{"id":"714957de-ff1d-40de-84db-0cc51864d760","arxiv_id":"2501.13128","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A half-quadratic splitting algorithm with a CNN denoiser, trained memory-efficiently per iteration, gives higher PSNR/SSIM than FDK, MBIR, and U-Net on sparse-view Walnut CBCT data.","lead":"This paper proposes a learned iterative reconstruction for sparse-view cone-beam CT that alternates a convolutional neural network with a physics-based data-consistency step. On the public Walnut dataset it reports better PSNR and SSIM than FDK, MBIR, and U-Net, while reducing runtime by more than an order of magnitude versus MBIR.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training and evaluation both hinge on the same 1200-view FDK target; without an independent reference, the claimed margin over MBIR may reflect target choice rather than recovery quality.","rationale":"The reader identified the FDK-1200 training target as the weakest assumption, and I agree that the supervision signal is the most load-bearing part of the argument. I sharpen the concern by noting that the evaluation metrics in Fig. 3 are reported without stating their reference; if the reference coincides with the training target, the comparison is not merely vulnerable to FDK artifacts but circular. This makes the central 'outperforms MBIR' claim underdetermined by the reported evidence. I do not treat this as an internal inconsistency in the algorithm — the HQS iteration and conjugate-gradient step are mathematically standard, and the memory-efficient training scheme is a plausible contribution. The issue is that the empirical evidence, as presented, does not establish that the quality advantage is about recovering the object rather than about mimicking a chosen target. A single controlled comparison against an independent reference, or a retraining with a different target, would resolve it. Since the reader's verdict was already CONDITIONAL and this concern is consistent with that condition, I recommend leaving the verdict unchanged rather than moving it to REJECT, because the flaw is empirical and fixable, not a demonstrated failure.","tokens_in":12343,"tokens_out":7158,"duration_ms":76249,"concrete_test":"Recompute the Fig. 3 comparison for all methods using an independent reference that is not the training target — e.g., a 1200-view MBIR reconstruction — over the full 450×380×380 volume, and separately retrain the proposed method with Eq. (9)'s ground truth changed from xFDK to xMBIR while keeping the same sparse-view inputs and evaluation protocol. The concern is settled if the proposed method still beats MBIR by a comparable margin under both changes; it lands if the advantage collapses, reverses, or is confined to a few manually selected patches.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that the proposed unshared-weight HQS method 'outperforms all the other methods' on the Walnut sparse-view CBCT benchmark. The evidence for this is the PSNR/SSIM comparison in Fig. 3, but the paper never states what reference image those metrics are computed against. In §III-A the training loss in Eq. (9) uses the 1200-view FDK reconstruction xFDK as ground truth, and in Fig. 3 the metrics are reported only on zoomed patches. If the evaluation reference is the same xFDK, the comparison is circular: the CNN is explicitly trained to output xFDK, while MBIR minimizes a different objective (data fidelity plus a hand-designed regularizer). The roughly 0.5 dB advantage over MBIR then establishes only that the learned method mimics the training target better, not that it recovers the object more faithfully. Moreover, 1200-view FDK is not artifact-free: cone-beam artifacts away from the midplane and possible cupping artifacts would be learned and reproduced by the CNN, and the data-consistency step in Eq. (7) cannot remove errors that the denoiser consistently injects. This concern is distinct from missing comparisons with deep-unrolled methods: it attacks the validity of the supervision/evaluation signal itself, and therefore the interpretation of the headline comparison. The run-time claim in Table I is less vulnerable, since it is a measured computational cost, but the quality advantage that makes the speed meaningful is exactly what is in question.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a learnt half-quadratic splitting (HQS) algorithm for sparse-view industrial cone-beam CT reconstruction. The method alternates a 2D CNN denoiser, trained separately per iteration, with a conjugate-gradient data-consistency step, and uses unshared per-iteration weights. Experiments on the public Walnut CBCT dataset compare the proposed method against FDK, MBIR, a single-step U-Net, and a shared-weights variant, reporting PSNR/SSIM on zoomed patches and run-times. The paper claims the proposed method outperforms all compared methods and is more than an order of magnitude faster than MBIR.","tokens_in":12679,"tokens_out":6766,"duration_ms":66859,"significance":"If the empirical claims hold, the method offers a practical middle ground between computationally expensive MBIR and single-step deep learning: it retains a physics-based data-consistency block, yet requires few outer iterations and avoids the large memory footprint of end-to-end deep unrolling. The separate per-iteration training of 2D CNNs is a sensible strategy for large 3D volumes, and the use of a public benchmark is a strength. However, the central claim is currently supported by an evaluation whose reference image is not stated, and