{"id":"9bd935b4-6c15-461f-bde1-a7bd24f135a0","arxiv_id":"2501.01013","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A diffusion-prior algorithm with affine projection and implicit neural representation improves sparse-view volume reconstruction for a multi-source static CT system.","lead":"DIP-ASPINS, a new algorithm, combines diffusion models, neural networks, and self-supervised refinement to reconstruct 3D CT volumes from sparse and incomplete multi-source X-ray scans. It is aimed at faster, lower-dose CT imaging without moving gantries.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pseudo-forward state transition in Section 3.1.4 is unproved: setting x_{t-1} = xbar_0 + sigma_{t-1} z replaces the reverse conditional marginal by a point estimate plus noise, with no correction or bound, so the method's status as an approximate posterior sampler is not established.","rationale":"The paper's central claim is empirical ('works well'), but its methodological novelty is the conditional sampling scheme. The pseudo-forward transition is the only place where the reverse-time diffusion chain is modified to return to the noisy state space; without a distributional justification, the chain is not anchored to a well-defined stochastic process. This matters more than the AAPM training-set overlap or the missing limited-angle experiments because those affect the strength of particular comparisons, whereas an incorrect state transition affects every reconstruction generated by the method. The concern is testable without access to the authors' code: in a Gaussian linear model the exact score and conditional marginals are known, so the proposed transition can be checked directly. If it fails there, it cannot be claimed as a general posterior sampler; if it passes, the heuristic is less worrying. I therefore keep the CONDITIONAL verdict: the method may be useful, but the central algorithmic step needs justification or replacement, in addition to the reproducibility items listed by the reader. There is no reason to move to REJECT, because the empirical tables are consistent and the concern is about theoretical status rather than an observed failure.","tokens_in":21910,"tokens_out":8612,"duration_ms":84990,"concrete_test":"Use Algorithm 3.2 (DIP-ASP, without INR/SSL) on a linear inverse problem with a Gaussian prior, where the true score and the true conditional marginal p_{t-1}(x_{t-1}|x_t, y) are available in closed form. From a known x_t, apply Tweedie's formula, affine projection, and the pseudo-forward rule x_{t-1} = x_tilde_0 + sigma_{t-1} z; compute the exact resulting Gaussian distribution of x_{t-1} and compare it to the exact conditional distribution using Wasserstein distance or KL divergence. Repeat for several times t along the noise schedule. If the divergence is non-negligible and does not vanish as sigma_t -> 0, the pseudo-forward transition is not a faithful approximation of the reverse marginal; if it matches, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is the pseudo-forward state transition in Section 3.1.4 (Algorithm 3.1, 'Next image prior'). For the reverse SDE/DDPM chain to continue correctly, x_{t-1} must be a sample from the conditional reverse marginal p_{t-1}(x_{t-1}|x_t, y), or at least from a distribution whose mismatch is controlled. Instead, the paper sets x_{t-1} = xbar_0 + sigma_{t-1} z, where xbar_0 = F(epsilon_0; Phi**) is the output of an INR/SSL minimization of ||P F(epsilon_0; Phi) - Y||^2 + R_lambda(Phi), initialized from the projected Tweedie estimate. This is a point estimate of the clean image, not a draw from p(x_0|x_t, y); no importance weight or correction accounts for the discrepancy between the pseudo-transition and the true reverse kernel. The SSL objective uses the same measurement Y, so xbar_0 depends on y in a non-Bayesian way; repeatedly replacing the state every K steps injects the same data-consistent point estimate at multiple noise levels, which can overemphasize the likelihood relative to the prior. The paper explicitly calls the scheme 'approximate' but gives neither a bound on the introduced error nor a convergence argument, so the central methodological claim that this is a posterior sampling strategy for MSCT is not established. If the transition is biased, the favorable numbers in Table 1 may reflect the particular phantoms and settings rather than a reliable property of the sampler.