{"id":"c153aef0-5063-4038-9191-b5fbc08de133","arxiv_id":"2508.03997","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Slice-block shuffle plus confidence-guided displacement reports 72.67% average Dice on Synapse with 20% labels and 63.99% on AMOS with 5% labels.","lead":"JanusNet is a data augmentation method for semi-supervised 3D organ segmentation: it shuffles same-index slice blocks across patient volumes, then swaps low-confidence blocks for confident ones. The authors report state-of-the-art Dice on Synapse and AMOS, but the preprint lacks code, error bars, and a full method specification.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never tests whether same-index slice-block alignment across patients is causally responsible for the reported gains; without a random-swap control, the central contribution is underdetermined.","rationale":"The reader's weakest_assumption—that same-index slices across patients correspond anatomically—captures a real risk, but the more precise and testable issue is that the paper never compares aligned to unaligned shuffling. Without that control, the mechanism is undetermined. This is load-bearing because if the same-index alignment is not the driver of the +2.84% SBS gain, the paper's theoretical narrative collapses and the contribution reduces to a minor variant of existing block-mixing augmentations (MagicNet, BCP). The concern is not that the alignment is definitely wrong; it is that the evidence does not discriminate. This is especially pertinent because across abdominal CT volumes, raw axial indices are unregistered, so the alignment may be noisy; the paper provides no preprocessing details (resampling/padding) to assess it. I also note the manuscript omits Sections 3.2–3.3 equations and Table 1, which further limits auditability, but the alignment ablation is the single most concrete way to test the core mechanism. A random-index shuffle control is cheap and decisive. I therefore agree with the reader's conditional verdict; the concern reinforces it without changing it.","tokens_in":10748,"tokens_out":8015,"duration_ms":92595,"concrete_test":"Run the ablation on 20% labeled Synapse with the SBS step modified to swap block i from volume A with block j (j ≠ i) from volume B (random index), keeping block size p=16 and all other settings identical. Compare average DSC to the aligned same-index version (Aug+SBS row: 72.51±0.54 vs full model 72.67±1.18). If the random-index shuffle achieves DSC within ~1 point of the aligned version, alignment is not load-bearing and the anatomical-continuity story is unsupported; if it drops >2 points, the alignment matters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that aligned same-index slice-block shuffling across volumes improves semi-supervised 3D segmentation by preserving anatomical context (Sec. 3.1, Fig. 1) is supported only by full-system and component ablations (Tables 2-3). No experiment isolates the alignment mechanism. The natural control is to shuffle blocks of the same thickness p across volumes at random indices (or with a random offset) instead of same-index blocks. Without this control, the +2.84% SBS gain over the weak-strong baseline (Table 3) could come from generic block mixing (diversity, local coherence) rather than from the assumed anatomical correspondence between same-index slices. In abdominal CT, axial indices are not registered across patients (varying body height, scan range, table position); if alignment is effectively noisy, the 'aligned' terminology may be misleading and the method reduces to a variant of CutMix/MixUp. Moreover, the paper does not specify how volumes with differing depth are resampled/padded to form same-index blocks, which is essential to assess the alignment. Thus the unique mechanism is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes JanusNet, a data-augmentation framework for semi-supervised 3D multi-organ segmentation. It introduces two stages built on a teacher–student setup: Slice-Block Shuffle (SBS), which mixes blocks of slices at the same index across volumes along a randomly chosen axis, and Confidence-Guided Displacement (CGD), which replaces low-confidence blocks with high-confidence counterparts within a slice. The claimed contribution is that this axis-aligned, slice-block-level perturbation preserves anatomical continuity better than cube-level mixing, improving learning on small and hard-to-segment organs. Experiments on Synapse (20% labeled) and AMOS (5% labeled) report state-of-the-art average Dice scores, with component ablations on Synapse showing stepwise gains for SBS and CGD over a weak–strong augmentation baseline.","tokens_in":10978,"tokens_out":3728,"duration_ms":46552,"significance":"If the reported results hold, JanusNet would be a simple, plug-and-play augmentation that improves several semi-supervised 3D segmentation baselines, which is practically valuable. The paper's component ablations (Table 3) show consistent positive contributions from both proposed steps, and the gains on small organs such as the stomach and pancreas are