the comparison omits deep-unrolled baselines that the paper itself identifies as the relevant alternative class.","major_comments":[{"comment":"The reference image used to compute PSNR/SSIM is never specified. In Eq. (9) the training target is the 1200-view FDK reconstruction xFDK; if the same xFDK is also the evaluation reference, the comparison is circular for the learnt component because the CNN is explicitly trained to output that image, while MBIR minimizes a different objective. If a different reference is used, it must be identified. The reported PSNR for FDK-1200 (about 10 dB) is implausibly low for a dense-view reconstruction against any sensible reference, which suggests either an error or an unusual reference; this must be corrected. The central claim of superior reconstruction quality requires evaluation against an independent reference (for example, a high-quality iterative reconstruction from all 1200 views, or a simulation with known ground truth) and full-volume metrics, not only zoomed patches.","section":"Section III-B, Fig. 3"},{"comment":"The comparison set contains only FDK, MBIR, U-Net, and the shared-weights variant. The introduction explicitly discusses deep-unrolled (DU) networks [7, 8, 9] as a competing class and motivates the proposed method as a memory-efficient alternative to them. The abstract and conclusion nevertheless claim that the proposed method \"outperforms all the other methods\" or \"outperforms the state-of-the-art methods.\" At least one DU baseline (for example, FISTA-Net [7] or the method in [9]) should be included in the comparison, or the claim should be restricted to the methods actually evaluated.","section":"Section III-B and abstract"},{"comment":"The empirical evidence consists of six test scans with a single PSNR/SSIM number per method and no error bars or statistical analysis. The reported advantages over MBIR are on the order of 0.2 to 0.5 dB, so without an assessment of variance it is difficult to judge whether these differences are significant. The authors should report error bars over the test volumes (and, if feasible, over multiple training runs) and should also report full-volume metrics, not only metrics on unspecified zoomed patches.","section":"Section III-B"}],"minor_comments":[{"comment":"Equation (4) uses A^T while Eq. (7) uses A^H; the notation should be made consistent.","section":"Eqs. (4) and (7)"},{"comment":"The choice K = 3 is empirical and no convergence analysis is provided. Since the shared-weights variant degrades with iterations (Fig. 4), the paper should clarify that the method is a learned unrolled scheme and that no convergence to a fixed point of the HQS problem is claimed.","section":"Section III-A"},{"comment":"The table lists FDK at 4 s and U-Net at 5 s, but the text says \"U-Net is much faster compared to the other two\" (referring to MBIR and the proposed method). FDK is also fast, so the wording should be adjusted to avoid confusion.","section":"Table I"},{"comment":"The paper does not state whether code or trained models will be released; please add a reproducibility statement.","section":"General"},{"comment":"The description of the figure says PSNR/SSIM are reported for \"zoomed patches,\" but the criteria for selecting those patches are not given. This should be specified so the reader can interpret the quantitative comparison.","section":"Fig. 3"},{"comment":"The phrase \"outperforms other methods\" should be qualified by naming the specific comparison set and dataset, as the current wording implies a broader claim than the experiments support.","section":"Abstract and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The most pressing issue is the missing definition of the evaluation reference. If the reference is indeed the FDK-1200 training target, the central comparison would be circular, which would substantially weaken the paper. I would recommend asking the authors to clarify this explicitly and to provide at least one independent validation (e.g., a known-ground-truth phantom experiment) before any reconsideration. The omission of deep-unrolled baselines is also important given the paper's positioning."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a competent plug-and-play half-quadratic splitting paper with a practical memory-efficient twist, but the headline empirical claim is not yet supported because the paper never says what reference image the PSNR/SSIM numbers are computed against.\n\nWhat is genuinely new: training each iteration's CNN separately with unshared weights on 2D patches, applied to large-scale CBCT. That is a sensible way to avoid the O(K) memory cost of deep unrolling, and the runtime table is useful: 30 seconds per volume versus 1440 for MBIR. The HQS math is standard and correctly presented.\n\nThe soft spot is the evaluation. The training loss in Eq. (9) uses the 1200-view FDK reconstruction as ground truth. The metrics in Fig. 3 are reported on zoomed patches, but the text never states what the reference image is. If that reference is the same FDK-1200 reconstruction, then the proposed method is being scored on how well it mimics its own training target, while MBIR is scored against a different objective. The ~0.5 dB margin over MBIR would then mean little. The authors need to state the reference explicitly and ideally evaluate against an independent ground truth, such as a phantom or a higher-quality reconstruction. Metrics on full volumes and error bars across the six walnuts would also help; right now we only see zoomed patches. No code is released, which makes it harder to check the numbers.