{"id":"d936eb22-f998-4231-b6dc-d3646c43111e","arxiv_id":"2411.19224","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A differentiable voxelgrid optimizer with the exact Siddon forward model reconstructs higher-quality CBCT volumes from sparse views than iterative and neural-field baselines on real walnut X-ray data.","lead":"This study shows that using an exact physics-based model of X-ray absorption, rather than a fast approximation, improves 3D cone-beam CT reconstruction from very few X-ray images. The method is self-supervised, runs in minutes rather than hours, and could reduce patient radiation exposure if it holds up in clinical settings.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TV regularization, not the exact Beer-Lambert renderer, appears to drive DiffVox's advantage over the classical baselines; matched TV-regularized classical baselines are needed before the headline comparison is secure.","rationale":"The paper's central claim requires that the differentiable Beer-Lambert renderer or the overall DiffVox framework, not an uncontrolled auxiliary component, explains the reported gains over classical iterative methods. The authors' own ablation in Table 1 makes this condition insecure: without TV regularization, DiffVox is worse than SIRT and Nesterov by roughly 4-7 dB PSNR, so TV is the dominant contributor to the headline margin. The classical baselines are unregularized, meaning Table 1 and Figure 3 compare a TV-regularized voxel optimizer with unregularized classical solvers. This does not make DiffVox useless; TV-regularized classical reconstruction is standard, and DiffVox may legitimately claim that its differentiable voxel representation makes such regularization easy to apply. But the abstract's 'outperforms widely used iterative algorithms' is currently too strong without a matched TV baseline. The proposed test directly separates renderer choice from regularization by giving the classical methods the same regularizer. If TV-matched classical baselines close the gap, the comparative claim reduces to 'TV helps'; if they do not, the renderer's contribution is confirmed. I do not see this as grounds for rejection, because the DiffVox-Siddon vs DiffVox-Trilinear comparison and the runtime advantage over neural fields remain useful, and the issue is addressable with released code and data. The reader's concern about the unspecified ground-truth reconstruction algorithm is valid and should be fixed by citing or stating the reconstruction protocol from reference [14], but it is secondary here because it does not explain the internal ablation pattern in Table 1. The verdict therefore remains conditional; no change from the reader's assessment is needed.","tokens_in":6800,"tokens_out":10515,"duration_ms":93132,"concrete_test":"Run matched TV-regularized classical baselines (e.g., SIRT or Nesterov with a proximal TV step, or a Chambolle-Pock TV solver) on the same two tuning walnuts, grid-searching lambda_TV and iteration count, then evaluate on the same 40-walnut test set at 5, 10, and 15 views. If any TV-regularized classical baseline reaches DiffVox-Siddon's PSNR within 1 dB or SSIM within 0.02, the claimed advantage over iterative algorithms is not attributable to the differentiable renderer; if DiffVox still leads, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1's internal ablation is the most direct threat to the central comparison. Removing TV from DiffVox lowers PSNR from 37.23/37.69 (trilinear/Siddon) to 29.25/27.41 and SSIM from 0.912/0.923 to 0.827/0.812, leaving both variants below the unregularized classical baselines SIRT (PSNR 33.0, SSIM 0.809) and Nesterov (PSNR 34.4, SSIM 0.834). The reported superiority over iterative algorithms is therefore carried largely by the TV regularizer in Eq. (4), not by the differentiable Beer-Lambert renderer. Because the classical baselines are run without TV, Table 1 conflates renderer/framework choice with regularization; adding TV to SIRT, CGLS, or Nesterov could close or invert the gap. The renderer-specific comparison (DiffVox-Siddon vs DiffVox-Trilinear) shows only a small advantage (0.923 vs 0.912 SSIM; 37.69 vs 37.23 PSNR), which underscores how much of the headline margin comes from TV. A matched TV-regularized classical baseline is the missing experiment needed to support the abstract's claim. The unspecified ground-truth reconstruction algorithm is a real evaluation risk, but it does not explain the within-paper ablation pattern and is secondary to this confound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DiffVox, a self-supervised CBCT reconstruction method that directly optimizes a voxelgrid using a differentiable X-ray renderer. It compares two forward models: Siddon's exact ray-voxel intersection