{"id":"68b0b100-599a-4919-9fcd-47a3ceb8aded","arxiv_id":"2411.18290","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A Siamese contrastive-loss network exploiting tumor-induced left-right asymmetry improves nasopharyngeal carcinoma gross tumor volume segmentation in non-contrast planning CT.","lead":"This paper trains a neural network to outline nasopharyngeal tumors directly on non-contrast planning CT scans, using the idea that tumors break the natural left-right symmetry of the head. It reports better outline accuracy than ten existing methods on both hospital and public datasets, which could shorten the manual tumor-drawing step in radiation therapy planning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (1) appears to invert the paper's core mechanism: for asymmetric tumor voxels (m=1) it minimizes feature distance instead of maximizing it, so as written the loss cannot implement the claimed semantic-asymmetry learning.","rationale":"The reader's weakest assumption was the anatomical symmetry premise and the risk that alignment failure or non-pathological asymmetry could misdirect the loss. That is a legitimate concern, but the more immediate and decisive issue is internal consistency: the margin loss as specified in Eq. (1) is opposite to the paper's stated objective. For the central claim to hold as a semantic-asymmetry method, the loss must push apart features at m=1. It instead pulls them together, and the m≠1 term also only enforces similarity. This is not a disagreement with an external consensus or a matter of statistical nuance; it is a mathematical contradiction within the method section. The empirical performance numbers could still be correct, and the method might improve segmentation through the Siamese architecture, the projection head, or extra fine-tuning, but then the paper's central explanatory claim is unsupported. Because the discrepancy could be a typographical error in the equation rather than in the implementation, the appropriate verdict is conditional: the paper should not be accepted in its current form until the authors provide the exact loss used in training, correct Eq. (1) if needed, and show an ablation with the corrected loss. This differs from the reader's identified weakest assumption, so my agreement is 'disagree' even though the overall conditional recommendation is similar.","tokens_in":8349,"tokens_out":10175,"duration_ms":99676,"concrete_test":"Obtain the authors' implementation or reimplement Section 2.2 exactly as written: Eq. (1) with squared L2 on m=1 and hinge on m≠1. Train on the in-house set and evaluate on the external test split using the same schedule, patch size, and hyperparameters, then compare against the intended variant: hinge on m=1 and squared L2 on m≠1. If the as-written loss reproduces Table 1's external DSC 66.80 or Table 2's 66.32, the reported improvement is produced by a loss that penalizes asymmetry, so the semantic-asymmetry narrative fails. If the intended variant is required to reproduce the numbers, Eq. (1) is a typo and the mechanism claim can be restored after correction. Also inspect the gradient of the margin-loss term at m=1 to verify whether training increases or decreases ||E−E′||^2.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central contribution rests on the claim, stated in the abstract and Section 1, that the method 'minimizes the voxel-wise distance between original and flipped areas without tumor and encourages a larger distance between original and flipped areas with tumor.' However, the margin loss in Eq. (1) does not do this. In Eq. (1), m = s − s∩s′ marks the asymmetric abnormal region, so m=1 is exactly the tumor-asymmetry region that should be pushed apart. The m=1 term is ||E−E′||^2, which is minimized, pulling tumor features toward their mirrored counterparts. The m≠1 term is max(t−||E−E′||^2,0), a hinge that only penalizes distances above t, i.e., it enforces similarity for non-tumor voxels. No term in the loss increases distance for m=1; as written, the loss suppresses feature asymmetry everywhere. This is load-bearing because Table 2 attributes a +4.98 DSC gain to adding the projection head and margin loss, and the paper's novelty is specifically the semantic-asymmetry mechanism. If the experiments used the loss exactly as printed, the reported gain cannot be explained by the advertised asymmetry mechanism; if the experiments used a different (corrected) loss, Eq. (1) is a material misstatement of the method that must be fixed before the central claim can be evaluated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SATS, a Siamese encoder-decoder method for segmenting nasopharyngeal carcinoma gross tumor volume (GTV) directly from non-contrast planning CT. The method relies on the premise that the healthy nasopharynx is bilaterally symmetric and that tumors break this symmetry. A voxel-wise margin loss is introduced to minimize feature distance between original and flipped images in symmetric regions and, according to the abstract and introduction, to