{"id":"d0849116-8b32-42d7-b6a8-903e20b41fbb","arxiv_id":"1908.09515","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A joint PET reconstruction method alternating ML-EM with a learned diffeomorphic registration network achieves a 1.0 dB PSNR gain over gate-by-gate reconstruction on a synthetic 2D Derenzo phantom at near-ML-EM cost.","lead":"The authors combine the standard ML-EM PET reconstruction algorithm with a deep learning based deformable image registration network (voxelmorph) to correct for patient motion across gated PET data. On a synthetic 2D Derenzo phantom, their joint method gains 1.0 dB PSNR over single-gate reconstruction, about 46% of the theoretical maximum, at a computational cost close to ordinary ML-EM.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Voxelmorph is trained on clean synthetic images but applied to noisy ML-EM reconstructions; the reported 1.0 dB gain may not survive this input distribution shift.","rationale":"The reader's verdict is CONDITIONAL and I largely agree. The strongest reason for caution is not that the ML-EM update is wrong (the derivation of M-ML-EM and the adjoint identity W_phi^T = tilde W_{phi^{-1}} are standard and correct), but that the learned registration step is a black box whose reliability is the linchpin of the whole pipeline. My concern sharpens the reader's weakest assumption: even before considering clinical data, the network is asked to register images that are noisier and more artifact-laden than anything it saw during training. The paper's robustness claim (Section 3.3) addresses only random variation at a fixed noise level; it does not test sensitivity to ninit or to the noise level. The proposed concrete test would settle whether the 1.0 dB gain is robust to the intended deployment condition. I note two secondary concerns that do not change the verdict: (1) the iteration counts (6, 42) are selected on the test phantom, so the headline gain is an upper envelope; a fixed-protocol user might see less. (2) The abstract's claim of 'only marginally greater' cost is hard to reconcile with the reported settings: 4 gates × 6 initial ML-EM iterations plus 42 M-ML-EM iterations at 4× per-iteration cost gives ~192 operator evaluations versus 29 for the optimal gate-zero ML-EM, a factor of ~6.6. Still, the core algorithmic contribution remains plausible as a proof of concept, and the conditions already imposed by the reader (out-of-distribution testing, error bars, code release) would address the main risk.","tokens_in":7476,"tokens_out":7997,"duration_ms":80022,"concrete_test":"Train (or fine-tune) the voxelmorph network on pairs of noisy images generated by applying Poisson noise and ninit = 6 ML-EM iterations to the same synthetic ellipsoid distribution used in Section 2.3, then rerun the Derenzo experiment with the reported iteration counts (6 ML-EM + 42 M-ML-EM) and compare the PSNR gain to the claimed 1.0 dB. As a control, also evaluate the original clean-trained network on noise-free phantom inputs to isolate the effect of the input distribution shift.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing link in the pipeline is the pre-trained registration network: the motion estimates it produces are inserted into M-ML-EM (10), and the reconstruction then trusts them completely. Section 2.3 states that the network is trained on tuples (f1, f2) where f1 is a clean synthetic ellipsoid image and f2 = W_{exp(v)} f1. However, in Algorithm 1 the network is deployed on gate images that are outputs of only ninit = 6 ML-EM iterations on Poisson-corrupted data (Section 3.3 and Figure 4). These inputs contain noise and reconstruction artifacts that are absent from the training distribution. The paper reports results for a single noise level (t = 60) and a single phantom, and claims robustness only with respect to random draws of deformations and Poisson noise, not with respect to the image-quality regime in which the network operates. If the network registers noise patterns rather than anatomical structure, the estimated deformations are biased, and the subsequent M-ML-EM update will propagate that bias into f0. The reported 1.0 dB gain (46% of the 2.2 dB maximal gain) could then shrink or vanish precisely in the low-count setting that motivates the paper. This is a train/test input-distribution mismatch internal to the method, not just an external generalization question.