directionally encouraging. The method is also agnostic to the backbone and thus potentially easy to adopt. However, the central claim rests on the anatomical alignment being beneficial; the manuscript does not isolate this mechanism, and the main quantitative comparisons lack statistical support. These gaps currently limit the strength of the conclusions and the confidence one can place in the reported SOTA numbers.","major_comments":[{"comment":"The central mechanism of Slice-Block Shuffle is that aligned same-index slice blocks across volumes preserve anatomical context. However, the paper never tests whether this alignment, rather than generic block mixing, causes the +2.84% gain over the weak-strong baseline (Table 3). A control experiment mixing blocks at random indices or random offsets is necessary; otherwise the method could be another variant of CutMix/MixUp with no unique anatomical contribution. Additionally, the paper does not state how volumes with differing depth, spacing, or scan range are resampled or padded to form same-index blocks, which is essential to interpret the alignment claim.","section":"§3.1, Fig. 1, Tables 3"},{"comment":"The main comparisons against prior methods report only point estimates of Avg. Dice and ASD, with no standard deviations or significance tests. For example, JanusNet's 63.99 vs. GA-MagicNet's 63.51 on AMOS (Table 2) is a difference of 0.48 percentage points, which may be within run-to-run variability. The abstract's phrase 'significantly surpasses' is therefore not supported by the evidence as presented. The authors should report means and standard deviations over multiple runs (as is done for ablations in Table 3) and, where appropriate, significance tests or at least paired per-class comparisons.","section":"Tables 1 and 2"},{"comment":"The default hyperparameters p=16, lambda_disp=0.25, and Top-K=2 are selected by ablations on the same Synapse 20% labeled benchmark that is used for the headline SOTA comparison (Table 1). Such test-set-based selection can overstate gains. The authors should either use a separate validation split for hyperparameter selection, report results for a sensible range of hyperparameters, or provide a nested evaluation to demonstrate that the chosen settings generalize. Without this, the reported Synapse gain may be optimistic.","section":"Tables 3–6 versus Table 1"}],"minor_comments":[{"comment":"The abstract describes the setting as 'weakly supervised' while the paper is semi-supervised; please use consistent terminology throughout.","section":"Abstract"},{"comment":"The number of runs used to compute mean ± std is not stated; please specify the number of independent repetitions.","section":"Tables 4–6"},{"comment":"The text cites 'MagicNet(Qi, Wu, and Chan 2024)' but the reference list entry is incomplete; also check that all cited works (e.g., Wang and Li 2023, Wu et al. 2022) appear in the references.","section":"References"},{"comment":"The figures are difficult to read in the submitted version, especially the small inset labels in Fig. 1 and the data flow annotations in Fig. 2. Please provide higher-resolution versions and enlarge key sub-captions.","section":"Figures 1–2"},{"comment":"The method description for the two proposed steps is sparse; pseudocode or a more detailed algorithmic description would improve reproducibility, particularly for CGD's Top-K selection and the exact formation of displaced targets.","section":"§3.2/3.3"}],"recommendation":"major_revision","confidential_remarks":"The concern raised in the stress-test note about the missing random-mixing control is well founded. Before this paper is accepted, I would want to see a direct comparison between aligned same-index shuffling and random-index/offset shuffling, plus a clear explanation of how volumes of different depths are matched. The lack of standard deviations in the main tables is also a substantive issue for a claim of 'significant' improvement. These are fixable within the manuscript's scope, but they affect the validity of the central contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a real, practical idea: instead of mixing whole cubes, JanusNet shuffles same-index slice blocks along a random axis, preserving orthogonal-plane anatomy, and then displaces hard blocks within a slice using pseudo-label confidence. That is a sensible extension of BCP/MagicNet, not a restatement. The ablations on Synapse 20% show consistent gains (SBS alone +2.84 DSC, CGD alone +2.15, full system 72.67 vs 64.34 baseline) and report standard deviations. On AMOS 5%, it edges GA-MagicNet by +0.48 Avg Dice with visible gains on esophagus, stomach, and duodenum. The method being plug-and-play is a genuine plus.