{"id":"7724e992-b6d3-4117-ac82-08cbc7cd38af","arxiv_id":"2506.19742","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding layer normalization between the hash encoder and the MLP, plus initializing the network from a dense-volume pretrained model, improves NeRF-based CBCT reconstruction speed and quality.","lead":"Sparse-view cone-beam CT scans often produce blurry 3D images; this paper adds two training fixes to NeRF-based reconstruction: normalizing hash features and pretraining the network on one annotated CT volume. The result is faster convergence and higher image quality on 128 cases across four datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unablated FFT channel-masking step (Sec 2.3) is part of the evaluated method but absent from the claimed LN+MCI recipe; the central attribution is unproven.","rationale":"The reader's CONDITIONAL verdict is based on possible MCI pretraining leakage. I share that concern, but the more immediate threat to the central claim is the unablated third component. The paper explicitly states that FFT channel masking improves performance, and this step is part of the proposed Normalized Hash Encoder in Sec 2.3, yet the abstract and conclusion frame the contribution as solely Normalization and Initialization. The ablation in Fig 8 is described as evaluating each modification, but no FFT-mask-only condition appears, and Table 1's method label omits FFT. Without an isolation experiment, a reader cannot tell whether LN+MCI alone reproduces the reported gains. This is an internal completeness problem rather than a dispute with external consensus, and it is directly testable by a code-level ablation. If the ablation shows the FFT mask is negligible, the paper's claim stands; if not, the recipe and title need revision. I therefore keep the reader's CONDITIONAL verdict without moving it.","tokens_in":6678,"tokens_out":6319,"duration_ms":67640,"concrete_test":"Re-run the Table 1 evaluation (at least abdomen and chest) for NAF+LN+MCI with the FFT channel-mask branch of Sec 2.3 disabled, keeping all other settings identical, and add this as a third ablation column in Fig 8. If PSNR/SSIM change by more than a small tolerance (e.g., >0.1 dB or >0.005 SSIM) on either dataset, the FFT mask is load-bearing and the central claim must be revised to include it; if not, the attribution to LN+MCI survives this check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The title and abstract claim that Layer Normalization and Mapping Consistency Initialization are what improve NAF. However, Sec 2.3 introduces a third component: \"at the channel level of the hash features, we find some channels exhibit meaningless noise. Via Fast Fourier Transform (FFT), we distinguish and mask these channels during training to improve model's performance.\" This FFT masking is not mentioned in the abstract, in the method name NAF+LN+MCI, or in the ablation of Fig 8, which reports only LN and MCI as the modifications. If the final model in Table 1 includes this masking, the comparison to NAF conflates three interventions, not two. The central causal claim that normalization and initialization alone produce the gains is therefore not established by the reported experiments. This is an internal, concrete concern: the passage itself asserts that FFT masking improves performance, yet no experiment isolates its contribution. It must be resolved before accepting the attribution, independent of the separate and also valid question of MCI pretraining-distribution leakage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses sparse-view CBCT reconstruction with hash-encoded NeRF. It attributes training instability to a local-global optimization mismatch between the hash grid and the MLP, and proposes two remedies: a Layer Normalization layer between hash features and the MLP, and a Mapping Consistency Initialization (MCI) that pretrains the MLP on a single ground-truth abdomen volume using a voxel-to-voxel loss. The method is evaluated against FDK, SART, NAF, SAX-NeRF, and R2-Gaussian on 128 CT cases from four datasets spanning seven anatomical regions, with PSNR/SSIM and segmentation Dice. The paper reports that NAF+LN+MCI converges faster and achieves higher image quality than NAF, and surpasses R2-Gaussian on most datasets with comparable runtime. The authors claim that normalization and initialization are sufficient to alleviate the local-global mismatch.","tokens_in":6902,"tokens_out":8624,"duration_ms":81805,"significance":"The proposed recipe is simple, easy to implement, and accompanied by a public code repository, which are genuine strengths. The evaluation is broad: 128 cases, four datasets, seven anatomical regions, and a clinical segmentation metric. If the attribution to LN and MCI were established, this would be a practically useful and general improvement for hash-NeRF sparse-view CBCT reconstruction. However, the current manuscript does not isolate the two claimed components because Section 2.3 introduces an additional