maximize distance in asymmetric tumor regions. The model is trained on an in-house dataset of 163 patients and evaluated on an internal test split and the external SegRap2023 dataset, with comparisons against ten CNN, Transformer, and Mamba baselines. The paper reports the best DSC and HD95 on both internal and external test sets, with ablation studies on the projection head and margin loss, plus a failure-case analysis.","tokens_in":8649,"tokens_out":5346,"duration_ms":47358,"significance":"The clinical problem is well motivated: direct segmentation in non-contrast planning CT could avoid MRI-to-CT registration errors in radiotherapy workflow. If the proposed method were correctly specified and the reported numbers reproducible, this would be a meaningful contribution: the first NPC GTV segmentation on non-contrast pCT, external validation on a public dataset, and a systematic comparison against ten strong baselines. The paper also honestly discusses failure cases. However, the central equation describing the margin loss inverts the paper's core mechanism, and several quantitative claims in the text do not match the tables. As written, the validity of the central claim is not established, so the significance is conditional on a substantial correction and verification.","major_comments":[{"comment":"Equation (1) is the core of the method and the current text is a material misstatement of the algorithm.","section":"Section 2.2, Eq. (1)"},{"comment":"Please verify all improvement figures against the underlying data and correct the text and Abstract.","section":"Section 3.2, Table 1"},{"comment":"The ablation study is a central piece of evidence for the contribution of the proposed loss; the current numbers are not reproducible from the table.","section":"Section 3.3, Table 2"},{"comment":"The internal-testing 'leading performance' claim is not statistically supported against the closest baselines.","section":"Section 3.2, Table 1"},{"comment":"Without sensitivity analysis, the reader cannot assess how robust the central mechanism is to hyperparameter choices.","section":"Section 3.1, Implementation"}],"minor_comments":[{"comment":"The reported metric is not present in the corresponding table.","section":"Abstract and Section 3.2"},{"comment":"The preprocessing is under-specified.","section":"Section 2.1"},{"comment":"Clarify that the 'detection' is done with ground-truth labels during training.","section":"Section 2.1"},{"comment":"Discrepancy in the number of external test cases.","section":"Section 3.3, Figure 6"},{"comment":"Minor formatting issue.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central issue is that Eq. (1) is the opposite of the advertised mechanism. This is not a mere presentation glitch: the abstract, introduction, and ablation discussion all describe a loss that maximizes distance in tumor regions, while the printed equation minimizes it. The authors must either correct the equation and confirm the experiments match the corrected loss, or, if the printed loss was actually used, explain how the reported performance gain can arise from suppressing asymmetry in tumor regions. Until this is resolved, the paper's scientific contribution cannot be evaluated. The numerical inconsistencies in Section 3.2 and Table 2 further undermine trust in the reported results, though they are fixable in revision. I would recommend the editor ask for a thorough revision with rerun or verification of the core experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick read: the paper has a genuinely useful empirical result, but the central loss in Equation (1) is mis-signed. As written, the \"asymmetric\" term 1(m=1)||E−E'||^2 is minimized, pulling tumor voxel features toward their mirrored partners. The non-tumor term is a hinge that also penalizes distances above t, so the whole loss enforces feature similarity. The abstract and Section 1 claim the opposite: larger distance for tumor regions. The stress-test note is right. This is load-bearing because the ablation in Table 2 attributes +4.98 DSC to adding this loss plus the projection head. If the experiments used the printed loss, the gain cannot come from semantic asymmetry learning; if they used a different loss, the paper misstates its own method. Either way, the central claim cannot be evaluated as written.\n\nWhat's genuinely new: using bilateral symmetry of the nasopharyngeal region as a cue for GTV segmentation in non-contrast planning CT, with a Siamese encoder and a supervised asymmetry mask derived from ground truth. The external evaluation on SegRap2023 is a real strength: +2.40 DSC over nnUNet (66.80 vs 64.40) on 118 cases, and the per-case comparison in Figure 6 is useful. The idea is plausible and the clinical motivation (avoiding MRI registration, saving 1–2 hours) is solid.\n\nOther soft spots: the text's improvement figures don't match Table 1 (0.92% vs the actual 0.81% over SwinUNETR-V2 internally; 4.4% vs 2.40% over nnUNet externally). The \"first work on non-contrast CT\" claim is contradicted by their own reference [30] (Men et al., planning CT). Hyperparameters t=20 and beta=1 have no sensitivity analysis, and significance markers are missing for several baselines. These are fixable but need attention.