{"id":"4b302a47-60ea-4ad7-9fbe-fc12747e1b25","arxiv_id":"2502.07777","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Analog multiplexing of SiPM signals from 32 to 16 channels preserves average FWHM of about 0.5 mm in a simulated single-layer monolithic PET detector.","lead":"A simulation study tests whether combining signals from 32 silicon photomultipliers into fewer readout channels preserves the position accuracy of a monolithic PET detector crystal. It finds that cutting from 32 to 16 channels keeps average spatial resolution near 0.5 mm, which could lower detector cost and power use.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-layer-to-multilayer transfer claim and the single-layer assumption in §2.1 are the weakest load-bearing links; the headline conclusion is otherwise internally supported.","rationale":"I read the paper as a simulation-only feasibility study whose central engineering conclusion is: 16-channel analog multiplexing preserves the 0.50 mm average spatial resolution of the 32-channel readout, so a 16-channel readout is the recommended lowest-cost configuration for the semi-monolithic detector concept. The strongest evidence is internal: Table 2 and Figure 12 show 0.50–0.52 mm average FWHM for 32 through 16 channels, with a clear degradation at 12 channels and below (0.52–0.53 mm at 12, 0.58 mm at 8, 0.69–0.71 mm at 4). This trend is consistent across configurations, and the authors examined multiple multiplexing geometries for 16, 12, 8, and 4 channels, which strengthens the claim that scheme choice matters but the channel-count trend is robust within the tested subset. The reader's weakest_assumption correctly identifies the single-layer-to-multilayer transfer: if the light distribution is not layer-count independent, the multiplexing conclusions do not automatically transfer to the stacked detector the authors ultimately propose. I agree with that concern and with the conditional verdict. The other flagged issues are secondary: the bias-corrected 0.34 mm FWHM is not a predictive metric because it uses the test-set ground-truth bias to correct the same test predictions; the paper does acknowledge this by presenting it separately and by emphasizing the uncorrected average FWHM in the conclusion. The extrapolation to image quality in the Discussion is explicitly speculative ('we estimate') and is not central to the headline claim. The interleaved training/test grid is a reasonable worst-case sampling choice, and the absence of electronic noise is defended with a plausible argument that photon statistics dominate; that is a mild limitation rather than a correctness flaw. No internal inconsistency undermines the relative 32-to-16 comparison. The single-layer transfer assumption is therefore the single most load-bearing concern, and it is testable with a straightforward multilayer simulation.","tokens_in":15333,"tokens_out":1911,"duration_ms":16372,"concrete_test":"Run the same GATE simulation with 2, 4, and 8 stacked layers (same 4 mm LSO layer, same ESR interfaces, same four-sided SiPM readout) at a fixed central interaction position and a fixed corner position, and compare the per-SiPM light distributions and the CNN-decoded FWHM for the 32-, 16-, and 8-channel readouts. If the average FWHM and the ranking of 16-channel vs 8-channel readouts change by more than the reported σ (≈0.20 mm), the single-layer transfer assumption in §2.1 and the Conclusion are falsified for the multilayer design.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's engineering conclusion targets a real semi-monolithic PET module built from stacked thin monolithic layers, but all simulations are single-layer (§2.1). The asserted justification — 'the light collection efficiency and distribution are independent of the number of layers' — is stated without simulation or analytic support. In a stacked module, optical photons produced in a layer can cross inter-layer ESR interfaces, be absorbed or reflected by adjacent layers, and the side-readout pattern changes because boundary conditions (dielectric-dielectric with reflectance 0.98) are replicated at every interface; escape, cross-talk, and inter-crystal Compton scatter (acknowledged in §4 as not considered) all alter the per-layer light distributions that feed the CNN. If those distributions change with layer number or with depth within a stack, the multiplexing ranking (e.g., 16 channels preserving 0.50 mm FWHM) may not transfer. This is not an internal inconsistency, but it is an unvalidated extrapolation that the paper's own framing ('a series of stacked thin monolithic scintillator plates') makes load-bearing. Additionally, the bias-correction result (0.34 mm after correction, §3.6) is not a predictive performance metric because it uses the test-set ground-truth bias, evaluated on the same interleaved grid, to subtract the bias; this inflates confidence in the absolute FWHM numbers, though the relative comparison across multiplexing schemes is less affected since the same procedure is applied everywhere.