{"id":"b5c0932d-fcbf-4e2f-ac6d-35148c49ca99","arxiv_id":"2607.14415","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"K-NeAS extends NeAS to an arbitrary number of materials via a shared latent backbone, a differentiable sequential occupancy selector, and GMM-derived attenuation bounds, improving 3D PSNR on multiple CBCT datasets.","lead":"This paper presents K-NeAS, an extension of neural-attenuation-surface CT reconstruction that models multiple tissue materials with a shared neural backbone and a soft, differentiable material selector, plus GMM-based automatic attenuation bounds. It reports higher 3D reconstruction fidelity than the single/two-material NeAS baseline on abdominal, chest, and foot CBCT scans, especially at 5–10 projection views, though it fails on jaw scans.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The +1.88 dB 3D PSNR headline claim is unverifiable: Section 3 says evaluation is limited to held-out projections, yet no reference volume for 3D PSNR is specified.","rationale":"The reader's CONDITIONAL verdict is reasonable: the paper is promising but needs more evidence. However, the reader's weakest_assumption focused on GMM attenuation bounds, which is a real limitation but one the authors explicitly acknowledge and which only clearly affects the Jaw scene. I find a more load-bearing issue: the principal advertised metric (3D PSNR) is never tied to a defined reference volume. This is an internal inconsistency—Section 3 limits evaluation to held-out projections, yet Tables 1 and 2 report 3D volume-domain metrics. The +1.88 dB improvement is the headline number and depends entirely on this undefined 3D PSNR. The 2D projection metrics are better defined and may well support the method, so I do not recommend rejection; rather, the condition on the paper should explicitly require a specification and independent recomputation of the 3D PSNR reference. Since the verdict remains CONDITIONAL, I mark it UNCHANGED.","tokens_in":6810,"tokens_out":8116,"duration_ms":78973,"concrete_test":"Ask the authors to specify exactly how 3D PSNR is computed: the reference volume (scanner CT, NAF full-view reconstruction, FDK, etc.), coordinate alignment, masking, and code path. Then, from released checkpoints, recompute Table 1's Abdomen 3M vs NeAS 1M 3D PSNR using that protocol, and also recompute it against an independent reference (e.g., 50-view FDK). If the reference volume cannot be specified, or if the +1.88 dB gain does not survive an independent reference, the central scalability claim should be downgraded or restricted to 2D projection metrics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claimed result is the +1.88 dB 3D PSNR improvement on Abdomen (abstract, Table 1). But Section 3 states: 'As real clinical scans lack ground-truth geometry, evaluation is limited to image quality metrics on held-out projections.' No definition is given for the reference volume used to compute 3D PSNR/SSIM. A 3D PSNR requires a ground-truth/reference volume; without naming it, the number is not reproducible. If the reference is a 50-view NAF reconstruction, the comparison is circular (the same projections train the model and define the target). If it is some other reconstruction, the choice of reference algorithm can dominate the metric. This is more fundamental than the acknowledged GMM miscalibration: even in Abdomen, where the GMM is not the reported failure mode, the central quantitative claim cannot be independently checked. The 2D projection-domain metrics are well-defined, but the headline +1.88 dB is a 3D volume claim, so the paper's key advertised result currently rests on an unspecified endpoint.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes K-NeAS, an extension of the NeAS implicit neural representation for sparse-view cone-beam CT reconstruction. It replaces NeAS's two-material hard selector with a shared latent backbone and K lightweight prediction heads, aggregates per-material attenuation via a soft sequential occupancy filter, automates attenuation bounds by fitting a Gaussian mixture model to a single-material prior, and adds a scheduled floater regularization loss. Experiments on four clinical CBCT regions (Abdomen, Chest, Foot, Jaw) report 2D and 3D PSNR/SSIM for K=1..4 and for 5/10/20/50 views, claiming improvements over NeAS, notably a +1.88 dB 3D PSNR gain on Abdomen at K=3.","tokens_in":7186,"tokens_out":4542,"duration_ms":46835,"significance":"The architectural ideas are reasonable and address a real limitation of NeAS: scaling beyond two materials with a differentiable selector and a shared backbone is a natural step, and the GMM-based bound automation reduces manual tuning. The floater regularization is a sensible addition for sparse-view artifacts. However, the central quantitative claims are not currently verifiable because 3D PSNR/SSIM are reported without specifying the reference volume, despite the manuscript stating that