\n\nA second issue: the comparison omits deep-unrolled baselines that the paper itself cites (e.g., Lahiri et al.). Even if memory efficiency is the motivation, a comparison on the same Walnut data would contextualize the quality claims. The unqualified \"state-of-the-art\" claim in the conclusion overreaches on the basis of six walnuts.\n\nThe out-of-distribution test at 50 views is a good addition, but it inherits the same potential circularity if the reference is still FDK-1200.\n\nThese are addressable. The core idea has merit and the runtime advantage is real. As written, the empirical support for \"outperforms all other methods\" is not convincing. This paper deserves a serious referee, but the referee should push for a clarified evaluation protocol, full-volume metrics, and an independent reference. I would not cite the quality results until those are fixed, though I might cite the memory-efficient training idea.","headline":"A memory-efficient PnP/HQS method with unshared per-iteration CNN weights, but the headline quality claim hinges on an unspecified evaluation reference that may be the same FDK target used in training.","tokens_in":13195,"tokens_out":2539,"would_cite":false,"duration_ms":25887,"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 learnt half-quadratic splitting algorithm that alternates a CNN denoiser with a conjugate-gradient data-consistency step delivers higher-quality sparse-view cone-beam CT reconstructions than MBIR, U-Net, and a shared-weight variant, in…","keywords":["Convolutional neural network","Cone-beam CT","sparse-view reconstruction","half-quadratic splitting","plug-and-play priors","data consistency","industrial XCT","deep learning"],"falsifier":"Train the same algorithm with a dense-view MBIR reconstruction as ground truth (or on a numerical phantom with known density) instead of FDK-1200, and compare the resulting PSNR/SSIM against MBIR-75. If the advantage over MBIR disappears, the claim that the learned scheme is inherently superior is falsified; it was an artifact of the supervision target.","tokens_in":12164,"feed_emoji":"🩻","tokens_out":6084,"duration_ms":51648,"temperature":0.7,"pith_summary":"This paper tries to establish that a hybrid reconstruction algorithm, which alternates a learnt CNN denoiser with a data-consistency step derived from cone-beam CT physics, can reconstruct high-quality volumes from sparse-view industrial scans (75 views instead of 1200) in about 30 seconds, with quality better than model-based iterative reconstruction (MBIR) and single-step deep learning baselines. The authors argue that this closes a practical gap: MBIR is too slow, single-step deep learning generalizes poorly, and deep unrolled networks are memory-hungry. They demonstrate the algorithm on the public Walnuts CBCT dataset, using only one scan for training and six for testing. If correct, the method makes sparse-view high-quality XCT practical for large industrial volumes.","feed_headline":"Learned CT reconstruction beats MBIR, runs in 30 seconds","feed_subtitle":"Alternates a neural denoiser with a data-consistency step, beating MBIR and U-Net on sparse-view walnut scans.","key_machinery":"The central mechanism is a learnt half-quadratic splitting (HQS) scheme, in which the regularized least-squares problem is split into a Gaussian denoising subproblem and a data-consistency subproblem, solved alternately for a fixed K iterations. The denoising subproblem is approximated by a CNN D_theta, making it a learned proximal operator in the plug-and-play style, and the data-consistency subproblem is solved approximately by a fixed number of conjugate-gradient steps using the ASTRA forward operator. The paper's key design choice is training a separate CNN for each outer iteration (unshared weights) rather than one shared network, and doing so memory-efficiently with 2D patches extracted from the 3D volume. The regularization parameter beta balances denoising against data fidelity and can be adjusted at inference time, and for K=3 this alternation iteratively refines the reconstruction with the CG step anchoring the result to the measured projections.","core_discovery":"On the paper's own terms, the central discovery is that replacing the proximal operator in half-quadratic splitting with a per-iteration convolutional neural network, and enforcing data consistency with a conjugate-gradient solve, yields reconstructions that outperform all the other methods on sparse-view Walnuts data. With K=3 outer iterations and unshared CNN weights, the method reaches PSNR/SSIM values above MBIR, U-Net, and a shared-weights variant, while taking 30 seconds per volume compared with MBIR's 1440 seconds. The paper also finds that unshared per-iteration weights matter: a shared-weight version over-smooths and degrades with iterations, whereas unshared weights sharpen edges and improve metrics. The method is memory-efficient because each iteration's CNN is trained separately on 2D patches, and the regularization parameter beta acts as a test-time knob for unseen imaging conditions.","pith_inferences":["The reported superiority may partly derive from the choice of training target (dense-view FDK); if FDK-1200 contains systematic cone-beam artifacts, the CNN will reproduce them, so the comparison against MBIR could change if a different ground truth were used.","A natural test is to retrain the pipeline with MBIR-1200 as ground truth and compare; if the ranking flips, the current results reflect supervision