\n\nThe soft spots, in order of severity. First, the core assumption—same-index layers are anatomically aligned across patients—is never tested. There is no random-swap or random-offset control. I agree with the stress-test point: the gains could come from generic block mixing (local coherence, diversity) rather than from the claimed anatomical pairing. The paper also never states how volumes of differing depth are padded or resampled before same-index block formation. Second, the version I read omits Sections 3.2–3.3 (equations) and Table 1 (the main Synapse comparison), so a full soundness check is impossible. Third, hyperparameters p=16, lambda_disp=0.25, Top-K=2 are chosen by ablation on the same Synapse 20% setting used for the headline number—that is borderline tuning on the test set. No code or data are released. Fourth, the AMOS margin of +0.48 is within run-to-run noise; calling that 'significant' without error bars is a stretch.\n\nThat said, the core idea is coherent and the ablation pattern is consistent. This is not a crank paper; it is a testable augmentation proposal with a plausible mechanism. It deserves serious refereeing, but the authors should be required to provide the missing sections, a random-alignment control, error bars on the main comparisons, and a code release.\n\nWho this is for: researchers working on semi-supervised 3D medical segmentation, especially those using teacher-student or mixing-based augmentations. My recommendation: send to peer review, major revision.","headline":"A plausible, practical augmentation that deserves a serious referee, but the central alignment claim is untested and the manuscript as provided is incomplete.","tokens_in":11498,"tokens_out":2643,"would_cite":false,"duration_ms":32338,"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":"The paper claims that replacing cube-level mixing with an anatomy-preserving, same-index slice-block shuffle—plus a confidence-guided in-slice displacement—raises semi-supervised 3D multi-organ segmentation, e.g., +4% Dice on Synapse with 2","keywords":["semi-supervised learning","3D medical image segmentation","multi-organ segmentation","data augmentation","slice-block shuffle","confidence-guided displacement","teacher-student","anatomical continuity"],"falsifier":"Run the same 20%-labeled Synapse protocol on a multi-center CT dataset where axial slice indices are known to vary across patients (e.g., different start positions). If JanusNet's average Dice drops to or below GA-MagicNet's, the index-alignment assumption is the load-bearing part. A direct control is to repeat JanusNet with index alignment replaced by random slice-block mixing: the gain should vanish.","tokens_in":10608,"feed_emoji":"🩻","tokens_out":8365,"duration_ms":82174,"temperature":0.7,"pith_summary":"Semi-supervised 3D multi-organ segmentation needs to exploit unlabeled CT volumes, and data augmentation is a key lever. JanusNet claims that the common trick of mixing random 3D cubes throws away exactly the information that matters—the smooth variation of organ shape and position along a volume's axes. So it shuffles aligned slice-blocks across labeled and unlabeled volumes along one randomly chosen axis, preserving in-plane anatomy, and then uses a teacher's confidence to move reliable patches into unreliable spots, focusing the model on hard regions. On the Synapse benchmark at 20% label ratio this produces 72.67% average Dice, about 4 points above the strongest prior method; on AMOS at 5% it produces 63.99%.","feed_headline":"Slice-shuffle augmentation adds 4% Dice in semi-supervised CT","feed_subtitle":"Aligned slice-block mixing preserves organ anatomy and beats cube-based mixing on Synapse and AMOS.","key_machinery":"Slice-Block Shuffle (SBS): partition each volume along a randomly chosen orthogonal axis and shuffle same-index slice blocks across a batch's labeled and unlabeled volumes, so the orthogonal planes retain anatomical continuity. Confidence-Guided Displacement (CGD): within each produced slice, replace low-reliability blocks (from the teacher's softmax confidence) with high-reliability blocks to emphasize hard-to-segment regions. Together they form a two-stage, layer-aware augmentation inside a teacher-student loop with an EMA teacher, pseudo-labels, and a displacement loss.","core_discovery":"The paper's central claim is that axis-aligned, slice-block-level mixing is a better perturbation for semi-supervised 3D multi-organ segmentation than cube-level mixing. JanusNet picks one of the three volume axes, splits labeled and unlabeled volumes into blocks along that axis, and shuffles same-index blocks across volumes, preserving the anatomical layout on the planes perpendicular to the chosen axis. A second step, Confidence-Guided Displacement, replaces low-confidence regions inside each slice with high-confidence counterparts from the same slice, using the teacher's predictions to home in on hard organs. In the reported experiments, the two-stage augmentation lifts average Dice on 20","pith_inferences":["The method's reliance on same-index slice correspondence is the main untested risk. In multi-center CT, axial slice index does not necessarily align to anatomy, so an obvious next experiment is to test JanusNet on scans with different table positions or to prepend a simple anatomical-level normalization; the paper does not do this.","Because the augmentation only needs intensity volumes and pseudo-labels, it should transfer