evaluation is limited to held-out projections. The soft selector weights are not normalized and thus do not form a probability distribution as claimed. Moreover, the reported results contradict several blanket claims in the abstract: Jaw is a consistent failure case and the 50-view Chest result favors the baseline. If the empirical comparisons were fully specified and claims appropriately qualified, the paper would make a useful contribution, but the current evidence base is insufficient for acceptance.","major_comments":[{"comment":"3D PSNR and 3D SSIM are reported, but the reference volume is never defined. Section 3 explicitly says 'As real clinical scans lack ground-truth geometry, evaluation is limited to image quality metrics on held-out projections.' A 3D PSNR requires a reference volume; without naming it, the +1.88 dB Abdomen headline is not reproducible. If the reference is a 50-view reconstruction, the comparison can be circular; if it is some other volume, the choice can dominate the metric. Specify the exact reference volume and the 3D metric computation.","section":"Section 3, Table 1"},{"comment":"The weights w_i(x) in Eq. (2) are not normalized. For K=2, w_0 = Ω_0(1−Ω_1) and w_1 = Ω_1, so the sum is Ω_0 + Ω_1 − Ω_0Ω_1, which is not 1 in general. Therefore Eq. (3) is not an expected attenuation under a probability distribution, despite the text calling it a 'probability distribution' and 'expected attenuation'. This is load-bearing because the selector's semantics and its comparison to the hard selector rely on the weighting. Provide a normalization or justify the unnormalized 'occupancy' interpretation.","section":"Eq. (2)-(3)"},{"comment":"Table 1 states 'Results are the average of 3 training runs' but no variance, standard deviation, or error bars are reported. The headline +1.88 dB gain and the comparisons across K values therefore cannot be assessed for statistical significance, especially since Fig. 3 is described as 'evaluated on a single run' and Table 3 uses a single run. Add error bars or per-run values for all reported metrics.","section":"Table 1"},{"comment":"The abstract claims K-NeAS 'successfully scales to arbitrary material counts' and 'exhibits enhanced robustness under sparse-sampling conditions, outperforming baseline 3D PSNR by up to 1.17 dB under 5- and 10-view constraints.' These claims are too broad. Table 1 shows K-NeAS underperforms NeAS on all metrics in the Jaw scene for every K, which the paper itself attributes to GMM miscalibration. Table 2 shows at 50 views K-NeAS is worse than NeAS in 2D PSNR (45.672 vs. 46.609) and 3D PSNR (31.784 vs. 32.171). The robustness claim cherry-picks 5- and 10-view results. Qualify the abstract and conclusion to reflect these failure cases.","section":"Abstract vs. Tables 1 and 2"}],"minor_comments":[{"comment":"The GMM fitting procedure is underspecified: how is the number of Gaussian components selected for each K (1..4)? How is the attenuation histogram sampled from the converged single-material prior? Provide implementation details or pseudocode so the step is reproducible.","section":"Section 2.2"},{"comment":"The loss weights λ_aux=0.1 and λ_reg=0.01 and the 20% floater-loss schedule are given without sensitivity analysis. Since the floater loss is a contribution, report at least a small ablation over these hyperparameters.","section":"Section 2.2, Eq. (5)"},{"comment":"The component ablation on Foot reports 'a single representative training run,' so the 0.1–0.5 dB differences between configurations may be within run-to-run noise. Report multiple runs or at least confidence intervals.","section":"Table 3"},{"comment":"The code availability section only promises future release. Provide a repository link or a detailed configuration file at submission time to enable independent verification.","section":"Code and Data Availability"},{"comment":"Qualitative comparisons in Fig. 3 are shown only for 2-material configurations. Since the main claim concerns K=3 on Abdomen, include a qualitative comparison at K=3.","section":"Section 4.1, Fig. 3"},{"comment":"Minor typographical issues: 'Weevaluateontheclinical' in Section 3 should be split into words; 'arbitrary material counts' should be 'configurable material counts' since only K=1..4 are tested; and Eq. (1) would benefit from an explicit statement of the sigmoid-domain convention for d.","section":"General"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"K-NeAS is a reasonable, incremental step beyond NeAS: it replaces the two-material hard selector with a differentiable sequential soft selector, shares a latent backbone across K lightweight heads, and adds a GMM to set attenuation intervals automatically. The parameter growth from K=1 to K=4 is negligible (17k parameters), which is a nice practical point. The flolater loss is a pragmatic addition. On the Abdomen, Chest, and Foot scenes, K-NeAS at K=3 does beat single-material NeAS by a useful margin, and the ablations show the components contribute positively. That part is genuinely worth publishing if the evaluation is solid.