{"id":"ef4abad7-7f5a-4c5d-b425-f2d8672cf27a","arxiv_id":"2411.17342","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A learned symmetry axis, enforced as a loss or through deformable registration, improves cranial defect reconstruction to near state-of-the-art at a fraction of the training cost.","lead":"Cranial implant modeling is slow and costly, and deep learning reconstruction methods often need huge training compute. This paper adds a neural network that detects a skull's symmetry plane and uses it as a training loss or as a post-hoc refinement step, matching near-state-of-the-art accuracy with under 500 GPU hours.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The best reported results rely on the symmetry network predicting an accurate plane from reconstructed, defective volumes (VRec), but Table 1 only evaluates it on healthy skulls and implants; if the plane is biased by an imperfect reconstruction, the registration objective in Eq.","rationale":"The paper's central claim is that learnable symmetry enforcement, as a training loss or registration objective, yields state-of-the-art-comparable skull reconstruction at a fraction of training cost. For that claim, two things must hold: (a) the symmetry plane of a healthy skull can be learned, and (b) the same network outputs a sufficiently accurate plane when its input is a defective skull plus a partially reconstructed implant. The paper provides strong evidence for (a): Table 1 shows ViT-SN reaches DSC ~0.956/0.969 on healthy skulls, and the symmetry loss consistently improves all tested RN architectures in Table 2. The weak point is (b). Eq. (2) and Eq. (3) call SN on VRec, a volume that contains the defect and the current reconstruction. Table 1 does not evaluate this distribution. The registration refinement is the more serious case because π is recomputed from the volume being optimized; a biased plane can be reinforced by the symmetry objective rather than corrected. The paper's own limitation discussion admits failure when the initial reconstruction is completely incorrect, but does not quantify how wrong Rec can be before the plane becomes unreliable. Since the best reported numbers (DSC/SDSC/HD95 of 0.94/0.94/1.31 and 0.95/0.95/1.28) come from the Reg-SN and Seg-SN+Reg-SN configurations, this untested generalization is load-bearing. The proposed test would settle it directly by comparing SN planes on VRec to a healthy-skull reference and by replacing the recomputed plane with an oracle plane during refinement. If the oracle-plane version performs similarly, the concern fails; if it performs better, the paper's central quantitative claim should be conditioned on plane-generalization evidence. I therefore agree with the reader's weakest-assumption analysis, and the appropriate verdict remains CONDITIONAL/UNCHANGED; the concern is real but addressable and does not invalidate the demonstrated consistency of the symmetry-loss training signal.","tokens_in":36,"tokens_out":6486,"duration_ms":178215,"concrete_test":"On SkullBreak/SkullFix test cases, take the baseline RUNet prediction Rec, form VRec = V ∨ Rec, and compare the SN plane π_SN = SN(VRec) to a reference plane π_ref obtained from the corresponding healthy skull (e.g., SN applied to the complete skull, or the plane maximizing healthy-skull symmetry). Report the plane error and its correlation with the Reg-SN DSC change. Then rerun Reg-SN twice: once with π recomputed as in Eq. (3), and once with π fixed to π_ref. If fixing π_ref improves mean DSC/HD95 by more than a small margin (e.g., ΔDSC > 0.01), the reported refinement results depend on the untested generalization of SN to VRec; if not, the concern is not the bottleneck.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest numerical results (Table 3/4, Reg-SN and Seg-SN+Reg-SN) depend on SN producing a reliable symmetry plane not from healthy skulls, but from VRec = V ∨ Rec during the refinement loop. Table 1 validates SN on complete skulls and implants; it does not measure plane accuracy on the VRec volumes that Eq. (3) actually feeds to SN. Because π = SN(VRec) is recomputed at every optimization step, the objective is a function of the current reconstruction Rec. If Rec is initially imperfect, the remaining asymmetry of V can shift the predicted plane; minimizing DSL(VRec, R(VRec, π)) then encourages deforming Rec to match reflection of the defective skull about a biased axis, potentially moving the implant toward the wrong location. The paper acknowledges failure only in the extreme case ('if the initial reconstruction is completely incorrect...') and asserts that SN 'has great generalizability,' but the supporting evidence is healthy-skull performance, not performance on partially reconstructed defects. The Seg-SN vs Reg-SN ablation (Table 3) compares end metrics and does not isolate plane error on VRec, so it cannot show whether the refinement gain is robust to plane bias. This is the load-bearing unsupported step for the headline accuracy figures.