method and trilinear-interpolation quadrature, trained with an L1 photometric loss plus total-variation regularization (Eq. 4). Experiments on the open walnut dataset use 40 held-out subjects, with Table 1 reporting 15-view results and Figure 3 showing 5-60 views, against classical ASTRA solvers (FDK, CGLS, SIRT, Nesterov) and two neural field baselines (NAF, SAX-NeRF). The paper reports that DiffVox, especially with Siddon's method, achieves the best 3D reconstruction and novel-view metrics in the sparse-view regime while being an order of magnitude faster than neural field approaches.","tokens_in":7061,"tokens_out":5505,"duration_ms":51155,"significance":"If confirmed, the central claim would be valuable: a simple differentiable voxelgrid with an exact ray-tracing forward model could outperform both classical iterative solvers and modern neural fields on real clinical-resolution CBCT data, with large speedups over neural fields. The paper has genuine strengths: it evaluates on real X-ray images rather than synthetic renderings, uses 40 held-out scans, reports standard errors across multiple metrics, includes an ablation, and releases code. However, the headline attribution of the gains to the exact Beer-Lambert renderer is not currently supported by the experimental design. The reported margin over classical baselines is dominated by the TV regularizer, and the ground-truth reconstruction protocol is unspecified; both issues are load-bearing for the paper's main empirical claim.","major_comments":[{"comment":"The manuscript states that 'Ground truth CBCT volumes were reconstructed using all images to produce high-resolution 501 x 501 x 501 volumes' but does not specify which algorithm, regularizer, or software produced these reference volumes. Since every SSIM, PSNR, MSE, and PCC value in the paper is computed against these volumes, unknown artifacts or forward-model bias in the reference reconstruction could systematically favor or penalize particular methods. The authors should report the reconstruction algorithm and its settings, and ideally evaluate sensitivity of the main comparisons to an alternative reference reconstruction.","section":"Section 3 (Dataset)"},{"comment":"The ablation rows in Table 1 show that removing the TV regularizer from DiffVox lowers PSNR from 37.23/37.69 to 29.25/27.41 and SSIM from 0.912/0.923 to 0.827/0.812, putting both variants below the unregularized classical baselines SIRT (PSNR 33.0, SSIM 0.809) and Nesterov (PSNR 34.4, SSIM 0.834). This indicates that the large advantage over classical baselines in Table 1 is carried primarily by the TV term in Eq. (4), not by the differentiable renderer or the voxel-based learning framework. Because the classical baselines are run without TV, the comparison conflates regularization with renderer/framework choice. I request TV-regularized versions of SIRT, CGLS, and Nesterov, with matched or tuned regularization strength and comparable iteration budgets, before the abstract's claim can be accepted.","section":"Table 1 and Eq. (4)"},{"comment":"The renderer-specific comparison, which is the cleanest test of the paper's secondary claim, shows only a modest difference between DiffVox-Siddon and DiffVox-Trilinear at 15 views (PSNR 37.69 vs 37.23; SSIM 0.923 vs 0.912), and Figure 3A shows that SAX-NeRF achieves higher PSNR at 30 and 60 views. The paper's claim that the exact Beer-Lambert implementation 'outperforms' other methods should therefore be scoped to the sparse-view regime and to the specific metrics. In addition, paired significance tests or confidence intervals for the Siddon-versus-trilinear difference would strengthen the claim, since the reported standard errors are non-negligible relative to the observed gap.","section":"Table 1 and Fig. 3A"}],"minor_comments":[{"comment":"The caption states that 'Novel views were rendered using the forward model implemented in each method'; this means the novel-view comparison partly reflects forward-model differences rather than purely reconstruction quality. Please clarify whether a common forward model was used for any of the comparisons, and discuss how this choice affects the ranking.","section":"Figure 2 caption"},{"comment":"Please clarify the number of X-ray images per orbit: the text says 3,600 images were acquired 'across three circular orbits,' and sparse subsets are taken 'from only the middle orbit,' but the per-orbit count is not stated explicitly.","section":"Section 3 (Dataset)"},{"comment":"The sentence describing trilinear interpolation as 'an approximation of second-order effects such as X-ray scatter' is speculative; the manuscript should label this as a hypothesis rather than a demonstrated property of the model.","section":"Section 2 (Methods)"},{"comment":"The 