FFT-based channel-masking step that is absent from the ablation and abstract, and the main comparison lacks statistical uncertainty quantification and leaves the pretraining-data overlap with the test distribution unclear. These issues prevent the central causal claim from being accepted as stated.","major_comments":[{"comment":"The paper introduces a third intervention that the attribution ignores. The text states that at the channel level of the hash features, FFT is used to distinguish and mask noisy channels during training to improve performance, but the abstract, the method name NAF+LN+MCI, and the ablation in Fig. 8 only mention LN and MCI. If the final model in Table 1 includes FFT masking, the comparison to NAF conflates three changes rather than two, and the claim that normalization and initialization alone produce the reported gains is not supported. Please add ablation arms that isolate FFT masking (e.g., NAF+LN, NAF+LN+FFT, NAF+LN+MCI, NAF+LN+MCI+FFT), report the corresponding Table 1 numbers, and either include FFT masking in the method name and abstract or remove it from the method description.","section":"Sec. 2.3 / Fig. 8 / Table 1"},{"comment":"The main quantitative comparison reports point estimates without error bars, confidence intervals, or significance tests. Several differences are small relative to likely run-to-run variability (e.g., Head PSNR 34.53 vs 34.09 for NAF, and GS_foot PSNR 31.64 vs 31.3 for NAF), and the conclusion that the method surpasses 3DGS depends on these means. Please report means and standard deviations over multiple runs or over the individual cases in each dataset, and perform paired significance tests for the headline comparisons.","section":"Table 1 / Fig. 6"},{"comment":"MCI pretraining uses a ground-truth attenuation volume of one abdomen case while the baselines receive no such pretraining, and the manuscript does not specify whether this case belongs to the 82 Pancreas_CT abdomen cases used in evaluation or is excluded from all test sets. If the pretraining case is drawn from the test distribution, the comparison leaks information; even if it is held out, the choice of an abdomen case from the same public dataset as the abdomen test set needs a transferability control. Please state the source of the pretraining case, verify that it is not in any test set, and report results when the pretraining volume is taken from a different anatomical region.","section":"Sec. 2.4 / Sec. 3.1 / Table 1"}],"minor_comments":[{"comment":"The word 'Sepcifically' should be 'Specifically'.","section":"Sec. 1"},{"comment":"The caption contains the typo 'Hash Gird' instead of 'Hash Grid'.","section":"Fig. 1"},{"comment":"The ablation figure reports SSIM curves for three images but does not state which cases or anatomical regions are used, nor does it provide the final SSIM values; please add this information.","section":"Fig. 8"},{"comment":"The paper uses the names R2_Gaussian and R2-Gaussian interchangeably; please unify the notation.","section":"Throughout"},{"comment":"The implementation details omit the hash-grid configuration (number of levels, feature dimension, resolution) and MLP architecture; providing these is important because the hash-grid size affects the local-global mismatch.","section":"Sec. 3.1"},{"comment":"The segment Dice scores are shown only graphically; please include the numerical values in the text or a table. Also, the phrase 'on three cases7' in Sec. 3.2 contains an unresolved footnote marker, and Fig. 7 is not explicitly discussed; please fix the reference.","section":"Fig. 6 / Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWorth a look if you follow hash-NeRF methods in sparse-view CT. The paper's framing—that the hash encoder's local sparse updates and the MLP's global dense updates create a train-time mismatch—is a genuinely useful way to talk about why hash-NeRF training is unstable. Adding LayerNorm between the hash grid and the MLP, plus initializing the MLP from a pretrained dense-volume model, speeds up NAF convergence and lifts reconstruction quality on their test bed. The gains look real, and the experiments cover 128 cases across four datasets, with public code.\n\nBut the paper's central attribution is not yet proven. The stress-test note is accurate: Sec. 2.3 introduces an FFT-based channel-masking step that is part of the evaluated method but is absent from the title, the abstract, the method name (NAF+LN+MCI), and the ablation in Fig. 8. If the final model includes FFT masking, the comparison in Table 1 conflates three interventions, not two. You cannot tell whether LN, MCI, or the masking drives the improvement. This is not a footnote; it is the load-bearing claim.\n\nSecond, MCI pretraining uses a ground-truth CT volume of one abdomen case. The paper even says the pretraining 'uses ground truth information' (Sec. 2.1) and the pretraining case comes from Pancreas_CT, which is also one of the test datasets. That is not the same as test-case leakage, but it is a supervised pretraining signal from the evaluation distribution, and no baseline is granted an equivalent. A cleaner baseline would give all methods the same initialization, or show that a random-case pretraining produces the same gains.