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a two-network pipeline for cranial defect reconstruction. A symmetry network (SN) is trained unsupervised on healthy skulls to regress a reflection plane pi using a Dice-based symmetry loss; the plane is then used (i) as an auxiliary loss during training of a reconstruction network (RN), Eq. (2), and (ii) in an optional deformable registration refinement after inference, Eq. (3). The method is evaluated on SkullBreak and SkullFix, with ablations over RN and SN architectures, and compared with prior work. The best configuration (Seg-SN plus Reg-SN) reports DSC/SDSC/HD95 of 0.936/0.932/1.312 on SkullBreak and 0.945/0.951/1.278 on SkullFix, close to a diffusion-augmented state of the art while claiming under 500 GPU-hours of training. Qualitative results are shown for 11 real clinical defects.","tokens_in":12141,"tokens_out":6164,"duration_ms":59008,"significance":"If the reported numbers hold, the paper makes a useful practical contribution: it demonstrates that a lightweight learnable symmetry prior can bring CNN-based reconstruction close to much more expensive generative augmentation methods, with a plausible clinical runtime. The paper has several strengths: it uses two public benchmarks, includes architecture ablations, separates segmentation-based and registration-based symmetry enforcement, reports inference times and training cost, and explicitly acknowledges that symmetry enforcement can hurt quantitative metrics and cannot fix completely incorrect initializations. The remaining weaknesses are centered on the validation of the symmetry network under the distribution it actually faces during refinement, and on statistical and hyperparameter reporting. I regard the contribution as incremental but potentially valuable for a clinical AI audience; the central claim is currently supported only by single-setting metrics on synthetic defects.","major_comments":[{"comment":"The headline refinement result relies on SN(VRec), where VRec is the volume being optimized, but Table 1 validates SN only on healthy complete skulls and on implants. The paper states in the Discussion that 'the SN has great generalizability,' yet no experiment measures plane accuracy on VRec during the iterative registration loop. Since the plane is recomputed from the current reconstruction, a biased plane could be reinforced rather than corrected, and the improvement of Reg-SN over Seg-SN (for example DSC 0.931 vs 0.904 on SkullBreak) is exactly the claim that needs that missing evidence. Please add an experiment that either (a) compares SN(VRec) with the plane obtained from the corresponding intact skull, reporting angle and distance error, or (b) runs the refinement with a fixed plane from the intact skull and compares final metrics to the proposed feedback scheme. The Discussion acknowledges failure when the initial reconstruction is completely incorrect, but the intermediate partial-error regime is not characterized, and that regime is the load-bearing one for the accuracy figures in Table 4.","section":"Section 2.4, Eq. (3), Table 3"},{"comment":"The paper promises that any claim of statistical significance is supported by a Wilcoxon signed-rank test with p-value below 0.01, but no p-values, test statistics, or confidence intervals appear anywhere in the Results or in Tables 2-4. Several relevant differences are small (for example Seg-SN plus Reg-SN versus Reg-SN alone differs by 0.005 DSC on SkullBreak and 0.004 on SkullFix), so significance is genuinely load-bearing. Please report the test statistics and p-values for the paired comparisons, state whether the comparisons are across cases or across architectures, and address multiple testing if several paired tests are performed.","section":"Section 2.6 and Section 3"},{"comment":"The method's two hyperparameters, alpha in Eq. (2) and lambda in Eq. (3), are fixed at alpha=1 and lambda=1e5 without sensitivity analysis or an explicit selection rule. A reader cannot tell whether the reported gains are due to the symmetry prior itself or to a favorable choice of these two constants. Please include a small sweep (at least over lambda, and preferably alpha) on the validation split, and report the dependence of DSC and HD95 on these parameters. This is also needed to support the claim that registration-based refinement improves over segmentation-based symmetry enforcement, which currently rests on a single lambda value.","section":"Section 2.3, Section 2.4, Tables 2-3"}],"minor_comments":[{"comment":"The phrase 'signed-rak test' should be 'signed-rank test'.","section":"Section 2.6"},{"comment":"The abstract uses 'bDSC' while the tables use 'SDSC' for the same quantity; please unify the terminology.","section":"Abstract and Tables 2-4"},{"comment":"There are citation and formatting errors: the entry for 'Yang et al. [22]' does not appear to match reference [22], and 'Mazzocchetti et al. et al.' and 'Kesornsri et al. et al.' contain duplicated 'et al.'