'DiffVox – Trilinear w/o TV' and 'DiffVox – Siddon's w/o TV' rows are reported only at 15 views; reporting the same ablation across the full view range would clarify whether the regularizer's contribution persists in less sparse settings.","section":"Table 1"},{"comment":"There are minor typographical issues, such as 'returns of the LAC' in the description of Eq. (2), and the inconsistent use of 'DiffV ox' with a space in the compiled text; these should be corrected in a revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a workshop-length empirical study with a strong claim about the source of its performance gains. The novelty beyond the authors' prior differentiable renderer [10] and Plenoxels-style voxelgrid optimization [9] lies mostly in the real-data evaluation and the renderer comparison; that is legitimate, but the contribution is incremental. The most pressing issue is the TV confound in Table 1, which must be resolved with matched TV-regularized classical baselines before the headline comparison is secure. The unspecified ground-truth reconstruction algorithm is also important for reproducibility and fair comparison. I encourage the editor to require these revisions rather than reject, because the core methodology is sound and the empirical setup is otherwise well executed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a well-run empirical study that does one thing cleanly and then overreaches on the headline. The clean part is the within-framework comparison of two forward models for differentiable X-ray rendering on real walnut CBCT data—exact Siddon's method versus trilinear interpolation, same optimizer, same TV regularizer. Siddon wins by a small but consistent margin, and the advantage is clearest for novel-view rendering at 5–20 views. That's a legitimate, modest result, and the paper is honest enough to report that SAX-NeRF beats both variants on PSNR at 30 and 60 views.\n\nThe paper also deserves credit for using 40 held-out real walnut scans, reporting standard errors, and shipping code. The classical baselines (FDK, CGLS, SIRT, Nesterov) are standard ASTRA implementations, and the runtime comparison against neural fields is useful.\n\nThe soft spot is real, and it's the one the stress-test flags. The abstract says DiffVox outperforms 'widely used iterative CBCT reconstruction algorithms.' But Table 1's ablation shows that removing TV drops both DiffVox variants below the unregularized classical baselines—PSNR goes from 37.23/37.69 to 29.25/27.41, while SIRT sits at 32.99 and Nesterov at 34.35. The classical baselines were run without TV. So the headline margin is carried by the TV regularizer, not by the differentiable Beer-Lambert renderer. The renderer-specific gain (0.923 vs 0.912 SSIM; 37.69 vs 37.23 PSNR) is small. To support the claim, you need matched TV-regularized classical baselines—SIRT+TV or Nesterov+TV. The paper itself acknowledges that an initialization with Nesterov could improve DiffVox, which makes the omission more puzzling.\n\nThe second issue is the ground truth. The 501^3 reference volumes are described only as 'reconstructed using all images,' with no algorithm named. Every metric is computed against those volumes, so an unknown reconstruction bias could affect the comparisons. This is a real evaluation risk, though secondary to the TV confound. Minor: training details for the neural-field baselines are thin, but the reported differences are large enough that this is unlikely to change the ordering.\n\nWho should read it? People working on sparse-view CBCT or differentiable rendering for tomography. It deserves a serious referee, but the revision needs matched regularized baselines and a specified ground-truth reconstruction before the abstract's claim can be taken at face value.","headline":"A clean within-framework comparison of Siddon vs trilinear rendering on real CBCT data, but the headline advantage over classical algorithms is mostly TV regularization, not the renderer.","tokens_in":7644,"tokens_out":2636,"would_cite":true,"duration_ms":21232,"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":"DiffVox claims that directly optimizing a regularized voxel grid with an exact differentiable Beer-Lambert X-ray renderer reconstructs higher-fidelity cone-beam CT volumes from fewer views than iterative algorithms and neural fields…","keywords":["cone-beam CT","sparse-view reconstruction","differentiable X-ray rendering","voxel grid optimization","Beer-Lambert law","Siddon's method","total variation regularization","real X-ray dataset"],"falsifier":"Take a phantom with a known ground-truth attenuation volume, render sparse X-ray views with the same acquisition geometry used for the walnuts, and reconstruct with DiffVox and the baselines; if