{"id":"02d0a868-0268-40ed-9cf6-c991492b67da","arxiv_id":"2412.11458","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HResFormer fuses 2D slice-level and 3D volume-level transformer features and adds the 2D prediction to the 3D output, reporting state-of-the-art Dice scores on Synapse, BraTS, and ACDC.","lead":"This paper introduces HResFormer, a network that combines 2D and 3D transformers to segment organs and tumors in CT and MRI scans. It reports higher accuracy than prior methods on three public medical imaging benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparisons are adopted from prior papers instead of rerun under HResFormer's exact protocol, so the reported SOTA margins may be protocol artifacts; the identical Table VI 'only 3D' row and the false 'all 13 organs' claim reinforce this concern.","rationale":"I read the paper in good faith. The architecture is coherent and the ablation is internally plausible except for the duplicated row. The strongest claim is empirical SOTA; the load-bearing condition is that the comparison protocols are matched. The paper's own footnotes admit baselines are adopted, and the authors do not provide code or rerun baselines, so this condition is not established. The Table VI duplication and Table II overclaim are independent red flags but secondary; they reinforce the need for controlled reruns rather than overturning the architecture on their own. Credit where due: the method section is detailed enough to reimplement, the HLGM and residual designs are clearly specified, and the lightweight variant gives a concrete efficiency check. My recommendation is unchanged: CONDITIONAL, pending the controlled comparison.","tokens_in":17707,"tokens_out":7602,"duration_ms":67209,"concrete_test":"Release the HResFormer code (or a faithful reimplementation from Sec. III) and run nnFormer and nnU-Net on the Synapse-8 split using exactly the Sec. IV.B protocol: same preprocessing/augmentation, 1000-epoch SGD with poly decay, same loss, same 18/12 cases. Recompute Table I. If nnFormer or nnU-Net rises to within ~0.5 DSC of 89.10, the claimed 2.5/2.1 margins are protocol artifacts rather than architectural gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that HResFormer is a new SOTA for volumetric segmentation—depends on cross-paper baseline numbers rather than controlled re-runs. Table I states 'The results are adopted from nnFormer' and Table III adopts baselines from Hatamizadeh et al.; Table II lists only CoTR and nnUnet with no indication that they were rerun under Sec. IV.B's pipeline (1000 epochs, SGD/poly, specific augmentations, ImageNet-pretrained 2D encoder). If those baselines used weaker protocols, the reported margins (2.5/2.1 DSC on Synapse, 2.1/1.8 on BraTS, 0.52 on ACDC) overstate the architecture's benefit. This concern is sharpened by internal evidence: in Table VI, the 'HResFormer (only 3D)' row is numerically identical to the nnFormer row of Table I across every organ (e.g., Ao 92.04, Gb 70.17, Ki(L) 86.57, ... average 86.83), which is implausible for an independently trained ablation and suggests the 3D branch is either literally nnFormer or the row was copied; either way it weakens the ablation as evidence about HResFormer's own components. Separately, Sec. IV.C claims HResFormer 'consistently outperforms nnUnet on all 13 organs,' but Table II shows lower DSC for Sp (96.40 vs 96.64), Ki(R) (94.58 vs 94.74), and Li (96.85 vs 97.05). This false claim does not refute the average gain, but it makes the evaluation/reporting unreliable enough that the SOTA assertion cannot be accepted without a controlled comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HResFormer, a hybrid 2D+3D Transformer for volumetric medical image segmentation. The 2D branch processes axial slices to capture fine-grained inner-slice information, and the 3D branch processes volumetric data with a Hybrid Local-Global fusion Module (HLGM) that mutually fuses the two streams. Residual learning adds the 2D predictions to the 3D output. The method is evaluated on Synapse multi-organ CT, BraTS brain tumor MRI, and ACDC cardiac MRI, reporting state-of-the-art DSC compared to published Transformer and CNN baselines.","tokens_in":18064,"tokens_out":7673,"duration_ms":54142,"significance":"The paper introduces a novel hybrid architecture that combines a 2D Transformer for fine-grained axial-slice features with a 3D Transformer for volumetric context, using a mutual fusion module (HLGM) and residual prediction addition. If the reported results are reproducible under controlled conditions, this is a valuable contribution with practical relevance, as it directly addresses a known weakness of 2D and 3D segmentation models. The manuscript provides extensive experiments on three benchmarks, includes parameter and FLOPs comparisons, and reports a lightweight variant with a favorable accuracy/cost trade-off; the authors also commit to releasing code. However, the evidence for the headline SOTA claim is weakened by the reliance on baselines adopted from other papers, by an internally inconsistent ablation row in Table VI, and by a false per-organ claim in Section IV.C. These issues need to be resolved before the results can be fully credited.","major_comments":[{"comment":"The state-of-the-art