.","section":"Table 4"},{"comment":"The claim that the refinement leaves only a small hole is a subjective visual reading; please add a quantitative surface-distance map or specify the evaluation conditions.","section":"Figure 4"},{"comment":"The statement that networks were 'trained until convergence' lacks a convergence criterion; please report the number of epochs or iterations and the early-stopping rule.","section":"Section 2.6"},{"comment":"No code or data availability statement is provided; since the method is evaluated on public datasets, releasing the symmetry network weights or an implementation would materially aid reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a medical imaging/AI journal. The main reason for major revision is the missing validation of the symmetry network on the volumes that actually appear during the registration-based refinement; this is a testable, localized gap rather than a fundamental flaw. I would also want the promised statistical tests to be reported before publication. I do not see a reason to suspect the primary training signal is circular, because the SN is trained on healthy skulls; the circularity concern is confined to the refinement loop and can be addressed experimentally. When the authors compare training costs with the diffusion-augmented method, they should specify exactly what is included in the 'less than 500 GPU hours' figure (SN training, RN training, parameter sweeps, refinement runs) to avoid an apples-to-oranges comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the core result is real. The authors integrate a learned symmetry-plane regressor into cranial implant reconstruction in two ways—as an extra Dice-based loss during training and as an objective for deformable registration refinement at test time. Prior mirroring methods needed a manually chosen axis; this removes that constraint. On SkullBreak/SkullFix the method reaches DSC around 0.94/0.95, within a hair of the diffusion-augmented state of the art, with two orders of magnitude less training compute. That claim is the main contribution and it holds up as far as the paper reports.\n\nCredit where due: the comparison is against public benchmarks, the ablation across five reconstruction architectures is informative, and the compute numbers are specific enough to check. The finding that a ViT is better for symmetry regression while a U-Net remains better for reconstruction is a useful data point. The failure mode for completely wrong initial reconstructions is acknowledged in the text.\n\nSoft spots, in order of importance. First, the registration refinement recomputes the symmetry plane from the partially reconstructed volume at every step, but Table 1 only validates the symmetry network on healthy skulls and implants, not on those VRec volumes. That is a genuine gap. It is not evidence the method is wrong—the test-set metrics improved after refinement, so a biased plane would likely have shown up there—but the paper should add an ablation measuring plane error on VRec and showing refinement is robust to initial reconstruction quality. Second, the paper promises that every significance claim is backed by a Wilcoxon test with p<0.01, but no p-values appear anywhere. That is a broken reporting promise, not a statistical argument. Third, no code or detailed configuration is released; for a method whose value is partly practical reproducibility, that matters. Fourth, the 11 real clinical cases are called successfully reconstructed with only visual evidence; that is qualitative and should be labeled as such.\n\nThe stress-test note about circularity overstates the risk. The symmetry network is trained unsupervised on healthy skulls, and the test-set results are measured against ground truth. If refinement were chasing a biased plane, the reported DSC/HD95 would likely worsen, not improve. The concern is legitimate as a missing evaluation, not as a demonstrated flaw.