{"id":"0afb77a2-906a-4d4c-8c91-f2499577bd49","arxiv_id":"2506.04976","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A harmonic-mean preconditioner that combines data-fidelity and prior Hessian information makes subset-based PET reconstruction with the relative difference prior converge faster and more reliably.","lead":"The paper tests fast stochastic optimization methods for PET scan image reconstruction and finds that a preconditioner that includes information from the smoothing prior speeds up convergence and improves stability. The work underlies the winning algorithm in the 2024 PETRIC reconstruction challenge, so it shows how careful algorithm design can make clinical PET reconstruction faster in practice.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Key claim that harmonic-mean preconditioner is essential rests on single-run inverse-crime curves and has no real-data ablation; a tuned MLEM preconditioner may behave differently.","rationale":"The reader's weakest assumption is that the inverse-crime simulations transfer to real PETRIC data. I agree that this is a real risk, but the more immediate problem is that even the in-simulation evidence for the central claim is thinner than the text suggests: the key MLEM-versus-harmonic comparison in Figure 2 is single-run, and no real-data ablation isolates the preconditioner. The harmonic preconditioner could still be genuinely important; the evidence presented does not yet rule out the alternative that a smaller, tuned stepsize would make the MLEM preconditioner competitive, or that the reported divergence is seed-specific. The proposed test directly targets the claim by swapping preconditioners on the actual PETRIC data and by checking whether MLEM simply needs a smaller tau(0). Because the paper is otherwise clear, methods are standard, and the PETRIC results are real-data evidence that the submitted algorithms work, I would not change the reader's CONDITIONAL verdict; I would only add this specific condition to the list of what needs to be demonstrated.","tokens_in":15213,"tokens_out":6252,"duration_ms":72636,"concrete_test":"Run the submitted ALG2 on the PETRIC Vision600 Hoffman training dataset with the harmonic-mean preconditioner replaced by the MLEM preconditioner and all other settings (SVRG, subsets, BB stepsize schedule, preconditioner update epochs) identical, across at least five random subset-order seeds. If the MLEM variant reaches all PETRIC convergence thresholds in comparable walltime, the claim that the prior-aware preconditioner is essential is not supported. As a secondary control, repeat the Figure 2 comparison at beta=16, tau(0)=1 for SVRG with a finer MLEM stepsize grid (e.g., tau(0)=0.03, 0.1, 0.3) and 5 seeds; if a smaller MLEM stepsize converges at competitive speed, the original comparison conflates preconditioner choice with stepsize tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3's central claim, that the harmonic-mean preconditioner is essential for stable convergence at tau(0) approximately 1, is built on Figure 2, whose caption states that each condition is the outcome of 1 run. The MLEM-preconditioner curves are therefore not established as representative: 'diverges or converges extremely slowly' at beta=16 could be a property of the particular subset order or of an untuned stepsize rather than a robust failure of the preconditioner. Figure 4 provides a five-seed stability check only for the harmonic preconditioner, not for the MLEM comparison that the claim depends on. In addition, the PETRIC results in Figures 7 and 8 never replace the harmonic preconditioner with the MLEM one, so the decisive design choice is never tested on the actual target domain. This matters because the paper's headline finding and the explanation for winning PETRIC are exactly this comparative claim. The inverse-crime setup (same projector for simulation and reconstruction) further means the comparison cannot reveal model-mismatch effects, and the unstated values of alpha in Eq. (9), delta, and the RDP parameters prevent an independent check of whether the harmonic preconditioner is compared at a fairly matched setting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports the winning contribution of the MaGeZ team to the 2024 PET Rapid Image Reconstruction Challenge (PETRIC). The authors study stochastic subset-based optimization methods (SGD, SAGA, SVRG) for RDP-regularized PET reconstruction, with emphasis on a diagonal preconditioner combining the classical MLEM preconditioner and the inverse diagonal Hessian of the RDP via a scaled harmonic mean (Eq. 9). Through fast inverse-crime simulations, they