comparisons in Tables I and III are adopted from prior publications (nnFormer for Table I, Hatamizadeh et al. for Table III) rather than rerun under the training protocol described in Section IV.B (1000 epochs, SGD with poly decay, specific augmentations, ImageNet-pretrained 2D encoder). Because HResFormer is trained under this specific protocol, the reported margins (e.g., 2.5% DSC over nnFormer on Synapse) may partly reflect differences in training schedules, data preprocessing, and hyperparameters rather than architectural superiority. To support the central SOTA claim, the authors should either rerun at least nnFormer and nnU-Net under the identical protocol (same data splits, preprocessing, optimizer, and training schedule) or substantially temper the claim and clearly state that all baselines are taken from their original papers with potentially different protocols.","section":"Section IV.C, Tables I and III"},{"comment":"The row labeled 'HResFormer (only 3D)' in Table VI reports per-organ DSC values identical to the nnFormer row in Table I (Ao 92.04, Gb 70.17, Ki(L) 86.57, Ki(R) 86.25, Li 96.84, Pa 83.35, Sp 90.51, St 86.83), yet it lists an average of 85.39 instead of the correct average 86.57. This is internally inconsistent and suggests the row was copied from the nnFormer results rather than being an independently trained HResFormer 3D branch. This undermines the ablation discussion in Section IV.D, where the authors argue that their 3D model alone is comparable to nnFormer and that the hybrid design is responsible for the gains.","section":"Table VI, Section IV.D"},{"comment":"The text states that 'our HResFormer consistently outperforms nnUnet on all 13 organs', but Table II shows that nnUnet achieves higher DSC than HResFormer on Spleen (96.64 vs 96.40), right kidney (94.74 vs 94.58), and liver (97.05 vs 96.85). This factual error must be corrected; the average gain over nnUnet is still positive, but the claim of universal per-organ superiority is false.","section":"Section IV.C, Table II"},{"comment":"The ablation study would benefit from variance estimates. Several differences used to support the design choices are small (e.g., 0.38 DSC between Ours-FF and Ours-DFF in Table VII; 0.98 DSC between the full model and the model without HLGM in Table VI), and no multiple-seed runs, confidence intervals, or significance tests are reported. Without these, it is difficult to determine whether the improvements are systematic or within run-to-run noise, and the ablation conclusions are not fully supported.","section":"Section IV.D, Tables VI and VII"}],"minor_comments":[{"comment":"In the sentence 'With residual learning, we can see that the performance consistently improves from 88.21% and 89.10%', the word 'and' should be 'to'.","section":"Section IV.D"},{"comment":"The module name is given as HLGM, but it appears as 'HGLM' in the configuration column of Table VI and in the ablation text of Section IV.D; please use a consistent acronym.","section":"Abstract/Throughout"},{"comment":"In the third contribution bullet, 'Our method consistently previous outperforms' should read 'Our method consistently outperforms previous'.","section":"Section I"},{"comment":"The typo 'HRFomrer' appears in the comparison text on the Synapse results; it should be 'HResFormer'.","section":"Section IV.C"},{"comment":"The equation P'_3d,i = P_2d,0 + P_3d,i contains a stray 'S2d' after the equation that should be removed.","section":"Section III.D, Equation (10)"},{"comment":"The paper claims to be 'the first hybrid Transformer' for 3D medical image segmentation, but PHTrans (Liu et al., 2022a), discussed in the Related Work, is also a hybrid architecture (parallel Transformer and CNN). Please qualify the novelty claim, e.g., as the first hybrid 2D+3D Transformer.","section":"Section I / Related Work"},{"comment":"The conclusion lists 'a) Limitations.' and 'b) Future work.' as sub-items; this formatting is unconventional and should be integrated into the conclusion or moved to a separate section.","section":"Section V"},{"comment":"The reported average DSC for HResFormer (87.46) differs from the arithmetic mean of the listed per-organ values (approximately 87.61). Please check whether the per-organ values or the average are rounded and clarify.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"I concur with the conditional assessment: the architecture is promising, but the paper's internal inconsistencies and the lack of controlled baseline comparisons are substantial. I recommend requiring the authors to (a) provide a corrected Table VI or a genuine ablation of the 3D model, (b) correct the false claim in Section IV.C, and (c) either rerun key baselines or restrict the SOTA claim. A replication package would strengthen a revised submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper proposes HResFormer, a hybrid of a 2D pyramid transformer and a 3D shifted-window transformer for volumetric medical image segmentation. The 2D branch produces slice-wise predictions, the HLGM module fuses those predictions with the 3D volume via mutual local-global cross-attention, and the 3D branch's output is added to the 2D predictions as a residual. That combination is genuinely new relative to the cited work; PHTrans is a CNN-transformer hybrid and 2.5D methods don't do this. The method section is clear and the idea is plausible.