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a reconstruction method for multi-source static CT (MSCT) from incomplete measurements, combining a pre-trained diffusion score model, affine-set projection for data consistency, an implicit neural representation (INR) for continuous 3D phantom representation, and a self-supervised learning (SSL) refinement step. The resulting algorithm, DIP-ASPINS, is evaluated against classical iterative methods, NAF, and DPER on non-uniform sparse-view projections of three phantoms (Abdomen, Pancreas, SAA), at multiple noise levels and SDE discretization steps. The authors claim that this is a new approximate posterior sampling strategy for MSCT volume reconstruction.","tokens_in":22248,"tokens_out":4257,"duration_ms":37117,"significance":"If the method's effectiveness were fully established, the combination of diffusion priors with INR and SSL would be a useful contribution to incomplete-data CT reconstruction, particularly because the MSCT hardware system with 24 static sources is a concrete application context. The paper includes several strengths: the imaging geometry of the CompoundEyeCT system is described in detail; the method is evaluated against a reasonable set of baselines; and the experiments cover sparse-view counts, noise levels, and SDE step sizes. However, the central methodological claim about approximate posterior sampling rests on an unproved state-transition heuristic, and the evaluation omits the limited-angle setting advertised in the abstract and includes a test phantom drawn from the same dataset used to train the diffusion prior. These issues need to be addressed before the claims can be accepted.","major_comments":[{"comment":"The abstract and introduction state that the method is designed for 'sparse view and limited angle data', but Section 4 contains no limited-angle experiment. All sparse-view experiments use non-uniformly distributed views around the full circle (#views = 48, 72, 120, 240), and the noise-level and SDE-step experiments also use a 120-view full-circle setting. The limited-angle claim is therefore unsupported by the reported results and should either be backed by an actual limited-angle simulation or removed from the claims.","section":"Abstract and Section 4"},{"comment":"The pseudo-forward state transition x_{t-1} = \\bar{x}_0 + \\sigma_{t-1} z is introduced without derivation or error bound. Here \\bar{x}_0 is the result of an INR/SSL minimization of \\|P F(\\epsilon_0;\\Phi)-Y\\|^2, so it is a data-dependent point estimate, not a sample from the reverse conditional marginal p_{t-1}(x_{t-1}|x_t,Y). Because this replacement is applied every K steps, the same measurement Y is injected at multiple noise levels, which can over-emphasize the likelihood relative to the prior. The paper explicitly calls the scheme 'approximate' but gives no convergence argument and no quantitative control of the introduced mismatch, so the central claim that DIP-ASPINS is a posterior sampling strategy is not established.","section":"Section 3.1.4 and Algorithm 3.1"},{"comment":"The Abdomen test phantom is taken from the same '2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge' dataset on which the diffusion prior (score function) was trained in the first paragraph of Section 4. No train/test separation is described. Consequently, the Abdomen rows in Table 1 are partially self-referential and the reported SSIM improvements over the baselines on Abdomen may reflect the diffusion prior having memorized the test distribution. The Pancreas and SAA phantoms are external and provide more reliable evidence; the paper should retrain on a disjoint dataset or report results on a hold-out set.","section":"Section 4.3 and Section 4 (first paragraph)"},{"comment":"The INR input tensor \\epsilon_0 is defined as a 3D mesh grid in Section 3.1.2, but Algorithm 3.1 sets \\epsilon_0 \\sim N(0,I) and draws a fresh noise tensor \\epsilon_{00} for the evaluation F(\\epsilon_{00};\\Phi^*). This discrepancy makes it unclear which encoding is actually used in the experiments and affects reproducibility. The paper should state the exact input encoding and whether the reported results use the grid encoding or random noise.","section":"Algorithm 3.1 and Section 3.1.2"}],"minor_comments":[{"comment":"Algorithm 3.2 uses the variable \\bar{x}_0 in the line 'Next image prior x_{t-1} = \\bar{x}_0 + \\sigma_{t-1} \\cdot z', but \\bar{x}_0 is not defined in that algorithm; it should likely be \\tilde{x}_0 from the affine projection step.","section":"Algorithm 3.2"},{"comment":"The text in Section 4.4.3 says the test phantom is 'a 10 slices of Pancreas', while the caption of Figure 9 first repeats