\n\nFor whom: people working on cranioplasty, implant design, or symmetry-based shape completion. It deserves a serious referee; ask for the p-values, the VRec plane-error ablation, and code/configuration, then it can be accepted.","headline":"A practical, low-compute symmetry-enforced skull reconstruction method that mostly delivers on its claims; the main open question is a missing ablation on the symmetry network's behavior with partially reconstructed volumes.","tokens_in":12695,"tokens_out":2515,"would_cite":true,"duration_ms":27408,"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":"This paper shows that a neural network trained to predict a skull's symmetry plane can be used both as a training loss and as a registration objective to substantially improve automatic cranial defect reconstruction, reaching near…","keywords":["skull reconstruction","cranial implant design","learnable symmetry plane","symmetry loss","deformable registration","deep learning","medical image analysis","vision transformer"],"falsifier":"For the SkullBreak test set, compute the symmetry network's predicted plane on the intermediate reconstructed volume VRec (defective skull plus initial reconstruction) and measure its deviation from the plane predicted on the healthy ground-truth skull. If the deviation is large on cases where registration refinement fails, or if using the ground-truth plane in the refinement does not produce better Dice than using the predicted plane, the claim that the symmetry network generalizes to reconstructed volumes is not supported.","tokens_in":11702,"feed_emoji":"🦴","tokens_out":5752,"duration_ms":48497,"temperature":0.7,"pith_summary":"The paper proposes to improve automatic reconstruction of cranial defects by enforcing skull symmetry in a learnable way. The authors train a symmetry network to predict the symmetry plane of a skull, then use that plane in two ways: as an extra loss during training of the reconstruction network, and as an objective for a deformable registration refinement after reconstruction. On the SkullBreak and SkullFix benchmarks, the symmetry-preserving reconstruction reaches DSC of 0.94 and 0.95, surface Dice of 0.94 and 0.95, and HD95 of 1.31 and 1.28 mm, substantially better than the baseline without symmetry and close to the best published method, while training with under 500 GPU hours instead of over 100,000. The authors argue this makes accurate, scalable, near-real-time skull reconstruction feasible and potentially extendable to other symmetric structures.","feed_headline":"Learnable symmetry brings skull reconstruction to top accuracy","feed_subtitle":"Near best-in-class cranial implant Dice at under 500 GPU hours, versus 100,000.","key_machinery":"The symmetry network (SN) and the symmetry loss (SL). The SN is an encoder-regressor (volumetric ResNet or vision transformer) that maps a skull volume V to a plane equation π; the SL equals DSL(V, R(V, π)), the Dice loss between the volume and its reflection about the predicted plane, which makes the plane trainable without manual labels. The same SL is reused after reconstruction as a differentiable objective for a deformable registration with diffusive regularization that deforms the reconstructed defect to maximize the whole skull's symmetry. This dual use, as a training loss and as a refinement objective, is the mechanism that transfers healthy-skull symmetry knowledge to defect reconstruction.","core_discovery":"The central claim is that skull symmetry can be learned by a vision-transformer-based encoder-regressor trained with a fully differentiable symmetry loss, defined as SL(V, π) = DSL(V, R(V, π)) where V is the skull volume, π is the predicted plane equation, and R is the reflection operator. The same loss is then used in two integration modes: as an additional objective during reconstruction-network training (Seg-SN), and as an instance-optimization objective for a deformable registration refinement that warps only the reconstructed defect (Reg-SN). In the best configuration, Seg-SN plus Reg-SN, the method reaches DSC/SDSC/HD95 of 0.94/0.94/1.31 on SkullBreak and 0.95/0.95/1.28 on SkullFix, outperforming all previously published methods except the diffusion-augmented one, and it successfully reconstructs all 11 real clinical cases while the baseline manages only 6. A notable design finding is that a vision transformer is better for symmetry estimation while a convolutional RUNet is better for the reconstruction itself.","pith_inferences":["The symmetry plane's accuracy on the intermediate reconstructed volumes (defective skull plus initial implant) is the linchpin of the method; directly measuring plane error on those volumes would show whether failures stem from the symmetry network or from the registration optimizer.","A natural, untested extension is a multi-resolution or adaptive regularization schedule for the refinement step, since the authors note the strong diffusive regularization leaves small holes on the largest clinical defects.","Because real skulls are only approximately symmetric, the method's ceiling is limited by intrinsic asymmetry; adding a per-case asymmetry estimate could allow the refinement to weight symmetry only where it is