{"id":"a6b55452-e2d6-45d2-9a82-ce47149add2c","arxiv_id":"2411.18415","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"A neural field with multi-scale pairwise distance regularization flattens sparse 3D anatomical structures into 2D views with lower peak distortion than mesh-based baselines.","lead":"This paper trains a small neural network to deform a flat plane into a wavy sheet that passes through sparse 3D anatomical structures, such as blood vessel centerlines, so the structure can be seen in a single 2D image. The goal is a generic alternative to organ-specific methods for flattening vessels, bones, and ducts in CT scans.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Peak-distortion win may be an artifact of the fidelity trade-off: ours reports 2–10x larger target-to-mesh distances than baselines while claiming lower max distortion; a matched-fidelity comparison is needed.","rationale":"The reader's weakest assumption was that the target point set lies near a single smooth 2D manifold, which limits the method on hand tips and the rib cage. My concern is different but related: the quantitative comparison that supports the headline claim is confounded by the fidelity-versus-distortion trade-off. The reader did note the larger target distances for the proposed method and the rib-cage distance exceeding pixel spacing, but did not make this the load-bearing issue. I therefore partially agree. The concern is concrete and testable, and it does not disprove the method; it identifies a missing control in the evaluation. Since the reader's conditional verdict already reflects some caution about the quantitative evidence, this concern does not require changing the verdict, but it should be addressed before the comparative claim is accepted at face value.","tokens_in":14290,"tokens_out":3562,"duration_ms":35655,"concrete_test":"Re-run the cerebral-vessel and hand-bone experiments with the target weight wt increased until the mean target-to-mesh distance of the neural field matches the baselines (e.g., <=0.1 mm), then recompute the max pairwise distortion in the 1 cm target window for all three methods. If the neural-field max distortion rises to the 20–53 mm range seen for ARAP/CeVasMap, the reported peak-distortion advantage is an artifact of the unconstrained target distance; if it remains below roughly 5 mm, the claim is upheld.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is that the neural unfolding method outperforms mesh-based baselines with respect to peak distortion. The evidence for this rests on comparing distortion values while allowing very different target-adherence levels. In Sup. Table 2, the mean target-to-mesh distance for Ours is 0.253–0.651 mm across the four tasks, whereas ARAP and CeVasMap achieve 0.024–0.138 mm (often <0.05 mm). For the rib cage, Ours has mean distance 0.651 mm versus 0.035 mm (ARAP) and 0.024 mm (CeVasMap). The paper explicitly acknowledges that the baselines enforce target intersection and therefore achieve smaller distances. Because the optimization is a weighted sum (Eq. 1) of a Chamfer-like target loss (Eq. 2) and a pairwise distortion loss (Eq. 4), the method can reduce peak distortion by leaving targets farther from the fitted manifold. The reported distortion advantage is therefore not an apples-to-apples comparison: lower max distortion is partly bought by relaxing the primary goal of making target structures visible. The distance column quantifies this cost, but the headline claim is framed solely on distortion. This is an evidence-quality concern, not a conceptual flaw; the method may still be useful, but the comparative claim requires either matching target-distance budgets or jointly reporting distortion and fidelity in a way that penalizes missed targets.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neural field-based framework for unfolding sparse anatomical structures (e.g., vessel centerlines, bone centerlines) into a 2D image. A low-frequency MLP predicts a displacement field on top of a PCA-initialized plane, and is trained with a Chamfer-like target-distance loss, a multi-scale pairwise distortion loss, and an optional image-based loss. The authors demonstrate the approach on cerebral vessels, hand bones, pancreas with vessels, and rib cages, and claim that it outperforms the mesh-based baselines CeVasMap and ARAP with respect to peak distortion, and that the proposed multi-scale distortion regularizer produces smoother transformations than Jacobian-based neural field regularizers.","tokens_in":14692,"tokens_out":5542,"duration_ms":50385,"significance":"If the quantitative claims were robustly supported, this would be a useful contribution to medical image unfolding: the method is modality-independent, does not require a global point ordering, and integrates geometric and appearance objectives in a single optimization. The multi-scale pairwise distortion regularizer and the importance-sampling scheme are practical ideas that could benefit future work. The paper is also transparent about its experimental setup, including a stopping criterion and ablation details. However, the central comparative claims are currently weakened by evaluation issues that need to be addressed before the paper can be accepted.","major_comments":[{"comment":"The headline claim of outperforming mesh-based baselines with respect to peak distortion is confounded by a large fidelity gap. The proposed method reports mean target-to-mesh distances