{"id":"787a18ea-2ba6-4470-bec0-339f0e7a5a5d","arxiv_id":"2512.02771","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A trained neural network reconstructs LED-spot positions on a 2x2 LG-SiPM tile with 2.4-3.5x lower mean error than the standard linear formula, but this gain mostly corrects systematic distortion, not intrinsic noise.","lead":"A 2x2 tile of position-sensitive silicon photomultipliers was scanned with a small LED spot, and a small neural network was trained to reconstruct where light hit. Compared with the standard center-of-gravity formula, the network cut average position error roughly three-fold, though true single-photon resolution barely changed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Granularity claim rests on single-spot grid error, not two-point resolution; sub-grid separations and edge/gap regions are untested.","rationale":"The reader's weakest_assumption focused on LED-to-gamma transfer and long-term stability. My concern is different and more immediate: even under the same LED/motor setup, the paper has not demonstrated that the reported 0.2 mm granularity corresponds to a true two-point resolution. The metric 2⟨d⟩ is asserted without derivation as the separation at which neighboring regions become indistinguishable. For a Rice-distributed radial error, two sources separated by 2⟨d⟩ would have substantially overlapping distributions; the criterion appears arbitrary. The DNN's improved ⟨d⟩ comes primarily from correcting average shifts, not reducing noise. A calibration-based shift correction can make isolated spots land closer to their true grid positions, but whether two spots 0.2 mm apart are resolvable is a separate question requiring sub-grid or two-source data. This is a load-bearing gap because the abstract and conclusion use the granularity number to claim a factor 5.7–12.1 improvement in resolved areas. The internal comparison between DNN and linear reconstruction on held-out positions is credible, and the splitting-C result helps rule out simple memorization; I therefore do not recommend rejection. But the headline quantitative claim should be re-stated as a linearity-correction improvement on the calibration grid, not a measured resolution limit. The proposed test—sub-0.5 mm stepping or two-source separation—would settle whether the granularity claim lands. If it fails, the central claim would need to be weakened to a calibration-based accuracy improvement rather than a pixel-count enhancement.","tokens_in":8770,"tokens_out":4579,"duration_ms":50427,"concrete_test":"On the same setup, move the LED in 0.1 mm steps over a 2×2 mm² subregion (or place two optical fibers separated by 0.2 mm) and reconstruct positions with the DNN trained on the original 0.5 mm grid. If reconstructed centroids track 0.1 mm offsets and two 0.2 mm-separated spots produce a bimodal distribution, the granularity claim is supported; if positions quantize to the 0.5 mm training lattice or the two spots merge into one peak, the 0.2 mm granularity and factor-10 pixel count are overstated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—0.2 mm granularity, 6400 distinguishable regions, factor 5.7–12.1 improvement—depends on interpreting 2⟨d⟩ in §3.4 as a two-point resolvability measure. But ⟨d⟩ in Table 1 is the mean Euclidean distance between reconstructed and motor positions for isolated LED spots on a 0.5 mm grid. The paper never measures two closely spaced sources or steps the stage at sub-0.5 mm increments. The DNN is trained with MSE loss on discrete 0.5 mm grid labels, so small errors at those nodes do not establish that positions between nodes (or two spots 0.2 mm apart) can be resolved. In fact, the DNN's improvement over the linear model is almost entirely a reduction in systematic shift ν (317→41–93 µm), not in statistical resolution σ (78→67 µm); a calibration map can remove smooth shifts, but that does not automatically create resolvability at sub-grid scales. Additionally, the Q-threshold cut in §2.1, Fig. 3 removes edge and inter-tile-gap events, so the quoted ‘6400 regions over 16×16 mm²’ extrapolates over areas excluded from the test sample. This is not an internal inconsistency, but it means the headline ‘granularity’ number is an extrapolation, not a directly measured resolution limit.