that phrase and then states 'The test phantom is the Abdomen from the AAPM dataset'. These statements contradict each other and should be corrected.","section":"Figure 9 and Section 4.4.3"},{"comment":"The statement that DIP-ASPINS has the best SSIM 'except the case #views = 48' is not accurate: at #views=48 on Abdomen, DIP-ASP has SSIM 0.7980 versus 0.7924 for DIP-ASPINS, and on SAA DIP-ASP has SSIM 0.8899 versus 0.8762 for DIP-ASPINS. The text should be revised to describe the actual pattern in the table.","section":"Section 4.4.1, Table 1"},{"comment":"No standard deviations or repeated runs are reported for the stochastic diffusion-based methods. Since the sampling process is random, a single run per setting is insufficient to support the ranking of methods, especially when the differences are small (e.g., Abdomen at #views=72: DIP-ASPINS 0.8226 vs DIP-ASP 0.8220).","section":"Table 1"},{"comment":"The noise term in Eq. (4.1) is written as '\\eta \\cdot \\epsilon, \\epsilon \\sim N(0, \\eta^2 I)', which implies a variance of \\eta^4. This is likely a typo; the intended Gaussian noise level and its scaling should be clarified.","section":"Equation (4.1)"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible system-level contribution, but the methodological claim about posterior sampling needs either a formal justification or an explicit reframing as a heuristic. The missing limited-angle experiment and the same-dataset test phantom weaken the empirical case. I would encourage the editor to request the limited-angle experiments and an external test set before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a plausible engineering contribution for a niche CT system, but the \"approximate posterior sampling\" claim rests on an unproved heuristic, and the experiments do not fully match the abstract. Worth a look if you work on sparse-view CT, but not a must-read.\n\nWhat is actually new: the specific pipeline DIP-ASPINS — VE-SDE diffusion prior, affine-set projection, hash-encoded INR, and SSL refinement — applied to the CompoundEyeCT multi-source static geometry. The ablation to DIP-ASP (without INR/SSL) helps show where the gains come from. The paper is clearly written and the algorithm is reproducible in principle.\n\nThe good: the comparison to DPER and NAF is reasonable, and the gains in SSIM at 120/240 views are believable. The noise study and SDE-step study are sensible. The authors honestly acknowledge that the affine projection is suboptimal under noise and that the whole scheme is approximate.\n\nThe soft spots, in order: (1) Section 3.1.4's pseudo-forward transition is the load-bearing assumption. Setting x_{t-1} = xbar_0 + sigma_{t-1} z is a point estimate plus noise, not a sample from the conditional reverse marginal, and no bound or correction is given. If this is biased, the reported numbers may be artifacts of the specific test phantoms. (2) The abstract claims limited-angle experiments, but the paper reports only sparse-view results. That mismatch matters. (3) The Abdomen phantom comes from the same AAPM dataset used to train the score function, so those results are partially self-referential; the Pancreas and SAA results are cleaner. (4) No error bars, no code/data release, which makes the numerical claims hard to verify.\n\nProportion: the core idea is not obviously wrong, and the experimental trends are plausible. But the paper overstates its theoretical status. If the authors release code, add limited-angle experiments, rerun Abdomen with a disjoint training set, and either justify or de-emphasize the pseudo-forward step, this could become a solid applied paper.\n\nRecommendation: send it to peer review, but with the expectation of heavy revision. The weaknesses are addressable, and the hardware context is real.","headline":"Plausible engineering combination for multi-source static CT, but the 'posterior sampling' claim leans on an unproved heuristic and the experiments are narrower than the abstract suggests.","tokens_in":22822,"tokens_out":2071,"would_cite":false,"duration_ms":18450,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","92C55","94A08","35R60"],"pacs":[],"model":"deepseek-v4-flash","headline":"An incomplete-data multi-source static CT reconstruction pipeline that combines a diffusion prior, affine-set projection, implicit neural representation, and self-supervised refinement is claimed to produce structurally accurate 3D…","keywords":["Multi-source static CT","Sparse-view reconstruction","Limited angle CT","Diffusion prior","Implicit neural