{"id":"2c7b9aec-3746-4e7b-a937-16d2324b6fbb","arxiv_id":"2502.05445","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Spener is an unsupervised iterative network that uses a self-generated prior image to stabilize sparse-view CT reconstruction, matching or beating supervised baselines.","lead":"The authors introduce Spener, an unsupervised method for sparse-view CT that iteratively refines a neural representation using its own previous reconstruction as an image prior. It combines implicit neural representations with plug-and-play denoising, and the authors report reconstructions comparable to or better than supervised methods without needing paired training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's headline accuracy claim is not yet supported because Figure 2's per-method PSNR/SSIM values contradict Table 1; if the figure is correct, Spener does not outperform baselines.","rationale":"The reader's weakest_assumption was the convergence of the alternating optimization, but the more immediate threat to the central claim is evidentiary: the main quantitative exhibits contradict each other. A missing convergence proof would not necessarily invalidate the method if the empirical curves were trustworthy; an internal mismatch in the reported metrics directly undermines the evidence. The reader's rationale did flag the Figure 2/Table 1 discrepancy as one of several secondary issues, so there is partial agreement, but it was not treated as the load-bearing concern. The available code is a genuine positive, making the proposed recomputation test feasible. I would keep the conditional verdict: the contribution may be valid, but the manuscript must correct the inconsistencies and provide reproducible numbers before the headline claim can be accepted.","tokens_in":10979,"tokens_out":7898,"duration_ms":73523,"concrete_test":"Recompute Table 1 and Figure 2 from the released code and checkpoints using the same AAPM/COVID-19/CMB test slices and fan-beam geometry. Specifically, regenerate the AAPM 60-view panel and verify which method yields 34.47/0.9163 and whether any method yields 23.42/0.4645. If Table 1 reproduces, correct and re-issue the figure; if not, the claimed SOTA-level performance is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that unsupervised Spener matches or beats supervised SOTA and generalizes out-of-domain—is carried by Tables 1 and 2. Those tables are not internally consistent with the qualitative figure. In the manuscript as supplied, Figure 2's AAPM 60-view row lists Spener as 23.42/0.4645, while Table 1 reports Spener at 34.47/0.9163 for the same setting; values attributed to neighboring methods also appear shifted (e.g., FBP is shown at 34.39/0.9274, close to Spener's tabulated PSNR). The COVID-19 and CMB rows show analogous mismatches. If the figure annotations are accurate, Spener performs near FBP and the central comparison fails; if Table 1 is accurate, the figure is mislabeled and the reported numbers need re-verification. Because the headline claim depends on these exact numbers, the current manuscript does not provide reproducible evidence for it. The RegFormer citation inconsistency (Methods lists Liu et al. 2023, a point-cloud registration paper, while the described CT transformer baseline is Xia et al. 2022) adds further uncertainty about whether the supervised comparator was configured as claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Spener, an unsupervised implicit neural representation (INR) method for sparse-view CT reconstruction. The method alternates between fitting an INR, conditioned on a prior image via a small CNN encoder and hash-encoded coordinates, and applying BM3D denoising within a plug-and-play half-quadratic splitting framework. The prior image for each iteration is the denoised reconstruction of the previous iteration. Experiments on AAPM, COVID-19, and CMB-CRC head datasets claim that Spener matches supervised state-of-the-art methods on in-domain data, outperforms them on out-of-domain data, and is more robust to low-dose noise than existing INR methods.","tokens_in":1442,"tokens_out":1574,"duration_ms":64019,"significance":"If the reported results are correct, the core idea is valuable: imperfect INR reconstructions can serve as effective local image-domain priors, and combining this self-prior embedding with an iterative denoiser can improve unsupervised SVCT reconstruction, especially in sparse-view and noisy settings. The method requires no paired training data and the authors provide a public code link. However, the central claims currently rest on quantitative results that are internally inconsistent between the main tables and qualitative figures, and on a supervised baseline whose citation is ambiguous. Until these inconsistencies are resolved, the significance of the contribution cannot be reliably assessed.","major_comments":[{"comment":"The quantitative annotations in Figure 2 directly contradict Table 1 for the same experimental settings. For AAPM 60 views, Figure 2 lists Spener at 23.42/0.4645 and FBP at 34.39/0.9274, whereas Table 1 lists Spener at 34.47/0.9163 and FBP at 23.61/0.4786. The COVID-19 90-view and CMB-CRC 90-view rows show similar reversals. These are not rounding differences; they invert the ranking of methods. Because the headline claim that Spener matches or outperforms supervised SOTA depends on Table 1, this inconsistency is load-bearing and must be corrected and