conclude that this prior-aware preconditioner is essential for stable convergence at initial stepsize tau^(0) approximately 1, that SVRG is slightly preferable to SAGA, and that n approximately 27 subsets is near-optimal. These insights are translated into three submitted algorithms (ALG1-3), all using SVRG with the harmonic preconditioner, which are validated on real PETRIC datasets and achieve the fastest convergence overall.","tokens_in":15488,"tokens_out":5907,"duration_ms":64756,"significance":"The paper is a strong empirical contribution: it documents a reproducible winning challenge entry, provides extensive ablation experiments, and proposes a simple, computationally cheap preconditioning heuristic that addresses a real practical problem (stepsize tuning under strong regularization). The availability of code and data is a notable strength. However, the headline claim—that the harmonic preconditioner is crucial—rests on single-run comparisons in simulations and is never ablated on the real PETRIC data, so the evidence is not yet fully commensurate with the strength of the claim.","major_comments":[{"comment":"The claim that the MLEM preconditioner 'diverges or converges extremely slowly' at beta-tilde = 16 and that the harmonic-mean preconditioner is 'essential' is based on a single run per configuration, as stated in the Fig. 2 caption. Figure 4 provides five-seed stability only for the harmonic preconditioner, not for the MLEM comparison that the claim depends on. Please provide repeated runs (e.g., five seeds) for the MLEM preconditioner at least in the critical beta-tilde = 16, tau^(0) = 1 cases, with shaded ranges or error bars, so that the central comparative claim is statistically supported.","section":"§3.2, Fig. 2; §3.3"},{"comment":"The scaling parameter alpha in the harmonic-mean preconditioner and the stabilization constant delta in D_MLEM are never specified for the experiments. Because the behavior of the harmonic preconditioner depends critically on alpha, and the comparison to the MLEM preconditioner depends on delta, the reader cannot determine whether the comparison is tuned fairly or reproduce the results. Please report the values used for alpha, delta, and the RDP parameters gamma and epsilon.","section":"§2.3, Eq. (9)"},{"comment":"The PETRIC validation uses only the three submitted algorithms, all of which employ the harmonic preconditioner. The assertion that this preconditioner is 'essential' is never tested on the actual target domain: an ablation replacing the harmonic preconditioner with the MLEM preconditioner on at least one real dataset is needed to confirm the design choice in the setting where it matters.","section":"§4.1, Figs. 7-8"}],"minor_comments":[{"comment":"'nonnegativity constrains' should be 'nonnegativity constraints'.","section":"§1.1"},{"comment":"The expression 'D(x) = 1/2 h(D_MLEM, alpha^{-1}D_betaS)' is a nonstandard use of the harmonic mean; consider defining the scaled harmonic mean directly to avoid confusion with the conventional harmonic mean.","section":"§2.3, Eq. (9)"},{"comment":"The condition 'k mod (omega n) = 0' uses the iteration index k, while Table 1 refers to 'preconditioner update epochs'; please clarify whether update_pc_iters contains iteration indices or epoch numbers and align the notation with the text.","section":"Appendix A.2, Algorithm 1"},{"comment":"The cofactor order is introduced as a novel sampling rule, but Figure 5 shows it yields negligible differences; the paper should state explicitly whether it is retained only for its marginal benefit or for ALG3's robustness.","section":"§2.4"},{"comment":"The vertical lines indicating convergence times are hard to distinguish across algorithms; a table of wall-clock convergence times per dataset and run would improve readability.","section":"§4.1, Figs. 7-8"}],"recommendation":"major_revision","confidential_remarks":"The paper is more of a challenge report than a methodological advance, though it does propose a practically useful heuristic. The main claim needs the additional experiments described in the major comments, particularly multi-seed MLEM comparisons and a real-data ablation. As the authors are also the challenge winners, the evaluation is partly self-referential; the manuscript should make this explicit in the discussion. The inverse-crime simulation methodology is a limitation that should be acknowledged more directly, even though the real-data results partially mitigate it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the harmonic-mean preconditioner is a genuinely useful heuristic, and this paper gives a clear account of why it worked in the PETRIC challenge, but the key comparative claim is built on single-run inverse-crime curves and needs more evidence before it is a settled result.