{"id":"e85ce9f1-48d0-427f-b9de-e2c7900e3e72","arxiv_id":"2607.27785","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Compton camera reconstruction can be made event-count-invariant by encoding events into detector-centred spherical histograms and inverting a precomputed sparse operator from that bounded state.","lead":"A Compton-imaging group replaces the raw event list in reconstruction with fixed-size spherical histograms built from each detected event, then reconstructs from those histogram snapshots. If the latency measurements hold, reconstruction time would no longer grow with the number of detected gamma events, which matters for real-time imaging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Flat transaction cost is only shown under a fixed iteration budget; whether adequate image quality is reached within that budget as events accumulate is never tested, so the practical bounded-latency claim is not yet established.","rationale":"The reader's weakest assumption identifies exactly the condition on which Eq. (10)'s practical import depends: fixed I and E must be sufficient for acceptable images as the event count grows. I agree. The authors are appropriately careful not to claim GPU/CPU hardware parity and to describe the list-mode reference as internal, but the bounded-latency headline is only meaningful if the fixed 20-iteration budget continues to reconstruct well as the accumulated state grows. Table II's three consistent 3.5–3.8 mm centroid errors across N=24k–100k are suggestive but not a scaling analysis, and the bddef1c6 outlier demonstrates quality failure even within this range. A direct iteration-versus-quality sweep on stored snapshots would settle the point. No internal contradiction in the cost model is apparent, and the reported flat timings are credible as measurements of the transaction with the cap. The concern is about sufficiency of the evaluation rather than formalism or integrity, so the existing CONDITIONAL verdict is appropriate.","tokens_in":19752,"tokens_out":6933,"duration_ms":74842,"concrete_test":"Take the stored coherent snapshots from session da19e991 at accumulated counts near 55, 1e3, 1e4, 1e5, and 2.4e5; for each, run the iterative solver with progressively larger iteration budgets and record the first iteration at which centroid localization error drops below 4 mm (or another fixed quality metric). If the required iteration count (or the SNR-selected resolution/E) rises monotonically with accumulated events, the practical bounded-latency claim fails; if it stays flat, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (10) places event-count dependence only in the online N·C_cone term, while the snapshot-to-volume transaction depends on iteration count I and active sparse support E. The flat 2-s timings in §III-B and §III-D demonstrate only that the transaction is stable under the fixed 20-iteration cap and the currently selected angular/fly-eye resolution. The paper never tests whether I or E must grow with the accumulated event count to maintain a fixed image-quality target. The SNR-driven resolution selection described in §II-F makes this a live mechanism: higher statistics could trigger a finer CES resolution, enlarging E and raising C_FB(E). The localization evidence in Table II does not settle it: all four snapshots use 20 iterations, three show 3.5–3.8 mm centroid error over N=24k–100k, but the fourth has 16.0 mm centroid error, and no systematic analysis of quality versus accumulated events or versus iteration budget is provided. Without such an analysis, the bounded-latency claim is only a statement about the cost formula under a fixed cap, not about the latency of a practically usable reconstruction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spherical-histogram representation for Compton cameras: each event is encoded online into detector-centred angular histograms (a fly-eye structure), and volumetric reconstruction is performed from coherent histogram snapshots through a precomputed sparse projection operator. The central claim is that, after state formation, the snapshot-to-volume reconstruction transaction is effectively independent of the number of accumulated events, with cost dominated by the active sparse operator and iteration budget (Eq. 10). The authors support this with timing measurements across 305 snapshots from nine sessions (mean ≈2003 ms, CV 5.6%), showing a flat transaction cost while accumulated events grow from 55 to 243,543 raw events, and with a conventional list-mode MLEM benchmark exhibiting linear scaling (R²=0.9999987). Additional results include near-field point-source localization (3.5–3.8 mm centroid error for three of four snapshots) and a qualitative structured-phantom demonstration.","tokens_in":20160,"tokens_out":8037,"duration_ms":84397,"significance":"If the representational change holds, it provides a principled way to remove event-dependent cone/voxel processing from the iterative inversion loop, enabling bounded-latency reconstruction under continuous acquisition—a relevant goal for Compton imaging in medical and nuclear applications. The theoretical cost decomposition is sound and the extensive timing evidence directly supports the conditional claim: for fixed geometry, operator, and iteration budget, the transaction cost does not grow with event count. The paper is explicitly careful to scope its