\n\nBottom line: the paper is for researchers working on CT-only head/neck segmentation, and the external benchmark is valuable. But the equation issue is not cosmetic. I'd send it to peer review, but with a clear request to correct the loss and confirm the ablation. If the corrected loss matches the text, the result stands as a nice contribution; if not, the empirical gain is unanchored.","headline":"Useful external benchmark, but Equation (1) is mis-signed, so the paper's core asymmetry mechanism is currently unverifiable as written.","tokens_in":9215,"tokens_out":4349,"would_cite":false,"duration_ms":38164,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a 3D Siamese segmentation framework exploiting the left-right asymmetry of nasopharyngeal tumors can segment gross tumor volume directly on non-contrast planning CT, improving external Dice by 4.4% over the…","keywords":["nasopharyngeal carcinoma","gross tumor volume","planning CT","semantic asymmetry","Siamese network","contrastive learning","margin loss","medical image segmentation"],"falsifier":"Take a cohort of NPC patients, run SATS and nnUNet on scans that are intentionally rotated by a few degrees around the head axis (simulating pose asymmetry), and measure Dice; if the margin loss truly depends on perfect bilateral symmetry, SATS's Dice should fall disproportionately relative to nnUNet as the rotation angle grows.","tokens_in":8172,"feed_emoji":"🎯","tokens_out":6242,"duration_ms":48924,"temperature":0.7,"pith_summary":"The paper tries to establish that nasopharyngeal carcinoma tumors can be segmented directly from non-contrast planning CT, the modality actually used for radiotherapy planning, without relying on MRIs that must be registered to CT. The core idea is that a healthy nasopharynx is bilaterally symmetric, and the tumor is precisely the structure that breaks that symmetry. The authors build a Siamese segmentation network that compares each CT with its mirror image and trains it to pull symmetric (non-tumor) voxel features together and push asymmetric (tumor) voxel features apart. On their in-house test set the method reaches 81.22% Dice, and on the external SegRap2023 dataset 66.80% Dice, outperforming ten CNN, Transformer, and Mamba baselines. If true, this would give radiation oncologists a planning-CT-only delineation tool, avoiding the registration errors of MRI-to-CT transfer.","feed_headline":"Asymmetry loss beats nnUNet by 4.4% Dice on NPC CT tumors","feed_subtitle":"Directly segmenting nasopharyngeal tumors on non-contrast CT skips MRI registration errors and generalizes better externally.","key_machinery":"The load-bearing mechanism is a Siamese encoder-decoder paired with a voxel-wise margin loss over asymmetric regions. The margin loss (Eq. 1) acts on features after a non-linear projection head: for voxels flagged by the asymmetrical mask $m = s - s \\cap s'$, it minimizes the $ℓ^2$ distance between the original and flipped feature maps $E$ and $E'$; for all other voxels it maximizes that distance up to a margin $t$ (set to 20). The asymmetric mask is computed from the ground-truth segmentation by subtracting the intersection of the mask with its flip, isolating the portion of the tumor that breaks bilateral symmetry.","core_discovery":"The central discovery claimed is that tumor-induced semantic asymmetry is a sufficient and generalizable cue for gross tumor volume segmentation in non-contrast planning CT. The method, SATS, assumes the healthy nasopharynx is bilaterally symmetric. It uses a shared-weights Siamese encoder-decoder to process the CT and its left-right flip, and a voxel-wise margin loss (Eq. 1) that decreases the distance between corresponding features in non-tumor regions and increases it in tumor regions, where the asymmetrical mask $m = s - s \\cap s'$ locates tumor voxels that do not survive the flip. A non-linear projection head is required; the ablation shows margin loss alone hurts (61.50% vs 63.44% external Dice), and only with the projection head does it give +4.98% Dice. The authors report 81.22% internal and 66.80% external Dice, with at least 2% absolute improvement over all baselines in external testing, and interpret the external gain as evidence of better generalizability of the asymmetry signal.","pith_inferences":["The same bilateral-symmetry premise could transfer to other lateralized head-and-neck structures or to organs like the breast and pelvis, where prior symmetry work exists, but the margin loss would need to handle anatomy whose healthy state is not perfectly symmetric.","The reliance on a manually selected symmetric atlas and affine alignment is a practical bottleneck; a fully automated alignment could make the method more deployable, and one testable extension is to replace the atlas step with a learned canonical-space registration.","The