\n\nWhat's actually new: combining the MLEM diagonal preconditioner with the diagonal of the RDP Hessian via a harmonic mean (Eq. 9) is a simple, cheap idea that appears to stabilize convergence under strong regularization. The cofactor subset ordering is a minor new heuristic, though the simulations show it gives negligible improvement over random without replacement. The paper also provides a careful empirical exploration of SVRG vs SAGA vs SGD under different preconditioners, stepsize rules, number of subsets, and sampling strategies.\n\nWhat the paper does well: the writing is clear, the simulation study is systematic, the PETRIC data provides an external benchmark, and the winning result is a concrete achievement. The code is said to be available. The analysis of why ALG1 is slower on the Hoffman dataset is honest and informative.\n\nSoft spots: the central claim that the harmonic preconditioner is essential rests almost entirely on Figure 2, where each curve is a single run. The MLEM preconditioner failure at high regularization might be an artifact of an unlucky subset order or an untuned stepsize; there is no repeated-run check for the comparison that the claim depends on. Moreover, the real-data PETRIC results never replace the harmonic preconditioner with the MLEM one, so the decisive design choice is never tested on the target domain. The simulations are inverse-crime (same projector for forward and inverse), which is optimistic. Some parameters (alpha, delta, RDP epsilon and gamma) are not reported, and the code link is not given in the text. These are fixable concerns and do not destroy the paper's practical value, but they do mean the headline finding is not as firmly established as the abstract implies.\n\nWho this is for: anyone working on fast regularized PET reconstruction, and more broadly on stochastic optimization with preconditioning for inverse problems. A serious referee would want the single-run curves replaced by multiple runs and a real-data ablation of the preconditioner. The paper deserves peer review and, with those additions, would be a solid methods paper.\n\nRecommendation: engage with the work; send it to review with a request for more runs and the missing parameter values.","headline":"Useful heuristic and a clear account of the PETRIC win, but the key comparative claim rests on single-run inverse-crime curves and needs more evidence.","tokens_in":16012,"tokens_out":2428,"would_cite":true,"duration_ms":27330,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Regularized PET reconstruction can be made both fast and stable by folding the prior's curvature into the preconditioner.","keywords":["PET reconstruction","relative difference prior","stochastic variance reduced gradient","preconditioning","harmonic mean","ordered subsets","PETRIC challenge","inverse-crime simulation"],"falsifier":"A decisive check is to run the submitted SVRG algorithm on a PETRIC test dataset such as Vision600 Hoffman with the harmonic-mean preconditioner replaced by the classic MLEM preconditioner, tuning the initial stepsize and decay schedule to whatever values are best; if the MLEM-preconditioned run reaches all PETRIC convergence thresholds (whole-object NRMSE below $10^{-2}$ and the VOI error thresholds) in equal or less wall time than the harmonic-mean version, the paper's central claim that prior-aware preconditioning is the decisive factor would be contradicted.","tokens_in":14999,"feed_emoji":"🧠","tokens_out":9929,"duration_ms":108204,"temperature":0.7,"pith_summary":"The paper asks which algorithmic choices make PET reconstruction with a smooth regularizing prior fast and stable, and answers by testing stochastic gradient methods under a battery of design choices in fast simulations. Its central finding is that the preconditioner—the per-voxel scaling of each update—must absorb the curvature of the prior, not just the data-fidelity term. The authors build a harmonic-mean preconditioner that combines the classic MLEM scaling with the inverse diagonal Hessian of the relative difference prior, and find that this single change lets the algorithm run at a unit initial stepsize across count levels and regularization strengths, where the classic MLEM preconditioner diverges or stalls. The SVRG-based algorithms built on this design formed the basis of the winning contribution to the 2024 PETRIC reconstruction challenge, so the insight has been tested on real brain PET data, not only on simulations.","feed_headline":"Harmonic-mean