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper characterizes a 2x2 array of linearly-graded SiPMs read out through six channels and compares two position reconstruction methods: the standard linear formula of Section 2.2 (Eqs. 1–4) and a fully connected DNN with 6 inputs, 64-unit tanh hidden layers, and a 2-unit linear output (Section 2.3.2). The DNN is trained on 10k-event LED scans on a 37x37 grid with 0.5 mm steps, using the mean squared error to motor positions as the loss. Three train/test splittings are investigated (random, chessboard, and random-position, Section 2.1). Table 1 reports that the DNN reduces the mean systematic shift from ~310–317 µm to 41–93 µm and the quantity 2⟨d⟩ from ~675–688 µm to 198–284 µm, while the mean statistical resolution σ remains similar (~66–79 µm). The conclusion claims a granularity of about 0.2 mm and an increase in distinguishable regions by a factor 10.24.","tokens_in":9168,"tokens_out":3468,"duration_ms":37610,"significance":"If the central claim holds, the work is significant for compact gamma cameras and other position-sensitive photon detectors: a neural calibration would remove the geometric and gain nonlinearities of a 6-channel LG-SiPM tile, achieving sub-millimeter positioning without increasing channel count. The paper has real strengths: it uses three data splittings, including B and C which exclude training grid positions and therefore provide genuine generalization evidence; it reports quantitative per-position metrics from Rice fits; and it compares directly against the standard linear reconstruction. The main caveat is that the headline 'granularity' / 'pixel factor' claim rests on an indirect metric that has not been validated as a two-point resolution measure.","major_comments":[{"comment":"The 'granularity' quantity 2⟨d⟩ is not a two-point resolution. ⟨d⟩ in Eq. (7) is the mean Euclidean distance between reconstructed and motor positions for isolated LED spots on a 0.5 mm grid. Multiplying by two and converting to '6400 distinguishable regions' over the 16x16 mm2 tile (Section 3.4, conclusion) assumes that a calibration error of ⟨d⟩ implies resolvability of two spots separated by 2⟨d⟩, but no two-spot or sub-grid measurement is performed. The improvement in 2⟨d⟩ is dominated by the reduction of the systematic shift ν (317→41–93 µm), not of σ (78→66–68 µm), so it primarily demonstrates that a smooth distortion map is learned. Please either provide a direct two-point resolution measurement (e.g., two sources or sub-0.5-mm stepping) or rephrase the claims in terms of position reconstruction accuracy rather than resolved pixels.","section":"§3.4 and Table 1"},{"comment":"The total-charge quality cut in Figure 3 removes fiber positions near the edges and inter-tile gaps, and the table caption states that the outer rows and columns are excluded from the averages. The abstract and conclusion nevertheless quote '16x16 mm2' as the field of view for the 6400-region count. The tested sensitive area is smaller than the nominal window. Please report the effective area used after the cuts and avoid extrapolating pixel counts to the full 16x16 mm2 unless the edge/gap regions are included in the measurements.","section":"§2.1, Fig. 3, and Table 1"},{"comment":"The evaluation metric d is the same mean squared error used as the DNN loss, and in splitting A the test events share exact motor coordinates with training. This makes gains in A partly a measure of memorization of per-position output biases. Splittings B and C, which use held-out grid positions, provide the true generalization evidence and still show improvements (2⟨d⟩ = 228.7 µm and 283.6 µm vs 674.8–687.6 µm for the linear model). The abstract's upper factor of 12.1 comes from A; the paper should present B and C as the primary evidence for generalization, or explicitly weight the splittings when summarizing the improvement range.","section":"§3, Eq. (7), and Section 2.3.2"},{"comment":"The conclusion states that the DNN 'reduces the granularity to about 0.2 mm thus increasing the number of distinguishable regions by a factor 10.24.' This number does not match Table 1 or the abstract's range. From Table 1, the squared ratios 2⟨d⟩_lin / 2⟨d⟩_DNN are about 12.1 (A), 9.0 (B), and 5.7 (C). The abstract says 5.7 to 12.1, but the conclusion's 10.24 is unexplained. Please reconcile these numbers or specify which splitting (or average) the conclusion refers to.","section":"Conclusion vs. Abstract and Table 1"}],"minor_comments":[{"comment":"Typo: 'The the performance' should be 'The performance'.","section":"§2.3"},{"comment":"'DDNs' should be 'DNNs' for consistency with the text.","section":"Fig. 4 caption"},{"comment":"In the text, 'a system of to 2 linear equations' should be 'a system of two linear equations.'","section":"§2.3.2"},{"comment":"The linear transformation parameters lx, ly, x0, y0, φ are introduced in (3)–(4) and reused in the matrix A of (6), but the relation between (x, y) in the relative frame and the motor positions is not fully stated. An explicit definition of the relative coordinates would improve readability.","section":"Eqs. (3)–(6)"},{"comment":"The caption says 'run splitting' for (C); this should be 'random position splitting' or 'random splitting' as in the text.","section":"Fig. 7 caption"},{"comment":"The paper does not mention whether the acquisition code, trained models, or dataset are available. Please add a data/code availability statement, or note that they are available from the authors on request.","section":"Data/code availability"}],"recommendation":"major_revision","confidential_remarks":"The experimental