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a GATE v8.2 simulation study of a single-layer LSO monolithic PET detector plate (40 mm × 40 mm × 4 mm) with four side-mounted 1×8 SiPM arrays (32 SiPMs total). Scintillation light distributions are converted into grayscale images and fed into a CNN that predicts the gamma interaction position by a probability-weighted center-of-mass over a 40×40 training grid. The study then simulates analog multiplexing by summing SiPM signals into 28, 24, 20, 16, 12, 8, and 4 readout channels, evaluating spatial resolution (FWHM) and bias on an interleaved 39×39 test grid. The main finding is that average FWHM remains roughly constant at 0.50–0.51 mm as the channel count is reduced from 32 to 16, with degradation below 16 channels; bias is largest at edges and corners. A bias-correction procedure is also presented, yielding a corrected FWHM of 0.34 mm for the 32-channel case.","tokens_in":15618,"tokens_out":3022,"duration_ms":30071,"significance":"If the findings hold, they provide a practical path to reducing readout electronics cost and power consumption in semi-monolithic PET detectors, which are of current interest for high-resolution and DOI-capable systems. The study is valuable for its systematic comparison of multiple concrete multiplexing schemes, its detailed GATE simulation setup, and its explicit treatment of bias in addition to FWHM. The manuscript is also honest about several limitations, including the omission of inter-crystal scattering and electronic noise. However, two load-bearing aspects—the transferability of single-layer results to multilayer stacks and the interpretation of the bias-corrected 0.34 mm value—need to be addressed before the central claims can be fully accepted.","major_comments":[{"comment":"The assertion that 'the light collection efficiency and distribution are independent of the number of layers for the semi-monolithic design' is stated without simulation or analytic support. In a stacked multilayer module, scintillation photons can cross inter-layer ESR interfaces, be reflected or absorbed by adjacent layers, and the boundary conditions repeat at every interface; the paper itself acknowledges in §4 that inter-crystal Compton scattering is not considered. Because the stated motivation is a semi-monolithic detector built from stacked thin plates, the multiplexing ranking obtained from a single-layer simulation may not transfer to the actual detector. Please provide supporting evidence or explicitly reframe the conclusions as applying to the single-layer module only.","section":"§2.1, first paragraph"},{"comment":"The bias-corrected FWHM of 0.34 mm is obtained by estimating the bias from the same test set and then subtracting that estimated bias from the predictions. This is an in-sample correction, not a predictive performance metric: it uses ground-truth information from the test positions to remove a systematic error that would not be known in a real measurement. The comparison of this 0.34 mm value with the uncorrected 0.50 mm average therefore overstates the practically achievable resolution. The bias-correction result should be cross-validated or presented explicitly as a post-hoc illustration of precision after perfect bias knowledge, not as the expected detector resolution.","section":"§3.6, Eq. (2)"},{"comment":"The plateau claim that average FWHM remains constant from 32 down to 16 channels relies on single multiplexing schemes for the 28, 24, and 20 channel cases. Since the paper shows that for 16 channels different multiplexing schemes give average FWHM values ranging from 0.51 mm to 0.70 mm (Table 2), the single-scheme points at 28/24/20 may not be representative of the range of achievable performance at those channel counts. Without testing at least two or three schemes for each of these intermediate counts, the conclusion that performance is preserved from 32 to 16 channels is not robust.","section":"§3.3 and Table 2"}],"minor_comments":[{"comment":"Electronic noise is stated to be negligible based on modern SiPM specifications, but the manuscript does not quantify this expectation in the context of summing analog signals. Since multiplexing combines multiple SiPM outputs, correlated or uncorrelated electronic noise contributions could differ across schemes; a brief quantitative estimate or a more explicit limitation statement would help.","section":"§2.1"},{"comment":"The caption describes the displayed schemes as those 'found to have optimal detector performance,' but the optimality criterion is not precisely defined (e.g., best average FWHM, best bias bounds, or a combination). Please state the selection rule used.","section":"Figure 8 and §2.4"},{"comment":"The error bars in Figure 12 are the standard deviation of the per-position FWHM values, not the standard error of the mean. Clarifying this in the caption would prevent readers from misinterpreting the spread as uncertainty in the average.","section":"Figure 12 and Table 2"},{"comment":"The corner test-grid sizes vary from 3×3 to 5×5 across channel counts, so the MSE values are not directly comparable across rows. This should be acknowledged or the analysis should use a fixed region for all channel counts.","section":"Table 1"},{"comment":"The explanation that reduced feature maps make it 'easier for the CNN to train' and hence improve central-region FWHM is speculative. A quantitative analysis, such as a plot of FWHM versus local light-gradient magnitude, would strengthen this claim.