of 0.253-0.651 mm across the four tasks, whereas ARAP and CeVasMap achieve 0.024-0.138 mm. For the rib cage, the proposed method has mean distance 0.651 mm versus 0.035 mm (ARAP) and 0.024 mm (CeVasMap), i.e., above the 0.5 mm pixel spacing the authors themselves use as a threshold. Because Eq. (1) is a weighted sum of target distance and distortion, the lower peak distortion can be partly purchased by leaving targets farther from the fitted manifold. A matched-fidelity comparison (e.g., constraining target distance to a common budget) or a joint distortion-fidelity cost is needed to support the claim of superior distortion.","section":"Sec. 6.1, Sup. Table 2"},{"comment":"The evaluation of distortion is partly circular: the distortion values reported in Fig. 5 and Sup. Table 2 are computed with Eq. (4), which is exactly the loss term Ld optimized during fitting, and they are measured in a 1 cm radius around the unfolded target, i.e., the region that the importance weighting (Sec. 5.2) explicitly emphasizes via ws in Eq. (4). The baselines do not optimize this loss, so the comparison is not on a neutral common yardstick. Please report an independent metric (e.g., manual assessment by clinicians, landmark-based accuracy, or area distortion measured against a ground-truth mesh) and analyze the sensitivity of the conclusions to the evaluation window size.","section":"Sec. 5.2, Sec. 6"},{"comment":"The claim that the proposed multi-scale distortion regularizer yields 'smoother transformations compared to Jacobian formulations' is not supported by the quantitative results in Sup. Table 3. The Jacobian regularizers J1 and J2 achieve much lower maximum distortion (0.957 mm and 1.384 mm, respectively) than the proposed multi-scale baseline (6.057 mm). The authors argue that J1 changes the initial plane structure and J2 stretches vessel radii, but those observations are qualitative. Please provide quantitative morphology-preservation metrics (e.g., fold counts, vessel-radius error, or landmark displacement) to substantiate the superiority of the proposed regularizer over the Jacobian-based alternatives.","section":"Sec. 6.2, Sup. Table 3"},{"comment":"The paper's claim of a 'versatile framework' is only partially supported. The PCA initialization and low-frequency embedding assume that the target point set lies near a single smooth 2-manifold; this assumption is central to the method. The experiments show its limits: the hand-bone case misses 'the last part of the tips,' and the rib-cage target distance remains above the pixel spacing. The paper would be strengthened by an explicit statement of the conditions under which the manifold assumption is violated, and by a diagnostic (e.g., PCA residual or target-distance statistics) that could tell a user whether the method is applicable to a new structure.","section":"Sec. 5.1, Sec. 7.4"}],"minor_comments":[{"comment":"The boxplot uses a logarithmic scale, but the horizontal line at the pixel spacing (0.5 mm) is near the lower end and may be visually misleading; please clarify the axis or use a linear scale for a subset of the data.","section":"Fig. 5"},{"comment":"The importance map formula VE = (|min(e(I)-α,0)|+β)/(α+β) is under-specified: please define the units of e(I) (Euclidean distance transform) and state whether α and β are in millimeters or voxels.","section":"Sec. 5.2"},{"comment":"In Eq. (4), the loss uses ||us,1 - us,2||, whereas the text earlier defines Δ as the distance between neighboring pixels; please clarify how the random sampling distances are chosen and how the loss scale relates to the image resolution.","section":"Sec. 4.2"},{"comment":"The column headers 'Ours 10 mm' and 'Ours 50 mm' are ambiguous when read in isolation; consider renaming them to 'Small-scale (10 mm)' and 'Large-scale (50 mm)' to match the terminology in Sec. 6.2.","section":"Sup. Table 3"},{"comment":"The references to Sup. Eq. (7) and Sup. Eq. (8) are clear in the supplementary, but the main text would benefit from a brief description of the sink-like function used for vessel intensity retrieval.","section":"Sec. 6.3"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and presents a genuinely novel combination of neural fields with medical image unfolding. The main issues are in the evaluation: the distortion comparison is confounded by fidelity differences and by the coincidence of the evaluation metric with the optimized loss. I believe the authors can address these concerns with additional experiments and a more cautious presentation, but the current version does not yet establish the claimed superiority over baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a genuinely new way to flatten sparse anatomical structures with neural fields, and the qualitative results look convincing. But the main quantitative claim—lower peak distortion than mesh baselines—is weaker than it appears, because the comparison lets target-to-mesh distance float. Their own Table 2 shows mean distances 2–10x larger than both baselines (e.g., cerebral vessels 0.352 vs 0.040/0.071 mm; rib cage 0.651 vs 0.035/0.024 mm). They acknowledge the baselines enforce target intersection and that they stay below pixel spacing except for the rib cage, but that caveat sits in the text while the abstract and conclusion lead with the distortion win. The lower max distortion is partly bought by not hitting the target exactly, so a matched-fidelity comparison is needed before I'd trust the headline. The 