verified.","section":"Comparison with SOTA Methods, Figure 2 vs. Table 1"},{"comment":"The same type of contradiction appears in the noisy-dose experiments. For AAPM 90-view normal dose, Figure 3 reports Spener at 25.86/0.5720, while Table 2 reports Spener at 36.61/0.9424. For low dose, the figure reports 25.98/0.5627 and the table reports 36.17/0.9349. The qualitative figure and the quantitative table cannot both be accurate. Since the noise-robustness claim is a central contribution, the authors must reconcile these numbers and re-verify all reported values.","section":"SVCT Reconstructions in Normal and Low Dose, Figure 3 vs. Table 2"},{"comment":"The supervised baseline RegFormer is cited inconsistently: the Datasets section cites it as Xia et al. 2022, while Methods in Comparison cites it as Liu et al. 2023, and the reference list entry for Liu et al. 2023 is a point-cloud registration paper unrelated to CT reconstruction. This ambiguity prevents the reader from knowing which model was actually used as the supervised comparator. Please correct the citation and state the exact architecture and training configuration of the RegFormer baseline.","section":"Methods in Comparison & Metrics; Datasets"},{"comment":"The out-of-domain and noise-robustness claims are supported by a very small test set: 10 slices per external dataset, with no error bars or statistical significance tests in Tables 1 and 2. Given that the reported differences between methods are sometimes small, the authors should report variance across test slices or otherwise justify that the conclusions are stable beyond a single 10-slice sample.","section":"Comparison with SOTA Methods, Table 1 and Table 2"}],"minor_comments":[{"comment":"The abstract contains the typo \"SVCsT\" in \"solving SVCsT inverse problems\"; it should read \"SVCT\".","section":"Abstract"},{"comment":"The caption of Figure 1 is very terse and does not explain the meaning of the (a)-(d) panels or the symbols z_t, x_t, and F_Phi; please expand it to make the pipeline self-contained.","section":"Figure 1"},{"comment":"The heading contains the typo \"Reconsturction\"; please correct it.","section":"Ablation Studies, Effectiveness of Iterative Reconstruction"},{"comment":"The paper states that detailed fan-beam geometry settings are provided in the supplementary materials, but the arXiv version does not include supplementary material; please either include it or move the geometry description into the main text.","section":"Experimental Settings, Dataset Simulation"},{"comment":"The ablation tables report means with standard deviations, but the number of test slices used to compute these statistics is not stated; please clarify.","section":"Ablation Studies, Tables 3-6"},{"comment":"The performance curve in Figure 4 is not clearly labeled: it is unclear whether the vertical axis is PSNR, SSIM, or both, and which curve corresponds to which metric; please add axis labels and a legend.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The numerical contradictions between Figures 2 and 3 and Tables 1 and 2 are severe enough that I could not verify the central claims of the paper. I recommend asking the authors to provide corrected figures/tables and the exact configuration of the RegFormer baseline before further consideration. The underlying idea of using a self-generated prior image within an iterative INR framework is interesting and potentially publishable, but the current evidence is not reliable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing you should know: this paper has a genuinely interesting procedural idea, but as supplied it is not evidence for its own headline claim. The numbers in Figure 2 and Table 1 contradict each other for the same settings, and the paper's OOD narrative is undermined by its own Table 1. That is not a nitpick; it is the core comparison.\n\nWhat is new: Spener runs an INR with hash encoding in a PnP-HQS loop, using an image encoder to pull local features from the previous iteration's reconstruction and feeding them into the MLP, then applying BM3D as the denoiser. That is a sensible new combination for sparse-view CT, and the authors test it on three datasets and two dose levels. The iterative ablation shows steady improvement over iterations. The code is released. The method is not circular: the data consistency loss uses the measured sinogram, and the self-prior is just the previous reconstruction, not the target.\n\nThe soft spots are real and load-bearing. Figure 2 lists Spener at 23.42 dB / 0.4645 SSIM on AAPM 60-view; Table 1 says 34.47 / 0.9163 for the same row. If the figure is right, Spener sits near FBP and the whole comparison collapses; if the table is right, the figure is mislabeled. Either way, the paper does not give a reproducible number for its central result. Separately, the text says supervised methods drop out-of-domain, but Table 1 has RegFormer scoring higher on CMB-CRC head (34.15 dB at 60 views) than on AAPM (33.78 dB) under identical view counts. That directly contradicts the claim. And the RegFormer citation points to Liu et al. 2023, a point cloud registration paper, while the actual CT baseline is Xia et al. 2022. The baseline may be fine, but the paper doesn't show it.\n\nMinor: the main tables lack error bars (the ablations have them), and the convergence of the alternating scheme is shown only empirically. That is acceptable for this type of paper.