to 3D MRI and other modalities where slice-index correspondence is often more reliable, though the authors only evaluate on CT.","The success of axis-aligned mixing hints that the anatomical prior is what matters more than the specific mixing mechanism. A cleaner test would compare JanusNet against cube mixing under identical compute and schedule, isolating alignment rather than block size or loss weights as the source of the gain."],"forward_implications":["If the reported comparisons hold, JanusNet's two augmentation steps can be dropped into existing teacher-student semi-supervised segmentation frameworks with minimal code changes and still yield consistent gains.","The per-class results indicate that the method helps most where cube mixing hurts most: small or elongated organs. On Synapse, stomach, pancreas, and spleen improve by 12.3, 11.6, and 5.4 points; on AMOS, esophagus, stomach, and duodenum improve by 9.2, 7.4, and 6.3 points.","The ablations show a well-defined optimum in perturbation strength (block thickness p=16, displacement weight 0.25, Top-K=2): too little mixing leaves hard regions untouched, too much erodes the anatomical prior and degrades Dice, giving practical guidance for applying the method to new datasets.","Because the gains are measured on both 20%-labeled Synapse and 5%-labeled AMOS, the method appears to stay effective under severe annotation scarcity, the regime where semi-supervised segmentation is most needed."],"supporting_citations":[{"why":"MagicNet, the cube-level partition-mix-recover baseline that JanusNet explicitly improves on; supplies the strongest prior mixing strategy.","marker":"(Chen et al. 2023)"},{"why":"GA-MagicNet, the gradient-aware variant that serves as the main comparison baseline in the 20% Synapse and 5% AMOS experiments.","marker":"(Qi, Wu, and Chan 2024)"},{"why":"BCP, the bidirectional copy-paste method that sets the distribution-mismatch baseline for the paper's data-level augmentation argument.","marker":"(Bai et al. 2023)"},{"why":"UA-MT, the teacher-student uncertainty-aware backbone on which the JanusNet teacher-student pipeline is built.","marker":"(Yu et al. 2019)"},{"why":"FixMatch, the source of the weak/strong-augmentation and confidence-thresholding scheme that JanusNet augments.","marker":"(Sohn et al. 2020)"},{"why":"Mean Teacher, which supplies the EMA teacher update used by all teacher-student baselines and JanusNet itself.","marker":"(Tarvainen and Valpola 2017)"}],"fun_headline_variants":["Slice-block shuffle yields 4% Dice gain in semi-supervised 3D","JanusNet: axis-aligned mixing for semi-supervised multi-organ segmentation","Preserve anatomy with slice-block shuffle: 4% DSC gain on 20% labels","Smart block mixing boosts 3D organ segmentation with scarce labels","Anatomy-aware shuffle improves semi-supervised segmentation by 4%"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The strongest assumption is that the same slice index across different patient volumes refers to comparable anatomy, so shuffling same-index slice blocks mixes consistent content; unregistered CT scans with shifted table positions could break this.","fun_headline_variants_meta":{"raw":{"variants":["Slice-block shuffle yields 4% Dice gain in semi-supervised 3D","JanusNet: axis-aligned mixing for semi-supervised multi-organ segmentation","Preserve anatomy with slice-block shuffle: 4% DSC gain on 20% labels","Smart block mixing boosts 3D organ segmentation with scarce labels","Anatomy-aware shuffle improves semi-supervised segmentation by 4%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000277,"raw_usage":{"total_tokens":1491,"prompt_tokens":755,"completion_tokens":736,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":636}},"tokens_in":499,"tokens_out":736,"duration_ms":9366,"temperature":1.0,"reasoning_tokens":636,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:57:49.240908+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 20%-labeled Synapse protocol on a multi-center CT dataset where axial slice indices are known to vary across patients (e.g., different start positions). If JanusNet's average Dice drops to or below GA-MagicNet's, the index-alignment assumption is the load-bearing part. A direct control is to repeat JanusNet with index alignment replaced by random slice-block mixing: the gain should vanish.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MagicNet, the cube-level partition-mix-recover baseline that JanusNet explicitly improves on; supplies the strongest prior mixing strategy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BCP, the bidirectional copy-paste method that sets the distribution-mismatch baseline for the paper's data-level augmentation argument."},{"cited_title":"A.; Cubuk, E","cited_arxiv_id":null,"evidence_quote":"FixMatch, the source of the weak/strong-augmentation and confidence-thresholding scheme that JanusNet augments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mean Teacher, which supplies the EMA teacher update used by all teacher-student baselines and JanusNet itself."}],"review_version":1}