{"id":"ba90ee41-078c-456f-ae1c-17c384276b72","arxiv_id":"2411.16258","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Timepix4 electron MTF at Nyquist is 0.16 at 100 keV and 0.0046 at 200 keV; clustering with ToA and ToT raises these values by factors of 2.12 and 3.16.","lead":"This paper measures how sharply the Timepix4 hybrid pixel detector can image electrons in an electron microscope at 100 and 200 keV, and shows that using arrival time and pulse height to localize each electron improves sharpness by factors of 2.12 and 3.16. The method is a practical step toward sharper electron microscopy images without changing the detector hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 200 keV raw and clustered MTF values rest entirely on an error-function ESF fit that could not be cross-checked by direct differentiation; non-Gaussian LSF tails can shift the Nyquist MTF and the 3.16x improvement far beyond the quoted errors.","rationale":"The reader's weakest_assumption is the right one. The central 200 keV claim depends on an unvalidated functional form for the LSF, and the paper itself flags the missing cross-check: direct differentiation was impossible, so the error-function fit was used for consistency without independent verification. A secondary issue is that Table 1 labels the clustered value 'MTF at Nyquist' even though the clustered image's Nyquist frequency is four times higher (quarter-pixel bins); the comparison is at the physical detector's Nyquist, which is legitimate only if stated that way. This does not change the verdict: the paper is a useful detector characterization, but the 200 keV numbers should be treated as conditional on the ESF model. No code or data is provided, and the quoted uncertainties are statistical only, so the conditions the reader attached (quantify fit-model dependence, report sigma and confidence bands, release the analysis code and representative datasets) are exactly what is needed.","tokens_in":9489,"tokens_out":12733,"duration_ms":127374,"concrete_test":"Re-analyze the 200 keV raw and clustered ESFs without the Gaussian assumption: bin the oversampled edge with coarser or nonuniform bins and numerically differentiate after smoothing (e.g., Savitzky-Golay or Fourier low-pass), and also fit a two-component LSF (narrow core plus exponential or Rutherford tail) to the same ESF. Recompute MTF at Nyquist for both raw and clustered data; if either the absolute 200 keV MTF or the clustered/raw ratio moves by more than roughly 50%, the headline factors are fit-model artifacts and the claims should be rephrased with model-dependent confidence bands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central 200 keV results (raw MTF at Nyquist 0.0046, clustered 0.014, improvement factor 3.16) all come from fitting the ESF with a single error function, Eq. (3), and differentiating the fit. The authors state in Section 4 that for 200 keV 'obtaining the MTF curve without a fit becomes impossible' because of noise, so the 100 keV validation of the erf choice cannot be repeated there. This matters because an erf fit imposes a Gaussian LSF, and the MTF at Nyquist is then controlled entirely by the fitted width sigma; any non-Gaussian tail from backscattered or large-angle electrons (which the authors' own CASINO simulation in Fig. 2 shows are present) changes the high-frequency roll-off. A heavy-tailed LSF with the same variance can give a much larger MTF at Nyquist than the fitted Gaussian predicts, so the reported 0.0046 and the 3.16x factor may be substantially wrong in either direction. The quoted errors in Table 1 (e.g., +/-0.000069) include only sampling-region scatter, not this model uncertainty, so they understate the real uncertainty by a large margin.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports knife-edge MTF measurements of a 300 µm silicon Timepix4 detector for 100 keV and 200 keV electrons in a TEM. It compares a pseudo-particle-counting mode that uses only hit coordinates with a clustered mode that uses ToA and ToT information to compute ToT-weighted centroids, rebinning the data into 2×2 virtual pixels. The measured MTF at the physical-detector Nyquist frequency is 0.16 at 100 keV and 0.0046 at 200 keV in the raw mode, rising to 0.34 and 0.014 after clustering, corresponding to improvement factors of 2.12 and 3.16. The authors argue that temporal and amplitude information can partially correct for the blur caused by extended electron trajectories in the sensor.","tokens_in":9758,"tokens_out":8151,"duration_ms":75887,"significance":"If the results are robust, the paper is a useful contribution to TEM detector characterization: it demonstrates that Timepix4 per-hit timing and energy information can recover spatial resolution beyond the pixel pitch, and it provides a quantitative comparison with earlier Medipix detectors. The 100 keV internal check between fitted and unfitted ESF analysis, the explicit discussion of cluster-shape biases, and the honest identification of the 200 keV fitting difficulty are credits. However, the 200 keV headline numbers rest entirely on an error-function fit with no cross-check, the quoted uncertainties exclude model and algorithm-parameter dependence, and the comparison frequency for the clustered data is not defined unambiguously. With additional validation the paper would be a solid instrument characterization; at present the central 200 keV claim needs revision.","major_comments":[{"comment":"The 200 keV entries in Table 1 (raw MTF 0.0046, clustered MTF 0.014, improvement factor 3.16) are derived entirely from fitting the ESF with the error function of Eq. (3), because the text states that for 200 keV 'obtaining the MTF curve without a fit becomes impossible'. An erf fit imposes a Gaussian LSF, so the MTF at Nyquist is controlled by the fitted width sigma and cannot