\n\nWho should read it: anyone working on INR-based CT reconstruction or unsupervised inverse problems. The idea has merit and the code will help people test it. But the current manuscript would mislead a reader who trusts its tables.\n\nMy recommendation: send it to peer review, but with the explicit expectation of major revision. The referees should demand reconciliation of Figure 2 and Table 1, correction of the RegFormer citation, and error bars or per-slice variance for the headline comparisons. If the authors cannot fix those, the paper should not be accepted. The idea deserves a fair shot; the numbers as printed do not.","headline":"Interesting idea, but the paper's own tables and figures contradict its headline claims; needs major revision before it can be believed.","tokens_in":11770,"tokens_out":4059,"would_cite":false,"duration_ms":37425,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Imperfect prior images make implicit neural representations robust for sparse-view CT.","keywords":["sparse-view CT","implicit neural representation","self-prior embedding","unsupervised reconstruction","plug-and-play half-quadratic splitting","BM3D regularization","out-of-domain generalization","low-dose CT"],"falsifier":"Run Spener on a digital phantom with known ground truth at a fixed 60-view geometry and repeat with several random seeds and Poisson noise levels; if the iteration-by-iteration PSNR or SSIM curves do not stabilize to the same fixed point, or if starting from different initial priors (e.g., FBP vs a constant image) yields materially different final images, the central iteration assumption is falsified.","tokens_in":10741,"feed_emoji":"🩻","tokens_out":8424,"duration_ms":77340,"temperature":0.7,"pith_summary":"Spener is an unsupervised method for sparse-view CT that asks a simple question: can a reconstruction network learn from its own previous, imperfect output? The paper argues yes. At each iteration, an implicit neural representation (INR) is conditioned on the prior image from the last iteration through a local image encoder, trained to match the measured sinogram, and then the result is passed through a denoiser (BM3D) to produce the next prior. The authors claim that this self-prior loop makes the unsupervised method competitive with supervised state-of-the-art reconstruction on in-domain CT data, better than them on out-of-domain data (COVID-19 chest CT and head CT), and markedly more robust to noisy low-dose sinograms. The paper matters because it suggests paired training data and per-scanner retraining may not be necessary for high-quality sparse-view CT reconstruction.","feed_headline":"Unsupervised CT method matches supervised AI on sparse views","feed_subtitle":"Spener feeds each iteration's imperfect image back in as a prior, and beats out-of-domain supervised models.","key_machinery":"The engine of the method is a plug-and-play half-quadratic splitting (PnP-HQS) loop, in which the data-fidelity step is not a closed-form update but the training of an implicit neural representation (INR). The INR maps spatial coordinates p and the previous prior image z_{t-1} to image intensity; the prior is injected through a two-layer CNN image encoder that produces local feature vectors V(p), which are concatenated with multiresolution hash-encoded coordinates before being fed to an MLP. The network is trained by minimizing a sinogram data-consistency loss plus a regularization loss that keeps the output close to the prior image (Eq. 12), and the new reconstruction is then denoised by BM3D to form the next prior (Eq. 13). This self-prior embedding is what carries the claim: the imperfect prior image explicitly constrains the solution space at every iteration, and the alternating loop is the mechanism that refines it.","core_discovery":"On the paper's own terms, the central discovery is that imperfect INR reconstructions are themselves a usable image-domain prior, and that embedding them explicitly stabilizes the neural representation. Spener alternates two subproblems: the data-fidelity subproblem is solved by training a coordinate-MLP network, whose input combines hash-encoded coordinates with local features extracted by a CNN from the previous reconstruction; the regularization subproblem is solved by applying a BM3D denoiser to the freshly rendered image to produce the next prior. The reported experiments show that on the AAPM dataset at 60 and 90 views, Spener reaches 34.47/37.16 dB PSNR, matching or beating the supervised RegFormer, while on COVID-19 and CMB-CRC head data the supervised methods degrade and Spener leads. Under low-dose Poisson noise, the advantage over supervised methods grows, with about 2.8–3.3 dB PSNR improvements at 90 views.","pith_inferences":["Editorial inference: because the self-prior comes entirely from the INR's own output, the procedure can be viewed as a fixed-point iteration; a formal convergence condition on the denoiser and INR step could turn the reported empirical stabilization into a theorem.","Editorial inference: the gains from the image encoder suggest that the method is doing more than denoising—the network is reading local structure of the prior to decide where to trust or override it, so artifact-removal behaviour may emerge without being trained.","Editorial inference: replacing BM3D with a stronger learned denoiser could push performance further, but would reintroduce an external prior and weaken the unsupervised claim; testing that trade-off would clarify how much of the