representation","Stochastic differential equation","Self-supervised learning","Affine set projection"],"falsifier":"Keep the INR and self-supervised stages fixed but replace the pseudo-forward transition $x_{t-1} = \\bar{x}_0 + \\sigma_{t-1} z$ with the standard reverse-time SDE update; if the SSIM and PSNR advantages over DIP-ASP largely disappear, those reported gains depend on the unproven marginal transition rather than on the rest of the pipeline.","tokens_in":21661,"feed_emoji":"🩻","tokens_out":10232,"duration_ms":88968,"temperature":0.7,"pith_summary":"This paper tries to establish that incomplete-data 3D CT reconstruction on a multi-source static scanner can be solved by a conditional diffusion loop that never needs paired ground-truth training data. The loop denoises the current reverse-time sample with a pretrained score function, projects the estimate onto the set of volumes matching the measured projections, represents the volume continuously with an implicit neural network, and refines it by a self-supervised fit to the projection data. In numerical tests on three phantoms, the proposed DIP-ASPINS model attains the highest SSIM among compared methods in most sparse-view settings, for example 0.8801 at 240 views on the Abdomen phantom versus 0.8691 for the DPER baseline. If the claim holds, CT scans could use far fewer views and a fraction of the radiation dose while still producing structurally accurate volumes.","feed_headline":"Diffusion prior sharpens sparse-view CT without paired training data","feed_subtitle":"A 24-source static scanner reconstructs 3D volumes from sparse views, cutting dose and scan time.","key_machinery":"The central object is the conditional generation step that carries one reverse-time diffusion state $x_t$ to $x_{t-1}$. It combines four mechanisms: a closed-form MMSE clean estimate $\\tilde{x}_{0t} = x_t + \\sigma_t^2 s_{\\theta^*}(x_t,t)$; the affine projector $P_C(\\tilde{x}_{0t}) = \\tilde{x}_{0t} + P^\\top(PP^\\top)^{-1}(Y - P\\tilde{x}_{0t})$, implemented with conjugate gradients to enforce consistency with measured data; an implicit neural representation (a multilayer perceptron with hash-encoded position embedding) that parameterizes the volume and permits arbitrary-resolution resampling; and a self-supervised refinement that fits the network to the projection $Y$ directly. Without the INR and self-supervised stages the algorithm reduces to DIP-ASP, which only projects; with them, DIP-ASPINS is the version that reports the best structural similarity.","core_discovery":"On the paper's own terms, the discovery is an approximate posterior sampling strategy for the multi-source static CT inverse problem $Y = Pu + n$. Between ordinary reverse-time diffusion steps, the method inserts a data-consistency stage: a closed-form MMSE clean estimate $\\tilde{x}_{0t} = x_t + \\sigma_t^2 s_{\\theta^*}(x_t,t)$ is computed from the noisy diffusion sample, projection onto the affine set $\\{x : Px = Y\\}$ is carried out by conjugate gradient, an MLP with hash encoding fits the projected volume, and a self-supervised loss $\\min_{\\Phi} \\|P F(\\epsilon_0;\\Phi) - Y\\|^2 + R_\\lambda(\\Phi)$ refines it. The refined volume $\\bar{x}_0$ is returned to the diffusion chain through the pseudo-forward transition $x_{t-1} = \\bar{x}_0 + \\sigma_{t-1} z$. Reported numbers show DIP-ASPINS leading in SSIM for most tested sparse-view settings on the Abdomen, Pancreas, and SAA phantoms, with PSNR sometimes below NAF and DPER; quality improves as projection noise falls and as SDE discretization steps increase.","pith_inferences":["Beyond the paper, the pseudo-forward transition is the piece most worth attacking: replacing it with a principled approximate posterior update could give the method a convergence guarantee and better noisy-measurement behavior.","The same four-step loop should transfer to other linear inverse problems with explicit forward operators, such as MRI or coded-aperture imaging, since the diffusion prior acts on slices and the projector is generic.","The PSNR-versus-SSIM gap suggests the method emphasizes edge structure over high-frequency fidelity; a frequency-resolved error analysis would show whether the remaining error is noise-like or anatomical.","The paper uses a slice-wise 2D score; a genuinely 3D diffusion prior might improve inter-slice coherence, at the cost of much heavier pretraining."],"forward_implications":["If DIP-ASPINS is correct, sparse-view MSCT scans with 48 to 240 views can produce volumes with better structural similarity than tuned