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of physics.ins-det and addresses a timely engineering question for PET detectors. The core simulation methodology is sound and the multiplexing comparison is useful, but the two issues I flagged—the unsupported single-layer-to-multilayer transfer and the in-sample bias correction—are central to the paper's conclusions. In my view these are fixable with additional simulations and a more careful presentation, so I recommend major revision rather than rejection. The authors should also consider presenting the 0.34 mm result only as an upper-bound precision estimate, not as the headline resolution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline result holds up. For this 40×40×4 mm monolithic LSO layer with four-sided SiPM readout, dropping from 32 to 16 summed analog channels leaves the average FWHM essentially unchanged (~0.50 mm), and the degradation below 16 channels is real. The authors also show the choice of summing pattern matters: at 16 channels the best configurations give 0.51 mm while others give 0.70 mm. That is a useful, non-obvious result for anyone designing PET readout electronics. The GATE setup is described in enough detail to reproduce, and the FWHM/bias analysis is internally consistent.\n\nWhat is new is the systematic channel-reduction map for this geometry with CNN decoding. Multiplexing itself is well-trodden (refs [14]–[19] cover it), so the novelty is bounded, but the trade-off data is a legitimate extension, and the supplementary material covers far more schemes than the main figures show.\n\nSoft spots, in order:\n\n1. The single-layer-to-multilayer transfer claim in §2.1 is the weakest link. The stated motivation is a stacked semi-monolithic module, yet all simulations are single-layer. The sentence asserting that light collection and distribution are independent of the number of layers is unsupported. In a stack the ESR-dielectric boundaries repeat at every interface, and inter-layer Compton scatter becomes significant — the paper concedes the latter in §4. The 16-channel result may transfer, but that is not demonstrated. This needs either a multi-layer sanity check or a narrower claim.\n\n2. The bias-corrected 0.34 mm FWHM in §3.6 should not be quoted as a predictive resolution. It is obtained by subtracting the test-set bias, estimated on the same interleaved grid, from the predictions — an in-sample correction. The relative comparison across multiplexing schemes is largely unaffected because the same procedure is applied everywhere, but the absolute number flatters the detector.\n\n3. Minor: electronic noise is omitted, which is defensible given Poisson photon statistics dominate and the comparison is relative. Only one scheme each is tested for 28, 24, and 20 channels, so the curve in Figure 12 is thinly sampled there.\n\nThe citation pattern is honest. Prior multiplexing work is credited, and the self-citations [4][5] point to the group's own earlier layered-detector work that this study extends.\n\nThis is for detector engineers choosing readout electronics for monolithic or semi-monolithic PET modules. Not a physics-changer, but a competent, reproducible simulation study whose 16-channel conclusion deserves experimental testing — and the authors say a prototype has been built. I would send it to peer review with a conditional recommendation: require either a multi-layer simulation check or a clear statement that the single-layer result is not claimed to transfer without further study. The reader's conditional verdict is about right.","headline":"Solid simulation study: the 16-channel conclusion holds up internally, but the single-layer-to-stack transfer is asserted not shown, and the bias-corrected 0.34 mm is an in-sample number.","tokens_in":16165,"tokens_out":6809,"would_cite":true,"duration_ms":59542,"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 simulated PET detector can halve its readout channels from 32 to 16 without losing its roughly 0.50 mm spatial resolution, and only below 16 channels does performance degrade.","keywords":["PET","scintillation crystal","semi-monolithic detector","side readout","SiPM","CNN","signal multiplexing","molecular imaging"],"falsifier":"Measure the built single-layer module's average FWHM at 32 and 16 readout channels with a collimated 511 keV source; if the 16-channel value is not within the simulation's reported spread of the 32-channel value (about 0.50 mm with standard deviation near 0.2 mm), the central claim fails. Separately, simulate a realistic stacked module