gain comes from the loop vs the denoiser.","Editorial inference: the evaluation uses 2D fan-beam slices; the clinically decisive test is whether the same self-prior loop stabilizes helical or cone-beam 3D acquisitions with real patient noise, which is not demonstrated in the paper."],"forward_implications":["If correct, supervised-level sparse-view CT reconstruction can be obtained without paired training data, removing a major practical barrier for new scanners and protocols.","Because no training data is tied to a specific anatomy or geometry, the same unsupervised method transfers to out-of-domain data where supervised models lose accuracy.","The alternating INR-plus-denoiser design gives a concrete robustness path for noisy and low-dose acquisitions, the setting closest to real clinical CT.","The same design may be applied to other linear inverse problems with differentiable forward models, though the paper presents that as future work."],"supporting_citations":[{"why":"Supplies the BM3D denoiser used to solve the regularization subproblem and produce the next prior image.","marker":"(Dabov et al. 2007)"},{"why":"Provides the plug-and-play framework that turns a denoiser into the regularization step of the iteration.","marker":"(Zhang et al. 2021)"},{"why":"Supplies the multiresolution hash encoding used by the coordinate encoder to capture high-frequency image detail.","marker":"(Müller et al. 2022)"},{"why":"Acts as the unsupervised INR baseline (CoIL) that the paper compares against in the sparse-view experiments.","marker":"(Sun et al. 2021)"},{"why":"Acts as the self-supervised coordinate projection baseline (SCOPE) and an alternative INR formulation.","marker":"(Wu et al. 2023)"},{"why":"Supplies the FBPConvNet supervised baseline used to establish the claim of comparable supervised performance.","marker":"(Jin et al. 2017)"},{"why":"Supplies the RegFormer supervised baseline that Spener matches on in-domain data and beats out-of-domain.","marker":"(Liu et al. 2023)"},{"why":"Supplies the AAPM low-dose CT dataset used for in-domain training and evaluation.","marker":"(McCollough et al. 2017)"},{"why":"Supplies the COVID-19 chest CT external dataset used to demonstrate out-of-domain generalization.","marker":"(Shakouri et al. 2021)"},{"why":"Supplies the CMB-CRC head CT external dataset used to demonstrate cross-organ generalization.","marker":"(Cancer Moonshot Biobank 2022)"}],"fun_headline_variants":["Self-prior embedding lets unsupervised CT match supervised AI","Imperfect CT images become self-priors to rival supervised methods","Spener: use noisy reconstructions as priors to beat supervised on sparse views","Sparse-view CT gains from self-priors without any supervision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that restarting the INR at each iteration from the denoised previous reconstruction will keep improving the estimate, rather than accumulating errors; the paper shows this empirically but offers no proof of convergence or bound on error accumulation.","fun_headline_variants_meta":{"raw":{"variants":["Self-prior embedding lets unsupervised CT match supervised AI","Imperfect CT images become self-priors to rival supervised methods","Spener: use noisy reconstructions as priors to beat supervised on sparse views","Sparse-view CT gains from self-priors without any supervision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00073,"raw_usage":{"total_tokens":3289,"prompt_tokens":988,"completion_tokens":2301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":2228}},"tokens_in":604,"tokens_out":2301,"duration_ms":16208,"temperature":1.0,"reasoning_tokens":2228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:16:59.615364+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Spener on a digital phantom with known ground truth at a fixed 60-view geometry and repeat with several random seeds and Poisson noise levels; if the iteration-by-iteration PSNR or SSIM curves do not stabilize to the same fixed point, or if starting from different initial priors (e.g., FBP vs a constant image) yields materially different final images, the central iteration assumption is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BM3D denoiser used to solve the regularization subproblem and produce the next prior image."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the plug-and-play framework that turns a denoiser into the regularization step of the iteration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Acts as the unsupervised INR baseline (CoIL) that the paper compares against in the sparse-view experiments."},{"cited_title":"H.; McCann, M","cited_arxiv_id":null,"evidence_quote":"Supplies the FBPConvNet supervised baseline used to establish the claim of comparable supervised performance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RegFormer supervised baseline that Spener matches on in-domain data and beats out-of-domain."},{"cited_title":"H.; Bartley, A","cited_arxiv_id":null,"evidence_quote":"Supplies the AAPM low-dose CT dataset used for in-domain training and evaluation."},{"cited_title":"A.; Layegh, P.; Kiani, B.; Masoumi, F.; Ataei Nakhaei, S.; and Mostafavi, S","cited_arxiv_id":null,"evidence_quote":"Supplies the COVID-19 chest CT external dataset used to demonstrate out-of-domain generalization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CMB-CRC head CT external dataset used to demonstrate cross-organ generalization."}],"review_version":1}