margin loss implicitly assumes that any non-pathological asymmetry (vessels, lymph nodes, patient pose) is either removed by alignment or filtered by the projection head; a stress test with deliberately misaligned or asymmetric patients would reveal how much of the gain is architecture versus this premise."],"forward_implications":["Segmentation can be done directly on non-contrast planning CT, eliminating the need for MRI-to-CT registration during delineation and the errors it introduces.","The external test improvement (66.80% vs 64.40% Dice for nnUNet) suggests the symmetry signal transfers across institutions better than texture-only features.","The projection head is not optional: removing it turns the margin loss from a help into a harm, so any reuse of the method must keep the non-linear projection.","Tumors that are symmetric about the midline (small, central lesions) are the method's known weakness, which the paper's failure-case analysis acknowledges."],"supporting_citations":[{"why":"It supplies the affine alignment to a symmetric atlas that enforces the bilateral symmetry premise before learning.","marker":"[33]"},{"why":"It provides the rigid registration used to initially align MRI to CT for ground-truth generation.","marker":"[1]"},{"why":"It provides the deformable registration (deeds) used to transfer MRI-based delineations to planning CT for training labels.","marker":"[11]"},{"why":"It is the primary CNN baseline (nnUNet) whose external performance SATS exceeds by 4.4% Dice.","marker":"[16]"},{"why":"It is the MedNeXt baseline, the second-best external performer, that SATS surpasses by 2.03% Dice.","marker":"[32]"},{"why":"It is the SwinUNETR-v2 transformer baseline, the best internal competitor, that SATS beats by 0.81% Dice.","marker":"[10]"},{"why":"It represents the STU-Net baseline family used to compare against scalable CNN segmentation models.","marker":"[15]"},{"why":"It is the Mamba-based baseline (UMambaBot) used to represent state-space model methods.","marker":"[27]"},{"why":"It is the international guideline that defines the GTV delineation standard used for annotations.","marker":"[18]"}],"fun_headline_variants":["Asymmetry cue in CT boosts NPC tumor segmentation by 2% Dice","Semantic asymmetry drives precise NPC segmentation on non-contrast CT","Siamese contrastive net exploits tumor asymmetry for GTV on CT","Direct CT segmentation of NPC tumors beats MRI-guided methods","Projection head key to asymmetry-based NPC tumor segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that, once scans are aligned to a manually chosen symmetric atlas, any remaining asymmetry in the nasopharyngeal region is caused by the tumor rather than by normal anatomy, patient pose, or imaging artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Asymmetry cue in CT boosts NPC tumor segmentation by 2% Dice","Semantic asymmetry drives precise NPC segmentation on non-contrast CT","Siamese contrastive net exploits tumor asymmetry for GTV on CT","Direct CT segmentation of NPC tumors beats MRI-guided methods","Projection head key to asymmetry-based NPC tumor segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000929,"raw_usage":{"total_tokens":4025,"prompt_tokens":1041,"completion_tokens":2984,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2897}},"tokens_in":657,"tokens_out":2984,"duration_ms":23391,"temperature":1.0,"reasoning_tokens":2897,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:19:33.021468+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a cohort of NPC patients, run SATS and nnUNet on scans that are intentionally rotated by a few degrees around the head axis (simulating pose asymmetry), and measure Dice; if the margin loss truly depends on perfect bilateral symmetry, SATS's Dice should fall disproportionately relative to nnUNet as the rotation angle grows.","supporting_citations":[{"cited_title":"Matching in the Wild: Learning Anatomical Embeddings for Multi-Modality Images","cited_arxiv_id":"2307.03535","evidence_quote":"It provides the rigid registration used to initially align MRI to CT for ground-truth generation."},{"cited_title":"In: MICCAI","cited_arxiv_id":null,"evidence_quote":"It provides the deformable registration (deeds) used to transfer MRI-based delineations to planning CT for training labels."},{"cited_title":"In: MICCAI","cited_arxiv_id":null,"evidence_quote":"It is the primary CNN baseline (nnUNet) whose external performance SATS exceeds by 4.4% Dice."},{"cited_title":"In: MICCAI","cited_arxiv_id":null,"evidence_quote":"It is the MedNeXt baseline, the second-best external performer, that SATS surpasses by 2.03% Dice."},{"cited_title":"In: MICCAI","cited_arxiv_id":null,"evidence_quote":"It is the SwinUNETR-v2 transformer baseline, the best internal competitor, that SATS beats by 0.81% Dice."},{"cited_title":"Radiotherapy and Oncology126(1), 25–36 (2018) 6","cited_arxiv_id":null,"evidence_quote":"It is the international guideline that defines the GTV delineation standard used for annotations."}],"review_version":1}