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a joint motion estimation and reconstruction algorithm for gated PET. It alternates between a motion estimation step, in which deformations between gate images are obtained by a pretrained voxelmorph network (trained on synthetic ellipsoid images deformed by Gaussian random vector fields), and a reconstruction step, in which all gates are combined through a modified ML-EM update (M-ML-EM) that incorporates the estimated deformations via composite forward operators. The main claims are that the method yields a 1.0 dB PSNR improvement over gate-zero-only ML-EM on a Derenzo phantom, recovering about 46% of the theoretical 2.2 dB gain from using four times the acquisition time, at a computational cost only marginally greater than that of standard ML-EM.","tokens_in":7632,"tokens_out":3244,"duration_ms":31904,"significance":"The algorithmic derivation is a genuine contribution: the M-ML-EM update (10) is a valid extension of ML-EM, the adjoint identity W_phi^T = ~W_{phi^-1} is correct and clearly stated, and the complexity argument in Section 2.6 convincingly shows that the per-iteration cost remains comparable to standard ML-EM. These properties are important and, if the registration was reliable, would make the method attractive for scaling to 3D and clinical 4D PET. The paper also deserves credit for being self-contained in its derivation and for using unsupervised training on synthetic data, which avoids the need for paired clinical motion fields. However, the experimental evidence is currently too narrow to support the headline claims: it consists of a single synthetic phantom, a single noise level, and test deformations drawn from the same distribution used to train the network. The claimed robustness is stated but not quantified by error bars or repeated trials, and the input distribution shift between clean training images and noisy reconstructions is not analyzed.","major_comments":[{"comment":"The reported 1.0 dB gain is obtained by selecting the iteration counts (6 initial ML-EM iterations and 42 M-ML-EM iterations) that maximize PSNR on the test data. No error bars or repeated realizations are provided, despite the sentence in Section 3.3 claiming that the improvement and the optimal iteration counts are 'extremely robust' to the random deformations and Poisson noise. Without a distribution of PSNR values over multiple independent draws of deformations and noise, the 1.0 dB gain could be an optimistic point estimate. Please report mean and standard deviation, and if possible select the iteration counts on a separate validation set.","section":"Section 3.3, Figure 4"},{"comment":"The pretrained voxelmorph network is trained on clean synthetic ellipsoid images (f1 clean, f2 = W_exp(v) f1), but in Algorithm 1 it is deployed on images that are outputs of only n_init = 6 ML-EM iterations on Poisson-corrupted data. This clean-to-noisy input distribution shift is not addressed, and the deformations estimated from noisy inputs are inserted into the M-ML-EM update (10) and fully trusted. If the network registers noise patterns rather than anatomical structure, the resulting bias propagates directly into the reconstructed f0. Please provide evidence that registration remains accurate on noisy reconstructions, for example by evaluating registration error on ML-EM outputs at varying iteration counts or noise levels, and report results for at least one additional noise level (e.g., t = 30 or t = 120) to show that the gain does not vanish in the low-count regime that motivates the method.","section":"Section 2.3 and Algorithm 1"},{"comment":"The test deformations are drawn from the same distribution used to train the network, so the validation is entirely in-distribution. The paper acknowledges in Section 4 that clinical data and the mass-preserving action are future work, but the claimed 'significantly decrease the noise' (abstract) is based solely on this in-distribution test. Since the method's central component is a learned registration network, a demonstration on out-of-distribution deformations (for example, vector fields with different smoothness scales, or deformations mimicking cardiac or respiratory motion) is needed to assess whether the 1.0 dB gain is an artifact of matching the training distribution or a substantive property of the algorithm.","section":"Section 3.1 and Section 4"}],"minor_comments":[{"comment":"The adjoint identity W_phi^T = ~W_{phi^-1} is stated without proof or citation; a short derivation or a reference to a standard text would help the reader verify this load-bearing step.","section":"Section 2.4"},{"comment":"The notation 'W_i <- W_{psi_i} W_{i-1}' is ambiguous about