\n\nWhat it does well: the ablation broadly supports the two claimed components. Adding HLGM over naive fusion improves DSC from 87.23 to 88.21, and residual learning adds another 0.89. The lightweight version cuts FLOPs roughly in half and still beats nnFormer. These are meaningful signals.\n\nThe soft spots are real but not fatal. The baselines in Tables I and III are adopted from prior papers rather than rerun under HResFormer's exact protocol. That's common in this field, but it does leave the SOTA margins unverified, especially with no error bars. More concerning are the internal inconsistencies. Section IV.C claims HResFormer 'consistently outperforms nnUnet on all 13 organs,' but Table II shows nnUnet higher on spleen (96.64 vs 96.40), right kidney (94.74 vs 94.58), and liver (97.05 vs 96.85). That's just false. And in Table VI, the 'HResFormer (only 3D)' row lists per-organ values identical to nnFormer's from Table I, yet reports an average of 85.39 when those numbers average to 86.57. Either the row was copied or the average is miscalculated; either way it breaks the ablation as evidence. No code is released yet, so these numbers can't be checked.\n\nThe central architecture idea holds up, and I'm not inclined to dismiss it. But the paper needs a major revision: a controlled re-run of at least nnUnet and nnFormer under the same split and preprocessing, correction of the false organ-level claim, and the Table VI row fixed or explained. I'd send it to peer review, but the reviewers should push hard for evidence rather than adopted numbers.\n\nIt's worth a reading-group slot to discuss the fusion design, though it's not something I'd cite until the reporting errors are sorted.\n\nRecommendation: engage, with a serious referee.","headline":"A promising hybrid 2D/3D transformer for volumetric segmentation, but adopted baselines and internal reporting errors leave the SOTA claim unverified.","tokens_in":18606,"tokens_out":4347,"would_cite":false,"duration_ms":34098,"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 combining a 2D Transformer that reads each CT/MRI slice closely with a 3D Transformer that reads the whole volume, joined by a mutual local-global fusion module and residual prediction addition, sets a new state of…","keywords":["volumetric medical image segmentation","hybrid Transformer","2D-3D feature fusion","residual learning","Hybrid Local-Global fusion Module","self-attention","computed tomography","magnetic resonance imaging"],"falsifier":"Take the exact Synapse 8-class split used in Table I, retrain nnFormer and nnU-Net with the same augmentation, optimizer, and 1000-epoch schedule the paper uses for HResFormer, and compare average DSC; if the gaps shrink to near zero or change sign, the claimed benefit of the hybrid architecture is not established.","tokens_in":17485,"feed_emoji":"🩻","tokens_out":5427,"duration_ms":41814,"temperature":0.7,"pith_summary":"This paper tries to establish that the best way to segment 3D medical scans is to combine two Transformers rather than choose between them: a 2D Transformer that examines each slice closely and a 3D Transformer that understands how slices connect into anatomy. On three benchmarks (Synapse multi-organ CT, brain-tumor MRI, cardiac MRI), the proposed HResFormer reports the highest Dice scores, beating the previous Transformer-based method nnFormer by 2.5% on Synapse and the convolutional method nnU-Net by 2.1%. The argument is that radiologists themselves work this way—read the axial plane, then review coronal views to build a 3D picture—so a network that mirrors that workflow should segment better than one that only sees slices or only sees volumes. If true, it would shift design practice: hybrid architectures become a viable default for volumetric segmentation, and the specific fusion and residual mechanisms would be reusable components.","feed_headline":"2D + 3D transformers lift organ segmentation beyond prior art","feed_subtitle":"Fusing slice detail with volume context beats nnFormer and nnU-Net on CT and MRI.","key_machinery":"The Hybrid Local-Global Fusion Module (HLGM) is the load-bearing component. It takes features from the 2D prediction volume and the raw 3D volume and performs mutual cross-feature attention twice: once in non-overlapping 3D windows (local mutual fusion) and once with spatially reduced key and value projections (global mutual fusion), then passes the sum through a Cross-Position Feed-Forward layer that uses depth-wise convolution. A second mechanism, residual learning, sets the final prediction as the 2D prediction plus the 3D prediction, letting the 3D model refine rather than replace the 2D segmentation. Together they let the 3D Transformer use the 2D prior without discarding its own volumetric context.","core_discovery":"The central claim is that a hybrid residual Transformer, built from standard 2D and 3D Transformer backbones, outperforms both pure 2D and pure 3D methods on volumetric medical image segmentation. The 2D branch produces fine-grained per-slice predictions; the 3D branch consumes the original volume together with those predictions, fused through a Hybrid Local-Global fusion Module (HLGM) that mutually exchanges