comparison is internally consistent and the held-out splittings B and C support the core claim that a DNN reduces systematic position errors. The main obstacle is the interpretation of 2⟨d⟩ as a granularity/pixel-count metric. This is fixable by either adding a direct two-point resolution measurement or by revising the abstract/conclusion to avoid the pixel-factor language until such a measurement is made. I do not see a fundamental flaw that would require rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-executed calibration study with an overstated headline. The DNN genuinely removes the systematic distortion in the LG-SiPM tile—the shift drops from ~300 µm to 40–90 µm across all three held-out splittings—and the split design shows it isn't just memorizing the grid. But the “0.2 mm granularity” and “10.24x more pixels” claims are extrapolated from a single-spot error metric, not measured as two-point resolution, and the paper never shows a single scintillation event.\n\nWhat’s actually new: this is the first DNN applied to position reconstruction on a linearly-graded SiPM array. The three split schemes (random, chessboard, random-position) directly address memorization, and the loss curves show no overfitting. The comparison with a zero-hidden-layer linear model is a nice way to isolate what the nonlinearity adds. Data collection is careful: 10k waveforms per step on a 0.5 mm grid.\n\nWhere it gets soft. The abstract and conclusion claim both “resolution” and “linearity” are enhanced. In fact, the statistical resolution sigma barely moves (78→67 µm); the real gain is systematic shift, a nonlinearity correction. Calling that “resolution” is misleading. The granularity claim is the bigger problem. 2⟨d⟩ is a mean single-spot error, not a two-point resolvability criterion. No sub-0.5 mm steps, no two-source measurement. So “6400 distinguishable regions” is a heuristic extrapolation, not an empirical limit. The Q threshold also removes events near edges and inter-tile gaps, so the 16×16 mm² extrapolation silently excludes exactly the areas where the linear reconstruction struggles. And all of this is LED light; there is no gamma source, no scintillator, no temperature or count-rate stability check. The evaluation metric is the same MSE used in training, so splitting A partly reflects memorization; splittings B and C are the real evidence, and they still show a solid gain in shift. Those are natural next steps, not fatal flaws—the internal comparison is coherent and the DNN clearly removes the distortion it was trained to remove.\n\nBottom line: the paper is worth refereeing, but the authors should be pushed to either re-state the granularity claim as a calibration-based linearity measure or back it with a real two-point test. I’d want code/data released and at least one gamma-event dataset before believing the sub-millimeter gamma-camera promise. For a reading group, it’s a useful case study in how a metric can be reasonable yet over-interpreted.\n\nI’d send it to peer review with a request for major revision on the claims.","headline":"Solid DNN calibration study, but the granularity and resolution claims outrun what the experiment actually measures.","tokens_in":9671,"tokens_out":4012,"would_cite":true,"duration_ms":40111,"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 neural network trained on six channel charges from a position-sensitive silicon photomultiplier tile reconstructs light-spot positions with roughly 0.2 mm granularity, a 2.4–3.5x improvement over the standard linear formula.","keywords":["SiPM","linearly-graded SiPM","position-sensitive detector","deep neural network","position reconstruction","gamma camera","charge sharing","sub-millimeter resolution"],"falsifier":"A collimated gamma-ray source scanned across the full tile, including the gaps between the four chips, would directly test transferability: if the DNN's mean position error near edges and gaps exceeds the linear formula's, or if retraining on scintillation light fails to reproduce the 0.2 mm granularity, the central claim fails.","tokens_in":2374,"feed_emoji":"🔬","tokens_out":1596,"duration_ms":60916,"temperature":0.7,"pith_summary":"This paper sets out to show that a deep neural network can replace the standard linear readout formula for a 2x2 array of linearly-graded silicon photomultipliers (LG-SiPMs) and recover nearly distortion-free position maps. The motivation is practical: such detectors read out a 16x16 mm2 imaging tile with only six channels, and if the DNN approach holds, compact gamma cameras could achieve sub-millimeter position resolution without dense pixelation. The authors report that the DNN reduces the average position error from about 0.68 mm to about 0.20 mm, mainly by removing a systematic shift while keeping the statistical noise nearly unchanged. They argue this makes the effective pixel count of the tile an order of magnitude larger, which matters for medical imaging applications