\n\nThe problem is the evaluation. Section 3 explicitly says real clinical scans lack ground-truth geometry and evaluation is limited to held-out projections. Yet Table 1 and the abstract report 3D PSNR and 3D SSIM. No reference volume is named anywhere. You cannot compute a 3D PSNR without a reference volume; whichever reconstruction you use as the target dominates the result. The stress-test note is right: the +1.88 dB headline claim is unverifiable as written. This is not a minor omission; it affects the paper's central quantitative claim. The 2D projection metrics are solid, but the 3D claim needs either a clear reference (e.g., a high-view analytical reconstruction) or removal from the abstract until it is defined.\n\nSecond, the robustness claim is overstated. The abstract says K-NeAS outperforms baseline 3D PSNR by up to 1.17 dB under 5- and 10-view constraints. That is true for those two sparsity levels, but at 50 views (the main setting), K-NeAS is actually worse than NeAS on Chest in both 2D and 3D PSNR (Table 2). The Jaw scene is worse across the board, which the authors honestly attribute to GMM boundary miscalibration. So 'scales to arbitrary material counts' and 'enhanced robustness' are too broad. The GMM failure is acknowledged in the conclusion, and that is a real limitation, not a fatal flaw.\n\nMissing error bars on the 3-run averages are also a concern—especially given the variance implied by the single-run ablation table. The non-normalized soft selector may deserve a comment, but it is a standard occupancy-style weighting and not the main issue.\n\nWho should read this: researchers working on sparse-view CT reconstruction or neural implicit surfaces for medical imaging. It deserves a serious referee, but the referee should ask for the reference volume definition, error bars, and a toned-down abstract before it can be accepted.\n\nWould I bring it to the reading group? Maybe, if the group cares about medical imaging. I would not cite it yet, because the headline number is not reproducible and the robustness claim does not hold at the standard 50-view setting. The thinking is clear and honest, but the evidence is not yet solid enough.","headline":"Solid incremental extension of NeAS from 2 to K materials, with a useful soft selector and GMM bounds, but the headline 3D PSNR numbers rest on an unspecified reference volume and the robustness claims are contradicted by the paper's own tables.","tokens_in":7595,"tokens_out":2188,"would_cite":false,"duration_ms":23103,"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":"K-NeAS extends implicit neural CT reconstruction from two materials to arbitrary K, using a differentiable soft selector and GMM-derived attenuation bounds, and reports up to +1.88 dB 3D PSNR on abdominal scans.","keywords":["sparse-view CT reconstruction","implicit neural representation","signed distance field","cone-beam CT","multi-material segmentation","Gaussian mixture model","differentiable volume rendering","floater regularization"],"falsifier":"Take a CT volume whose attenuation histogram has three well-separated tissue peaks, train K-NeAS at K=3, and record 3D PSNR. Then recompute K-NeAS with GMM boundaries deliberately shifted a few percent off each valley. The central claim predicts fidelity should degrade with shifted boundaries; if it does not, the hard interval constraint is not the active mechanism. The Jaw results already provide one natural example where the unshifted GMM fails.","tokens_in":6728,"feed_emoji":"🩻","tokens_out":6074,"duration_ms":61317,"temperature":0.7,"pith_summary":"Sparse-view CT reconstruction must recover a 3D attenuation field from a handful of X-ray projections, and geometry-aware methods that couple neural signed distance fields with attenuation have been limited to two materials and manually tuned attenuation bounds. K-NeAS claims to remove both limits. It replaces per-material networks with a shared latent backbone plus one lightweight head per material, and resolves pointwise material membership with a fully differentiable sequential soft selector that works for any number of tissues. It also fits a Gaussian mixture model to a single-material prior to set each material's attenuation interval automatically, and adds a scheduled floater loss against spurious geometry. On four clinical cone-beam CT datasets, the paper reports that K=3 materials gives the best volumetric fidelity on abdomen, chest, and foot, with +1.88 dB 3D PSNR over the single-material baseline on abdomen, while the high-contrast Jaw scene remains a failure case traced to GMM boundary miscalibration.","feed_headline":"Multi-material CT from sparse X-rays gains 1.88 dB with K-NeAS","feed_subtitle":"A shared neural backbone and soft K-material selector reconstruct up to four tissues without manual attenuation