composition order; please clarify whether this means W_i = W_{psi_i} ∘ W_{i-1} and how it relates to the definition of φ_i = ψ_i ∘ ... ∘ ψ_1.","section":"Section 2.5, Algorithm 1"},{"comment":"The caption should state explicitly that PSNR is computed against the gate-zero phantom and should indicate the number of iterations used for the 'no-movement' reference reconstruction.","section":"Section 3.2, Figure 3"},{"comment":"The phrase 'significantly decrease the noise' is vague; consider quantifying the claim as 'reduce noise, as measured by PSNR, by 1.0 dB in a synthetic Derenzo phantom study' to avoid over-generalization.","section":"Abstract and Section 3.3"},{"comment":"The statement that 'one single outer iteration of our algorithm is close to the recently proposed approach [17]' is not elaborated; please specify what [17] does and how it differs from the proposed method.","section":"Section 1.2"},{"comment":"The intensity-preserving action W_ψ is defined in the text, but the mass-preserving action ~W_ψ is only introduced later in (2) within the adjoint discussion; consider defining both actions together in Section 2.1 for consistency.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The algorithmic contribution is solid: the M-ML-EM update is derived cleanly, the adjoint identity is correct, and the complexity argument is convincing. The main weakness is the experimental validation, which is too thin to support the strength of the claims. I would encourage the editor to require the authors to add error bars or repeated trials, address the clean-to-noisy input distribution shift, and ideally include out-of-distribution tests before the paper can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the math is solid, the learning part is off-the-shelf, and the one experiment is too thin to support the headline claim. I largely agree with the reader's conditional verdict, and the stress-test note sharpens the main weakness rather than refuting the paper.\n\nWhat's new: the alternating scheme that wraps M-ML-EM (known from Hinkle et al.) around a pre-trained voxelmorph network, with a variational interpretation and a clean complexity argument showing the cost stays close to ordinary ML-EM. The derivation of the compound ML-EM update is correct, the adjoint identity for the intensity-preserving action is handled properly, and the authors are honest that a single outer iteration is close to Li et al. (2019). That is a legitimate, useful framing.\n\nWhat's soft: the evaluation is a single 2D Derenzo phantom at one noise level, with test deformations drawn from the same distribution used to train the network. The reported 1.0 dB gain uses iteration counts tuned on the test data, with no error bars. The stress-test point is the real kicker: the network is trained on clean synthetic ellipsoid images but deployed on noisy ML-EM reconstructions with artifacts and Poisson noise. That is an internal train/test input-distribution shift, and the paper doesn't test robustness to it. If the network registers noise instead of anatomy, the motion estimates bias the M-ML-EM update and the gain could shrink or vanish in low-count settings. This is an evaluational gap, not a mathematical flaw in the EM update, but it is load-bearing for the claimed practical contribution.\n\nNo code, data, or network weights are released, which makes the robustness claim hard to check. The authors also concede that clinical data and the mass-preserving action are future work, so this is a proof-of-concept, not a turnkey method.\n\nWho this is for: researchers working on learned registration for 4D reconstruction or on integrating deep networks into iterative EM. A reading group would get a good discussion out of the train/test distribution mismatch. I wouldn't cite it as evidence of clinical utility, but it's a reasonable methodological reference.\n\nRecommendation: send it to peer review. It deserves referee time, but expect major revision: out-of-distribution testing, multiple noise levels, error bars, and ideally code release. The core idea is worth publishing in a stronger form.","headline":"A clean algorithmic contribution combining M-ML-EM with learned registration, but the single synthetic evaluation is too thin to support the scalability claim as stated.","tokens_in":8300,"tokens_out":1581,"would_cite":true,"duration_ms":18476,"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 PET motion-correction algorithm that costs about the same as standard ML-EM, recovering 46% of the maximal image-quality gain from using