claims, disclaiming hardware comparisons and qualitative phantom results. Strengths include reproducible session-level statistics, clear presentation of the scaling contrast, and the conceptual reframing of the reconstruction state as a fixed-dimensional object. The main weakness is that the practical interpretation of 'bounded-latency reconstruction' requires demonstrating that a fixed iteration budget delivers adequate image quality as event counts grow, which is not provided.","major_comments":[{"comment":"The paper demonstrates that the per-transaction reconstruction cost is flat under a fixed iteration cap (max 20 iterations, operational stop below 1e-4), but it never tests whether that fixed budget is sufficient to reach a fixed image-quality target as the accumulated event count grows. Table II contains only four snapshots, with one centroid error of 16.0 mm (session bddef1c6), and there is no systematic sweep of event count versus iteration budget or image quality. Section II-F leaves open the possibility that SNR-driven resolution selection enlarges the active sparse support E with statistics; if the required iteration count or E grows, the practical latency of a usable reconstruction would grow even though C_FB(E) is fixed. Please add a quality-versus-event-count analysis (e.g., localization error and convergence metric at 5/10/15/20 iterations across the reported event range), or e","section":"§III-D, Eq. (10), Table II"},{"comment":"The central scaling and localization results depend on an unreleased internal implementation, and the paper does not specify the parameters that determine the sparse operator: the number, positions, and diameters of the fly-eye spheres; the angular bin count N per sphere; the definition of the active sparse support E; and the construction of M(j) in Eq. (6). Without these, the reported 2-s transaction cost and the linear list-mode fit (R²=0.9999987) cannot be independently reproduced or compared. Please provide a complete parameter table (or release the code) for the reported sessions, including the fly-eye geometry and angular resolution used in Tables II–V.","section":"§II-C, §II-H, §III-C"},{"comment":"The headline scaling comparison is between a GPU list-mode implementation with 15 fixed iterations (Table IV) and a CPU bounded-state implementation with up to 20 iterations and an adaptive stop. Although the paper disclaims a hardware benchmark, the fit T_LM(N) ≈ 2.585 + 4.580×10^-3 N s is presented as the conventional baseline for the architectural contrast. Because the reference implementation is internal and described only as '15 fixed iterations,' it is difficult to assess whether the linear law is representative of list-mode MLEM generally or an artifact of that specific unoptimized code. A matched CPU/CPU comparison, or a reference to a publicly available list-mode implementation, would strengthen the central claim that the proposed representation changes the scaling law rather than merely reflecting implementation choices.","section":"§III-C, Fig. 8"}],"minor_comments":[{"comment":"The text states 'No thresholding or dedicated volumetric post-processing was applied; volumes are analysed directly in the state produced by the iterative solver,' but Table II notes 'Centroid computed over voxels above 50% of peak value.' This is a threshold and should be described consistently.","section":"§III-A, Table II"},{"comment":"The claim that 'resolution adjustments can be executed at a unit computational cost' is vague. Specify what 'unit computational cost' means (e.g., constant-time access to precomputed multi-resolution states) and how the SNR-based resolution selection would work in practice.","section":"§II-F"},{"comment":"The difference between the first five snapshots (1,977 ms) and the last five (2,105 ms) is not tested for significance. A simple t-test or confidence interval would substantiate the claim that the slight upward trend is within noise and not a systematic event-count dependence.","section":"§III-B"},{"comment":"The 'computational hologram' analogy is used repeatedly but never defined operationally. Consider adding a sentence clarifying that it is an architectural analogy only, to avoid overinterpretation or confusion with optical holography.","section":"§I, §IV"},{"comment":"The detector-response convolutor is reported to have no visible effect on the reconstructions. This is surprising and would benefit from a supplementary figure showing the same reconstruction with and without the convolutor, since the chosen voxel and fly-eye scales may make the effect genuinely negligible.","section":"§II-K"}],"recommendation":"major_revision","confidential_remarks":"This is a framework/methods paper with a clear representational contribution and honest scoping. The main obstacle to acceptance is not the conditional scaling claim, which is sound, but the gap between that conditional claim and the practical 'bounded-latency reconstruction' language in the title/abstract. The missing quality-versus-event-count analysis and the lack of detailed parameters/code are both fixable in revision. The authors' disclosure of the patent interest and funding is exemplary. The paper may be well suited to a medical-physics or instrumentation journal, but the reproducibility risk from the unreleased reference implementation should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is a genuine architectural change, not a benchmark game. They move Compton reconstruction from list-mode event replay to detector-centred spherical histogram snapshots plus a precomputed sparse operator. The scaling split is structurally baked into Eq. (10), and they measure it: 305 snapshots across nine sessions stay around 2.0 s (CV ~5.6%), while their list-mode MLEM reference scales linearly in event count (R^2 ~0.999999). That part is credible. Three of four near-field point-source runs localize to 3.5–3.8 mm, and the structured phantom shows separable three-sphere structure under a deliberately crude bias subtraction. The patent is disclosed, and the paper is appropriately modest about what it does not do.