\n\nMinor but not negligible: Table 1 reports single numbers with no error bars or significance tests, and some deltas over NAF are small (e.g., head PSNR 34.53 vs 34.09).\n\nWhat the paper does well: it names a real training dynamic, tests a simple recipe across several anatomies, and reports segmentation-based evaluation in addition to PSNR/SSIM. The implementation is public.\n\nBottom line: the hypothesis is plausible and the authors did the empirical work, but the reported experiments do not isolate the claimed mechanisms. I would still send this to peer review—the question is interesting and the paper is substantial enough to warrant referee time—but I would ask for an ablation that separates FFT masking, error bars, and a fairer initialization comparison before the central claim can be accepted.","headline":"The local-global mismatch framing is useful, but the gains are attributed to LN+MCI while an unablated FFT-masking step is folded into the evaluated method; the central claim needs an ablation rerun.","tokens_in":7416,"tokens_out":3351,"would_cite":false,"duration_ms":34201,"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":"Layer normalization and a one-case pretrained initialization fix the core training mismatch in NeRF-based CBCT reconstruction.","keywords":["CBCT reconstruction","NeRF","hash encoding","layer normalization","initialization","sparse-view CT","3D Gaussian splatting","transfer learning"],"falsifier":"Train NAF+LN+MCI with the MCI weights pretrained on a case sampled from a scanner or dataset entirely outside the four evaluation datasets, and compare against random initialization; if the quality gap mostly disappears, the transfer claim fails. Independently, record a fixed batch of hash features, run them through the MLP at two different training times, and measure the L1 drift in outputs: if adding Layer Normalization does not reduce that drift, the proposed mismatch mechanism is not the active cause.","tokens_in":1693,"feed_emoji":"🩻","tokens_out":2827,"duration_ms":72781,"temperature":0.7,"pith_summary":"The paper argues that NeRF-based cone-beam CT (CBCT) reconstruction suffers from a local-global optimization mismatch: each training step updates only a small subset of the hash encoder's parameters but all of the neural network's parameters, so the features the network sees keep shifting underneath it. To fix this, the authors insert a Layer Normalization between the hash encoder and the network, and initialize the network with weights pretrained on the dense volume of one abdomen case, calling this Mapping Consistency Initialization. They report that this few-line change makes training converge faster and raises reconstruction quality above the NAF baseline, reaching quality comparable to or better than 3D Gaussian Splatting methods while keeping a similar training time. Experiments cover 128 CT cases from four datasets and seven anatomical regions. The central claim is that normalization plus one-case initialization is a general cure for the mismatch in hash-NeRF sparse-view reconstruction.","feed_headline":"Layer norm plus one-case pretraining lifts CBCT quality beyond 3DGS","feed_subtitle":"A few lines of code make hash-NeRF CBCT training stable, faster, and more accurate across 128 cases and 7 body regions.","key_machinery":"The central mechanism is the pairing of a Normalized Hash Encoder with Mapping Consistency Initialization. The former is a Layer Normalization layer placed between the multiresolution hash grid and the MLP; it forces hash features from different subsets of the grid to share one mean and variance, so the MLP no longer sees a moving target. The latter bypasses rendering: a full NeRF is trained directly on the voxel attenuation values of one abdomen case with an $\\ell^1$ loss, and only its LN and MLP weights are copied into the reconstruction model. FFT-based channel masking of noisy hash channels is a secondary mechanism. The work of these components is to stabilize the early training dynamics that the paper identifies as the bottleneck.","core_discovery":"The paper's central discovery is that unstable training in hash-NeRF-based CBCT reconstruction is caused by a local-global optimization mismatch, and that two cheap interventions substantially reduce it: Layer Normalization on the hash features, and initialization of the MLP from a dense-volume pretraining on a single ground-truth case. The normalized hash encoder keeps the features' global mean and variance stable across training steps, and the Mapping Consistency Initialization gives the network a sensible global mapping from coordinates to attenuation values before sparse-view training begins. In the paper's experiments, the resulting method (NAF+LN+MCI) outperforms NAF on every dataset, outperforms