such as scintillation cameras.","feed_headline":"Neural network shrinks light-detector pixel size to 0.2 mm","feed_subtitle":"A six-channel silicon photomultiplier tile resolves 10x more spots once a DNN corrects its nonlinear response.","key_machinery":"The central mechanism is the trained deep neural network mapping from six channel amplitudes to planar coordinates. The network has a 6-unit input layer (charge amplitudes normalized by total charge), one or more hidden layers of 64 units with hyperbolic tangent activations, and a two-unit linear output layer for the reconstructed coordinates. It is trained by minimizing the mean squared error between reconstructed and motor-stage positions. The network's role is to learn and compensate the nonlinear, position-dependent distortions of the LG-SiPM tile, including inter-tile gain variations and charge-sharing asymmetries, which the hand-derived linear formula (a center-of-gravity calculation w","core_discovery":"The paper claims that a deep neural network taking the six charge amplitudes from a 2x2 array of linearly-graded SiPMs (normalized by total charge) as input can reconstruct the coordinates of a light spot on the 16x16 mm2 tile with an average granularity of about 0.2 mm, compared with about 0.68 mm for the standard linear center-of-gravity formula. The improvement comes almost entirely from correcting a systematic spatial shift, which drops from roughly 317 µm to 41–93 µm depending on the train/test split, while the random noise component (resolution) stays nearly unchanged at 66–68 µm versus 78–79 µm for the linear model. Across three different splitting strategies — random, chessboard, and","pith_inferences":["A natural next test is a gamma-irradiated scintillator read out by the same tile; if the DNN trained on LED spots does not transfer to the broader scintillation light profile, a domain-adapted training set will be needed — this is not addressed in the paper.","The total-charge cut used to remove edge and inter-tile-gap events implies the reported 6400 regions is an upper bound for the interior field of view; a real camera would need a strategy to handle or reject events near gaps, which the paper does not specify.","The paper's zero-hidden-layer model already shows that allowing a full 6x2 weight matrix (14 free parameters) captures a large fraction of the linear correction; a closed-form least-squares fit might recover much of the gain without a deep network.","The stability of the learned mapping over temperature, bias voltage, and count rate is untested; if it drifts, periodic recalibration would be required in the field."],"forward_implications":["If the result transfers to gamma scintillation detection, a compact handheld gamma camera could achieve sub-millimeter intrinsic resolution with only six readout channels per 16x16 mm2 tile.","Because the DNN mainly removes systematic shifts, detector noise remains the fundamental resolution limit; improving the sensor itself would further push the achievable granularity.","The same training procedure can be applied to larger arrays or other position-sensitive detector geometries, replacing hand-tuned linear gain matrices with a learned calibration.","The demonstrated 0.2 mm granularity is finer than the pixel pitch of many scintillator arrays, suggesting the detector, not the crystal, could set the imaging resolution in future systems."],"fun_headline_variants":["Neural net shrinks SiPM pixels to 0.2 mm","DNN sharpens photodetector spots 10x","SiPM array gains 12x pixels via DNN","Neural network fixes SiPM spatial shift","Deep learning boosts SiPM position accuracy"],"cache_read_input_tokens":10880,"weakest_assumption_plain":"The measured improvement applies to small LED spots on interior grid positions after a total-charge cut; the claim that the same accuracy holds for the broader light distributions, edges, and gaps encountered with real gamma-ray scintillation events is assumed, not demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["Neural net shrinks SiPM pixels to 0.2 mm","DNN sharpens photodetector spots 10x","SiPM array gains 12x pixels via DNN","Neural network fixes SiPM spatial shift","Deep learning boosts SiPM position accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1046,"prompt_tokens":715,"completion_tokens":331,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":254}},"tokens_in":459,"tokens_out":331,"duration_ms":10481,"temperature":1.0,"reasoning_tokens":254,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T18:54:47.838260+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A collimated gamma-ray source scanned across the full tile, including the gaps between the four chips, would directly test transferability: if the DNN's mean position error near edges and gaps exceeds the linear formula's, or if retraining on scintillation light fails to reproduce the 0.2 mm granularity, the central claim fails.","supporting_citations":[],"review_version":1}