iterative, NAF, and DPER baselines, supporting lower-dose clinical protocols.","The method needs no paired ground-truth volumes for the reconstruction itself; a pretrained score function plus a self-supervised projection fit is enough, so new scan geometries can be handled without retraining a conditional network.","Because the phantom is represented implicitly, the fitted network can be resampled to a finer grid to produce higher-resolution slices than the training representation.","The reported curves imply a direct trade-off: larger SDE discretization step counts improve PSNR and SSIM but increase runtime, so deployment would choose a step count by available compute."],"supporting_citations":[{"why":"It supplies the pretrained VE-SDE score function and architecture used as the diffusion image prior.","marker":"[56]"},{"why":"It supplies the closed-form MMSE denoising estimate that converts the noisy diffusion sample into a clean volume estimate.","marker":"[22]"},{"why":"It establishes the approximate posterior sampling idea for general noisy inverse problems that the conditional generation loop extends.","marker":"[12]"},{"why":"It provides the definition and properties of projection onto an affine set used for data consistency.","marker":"[5]"},{"why":"It contributes the hash-encoding position embedding used as the implicit neural representation of the 3D phantom.","marker":"[46]"},{"why":"It is one of the compared self-supervised baselines for sparse-view cone-beam CT reconstruction.","marker":"[70]"},{"why":"It is the diffusion-prior neural-representation baseline against which DIP-ASPINS is compared.","marker":"[21]"}],"fun_headline_variants":["Diffusion + neural nets sharpen sparse-view CT without paired data","New CT method uses diffusion priors for sharp sparse-view 3D scans","No training pairs needed: diffusion priors enhance sparse-angle CT","Implicit neural nets boost diffusion-based CT from limited views","Sparse-view CT gets sharper with diffusion prior and INR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the pseudo-forward transition: after self-supervised refinement, adding Gaussian noise at the scheduled level is treated as a true sample from the reverse-time diffusion marginal, but no derivation or convergence argument supports that equivalence.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion + neural nets sharpen sparse-view CT without paired data","New CT method uses diffusion priors for sharp sparse-view 3D scans","No training pairs needed: diffusion priors enhance sparse-angle CT","Implicit neural nets boost diffusion-based CT from limited views","Sparse-view CT gets sharper with diffusion prior and INR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000142,"raw_usage":{"total_tokens":1183,"prompt_tokens":973,"completion_tokens":210,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":123}},"tokens_in":589,"tokens_out":210,"duration_ms":3045,"temperature":1.0,"reasoning_tokens":123,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:36:42.350829+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Keep the INR and self-supervised stages fixed but replace the pseudo-forward transition $x_{t-1} = \\bar{x}_0 + \\sigma_{t-1} z$ with the standard reverse-time SDE update; if the SSIM and PSNR advantages over DIP-ASP largely disappear, those reported gains depend on the unproven marginal transition rather than on the rest of the pipeline.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the pretrained VE-SDE score function and architecture used as the diffusion image prior."},{"cited_title":"Efron, Tweedie’s formula and selection bias , Journal of the American Statistical Association, 106 (2011), pp","cited_arxiv_id":null,"evidence_quote":"It supplies the closed-form MMSE denoising estimate that converts the noisy diffusion sample into a clean volume estimate."},{"cited_title":"Chung, J","cited_arxiv_id":null,"evidence_quote":"It establishes the approximate posterior sampling idea for general noisy inverse problems that the conditional generation loop extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the definition and properties of projection onto an affine set used for data consistency."},{"cited_title":"M¨uller, A","cited_arxiv_id":null,"evidence_quote":"It contributes the hash-encoding position embedding used as the implicit neural representation of the 3D phantom."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It is one of the compared self-supervised baselines for sparse-view cone-beam CT reconstruction."}],"review_version":1}