R2_Gaussian on most datasets, and reaches convergence in about 18 minutes, matching the speed of 3DGS.","pith_inferences":["The pretraining is a pure coordinate-to-attenuation regression, so the same MCI weights should transfer to other forward models, such as different detector geometries or noise models, as long as the coordinate frame is normalized; the paper does not test this.","The paper's mechanism predicts that Layer Normalization alone should reduce the output drift of the MLP on fixed hash features in any hash-encoded scene regression, which could be verified outside CBCT, for instance in instant-NGP view synthesis.","The FFT channel-masking result suggests some hash channels carry dataset-specific noise; one could test whether the masked channels identified in the abdomen pretraining case generalize to chest and head, or whether each region needs its own mask.","If the transfer holds, this is effectively a pretraining recipe for inverse problems, and the one-case requirement could be stress-tested by measuring how reconstruction quality degrades as the pretraining case moves further in feature space from the target anatomy."],"forward_implications":["Under the same 50-view protocol, NAF+LN+MCI improves PSNR and SSIM over NAF on the chest, abdomen, and head datasets, and surpasses R2_Gaussian on most of the four R2_Gaussian cases.","The method reaches convergence in about 18 minutes, matching 3DGS speed while exceeding its reconstruction quality on most datasets.","Better Average Segment Dice scores than R2-Gaussian indicate the reconstructions preserve anatomical structure, not just pixel-level metrics.","The recipe requires only a few lines of code and no paired data at reconstruction time, so it can be dropped into existing hash-NeRF CBCT pipelines.","Ablation shows each component, Layer Normalization and Mapping Consistency Initialization, independently accelerates convergence and improves final quality.","The improvements hold across 128 cases spanning seven anatomical regions, suggesting the recipe is not tuned to one body part."],"supporting_citations":[{"why":"Supplies the NAF baseline, the hash-NeRF architecture that the paper modifies and the primary comparison method.","marker":"[26]"},{"why":"Introduces the multiresolution hash encoding, the component whose local sparse updates create the local-global mismatch.","marker":"[15]"},{"why":"Provides the R2_Gaussian 3DGS method and its dataset, the main speed competitor that the paper must match and beat in quality.","marker":"[25]"},{"why":"Defines the original NeRF architecture whose global-dense MLP mapping is the other half of the mismatch.","marker":"[14]"},{"why":"TIGRE is used to simulate the 50-view projection data for all experiments, grounding the evaluation protocol.","marker":"[3]"},{"why":"SAX-NeRF serves as a second NeRF baseline on the R2_Gaussian dataset, against which the method is compared.","marker":"[5]"}],"fun_headline_variants":["Normalize hash features and initialize MLP: stable CBCT NeRF","Local-global mismatch fixed: norm + init for hash NeRF CBCT","Two-line fix stabilizes CBCT reconstruction with NeRF","Hash encoder norm and MCI boost CBCT NeRF convergence"],"cache_read_input_tokens":9600,"weakest_assumption_plain":"The gains rest on the premise that an MLP pretrained on the ground-truth attenuation values of a single abdomen case transfers to every other case and body region, and that the improvement comes from the initialization itself rather than from the pretraining case sharing its distribution with the test datasets.","fun_headline_variants_meta":{"raw":{"variants":["Normalize hash features and initialize MLP: stable CBCT NeRF","Local-global mismatch fixed: norm + init for hash NeRF CBCT","Two-line fix stabilizes CBCT reconstruction with NeRF","Hash encoder norm and MCI boost CBCT NeRF convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000498,"raw_usage":{"total_tokens":2455,"prompt_tokens":978,"completion_tokens":1477,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":1402}},"tokens_in":594,"tokens_out":1477,"duration_ms":11256,"temperature":1.0,"reasoning_tokens":1402,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:25:55.958200+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train NAF+LN+MCI with the MCI weights pretrained on a case sampled from a scanner or dataset entirely outside the four evaluation datasets, and compare against random initialization; if the quality gap mostly disappears, the transfer claim fails. Independently, record a fixed batch of hash features, run them through the MLP at two different training times, and measure the L1 drift in outputs: if adding Layer Normalization does not reduce that drift, the proposed mismatch mechanism is not the active cause.","supporting_citations":[{"cited_title":"In: Eu- ropean conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Defines the original NeRF architecture whose global-dense MLP mapping is the other half of the mismatch."}],"review_version":2}