preconditioner keeps PET reconstruction fast","feed_subtitle":"Combining data and prior Hessians into one preconditioner stabilized the algorithm that won PETRIC 2024.","key_machinery":"The central object is the harmonic-mean preconditioner: a diagonal matrix $D(x)$ obtained from the harmonic mean $h(a,b)=2/(1/a+1/b)$ of the MLEM preconditioner $D_{\\mathrm{MLEM}}(x)=\\mathrm{diag}\\big((x+\\delta)/(A^\\top 1)\\big)$ and the inverse diagonal Hessian of the relative difference prior $D_{\\beta S}(x)=\\mathrm{diag}(1/\\mathrm{diag}(H_{\\beta S}(x)))$, with a strength parameter $\\alpha$ controlling how much prior curvature enters. It is cheap: only the diagonal of the Hessian of the RDP is needed, and the paper finds it need only be updated in the first three to six epochs. The mechanism by which it carries the argument is that it approximates the inverse Hessian of the full smooth objective, thereby stabilizing unit stepsizes across different counts and regularization strengths, while the MLEM-only preconditioner underestimates curvature when the prior is strong. The other ingredients—SVRG's variance reduction and a moderate number of subsets ($n\\approx 27$)—matter, but the paper identifies the preconditioner as the element that made the other choices robust.","core_discovery":"The paper's central claim is that the decisive design choice for fast, stable RDP-regularized PET reconstruction is a preconditioner that includes information about the regularizer. Concretely, it uses the (scaled) harmonic mean of the standard MLEM preconditioner and the inverse diagonal Hessian of the relative difference prior, $$D(x)=\\left(D_{\\mathrm{MLEM}}^{-1}(x)+\\$\\alpha$\\,D_{\\$\\beta$ S}^{-1}(x)\\right)^{-1},$$ so that regions where the prior contributes large curvature receive proportionally smaller updates. The paper reports that with this preconditioner, SVRG converges reliably at $\\tau^{(0)}=1$ with mild decay in every count and regularization regime it tested, whereas the MLEM-only preconditioner diverges or converges extremely slowly under strong regularization ($\\tilde{\\beta}=16$). It further claims that SVRG slightly outperforms SAGA, that both clearly outperform plain SGD, and that about 27 subsets give the fastest wall-time convergence; these choices, embedded in three submitted algorithms, consistently reached the PETRIC convergence thresholds on real test datasets. The conclusion is that variance reduction plus a prior-aware preconditioner removes the need for per-dataset stepsize tuning.","pith_inferences":["Editorial inference: the harmonic-mean construction is generic — any smooth regularizer with a cheap diagonal Hessian (total generalized variation, anatomical priors, learned regularizers) could be folded into the same preconditioner, potentially extending the result beyond RDP.","Editorial inference: because the preconditioner uses only the diagonal Hessian and the MLEM scaling, the recipe should transfer to other Poisson inverse problems where MLEM-type scaling is standard, such as single-photon emission CT or list-mode PET.","Editorial inference: a conservative reading of the paper's evidence is that updating the preconditioner only in the first few epochs is sufficient in inverse-crime settings; on very low-count or non-ToF real data, more frequent updates could in principle help, which is a testable extension the paper did not explore."],"forward_implications":["If the claim is right, RDP-regularized PET reconstruction can be run at a unit initial stepsize with only mild decay, without scanner-specific or count-specific stepsize tuning.","SVRG (or SAGA) with roughly 25–30 subsets will beat plain SGD in wall-time to practical convergence across clinically relevant count and regularization regimes.","The harmonic-mean preconditioner is the component that makes these speed-ups reliable: removing it reintroduces divergence or very slow convergence under strong regularization.","The three submitted algorithms reach the PETRIC convergence thresholds on all test datasets, with whole-object NRMSE being the last metric to converge, meaning practical runtime is bounded by that metric.","Fast inverse-crime simulations are a sufficient development loop: design choices found there transferred directly to real scanner data in the challenge."],"supporting_citations":[{"why":"Defines the relative difference prior whose diagonal Hessian is used in the harmonic-mean preconditioner.","marker":"[1]"},{"why":"Defines the PETRIC 2024 task, datasets, and convergence metrics that serve as the real-data benchmark.","marker":"[2]"},{"why":"Shows stochastic variance-reduced methods applied to RDP-regularized