tuning.","key_machinery":"The load-bearing mechanism is the K-material sequential soft selector: for each material i, the membership weight is the sigmoid occupancy of its SDF surface, Ω(d_i,s), multiplied by the probability that all denser materials are absent, ∏_{j>i}(1−Ω(d_j,s)). This keeps the final attenuation µ(x)=Σ_i µ_i w_i(x) fully differentiable and independent of K. Two supporting components carry the automation: a Gaussian mixture model over the attenuation histogram of a converged single-material model, whose valley boundaries define hard per-material attenuation intervals, and a scheduled floater loss that pushes air rays to zero during early training. The shared backbone plus per-material heads keeps a","core_discovery":"The central discovery is that the two-material ceiling of previous geometry-aware CT reconstruction is not a fundamental limit: by replacing independent material networks with one shared latent backbone and K lightweight prediction heads, and by resolving material membership with the fully differentiable sequential weight w_i(x)=Ω(d_i,s)∏_{j>i}(1−Ω(d_j,s)), a single pipeline can reconstruct an arbitrary number of tissues end-to-end. On clinical cone-beam CT, K-NeAS reports that K=3 gives the best 3D PSNR on abdomen (33.28 dB vs. 31.40 dB for the single-material baseline), matches or exceeds the baseline on chest and foot, and improves sparse-view performance at 5 and 10 views. The paper attr","pith_inferences":["The Jaw failure suggests a concrete test: replace the GMM's valley detection with a constrained or interactive interval estimator; if the architecture recovers on high-contrast scenes, the ceiling is in the bounding step, not the selector.","Because the selector orders materials by ascending attenuation, scenes with overlapping or alternating tissue attenuation ranges would need a different composition rule; the current sequential discount assumes a nested or ordered tissue structure.","The reported cranial CT collapse (volumetric noise and surface boundary collapse) implicates the hash encoding and ray sampling on extreme bone/soft-tissue contrast, pointing to encoding-level fixes rather than material-count scaling.","Applying K-NeAS to phantom datasets with known ground-truth geometry could quantify segmentation accuracy directly, since the clinical data used here only supports projection-domain image quality metrics."],"forward_implications":["If the central claim holds, manual scene-specific attenuation bounds are no longer required for multi-material sparse-view CT reconstruction.","Material count becomes a tunable hyperparameter: adding a K=3 head improves volumetric PSNR on abdomen, chest, and foot, and K=4 remains competitive without significant parameter or memory growth.","Because the selector is differentiable over overlapping SDFs, the same loss can drive per-material surface extraction, enabling tissue segmentation as a byproduct of reconstruction.","Sparse-view robustness improves at 5 and 10 views (up to 1.17 dB 3D PSNR on chest), which matters for dose reduction."],"fun_headline_variants":["K-NeAS lifts sparse CT to 33.28 dB with multi-material stack","Shared backbone and soft K-selector: CT beyond two materials","Automated CT reconstruction now handles any tissue count","Sparse-view CT improves 1.88 dB via neural SDF multi-material model"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the Gaussian mixture model fitted to a converged single-material model's attenuation histogram finds the true per-material attenuation intervals; K-NeAS hard-constrains every material head to those intervals, and the Jaw scene shows that when the GMM misses a density valley, all material configurations underperform the two-material baseline.","fun_headline_variants_meta":{"raw":{"variants":["K-NeAS lifts sparse CT to 33.28 dB with multi-material stack","Shared backbone and soft K-selector: CT beyond two materials","Automated CT reconstruction now handles any tissue count","Sparse-view CT improves 1.88 dB via neural SDF multi-material model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001128,"raw_usage":{"total_tokens":4571,"prompt_tokens":832,"completion_tokens":3739,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":3661}},"tokens_in":576,"tokens_out":3739,"duration_ms":26853,"temperature":1.0,"reasoning_tokens":3661,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T02:07:26.288729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a CT volume whose attenuation histogram has three well-separated tissue peaks, train K-NeAS at K=3, and record 3D PSNR. Then recompute K-NeAS with GMM boundaries deliberately shifted a few percent off each valley. The central claim predicts fidelity should degrade with shifted boundaries; if it does not, the hard interval constraint is not the active mechanism. The Jaw results already provide one natural example where the unshifted GMM fails.","supporting_citations":[],"review_version":1}