features in local windows and through globally reduced attention. The final output is the 3D model's prediction plus the 2D prediction, i.e. residual learning of the whole volume relative to the slice-level baseline. The paper reports state-of-the-art results on Synapse (89.10% average DSC for 8 classes and 87.46% for 13 organs), BraTS (79.5% average DSC), and ACDC (91.88% DSC), with an ablation study showing that the fusion module contributes about 1% and residual learning about 0.9% on Synapse.","pith_inferences":["The hybrid design may generalize outside segmentation: any 3D medical task where a 2D high-resolution view matters, such as small-lesion detection or classification, could benefit from the same 2D-prior-plus-3D-context structure.","A direct testable variant would swap the 2D Transformer for a convolutional 2D network and keep the fusion and residual machinery; that would isolate whether the gain comes from the hybrid topology itself or from Transformer attention.","Because the paper adopts baseline numbers from external publications rather than rerunning them in its own protocol, the honest comparison requires in-house reproduction of nnFormer and nnU-Net under identical splits and training schedules; until then the true margin may be smaller than reported.","Since the 3D attention is quadratic in volume size, the hybrid gain might be preserved at lower cost by substituting a linear-complexity sequence model for the 3D self-attention, a direction the authors themselves flag as future work."],"forward_implications":["On Synapse multi-organ CT with 8 classes, HResFormer reports 89.10% average DSC, 2.5 points above nnFormer and 2.1 points above nnU-Net.","On the 13-organ Synapse split, HResFormer outperforms both CoTR and nnU-Net on average DSC, and beats nnU-Net on all 13 individual organs.","HResFormer keeps parameters comparable to other Transformer methods (117M vs. 150M for nnFormer) and lowers FLOPs from 157.9G to 131.7G, with a lightweight variant at 63.4G still surpassing nnFormer.","On BraTS brain tumor MRI, HResFormer surpasses nnU-Net by 1.8% and nnFormer by 2.1% average DSC; on ACDC it surpasses nnU-Net by 0.52 points.","Ablation studies on Synapse show that both HLGM and residual learning contribute independently to the reported gain, with HLGM adding roughly 1% and residual learning roughly 0.9%."],"supporting_citations":[{"why":"nnFormer is the Transformer-based state of the art it claims to beat, and its setting and adopted results provide the Synapse comparison table.","marker":"(Zhou et al., 2023)"},{"why":"nnU-Net is the convolutional state of the art baseline it claims to surpass on all three benchmarks.","marker":"(Isensee et al., 2021)"},{"why":"UNETR supplies the baseline results adopted for BraTS and the 3D sequence-transformer design HResFormer builds on.","marker":"(Hatamizadeh et al., 2022a)"},{"why":"CoTR is the 3D CNN-plus-deformable-Transformer baseline compared on the 13-organ Synapse split.","marker":"(Xie et al., 2021)"},{"why":"Pyramid Vision Transformer provides the global reduction attention and pyramid encoder design used in the 2D branch.","marker":"(Wang et al., 2021a)"},{"why":"Swin Transformer supplies the shifted-window attention that the 3D branch adapts to non-overlapping 3D regions.","marker":"(Liu et al., 2021)"},{"why":"Video Swin Transformer extends shifted-window attention to 3D, informing the shift local 3D self-attention design.","marker":"(Liu et al., 2022b)"}],"fun_headline_variants":["Hybrid 2D-3D transformer beats prior art on organ segmentation","Residual transformer fuses slice and volume for SOTA segmentation","Two-branch transformer wins on CT and MRI organ segmentation","HResFormer: hybrid residual transformer tops medical segmentation","2D + 3D fusion boosts volumetric organ segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported margins assume the comparison baselines, especially nnFormer and nnU-Net, were trained under exactly the same data splits, preprocessing, and training schedule as HResFormer, but the paper adopts those baseline numbers from earlier publications rather than rerunning them in its own protocol.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid 2D-3D transformer beats prior art on organ segmentation","Residual transformer fuses slice and volume for SOTA segmentation","Two-branch transformer wins on CT and MRI organ segmentation","HResFormer: hybrid residual transformer tops medical segmentation","2D + 3D fusion boosts volumetric organ segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1652,"prompt_tokens":1109,"completion_tokens":543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":725,"completion_tokens_details":{"reasoning_tokens":458}},"tokens_in":725,"tokens_out":543,"duration_ms":4693,"temperature":1.0,"reasoning_tokens":458,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:54:45.098187+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the exact Synapse 8-class split used in Table I, retrain nnFormer and nnU-Net with the same augmentation, optimizer, and 1000-epoch schedule the paper uses for HResFormer, and compare average DSC; if the gaps shrink to near zero or change sign, the claimed benefit of the hybrid architecture is not established.","supporting_citations":[],"review_version":1}