including inter-crystal Compton scattering and random interaction positions; if 16-channel resolution degrades relative to 32 channels more than the single-layer study predicts, the transferability assumption fails.","tokens_in":15129,"feed_emoji":"⚛️","tokens_out":7684,"duration_ms":60924,"temperature":0.7,"pith_summary":"This paper asks whether a PET detector built from a monolithic scintillator plate read out by 32 silicon photomultipliers can have its analog signals summed before digitization without losing the ability to locate gamma-ray hits. Using GATE optical simulations and a convolutional neural network to decode interaction positions, it finds that reducing the readout from 32 to 16 channels leaves the average spatial resolution essentially unchanged, at about 0.50 mm full width at half maximum. Below 16 channels the resolution degrades and the positioning bias near the detector corners grows, so 16 is identified as the lowest channel count that preserves performance. The practical point is cost: fewer digitization channels means cheaper, lower-power electronics for a PET scanner module.","feed_headline":"Halving PET readout channels keeps detector resolution at 0.5 mm","feed_subtitle":"Simulation of a 32-SiPM PET module: 16 summed channels match the full readout, cutting electronics cost.","key_machinery":"The central mechanism is the comparison of average FWHM and corner bias across a ladder of multiplexing schemes, carried by three components: a GATE v8.2 optical simulation of a single LSO layer with 32 side SiPMs; a set of analog summation schemes that reduce 32 signals to 28, 24, 20, 16, 12, 8, or 4 channels; and a convolutional neural network that converts the summed light distribution into a grayscale image, assigns probabilities over 1600 calibration positions, and computes the interaction position as a probability-weighted center of mass. The load-bearing comparison is the average FWHM (and its standard deviation) over a 39 x 39 test grid, together with the mean squared error of corner bias.","core_discovery":"For a single-layer 40 mm x 40 mm x 4 mm LSO monolithic detector with 0.60 mm ESR reflector films on top and bottom and 32 SiPMs distributed around the four sides, the authors show that summing analog SiPM signals before digitization—from 32 channels down to 28, 24, 20, and 16—keeps the average FWHM of the CNN-decoded interaction position nearly constant at about 0.50 mm in both X and Y. The 16-channel readout that pairs adjacent SiPM signals is the lowest channel count that matches the no-multiplexing performance, including the magnitude of corner bias. Reducing to 12, 8, or 4 channels worsens average FWHM and, especially at the detector corners, produces positioning bias that can extend several millimeters away from the true interaction point. The study also finds that not all multiplexing schemes with the same channel count perform equally; combining signals from SiPMs at the corners of neighboring sides tends to help.","pith_inferences":["If the single-layer-to-multilayer transferability holds, the cost argument extends to a full scanner: a 16-channel-per-layer readout halves the digitizer count for every stacked layer, so the savings grow with the number of layers.","The same simulation pipeline could test more aggressive or non-uniform multiplexing patterns—such as row-column charge division or unequal grouping—to probe whether the 16-channel floor can be pushed lower while keeping bias correctable.","Because training and test grids are interleaved, the reported FWHM is a worst-case interpolation test; interactions at continuous random positions might yield different average resolution, and the corner-bias structure could shift.","The observation that central-region FWHM improves as channels decrease suggests aggressive multiplexing can regularize the CNN by reducing feature maps, implying a trade-off between information loss and network trainability that could guide co-design of hardware and machine learning."],"forward_implications":["PET detector modules can halve their digitized channel count (from 32 to 16) in this geometry without degrading average spatial resolution, lowering electronics cost and power consumption.","The choice of which SiPM signals to combine matters: at 16 channels, the best summing patterns give 0.51 mm average FWHM while the worst studied give 0.70 mm, so multiplexing schemes need to be designed deliberately.","At 12 or fewer channels, corner bias exceeds the 1 mm training-grid pitch, so simple bias correction no longer recovers position accuracy and image reconstruction must model the full point spread function.","For the 32-channel readout, applying bias correction across the whole detector plane improves the average FWHM from about 0.50 mm to about 0.34 mm, showing that much of the remaining positioning error is systematic and recoverable.","Because combining signals from corner-adjacent SiPMs repeatedly improves FWHM and bias, future multiplexing layouts should preserve light-distribution gradients near