DiffVox's advantage over SIRT and Nesterov shrinks or disappears on this exact ground truth, the real-data result depends on the undisclosed reference reconstruction rather than on the forward model.","tokens_in":6560,"feed_emoji":"🩻","tokens_out":8088,"duration_ms":62757,"temperature":0.7,"pith_summary":"DiffVox is a self-supervised framework that reconstructs cone-beam CT (CBCT) volumes by directly optimizing a voxel grid of linear attenuation coefficients, using a differentiable X-ray renderer as the forward model. The central claim is that using an exact implementation of the discrete Beer-Lambert law (Siddon's method) in the renderer produces higher-fidelity 3D reconstructions and novel 2D views than widely used iterative algorithms such as SIRT and Nesterov, and than modern neural-field approaches, in the sparse-view regime of 5 to 60 input X-rays. The appeal is practical: if true, the same diagnostic quality can be obtained from fewer X-rays, reducing radiation exposure, and the voxel representation is fast enough (tens of minutes) to be used on real clinical-resolution data, unlike neural fields that take hours to days. The paper evaluates on a dataset of 42 walnuts with 3,600 real X-ray images per walnut, reconstructing from sparse subsets and comparing against several classical and neural baselines.","feed_headline":"Exact X-ray line integrals beat neural fields for sparse CBCT","feed_subtitle":"DiffVox reconstructs high-fidelity volumes from a handful of views in minutes, versus hours for neural-field rivals.","key_machinery":"The central object is the differentiable X-ray renderer implementing Siddon's method: an exact discretization of the Beer-Lambert law in which each detector pixel's negative log intensity equals the sum, over all ray-voxel intersections, of the linear attenuation coefficient (LAC) times the intersection segment length. The renderer is contrasted with a faster approximation that samples evenly spaced points along the ray and uses trilinear interpolation of the LAC volume. Both renderers are embedded in a photometric L1 loss plus total-variation regularization, optimized with Adam directly on a zero-initialized voxel grid whose values are passed through a softplus activation to enforce nonnegativity. Siddon's exact line integrals carry the argument by giving the optimizer physically accurate gradients from sparse X-ray measurements.","core_discovery":"The paper's discovery is that the numerical integration scheme inside the differentiable renderer is the decisive factor for sparse-view CBCT reconstruction quality. Siddon's method, which exactly integrates the Beer-Lambert attenuation along each ray by weighting each intersected voxel's attenuation coefficient by the geometric intersection length, consistently produces sharper reconstructions and more accurate novel views than trilinear-interpolation quadrature, with the largest gains in the 5–20 view regime. On 15-view reconstructions of held-out walnuts, DiffVox with Siddon's method reaches SSIM 0.923 and PSNR 37.69, outperforming the best neural-field baseline (SSIM 0.835, PSNR 33.43) and all classical algorithms. The paper further shows that total-variation regularization is essential to both renderers, and that removing it degrades Siddon's method more than trilinear interpolation because the exact line integral confines each pixel's gradient to voxels on the ray, while trilinear interpolation lets supervision leak to neighboring voxels, acting as a crude scattering approximation.","pith_inferences":["It is a plausible but untested extension that the same exact Beer-Lambert renderer could improve neural-field reconstructions, since the paper's ablation isolates the forward model as the main driver rather than the voxel representation itself.","Because the renderer is differentiable with respect to geometry as well as the volume, the framework could jointly refine unknown X-ray source poses during reconstruction; the authors mention this as future work but do not demonstrate it.","The observation that trilinear interpolation behaves as a mild scatter-like regularizer suggests a tunable hybrid: weighting the exact Siddon gradient and the interpolated gradient could trade off sharpness and robustness in ultra-sparse settings.","The reported runtimes assume a large-batch GPU setting (up to 1.8 million rays per batch for trilinear); clinical adoption at higher resolution may require gradient checkpointing or streamed optimization."],"forward_implications":["At 15 input views, DiffVox with Siddon's method achieves SSIM 0.923 and PSNR 37.69 on real clinical-resolution walnut data, above all classical and neural baselines tested.","Reconstruction with Siddon's method takes roughly 11 minutes at 15 views and 45 minutes