trustworthy.","The same symmetry-loss refinement could transfer to other bilateral structures such as the pelvis or mandible, though the paper only evaluates cranial data."],"forward_implications":["Skull reconstruction can run in under 15 seconds per case (under 5 seconds without refinement), fast enough for potential intraoperative use, while staying close to state-of-the-art accuracy.","The method approaches the ceiling of what symmetry alone can provide, since the best achievable reflected-implant DSC (0.956 on SkullBreak, 0.969 on SkullFix) is close to the final reconstruction DSC, indicating symmetry is the dominant recoverable signal on these benchmarks.","Training both networks costs under 500 A100 GPU hours, roughly 200 times less than the diffusion-augmented alternative, making the approach feasible for smaller groups and retraining on other symmetric anatomical structures.","The registration-based refinement acts as per-case instance optimization, improving out-of-distribution real defects, which is where segmentation-only networks typically fail."],"supporting_citations":[{"why":"Supplies the open SkullBreak and SkullFix datasets with real skulls and synthetic defects used for all training, validation, and testing.","marker":"[13]"},{"why":"The state-of-the-art heavy-augmentation method this work compares against; the proposed method reaches comparable scores at far lower training cost.","marker":"[37]"},{"why":"PRS-Net, which learns planar reflective symmetry detection in 3D models, is the conceptual basis for the symmetry network.","marker":"[9]"},{"why":"Nerd, a neural 3D reflection symmetry detector, provides a second reference for learnable symmetry estimation.","marker":"[39]"},{"why":"An earlier method by the same group that the symmetry-enforced reconstruction improves on, used as a comparison baseline.","marker":"[34]"}],"fun_headline_variants":["Learnable symmetry boosts skull implant accuracy","Skull reconstruction: 200x less compute, near-best DSC","Learned symmetry cuts skull implant compute to 500 hours","Near-best skull implant accuracy from learned symmetry","Symmetry learning sharpens skull reconstruction results"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The symmetry network, trained only on healthy skulls, is assumed to predict a symmetry plane accurate enough on reconstructed volumes that contain the implant and defect, and the registration refinement must deform the implant toward the correct shape rather than reinforcing a wrong plane.","fun_headline_variants_meta":{"raw":{"variants":["Learnable symmetry boosts skull implant accuracy","Skull reconstruction: 200x less compute, near-best DSC","Learned symmetry cuts skull implant compute to 500 hours","Near-best skull implant accuracy from learned symmetry","Symmetry learning sharpens skull reconstruction results"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000832,"raw_usage":{"total_tokens":3696,"prompt_tokens":1076,"completion_tokens":2620,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":2547}},"tokens_in":692,"tokens_out":2620,"duration_ms":19583,"temperature":1.0,"reasoning_tokens":2547,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:13:00.476126+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the SkullBreak test set, compute the symmetry network's predicted plane on the intermediate reconstructed volume VRec (defective skull plus initial reconstruction) and measure its deviation from the plane predicted on the healthy ground-truth skull. If the deviation is large on cases where registration refinement fails, or if using the ground-truth plane in the refinement does not produce better Dice than using the predicted plane, the claim that the symmetry network generalizes to reconstructed volumes is not supported.","supporting_citations":[{"cited_title":"Kodym, J","cited_arxiv_id":null,"evidence_quote":"Supplies the open SkullBreak and SkullFix datasets with real skulls and synthetic defects used for all training, validation, and testing."},{"cited_title":"Wodzinski, K","cited_arxiv_id":null,"evidence_quote":"The state-of-the-art heavy-augmentation method this work compares against; the proposed method reaches comparable scores at far lower training cost."},{"cited_title":"Gao, L.-X","cited_arxiv_id":null,"evidence_quote":"PRS-Net, which learns planar reflective symmetry detection in 3D models, is the conceptual basis for the symmetry network."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Nerd, a neural 3D reflection symmetry detector, provides a second reference for learnable symmetry estimation."},{"cited_title":"Wodzinski, M","cited_arxiv_id":null,"evidence_quote":"An earlier method by the same group that the symmetry-enforced reconstruction improves on, used as a comparison baseline."}],"review_version":1}