\n\nThe main soft spot is exactly the one a careful reader should flag. The flat transaction cost is shown under a fixed iteration budget (max 20, typically 10–15 iterations). Nothing in the paper tests whether adequate image quality stays reachable within that budget as events accumulate. Eq. (10) still has I·C_FB(E); if more statistics push the SNR-based resolution selection to a finer CES resolution or require more iterations, the practical latency will grow. The localization table is too small to close that gap: four snapshots, one 16 mm centroid outlier, no repeated-trial error bars, and centroid computed over a 50%-of-peak threshold that is itself a post-selection. So the event-count invariance of the transaction is measured and real, but the bounded-latency claim for practically usable reconstruction is partly open.\n\nOther caveats are minor and mostly acknowledged: list-mode runs on GPU while the bounded-state runs on CPU, so the absolute comparison is not hardware-matched; the reference list-mode code is internal and unreleased; the phantom evidence is qualitative; and there is a long list of free parameters (fly-eye layout, angular bin count, voxel size, stopping threshold). None of this looks like fitting-to-the-result circularity. The novelty is modest but genuine: no one in the cited prior work uses a detector-centred spherical histogram as a fixed intermediate state.\n\nBottom line: for anyone working on Compton camera reconstruction, prompt-gamma imaging, or continuous-source localization, this paper deserves a serious referee. I would send it out. The referee should ask for code/data release or an independent reimplementation, for a quality-vs-event-count sweep at matched iteration budgets, and for repeated-trial localization statistics. With those, the core claim would be much firmer.","headline":"A real representational change for Compton reconstruction, backed by credible flat-latency timing; the main open question is whether image quality stays adequate within the fixed iteration budget as events accumulate.","tokens_in":20619,"tokens_out":2643,"would_cite":true,"duration_ms":27807,"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":"A spherical-histogram state makes Compton-camera reconstruction cost independent of the accumulated event count.","keywords":["Compton camera","gamma-ray imaging","spherical histograms","image reconstruction","list-mode MLEM","bounded latency","event-count invariance","sparse projection operator"],"falsifier":"Count the number of iterations, or total reconstruction time, needed to reach a fixed image-quality target — such as a set centroid error or contrast-to-noise ratio — as accumulated events range from 10^2 to 10^7 in the same geometry. If the required iteration count or active sparse support grows with event count at a fixed quality target, the snapshot-to-volume cost is not truly event-count independent; conversely, an unchanged cost at 10^7 events with quality held fixed would support the claim. A second test is to repeat the phantom experiment with a matched list-mode implementation and acce","tokens_in":19692,"feed_emoji":"⚛️","tokens_out":5183,"duration_ms":45645,"temperature":0.7,"pith_summary":"Compton gamma cameras reconstruct a source from many Compton cones, but conventional list-mode methods re-process every event inside the inversion loop, so reconstruction cost grows with event count. This paper proposes encoding each event, as it arrives, into detector-centred spherical histograms, then reconstructing the volume from a frozen snapshot using a precomputed sparse projection operator. The central claim is that this detaches event accumulation from iterative inversion: once geometry, the sparse operator, and the iteration budget are fixed, the snapshot-to-volume transaction cost is effectively independent of how many events have accumulated, measured at about two seconds per snapshot. The authors support this with a measured linear scaling law for list-mode MLEM, bounded-latency timings over nine acquisition sessions and 305 snapshots, near-field Monte Carlo point-source localization of a few millimetres, and a qualitative three-sphere phantom result where the encoded state separates coherent source structure from diffuse cone background.","feed_headline":"Reconstruction time stays flat as Compton events pile up","feed_subtitle":"Spherical-histogram encoding holds snapshot-to-volume cost near 2 seconds as event counts grow by orders of magnitude.","key_machinery":"The fly-eye spherical-histogram encoding: the detector surface is decomposed into spheres, and each Compton cone is projected as a circular support on each sphere, with anti-aliased bin updates. A