at 60 views, an order of magnitude faster than neural-field baselines that require hours to days.","Siddon's method yields the most accurate novel 2D X-ray views in the 5–20 view regime, which is the range most relevant for reducing patient radiation dose.","Total-variation regularization is necessary for both renderers, and its removal hurts the exact Siddon renderer more than the trilinear approximation.","The only regime where a baseline (SAX-NeRF) beats DiffVox is PSNR at 30 and 60 views for 3D reconstruction; DiffVox wins all other view counts and metrics."],"supporting_citations":[{"why":"supplies FDK, the classical analytical CBCT baseline that DiffVox must outperform.","marker":"[1]"},{"why":"supplies SIRT, a widely used iterative baseline in the comparison.","marker":"[3]"},{"why":"SAX-NeRF, the strongest neural-field baseline, which DiffVox outperforms on most view counts.","marker":"[8]"},{"why":"provides the differentiable implementation of Siddon's method that makes voxel-grid optimization possible.","marker":"[10]"},{"why":"gives the total-variation regularizer that shapes the reconstruction loss.","marker":"[11]"},{"why":"Adam optimizer used to update the voxel grid.","marker":"[12]"},{"why":"defines the exact discrete Beer-Lambert line-integral computation that is the paper's central forward model.","marker":"[13]"},{"why":"provides the real walnut X-ray dataset and the high-resolution reference volumes used for evaluation.","marker":"[14]"},{"why":"offers the GPU-accelerated implementations of the classical baseline algorithms.","marker":"[15]"}],"fun_headline_variants":["Exact ray integration wins for sparse-view CT","Siddon's ray tracing sharpens sparse CBCT","Fewer views, better CT with exact line integrals","Sparse CT boosted by exact Beer-Lambert integration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's quality metrics are computed against high-resolution reference volumes reconstructed from all 3,600 X-ray images per walnut, but the paper does not state which algorithm produced those references; if those references contain artifacts or forward-model bias, the measured advantage of DiffVox could be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Exact ray integration wins for sparse-view CT","Siddon's ray tracing sharpens sparse CBCT","Fewer views, better CT with exact line integrals","Sparse CT boosted by exact Beer-Lambert integration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000147,"raw_usage":{"total_tokens":1178,"prompt_tokens":928,"completion_tokens":250,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":186}},"tokens_in":544,"tokens_out":250,"duration_ms":3111,"temperature":1.0,"reasoning_tokens":186,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:23:59.413630+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a phantom with a known ground-truth attenuation volume, render sparse X-ray views with the same acquisition geometry used for the walnuts, and reconstruct with DiffVox and the baselines; if DiffVox's advantage over SIRT and Nesterov shrinks or disappears on this exact ground truth, the real-data result depends on the undisclosed reference reconstruction rather than on the forward model.","supporting_citations":[{"cited_title":"Practical cone-beam algorithm","cited_arxiv_id":null,"evidence_quote":"supplies FDK, the classical analytical CBCT baseline that DiffVox must outperform."},{"cited_title":"Computational analysis and improvement of SIRT","cited_arxiv_id":null,"evidence_quote":"supplies SIRT, a widely used iterative baseline in the comparison."},{"cited_title":"Structure-aware sparse-view X-ray 3D reconstruction","cited_arxiv_id":null,"evidence_quote":"SAX-NeRF, the strongest neural-field baseline, which DiffVox outperforms on most view counts."},{"cited_title":"Fast auto-differentiable digitally reconstructed radiographs for solving inverse problems in intraoperative imaging","cited_arxiv_id":null,"evidence_quote":"provides the differentiable implementation of Siddon's method that makes voxel-grid optimization possible."},{"cited_title":"Fast calculation of the exact radiological path for a three-dimensional CT array","cited_arxiv_id":null,"evidence_quote":"defines the exact discrete Beer-Lambert line-integral computation that is the paper's central forward model."},{"cited_title":"A cone-beam X-ray computed tomography data collection designed for machine learning","cited_arxiv_id":null,"evidence_quote":"provides the real walnut X-ray dataset and the high-resolution reference volumes used for evaluation."},{"cited_title":"Fast and flexible X-ray tomography using the ASTRA toolbox","cited_arxiv_id":null,"evidence_quote":"offers the GPU-accelerated implementations of the classical baseline algorithms."}],"review_version":1}