1 cm evaluation window also aligns with the importance map, which is a favorable choice for them. What is new and genuinely good: the multi-scale pairwise isometry regularizer (Eq. 4) is a clean idea that avoids the fold artifacts of Jacobian-based regularizers, and the ablation in Fig. 6 makes that point well. Combining geometric and intensity-based losses to recover missing annotations or unfold auxiliary targets is a useful extension, and the results on retrieving removed vessel segments are compelling. The framework is flexible across different anatomies (vessels, bones, pancreas, ribs) with essentially the same hyperparameters, which is a real practical advantage over organ-specific pipelines. Softer spots beyond the fidelity issue: the distortion metrics are the same functions being optimized, so part of the evaluation is a convergence check; there's no statistical testing across patients; sample sizes are small (5–10); and no code is released. These are addressable, not fatal. The paper is honestly written—they flag the distance trade-off and the missing fingertips in the hand case—and the method is clearly explained. Who should read it: anyone working on medical image unfolding or neural-field deformations. It's worth referee time. My recommendation: send to peer review, but demand a matched-fidelity evaluation (either match target-distance budgets or report distortion at equal fidelity) and add some statistical rigor to the patient-level claims.","headline":"Novel neural-field unfolding method with real clinical potential, but the headline distortion advantage is confounded by a fidelity trade-off and needs a matched-fidelity comparison before the claim holds.","tokens_in":662,"tokens_out":693,"would_cite":true,"duration_ms":24055,"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":"A neural field fitted to sparse 3D anatomical points can flatten them into a 2D overview with lower peak distortion than mesh-based unfolding baselines.","keywords":["neural fields","image unfolding","sparse structures","distortion regularization","medical visualization","tomographic imaging","point cloud parameterization","deformation field"],"falsifier":"Fit the model to a sparse point set sampled from a high-curvature 3D helix with radius comparable to the pixel spacing and the same loss weights and normalization used in the paper; if the median distance from target points to the fitted sheet exceeds the 0.5 mm pixel spacing, the single-smooth-manifold assumption—and with it the claimed versatility on complex sparse structures—fails.","tokens_in":14141,"feed_emoji":"🩻","tokens_out":8318,"duration_ms":65830,"temperature":0.7,"pith_summary":"The paper proposes a versatile, modality-independent method to unfold thin, sparse anatomical structures—such as cerebral vessels, hand bones, the pancreas with adjacent arteries, and the rib cage—into a single 2D overview image. Rather than requiring a densely sampled surface or a global ordering of points, it fits a small neural field that deforms an initial plane so that the plane intersects the sparse target points with minimal local distortion. The paper reports that this approach outperforms mesh-based baselines on maximum distortion for complex sparse structures, and that its multi-scale pairwise distortion regularizer produces smoother, fold-free transformations than Jacobian-based regularizers from neural field registration. If correct, the method gives clinicians a general tool for producing standardized, low-distortion views of structures that currently lack a versatile unfolding technique.","feed_headline":"Neural field unfolding beats mesh baselines on peak distortion","feed_subtitle":"One framework unfolds vessels, bones, pancreas, and ribs into low-distortion 2D views without point ordering.","key_machinery":"The central object is a deformation neural field: a three-layer MLP with Leaky ReLU activations and a low-frequency trigonometric embedding that maps points on an initial PCA-derived plane to positions in the tomographic volume. The load-bearing loss is the multi-scale distortion regularizer, which samples point pairs at distances spanning 0.5 to 40 mm and penalizes deviations between their 3D distance and their fixed 2D pixel distance, preventing folds and distributing inevitable distortion away from the target when weighted by an importance map derived from a Euclidean distance transform. A Chamfer-style target loss pulls the deformed plane onto the sparse point set, and an optional intensity-based loss optimizes appearance in the unfolded image.","core_discovery":"The central claim is that a pointwise-fitted neural field can serve as a general-purpose unfolding parametrization for sparse anatomical structures: starting from a PCA-aligned planar initialization, a low-frequency multi-layer perceptron learns a deformation field that pulls the plane onto the target centerline points while a multi-scale distance-preserving loss keeps the read-out image geometrically faithful. Alongside this geometric fit, an intensity-based loss recovers non-annotated vessels or maximizes the display of an auxiliary organ mask. The paper demonstrates on four CT applications that this framework yields lower maximum distortions than CeVasMap and ARAP, and that the proposed distortion regularizer leads to smoother results than the Jacobian regularizers used in prior neural field registration work.","pith_inferences":["Because the neural field is resolution-independent, the