot is the central causal claim. The paper says the gains come from orthogonality, not annotation count, leaning entirely on the budget-matched 3-axial control. That control is executed with the same distance-weighted fusion, sigma=30. At a seed plane, the chain carrying the exact ground-truth label gets weight 1, but the other two axial chains are only 10-25 slices away, so they contribute weights of roughly 0.5-0.7. The fused probability at that voxel gives the exact mask at most ~50-60% weight. The per-slice Dice dip to ~0.29 near the seed plane in Fig. 2 is exactly what this dilution would produce. So the 3-axial control may be failing because of the fusion rule, not because coplanar seeds are uninformative. A nearest-seed fusion or smaller sigma could plausibly make 3-axial improve, which would collapse the 'orthogonality rather than count' conclusion. The paper does not test this.\n\nAlso, Sec. 2 promises an equal-weight averaging ablation (omega_c = 1) for the correspondence backbone, but no such result appears in Sec. 3. That is a missing experiment, not just a presentation slip. It matters because without it, the contribution of the distance-weighting rule to the orthogonal gain is untested.\n\nNone of this kills the paper. The practical intervention—three orthogonal seeds fused with distance weighting—works, is cheap, and is reproducible. But the manuscript currently overinterprets its own control. A serious referee should ask for the missing ablation, a 3-axial run with nearest-seed or sigma-to-0 fusion, and an explicit analysis of fused weight shares around the seed planes. With those, the orthogonality claim would either be established or honestly replaced with the weaker claim that orthogonal seeds plus this fusion rule work well.\n\nI'd send it out, expecting major revision.","headline":"Real, cheap, reproducible gains from orthogonal seed propagation, but the 'orthogonality, not annotation count' claim rests on a 3-axial control that the fusion rule itself handicaps.","tokens_in":8960,"tokens_out":5161,"would_cite":true,"duration_ms":50719,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that seed geometry—one axial, one coronal, one sagittal—rather than the registration model or training, drives slice-propagation segmentation quality, improving Dice by 21.9% and cutting surface distance by 53.5%.","keywords":["slice propagation","orthogonal seeds","distance-weighted fusion","label-free registration","self-supervised segmentation","CT organ segmentation","Sli2Vol","multi-planar fusion"],"falsifier":"Run the same three-axial experiment with a fusion rule that gives each voxel to the chain whose own seed is closest (or otherwise prioritizes the chain carrying the exact mask at its seed), and check whether 3-axial then matches 3-orthogonal. If it does, the near-seed dip in Fig. 2 is a fusion artifact and the claim that orthogonality rather than count drives the gain fails; if it does not, the coplanarity explanation survives. A simpler observation: at the shared seed plane in 3-axial, the axial chain's prediction should equal the ground truth by construction, so any large error there directly indicts the fusion rule.","tokens_in":7919,"feed_emoji":"📐","tokens_out":8630,"duration_ms":73249,"temperature":0.7,"pith_summary":"Slice-propagation methods label a 3D volume by propagating one annotated slice through neighbors, but their accuracy degrades with distance from the seed. This paper tests whether the bottleneck is the registration model's training or the geometric placement of seed slices, and concludes it is the seed geometry. Propagating from one axial, one coronal, and one sagittal seed, then fusing the three label chains with a distance-weighted average, improves Dice by 21.9%, Normalized Surface Dice by 25.5%, and reduces Average Hausdorff Distance by 53.5% over the single-axis baseline on abdominal CT. A budget-matched control with three axial seeds provides no benefit and can degrade performance, so the paper attributes the gain to orthogonality rather than the number of annotated slices. If correct, the result is a purely inference-time upgrade: two extra annotations, no retraining, and no architectural change.","feed_headline":"Three orthogonal seeds beat one axial seed in organ segmentation","feed_subtitle":"Swapping two axial annotations for coronal and sagittal slices lifts Dice by 21.9% and cuts surface distance by 53.5%.","key_machinery":"The load-bearing object is the orthogonal seed set combined with the distance-weighted fusion rule of Eq. (5). Each propagation chain $c$ (axial, coronal, or sagittal) contributes at voxel $v$ with weight $\\omega_c(v)=\\exp(-|[v]_{a(c)}-s_c|/\\sigma)$, where $a(c)$ is the chain's axis and $s_c$ its seed position, and the fused soft prediction is the weighted average over chains, thresholded at 0.5. The weights depend only on seed geometry, not on the network's predictions, so the fusion needs no labeled tuning data; the paper uses $\\sigma=30$ slices. Orthogonal seeds work because every voxel is within a short effective distance of at least one seed along its own axis, shortening the propagation chains and reducing the error accumulation that the paper documents for single-axis propagation.","core_discovery":"The central claim is that, for slice-propagation segmentation, inference-time seed geometry—not the registration backbone or its self-supervised training—determines segmentation quality. The evidence is the consistent ordering across backbones: one axial seed and three coplanar axial seeds are both beaten by three orthogonal seeds, with the axial Sli2Vol backbone improving from Dice 0.595 to 0.725, NSD@2mm from 0.469 to 0.589, and AHD from 7.28 mm to 3.39 mm. The paper reads the failure of the budget-matched three-axial control as proof that the gain is orthogonality, not annotation count, and the per-slice curves in Fig. 2 as showing that orthogonal seeds keep both Dice and surface distance flat across the organ where single-axis and coplanar seeds decay or