sparse projection operator, precomputed once geometry and resolution are fixed, maps histogram bins to volume voxels using ray-tracing with voxel-intersection length, efficiency, and attenuation weights. Because the inversion runs as forward/backward passes over this fixed sparse operator rather than over event-specific cone/voxel interactions, the reconstruction state is bounded and event accumulation is decoupled from the iterative solver.","core_discovery":"The central claim is that the event stream of a Compton camera can be converted, online, into a bounded reconstruction state — a set of detector-centred spherical histograms — so that volumetric image reconstruction never has to replay the photon list. Each event's Compton cone is projected as a circle onto each sphere and its bins are incremented with normalized anti-aliased weights; a precomputed sparse operator then maps histogram bins to voxels, and MLEM runs as snapshot-based forward/backward passes over that operator. With geometry, sparse operator, and iteration budget fixed, the paper's cost decomposition gives reconstruction time as an online per-event encoding term plus a setup ter","pith_inferences":["If the bounded-latency property persists at much larger event counts and with quality-matched stopping rules, Compton imaging could shift from batch reconstruction to continuous near-real-time volumetric monitoring, subject to detector readout capacity.","The same histogram-state abstraction may apply to other cone-based imaging modalities, since any inverse problem with continuously varying cone axes could be re-expressed through a fixed angular basis attached to the detector surface.","The phantom observation suggests a testable extension: multi-view angular consistency across spheres could suppress incoherent background before volumetric inversion, not only after it.","Because the paper's GPU claim is architectural rather than measured, a matched GPU implementation of the sparse forward/backward operator may reduce the two-second transaction substantially, making sub-second reconstruction feasible."],"forward_implications":["Continuous acquisition can be decoupled from inversion: histogram states can be frozen and reconstructed asynchronously while new events keep being encoded.","Adding events changes only the online accumulation stage, so higher detection rates need not raise reconstruction latency, potentially enabling shorter acquisition windows or finer temporal sampling.","The encoded state is fixed-dimensional and tensor-compatible, so regularized or learning-based reconstruction methods could operate on the state rather than on raw event lists.","The precomputable sparse operator supports multi-view and multi-resolution operation with a natural per-sphere parallelization, localizing queue contention in online processing.","The representation separates coherent source structure from diffuse cone background, as shown qualitatively in the phantom sequence, a property not observed in the list-mode pipeline used for comparison."],"fun_headline_variants":["Bounded histogram state keeps Compton reconstruction time flat","Compton camera: online histogram encoding means constant reconstruction cost","From event list to histogram: reconstruction time independent of event count","Snapshots instead of photon replay: bounded-latency Compton imaging","Reconstruction cost decoupled from event count via spherical histograms"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The bounded-latency result rests on the assumption that a fixed iteration budget (at most 20 MLEM iterations, stopping below 1e-4) continues to deliver adequate image quality as the accumulated event count grows; if more events demand more iterations or a larger active sparse support to reach the same quality, practical latency would grow even though the per-transaction cost stays flat.","fun_headline_variants_meta":{"raw":{"variants":["Bounded histogram state keeps Compton reconstruction time flat","Compton camera: online histogram encoding means constant reconstruction cost","From event list to histogram: reconstruction time independent of event count","Snapshots instead of photon replay: bounded-latency Compton imaging","Reconstruction cost decoupled from event count via spherical histograms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000648,"raw_usage":{"total_tokens":2750,"prompt_tokens":621,"completion_tokens":2129,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":365,"completion_tokens_details":{"reasoning_tokens":2045}},"tokens_in":365,"tokens_out":2129,"duration_ms":15242,"temperature":1.0,"reasoning_tokens":2045,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:25:22.599032+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count the number of iterations, or total reconstruction time, needed to reach a fixed image-quality target — such as a set centroid error or contrast-to-noise ratio — as accumulated events range from 10^2 to 10^7 in the same geometry. If the required iteration count or active sparse support grows with event count at a fixed quality target, the snapshot-to-volume cost is not truly event-count independent; conversely, an unchanged cost at 10^7 events with quality held fixed would support the claim. A second test is to repeat the phantom experiment with a matched list-mode implementation and acce","supporting_citations":[],"review_version":1}