same fitted model could be re-evaluated at arbitrary output resolutions, enabling zoomable or progressive unfolding without retraining.","The importance map could also gate the image-based loss, focusing appearance optimization near the target or in peripheral areas depending on clinical need, a step the paper mentions but does not explore.","For structures that are genuinely volumetric rather than sheet-like, the single-manifold assumption will break; fitting multiple coupled neural patches or adding a thickness-aware cost would be a natural next step.","The reported trade-off of slightly larger target-to-sheet distances than mesh baselines suggests hybrid pipelines could use the neural field for coarse, low-distortion unfolding followed by a local refinement to pin exact landmarks."],"forward_implications":["A single framework can unfold sparse structures such as vascular trees, ducts, or bone systems without requiring global point ordering, removing a limitation of curved planar reformation.","The multi-scale pairwise distortion regularizer is a reusable component for any neural-field geometric task that needs smooth, fold-free deformations, independent of medical imaging.","The same default configuration works across four anatomies, so new unfolding tasks can start from a known-good setting and only adjust the normalization constant tied to target extent.","Combining geometric and intensity losses lets the unfolded image recover structures with missing annotations and display auxiliary targets, such as an organ mask, alongside the primary structure."],"supporting_citations":[{"why":"Supplies the CeVasMap baseline and the initial PCA-plane formulation that the neural field deforms.","marker":"[26]"},{"why":"Supplies the ARAP mesh-deformation baseline that the method is compared against for peak distortion.","marker":"[32]"},{"why":"Provides the Jacobian-based regularizer and neural field registration approach used as a distortion-regularization baseline.","marker":"[37]"},{"why":"Provides the second Jacobian regularizer and the hybrid sampling scheme compared in the ablation study.","marker":"[34]"},{"why":"Inspires the importance-weighting of distortion via a distance transform, adapted as the weighting factor in the distortion loss.","marker":"[13]"},{"why":"Demonstrates the point-ordering requirement of curved planar reformation, the prior standard for sparse vascular unfolding that the proposed method avoids.","marker":"[8]"}],"fun_headline_variants":["Neural field flattens sparse anatomy with less distortion","Unfold vessels, bones, pancreas with neural fields","Neural field unfolding lowers peak distortion vs meshes","Sparse structure to 2D: neural field wins on distortion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The target point set must lie close to a single smooth 2D manifold that a low-frequency neural field can parameterize with acceptable distortion; if the structure is thick, tightly folded, or consists of distant parallel sheets, the unfolded image will miss parts of it.","fun_headline_variants_meta":{"raw":{"variants":["Neural field flattens sparse anatomy with less distortion","Unfold vessels, bones, pancreas with neural fields","Neural field unfolding lowers peak distortion vs meshes","Sparse structure to 2D: neural field wins on distortion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1238,"prompt_tokens":857,"completion_tokens":381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":314}},"tokens_in":473,"tokens_out":381,"duration_ms":4038,"temperature":1.0,"reasoning_tokens":314,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:14:16.446961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the model to a sparse point set sampled from a high-curvature 3D helix with radius comparable to the pixel spacing and the same loss weights and normalization used in the paper; if the median distance from target points to the fitted sheet exceeds the 0.5 mm pixel spacing, the single-smooth-manifold assumption—and with it the claimed versatility on complex sparse structures—fails.","supporting_citations":[{"cited_title":"Flexible Unfolding of Circu- lar Structures for Rendering Textbook-Style Cerebrovascular Maps","cited_arxiv_id":null,"evidence_quote":"Supplies the CeVasMap baseline and the initial PCA-plane formulation that the neural field deforms."},{"cited_title":"As-rigid-as-possible surface modeling","cited_arxiv_id":null,"evidence_quote":"Supplies the ARAP mesh-deformation baseline that the method is compared against for peak distortion."},{"cited_title":"Implicit neural representations for deformable image registration","cited_arxiv_id":null,"evidence_quote":"Provides the Jacobian-based regularizer and neural field registration approach used as a distortion-regularization baseline."},{"cited_title":"Duncan, and Xiaohui Xie","cited_arxiv_id":null,"evidence_quote":"Provides the second Jacobian regularizer and the hybrid sampling scheme compared in the ablation study."},{"cited_title":"ADR- anatomy-driven reformation","cited_arxiv_id":null,"evidence_quote":"Inspires the importance-weighting of distortion via a distance transform, adapted as the weighting factor in the distortion loss."},{"cited_title":"CPR-curved planar refor- mation","cited_arxiv_id":null,"evidence_quote":"Demonstrates the point-ordering requirement of curved planar reformation, the prior standard for sparse vascular unfolding that the proposed method avoids."}],"review_version":1}