dip.","pith_inferences":["If the orthogonality result holds beyond abdominal CT, the same inference-only recipe—adding one coronal and one sagittal seed to any single-axis slice-propagation pipeline—could transfer to MRI or other modalities, provided the backbone's slice correspondence remains reliable off-axis.","The paper's own assumption that targets are compact with comparable extent across all three axes predicts a testable boundary: for thin, elongated, or branching structures such as vessels or bronchi, the orthogonal-seeding gain should shrink, and seed count might then matter more than orthogonality.","The near-seed dip of the three-axial control in Fig. 2 suggests an untested explanation: the distance-weighted fusion rule may be the culprit. A nearest-seed-priority fusion could close much of the gap, which would reframe the contribution as fusion design rather than seed geometry."],"forward_implications":["Switching from one axial seed to three orthogonal seeds improves Dice by 21.9%, NSD@2mm by 25.5%, and reduces AHD by 53.5% on the axial Sli2Vol backbone over the single-axis baseline.","Explicitly training the correspondence network on three axes adds little once orthogonal seeds are used; the axially trained network applied to off-axis seeds captures nearly all the gain.","A budget-matched three-axial seed control does not beat the single axial seed and can degrade performance, so annotation count alone is not the driver.","The pattern is consistent across correspondence and deformable backbones for Dice and AHD; the deformable backbone's lack of a soft prediction weakens its NSD gain under hard voting, which the paper attributes to fusion rather than to orthogonal seeding itself.","The improvement requires only two additional annotated slices and no architectural changes or retraining, making it a plug-and-play inference strategy for slice-propagation segmentation."],"supporting_citations":[{"why":"Supplies the Sli2Vol correspondence-based slice-propagation backbone and the axial-only baseline all gains are measured against.","marker":"[22]"},{"why":"Supplies the label-free slice-reconstruction meta-training task that trains the correspondence backbone.","marker":"[21]"},{"why":"Supplies the deformable TransMorph backbone used as the cross-paradigm comparison.","marker":"[7]"},{"why":"Supplies the distance-aware, spatially varying weighting idea on which the fusion rule is based.","marker":"[1]"},{"why":"Documents how slice-propagation error accumulates with distance from the seed, motivating orthogonal seeds.","marker":"[14]"},{"why":"Prior evidence that orthogonal annotations benefit barely-supervised segmentation; the paper extends this to inference-time seeds.","marker":"[6]"},{"why":"Supplies the Medical Segmentation Decathlon evaluation and training cohorts for pancreas and spleen.","marker":"[19]"}],"fun_headline_variants":["Orthogonal seeds beat axial-only seeds in 3D organ segmentation","Seed geometry, not training, drives slice-propagation segmentation gains","Three orthogonal slices lift organ segmentation Dice by 21.9%","Inference-time seed orthogonality boosts 3D organ segmentation","Why orthogonal seeds outperform more axial slices in 3D segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion depends on the three-axial control being a fair test: if the control's poor performance near the shared seed plane comes from the fusion rule washing out the chain that holds the exact annotation, rather than from the seeds lying in one plane, then the claim that orthogonality rather than seed count drives the gain is not established.","fun_headline_variants_meta":{"raw":{"variants":["Orthogonal seeds beat axial-only seeds in 3D organ segmentation","Seed geometry, not training, drives slice-propagation segmentation gains","Three orthogonal slices lift organ segmentation Dice by 21.9%","Inference-time seed orthogonality boosts 3D organ segmentation","Why orthogonal seeds outperform more axial slices in 3D segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1291,"prompt_tokens":1009,"completion_tokens":282,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":193}},"tokens_in":625,"tokens_out":282,"duration_ms":2805,"temperature":1.0,"reasoning_tokens":193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:02:11.855231+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same three-axial experiment with a fusion rule that gives each voxel to the chain whose own seed is closest (or otherwise prioritizes the chain carrying the exact mask at its seed), and check whether 3-axial then matches 3-orthogonal. If it does, the near-seed dip in Fig. 2 is a fusion artifact and the claim that orthogonality rather than count drives the gain fails; if it does not, the coplanarity explanation survives. A simpler observation: at the shared seed plane in 3-axial, the axial chain's prediction should equal the ground truth by construction, so any large error there directly indicts the fusion rule.","supporting_citations":[{"cited_title":"In: Medical Image Computing and Com- puter Assisted Intervention (MICCAI)","cited_arxiv_id":null,"evidence_quote":"Supplies the Sli2Vol correspondence-based slice-propagation backbone and the axial-only baseline all gains are measured against."},{"cited_title":"In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Supplies the label-free slice-reconstruction meta-training task that trains the correspondence backbone."},{"cited_title":"IEEE Transactions on Medical Imaging28(8), 1266–1277 (2009)","cited_arxiv_id":null,"evidence_quote":"Supplies the distance-aware, spatially varying weighting idea on which the fusion rule is based."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Documents how slice-propagation error accumulates with distance from the seed, motivating orthogonal seeds."},{"cited_title":"In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Prior evidence that orthogonal annotations benefit barely-supervised segmentation; the paper extends this to inference-time seeds."}],"review_version":1}