the detector corners."],"supporting_citations":[{"why":"Supplies the GATE v8.2 Monte Carlo engine used for all optical photon transport and detector response simulations.","marker":"[20]"},{"why":"Establishes the semi-monolithic detector design and the machine-learning position-decoding approach that this work extends to a single layer.","marker":"[4]"},{"why":"Presents the layered Compton PET detector concept whose single-layer readout schemes are evaluated here.","marker":"[5]"},{"why":"Previous study of how the number of readout channels affects intrinsic spatial resolution in a continuous crystal detector, providing the baseline this work extends to side-readout monolithic geometry.","marker":"[15]"},{"why":"Classic charge-division readout method for scintillator arrays, framing the analog multiplexing approach adopted here.","marker":"[14]"},{"why":"Reviews SiPM readout and multiplexing techniques, supplying the taxonomy of light sharing versus charge multiplexing.","marker":"[19]"},{"why":"Edge-readout multilayer detector study that explains the side-readout geometry and the expectation of larger bias where photon sensors are absent.","marker":"[2]"}],"fun_headline_variants":["Half the PET readouts, same 0.5 mm resolution","16 SiPM channels match full readout, cut costs","Monolithic PET: multiplex to 16 channels, keep accuracy","Cut PET electronics cost with 16-channel SiPM summing","SiPM multiplexing: 16 channels preserve PET resolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results for a single 4 mm LSO layer carry over to a full multi-layer stack because the paper assumes light collection efficiency and distribution are independent of the number of layers, so inter-crystal scattering and multilayer electronic effects are not simulated.","fun_headline_variants_meta":{"raw":{"variants":["Half the PET readouts, same 0.5 mm resolution","16 SiPM channels match full readout, cut costs","Monolithic PET: multiplex to 16 channels, keep accuracy","Cut PET electronics cost with 16-channel SiPM summing","SiPM multiplexing: 16 channels preserve PET resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1571,"prompt_tokens":1029,"completion_tokens":542,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":458}},"tokens_in":645,"tokens_out":542,"duration_ms":5579,"temperature":1.0,"reasoning_tokens":458,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T11:34:29.142809+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the built single-layer module's average FWHM at 32 and 16 readout channels with a collimated 511 keV source; if the 16-channel value is not within the simulation's reported spread of the 32-channel value (about 0.50 mm with standard deviation near 0.2 mm), the central claim fails. Separately, simulate a realistic stacked module including inter-crystal Compton scattering and random interaction positions; if 16-channel resolution degrades relative to 32 channels more than the single-layer study predicts, the transferability assumption fails.","supporting_citations":[{"cited_title":"GATE: a simulation toolkit for PET and SPECT,","cited_arxiv_id":null,"evidence_quote":"Supplies the GATE v8.2 Monte Carlo engine used for all optical photon transport and detector response simulations."},{"cited_title":"Compton PET: A Simulation Study for a PET Module with Novel Geometry and Machine Learning for Position Decoding,","cited_arxiv_id":null,"evidence_quote":"Establishes the semi-monolithic detector design and the machine-learning position-decoding approach that this work extends to a single layer."},{"cited_title":"Compton PET: a layered structure PET detector with high performance,","cited_arxiv_id":null,"evidence_quote":"Presents the layered Compton PET detector concept whose single-layer readout schemes are evaluated here."},{"cited_title":"Effect of Number of Readout Channels on the Intrinsic Spatial Resolution Performance of a Continuous Miniature Crystal Element (cMiCE) Detector,","cited_arxiv_id":null,"evidence_quote":"Previous study of how the number of readout channels affects intrinsic spatial resolution in a continuous crystal detector, providing the baseline this work extends to side-readout monolithic geometry."},{"cited_title":"Simple charge division readouts for imaging scintillator arrays using a multi-channel PMT,","cited_arxiv_id":null,"evidence_quote":"Classic charge-division readout method for scintillator arrays, framing the analog multiplexing approach adopted here."},{"cited_title":"Silicon photomultiplier signal readout and multiplexing techniques for positron emission tomography,","cited_arxiv_id":null,"evidence_quote":"Reviews SiPM readout and multiplexing techniques, supplying the taxonomy of light sharing versus charge multiplexing."},{"cited_title":"An edge-readout, multilayer detector for positron emission tomography.,","cited_arxiv_id":null,"evidence_quote":"Edge-readout multilayer detector study that explains the side-readout geometry and the expectation of larger bias where photon sensors are absent."}],"review_version":1}