rather than the iterative scheme itself.","The method still requires an accurate forward model A and its adjoint, so on industrial CBCT with large cone angles the ASTRA operator's accuracy is a hidden assumption that may limit generalization across scanner geometries.","One could extend the approach to 3D CNNs or learned denoisers with uncertainty estimates, but the memory-efficient 2D patch training would need rethinking for isotropic resolution."],"forward_implications":["Sparse-view XCT with 16x fewer projections becomes practical: reconstruction takes about 30 seconds per volume instead of about 1440 seconds for MBIR, enabling faster industrial scanning workflows.","The unshared-weights variant yields progressively sharper reconstructions across iterations, while shared weights over-smooth, so per-iteration training is a key contributor to quality.","The method generalizes to unseen sparse-view settings: at 50 views, a scenario not seen in training, it outperforms U-Net, suggesting better robustness to out-of-distribution measurement conditions.","Memory cost stays flat with respect to iterations (O(1)) rather than scaling with K as in deep unrolled end-to-end training, making the approach applicable to large industrial volumes.","The beta knob allows test-time trade-off between denoising and data consistency without retraining, giving users control over image quality when measurement conditions vary."],"supporting_citations":[{"why":"Defines the FDK analytic reconstruction that serves as the initial input and as the source of the training target.","marker":"[1]"},{"why":"Describes the model-based iterative reconstruction (MBIR) approach that the proposed method must beat in quality.","marker":"[3]"},{"why":"Supplies the public Walnuts CBCT dataset used for all training and evaluation experiments.","marker":"[10]"},{"why":"Provides the plug-and-play priors concept justifying replacement of the proximal operator with a CNN denoiser.","marker":"[12]"},{"why":"Implements the cone-beam forward projector used inside the conjugate-gradient data-consistency step.","marker":"[13]"},{"why":"Is the pyMBIR implementation used as the MBIR baseline for run-time and quality comparisons.","marker":"[14]"},{"why":"Defines the U-Net architecture used as the single-step deep-learning baseline.","marker":"[15]"},{"why":"Provides a prior memory-efficient deep unrolled approach that motivates the paper's memory-efficiency comparison.","marker":"[9]"}],"fun_headline_variants":["Neural CT recon beats MBIR, runs 48x faster","Sparse-view CT quality from 30-second neural recon","Learned half-quadratic splitting for fast sparse-view CT","Unshared CNN weights key to fast high-quality CT reconstruction","30-second neural CT beats MBIR and U-Net on sparse views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the dense-view FDK reconstruction used as the training target is a faithful, artifact-free proxy for the true object; if it contains systematic artifacts, the learned denoiser will preserve and amplify them, and the reported advantage over MBIR could be an artifact of supervision.","fun_headline_variants_meta":{"raw":{"variants":["Neural CT recon beats MBIR, runs 48x faster","Sparse-view CT quality from 30-second neural recon","Learned half-quadratic splitting for fast sparse-view CT","Unshared CNN weights key to fast high-quality CT reconstruction","30-second neural CT beats MBIR and U-Net on sparse views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001173,"raw_usage":{"total_tokens":4861,"prompt_tokens":966,"completion_tokens":3895,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":3808}},"tokens_in":582,"tokens_out":3895,"duration_ms":27722,"temperature":1.0,"reasoning_tokens":3808,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:10:38.899262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same algorithm with a dense-view MBIR reconstruction as ground truth (or on a numerical phantom with known density) instead of FDK-1200, and compare the resulting PSNR/SSIM against MBIR-75. If the advantage over MBIR disappears, the claim that the learned scheme is inherently superior is falsified; it was an artifact of the supervision target.","supporting_citations":[{"cited_title":"A three- dimensional statistical approach to improved image quality for multislice helical ct,","cited_arxiv_id":null,"evidence_quote":"Describes the model-based iterative reconstruction (MBIR) approach that the proposed method must beat in quality."},{"cited_title":"A cone-beam x-ray computed tomography data collection designed for machine learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the public Walnuts CBCT dataset used for all training and evaluation experiments."},{"cited_title":"The astra toolbox: A platform for advanced algorithm development in electron tomography,","cited_arxiv_id":null,"evidence_quote":"Implements the cone-beam forward projector used inside the conjugate-gradient data-consistency step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the pyMBIR implementation used as the MBIR baseline for run-time and quality comparisons."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Defines the U-Net architecture used as the single-step deep-learning baseline."},{"cited_title":"Sparse-view cone beam ct reconstruction using data-consistent supervised and adversarial learning from scarce training data,","cited_arxiv_id":null,"evidence_quote":"Provides a prior memory-efficient deep unrolled approach that motivates the paper's memory-efficiency comparison."}],"review_version":1}