PET, the line of work this paper extends and improves.","marker":"[7]"},{"why":"Introduces SAGA, one of the variance-reduced gradient estimators compared in the simulations.","marker":"[21]"},{"why":"Introduces SVRG, the gradient estimator on which all three submitted algorithms are built.","marker":"[22]"},{"why":"Supplies the fast GPU projection routines used in the inverse-crime simulation experiments.","marker":"[27]"},{"why":"Records the PETRIC 2024 leaderboard that supports the paper's claim that the submitted algorithms won.","marker":"[28]"}],"fun_headline_variants":["Harmonic-mean preconditioner accelerates PET","Prior-aware preconditioner removes PET tuning guesswork","PETRIC-winning PET uses prior-aware preconditioner","Variance reduction plus prior-aware preconditioner for fast PET","Stable fast PET via harmonic-mean prior preconditioner"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the inverse-crime simulations of Section 3.1, in which the same simulated scanner and projector are used to generate the data and to reconstruct it, are faithful enough to real PET acquisitions that the algorithm rankings and the best settings found there—SVRG over SGD, about 27 subsets, the harmonic-mean preconditioner, and the stepsize schedule—also hold on the clinical PETRIC datasets. If real scanner geometry, noise correlations, or resolution effects shift the ordering, the derived design choices may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Harmonic-mean preconditioner accelerates PET","Prior-aware preconditioner removes PET tuning guesswork","PETRIC-winning PET uses prior-aware preconditioner","Variance reduction plus prior-aware preconditioner for fast PET","Stable fast PET via harmonic-mean prior preconditioner"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000628,"raw_usage":{"total_tokens":2951,"prompt_tokens":1037,"completion_tokens":1914,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":1840}},"tokens_in":653,"tokens_out":1914,"duration_ms":17861,"temperature":1.0,"reasoning_tokens":1840,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:29:15.725102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is to run the submitted SVRG algorithm on a PETRIC test dataset such as Vision600 Hoffman with the harmonic-mean preconditioner replaced by the classic MLEM preconditioner, tuning the initial stepsize and decay schedule to whatever values are best; if the MLEM-preconditioned run reaches all PETRIC convergence thresholds (whole-object NRMSE below $10^{-2}$ and the VOI error thresholds) in equal or less wall time than the harmonic-mean version, the paper's central claim that prior-aware preconditioning is the decisive factor would be contradicted.","supporting_citations":[{"cited_title":"A concave prior penalizing relative differences for maximum-a-posteriori reconstruction in emission tomography.IEEE Transactions on nuclear science , 49(1):56–60, 2002","cited_arxiv_id":null,"evidence_quote":"Defines the relative difference prior whose diagonal Hessian is used in the harmonic-mean preconditioner."},{"cited_title":"Ehrhardt, Christoph Kolbitsch, Evgueni Ovtchinnikov, Edoardo Pasca, Kris Thielemans, and Charalampos Tsoumpas","cited_arxiv_id":null,"evidence_quote":"Defines the PETRIC 2024 task, datasets, and convergence metrics that serve as the real-data benchmark."},{"cited_title":"An inves- tigation of stochastic variance reduction algorithms for relative difference penalized 3d pet image reconstruction","cited_arxiv_id":null,"evidence_quote":"Shows stochastic variance-reduced methods applied to RDP-regularized PET, the line of work this paper extends and improves."},{"cited_title":"Saga: A fast incremental gradient method with support for non-strongly convex composite objectives","cited_arxiv_id":null,"evidence_quote":"Introduces SAGA, one of the variance-reduced gradient estimators compared in the simulations."},{"cited_title":"Accelerating stochastic gradient descent using predictive vari- ance reduction","cited_arxiv_id":null,"evidence_quote":"Introduces SVRG, the gradient estimator on which all three submitted algorithms are built."},{"cited_title":"PARALLELPROJ–an open-source framework for fast calculation of projections in tomography","cited_arxiv_id":null,"evidence_quote":"Supplies the fast GPU projection routines used in the inverse-crime simulation experiments."},{"cited_title":"Ehrhardt, Christoph Kolbitsch, Evgueni Ovtchinnikov, Edoardo Pasca, Kris Thielemans, and Charalampos Tsoumpas","cited_arxiv_id":null,"evidence_quote":"Records the PETRIC 2024 leaderboard that supports the paper's claim that the submitted algorithms won."}],"review_version":1}