all gated data.","keywords":["PET reconstruction","motion correction","gated imaging","ML-EM","diffeomorphic registration","deep learning","unsupervised training"],"falsifier":"Run the same pipeline on a digital phantom whose deformations come from a different distribution (e.g., a realistic breathing or cardiac motion model) and check whether the PSNR gain of roughly 1.0 dB persists; if the gain drops to near zero or the reconstruction degrades, the learned registration component is the bottleneck.","tokens_in":7148,"feed_emoji":"","tokens_out":1498,"duration_ms":14670,"temperature":0.7,"pith_summary":"This paper proposes a joint reconstruction-and-motion-estimation method for gated PET that combines an enhanced ML-EM update with a pre-trained deep-learning registration network. The central claim is that one can correct for inter-gate motion while keeping the per-iteration cost of the order of the standard ML-EM algorithm, and that this yields a significant reduction in noise compared to using only a single gate. On a Derenzo phantom with four gates, the method gains 1.0 dB in PSNR over the best uncorrected single-gate reconstruction, capturing about 46% of the 2.2 dB gain that would be achieved with no motion at all. The authors emphasize that the training is unsupervised, uses synthetic data, and that the approach is intended to scale to 3D and higher resolutions.","feed_headline":"One PET algorithm fixes motion for the cost of a standard ML-EM","feed_subtitle":"Gated reconstructions gain 1.0 dB in PSNR by combining learned diffeomorphic registration with a multi-gate ML-EM update.","key_machinery":"The central object is the compound ML-EM update for multiple gates, f^(n+1)_0 = f^(n)_0 / (Σ_i A_i^T 1) Σ_i A_i^T ( g_i / (A_i f^(n)_0 ) ), where A_i = A W_φ_i and φ_i is the composition of learned diffeomorphisms mapping gate 0 to gate i. The motion between consecutive gates is estimated by a pre-trained voxelmorph-style network that maps a pair of images to a stationary vector field v, with the diffeomorphism given by exp(v); the network is trained unsupervised on synthetic images of random ellipsoids deformed by Gaussian random vector fields. Crucially, the cost of this network evaluation is negligible compared to the ML-EM iterations, and the transposed operator A_i^T is computed through the identity W_φ^T = ~W_{$φ^{{-1}}$} using the mass-preserving action.","core_discovery":"The paper's central claim is that spatiotemporal PET reconstruction can be performed by alternating a motion-estimation step, where a learned diffeomorphic registration network matches consecutive gate images, with a generalized ML-EM reconstruction step, where all gated data are combined through a compound forward operator. The key finding is that this joint scheme, called M-ML-EM, recovers a substantial fraction of the information lost when reconstructing from a single gate, while its computational cost is only marginally greater than that of standard ML-EM. Concretely, with four gates of the Derenzo phantom, the optimal strategy of six initial ML-EM iterations followed by 42 M-ML-EM iterations yields a PSNR gain of 1.0 dB over the best single-gate reconstruction, which is about 46% of the 2.2 dB gain obtainable if there were no motion. The authors also note that one outer iteration of their algorithm closely resembles a previously proposed deep-learning-based motion-correction method, and that their framework generalizes it as an optimization problem.","pith_inferences":["The method's practical value depends on whether the synthetic-training regime—random ellipsoids deformed by Gaussian random fields—transfers to real anatomical motion; the paper validates only in-distribution deformations, and a natural test is to apply the same pipeline to clinical or realistic digital phantoms with known motion.","Because the network is trained on intensity-preserving deformations, applying the method to real PET data, where activity changes with tracer uptake as well as motion, may require switching to the mass-preserving action or retraining on time-varying activity; the paper identifies this as future work.","The reported 1.0 dB gain is measured against the best single-gate reconstruction; aggregating the same data with simple averaging can also recover some information, so comparing M-ML-EM against that alternative would clarify where the gain actually comes from."],"forward_implications":["The per-iteration cost of M-ML-EM scales linearly with the number of gates, without any additional burden from motion estimation, so the method could plausibly handle the large data volumes of clinical 3D gated PET.","The framework naturally accommodates other deformations, such as the mass-preserving action, and can incorporate attenuation correction, extending the approach to more physically realistic settings.","Because the method recovers a large fraction of the single-gate quality loss, it could reduce the need for long acquisitions or multiple gates, potentially lowering patient scan time or radiation dose without sacrificing image quality.","The fact that a single outer iteration captures most of the improvement suggests that a simple two-stage pipeline—reconstruct each gate, register, then jointly refine—is nearly optimal, simplifying implementation.","The variational formulation provides a principled interpretation of prior deep-learning-based motion-correction heuristics, which could guide further algorithmic improvements."],"supporting_citations":[{"why":"Supplies the unsupervised voxelmorph registration network, its architecture, hyperparameters, and regularisation functional, which the paper uses as the motion-estimation component.","marker":"[6]"},{"why":"Supplies the standard ML-EM algorithm that the paper extends to multiple gates as M-ML-EM.","marker":"[21]"},{"why":"Shows the M-ML-EM update for the intensity-preserving action, which the paper generalizes to learned diffeomorphisms.","marker":"[13]"},{"why":"Provides the previously proposed deep-learning-based motion-correction approach that one outer iteration of the paper's algorithm closely resembles.","marker":"[17]"},{"why":"Provides the Operator Discretization Library (ODL) used for manipulating the forward operator and implementing the PET model.","marker":"[2]"},{"why":"Supplies the warping utilities used in the implementation for applying diffeomorphisms to images.","marker":"[7]"}],"fun_headline_variants":["PET motion fix with ML-EM cost, +1 dB over gating","Learned diffeomorphic registration beats single-gate PET","Near-zero extra cost for motion-corrected PET via ML-EM","Combine all gates: 1 dB gain, no extra compute burden"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pre-trained registration network, trained on synthetic images of random ellipsoids deformed by Gaussian random vector fields, must register the actual test phantom accurately; the paper only validates deformations drawn from the same distribution used for training.","fun_headline_variants_meta":{"raw":{"variants":["PET motion fix with ML-EM cost, +1 dB over gating","Learned diffeomorphic registration beats single-gate PET","Near-zero extra cost for motion-corrected PET via ML-EM","Combine all gates: 1 dB gain, no extra compute burden"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001055,"raw_usage":{"total_tokens":4425,"prompt_tokens":936,"completion_tokens":3489,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":3413}},"tokens_in":552,"tokens_out":3489,"duration_ms":24007,"temperature":1.0,"reasoning_tokens":3413,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:09:21.041729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pipeline on a digital phantom whose deformations come from a different distribution (e.g., a realistic breathing or cardiac motion model) and check whether the PSNR gain of roughly 1.0 dB persists; if the gain drops to near zero or the reconstruction degrades, the learned registration component is the bottleneck.","supporting_citations":[{"cited_title":"IEEE transactions on medical imaging 1(2), 113–122 (1982)","cited_arxiv_id":null,"evidence_quote":"Supplies the standard ML-EM algorithm that the paper extends to multiple gates as M-ML-EM."},{"cited_title":"Medical image analysis 16(6), 1307–1316 (2012)","cited_arxiv_id":null,"evidence_quote":"Shows the M-ML-EM update for the intensity-preserving action, which the paper generalizes to learned diffeomorphisms."},{"cited_title":"In: 15th Inter- national Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine","cited_arxiv_id":null,"evidence_quote":"Provides the previously proposed deep-learning-based motion-correction approach that one outer iteration of the paper's algorithm closely resembles."},{"cited_title":"Royal Institute of Technology (2017)","cited_arxiv_id":null,"evidence_quote":"Provides the Operator Discretization Library (ODL) used for manipulating the forward operator and implementing the PET model."},{"cited_title":"In: Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the warping utilities used in the implementation for applying diffeomorphisms to images."}],"review_version":1}