reflect non-Gaussian tails from backscattered or large-angle electrons, which the authors' own CASINO simulation in Fig. 2 indicates are present. The quoted uncertainties in Table 1 (±0.000069, ±0.00015, ±0.018) reflect only the sampling-region spread and do not include this model dependence. I request reporting of the fit parameters and residuals, a Monte-Carlo or simulated-data validation of the erf assumption at 200 keV, or an explicit statement that the 200 keV improvement factor is model-dependent rather than a measured value.","section":"Section 4, Eq. (3), Table 1"},{"comment":"The clustered data are re-binned into 2×2 virtual pixels of 27.5 µm, so the Nyquist frequency of the clustered image is twice the physical-detector Nyquist frequency. The improvement factors in Table 1 are evaluated at the physical-detector Nyquist frequency, as stated in the Fig. 10 caption, not at the Nyquist frequency of the clustered image. The abstract's phrase 'improved the MTF at Nyquist' is therefore ambiguous. Please define unambiguously which Nyquist frequency is meant in the abstract and Table 1, and justify the comparison at the physical Nyquist frequency if that is the intended figure of merit.","section":"Section 4.2/4.3, Table 1, Fig. 10"},{"comment":"The clustering analysis relies on several user-chosen parameters: 100 ns ToA grouping, the 7×7 spatial window, the 2×2 virtual-pixel subdivision, and the random uniform assignment for the unresolved coordinate of 1×n clusters. No sensitivity study is presented. Because the improvement factors are the central quantitative result, the authors should demonstrate that the factors are stable against reasonable variations of these parameters, or at least quantify how much of the improvement depends on each choice.","section":"Section 4.2"},{"comment":"The paper acknowledges that a ToT-weighted centroid can be displaced from the electron entry point when energy is deposited along the track, as is expected for 200 keV electrons. The text argues that the centroid still improves resolution, but no quantitative estimate of this displacement or its effect on the MTF is given. Since the 200 keV improvement factor is the largest claimed, a simulation-based estimate of the centroid bias, or at least a discussion of its expected effect on the measured MTF, would materially strengthen the paper.","section":"Section 4.3"},{"comment":"The clustered image in Fig. 11 shows a visible line at rows 255/256, attributed to a timestamp misalignment between the two readout halves. This offset could split clusters that straddle the seam and corrupt centroid positions. The paper does not state whether the MTF sampling regions included this seam or whether a correction was applied. Since the MTF values are obtained from sampling regions of 200–350 rows, the seam is relevant to all reported numbers; please clarify whether the seam was excluded or corrected, and if necessary re-analyze the data without it.","section":"Figure 11, Section 4.3"}],"minor_comments":[{"comment":"In the sentence describing cluster spatial footprints, 'refer to the interaction volumes in Figure 3' should refer to Figure 2, which shows the Monte-Carlo interaction volumes.","section":"Section 4.3"},{"comment":"The sentence 'leads to a visible in the final image line' is missing a noun; it should read 'leads to a visible artifact in the final image' or similar.","section":"Section 4.3, Figure 11 caption"},{"comment":"The oversampling formula would be clearer with parentheses and a definition of n_bins: the oversampling factor is approximately n_bins / cos θ, where n_bins is the number of ESF bins per pixel.","section":"Eq. (2)"},{"comment":"Reference [20] lists '0304–3991' where page or article numbers are expected; please check and complete the bibliographic data.","section":"References"},{"comment":"No data or code availability statement is provided; for a measurement paper of this kind, making the analysis scripts or representative datasets available would aid reproducibility assessments.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"This is a straightforward instrument-characterization manuscript. The 100 keV result and the general clustering concept are plausible and publishable, but the 200 keV improvement factor is currently fit-derived with unquantified model uncertainty, and the readout-seam issue could affect the data. The revision does not necessarily require new beam time; synthetic-data validation of the erf assumption and a sensitivity study of the clustering parameters would likely be sufficient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take: the paper's real contribution is the first electron exposure of Timepix4, with MTF measurements at 100 and 200 keV, and a clean demonstration that offline ToT-weighted centroid clustering improves the Nyquist MTF by roughly 2.1x and 3.2x. That is genuinely useful for TEM detector users; it does not open a new capability beyond hybrid pixel detectors, but it quantifies what this chip can do.\n\nThe good parts first. The experimental setup is sensible: slanted knife edge, flat-field correction, Faraday cup calibration. The clustering algorithm is described in enough detail to be plausible: ToA grouping, spatial window, ToT centroid, 2x2 virtual pixel splitting. The authors check that centroid positions are uniformly distributed across the four virtual pixels, which rules out the obvious artifact where improvement just comes from shrinking the effective pixel area. The 100 keV result is internally validated because they compute the MTF both with and without the error-function fit and get consistent values. The comparison to Medipix2/3 results is helpful, and the citation pattern looks fine.\n\nNow the soft spots. The entire 200 keV analysis rests on a single error-function fit to the ESF, because direct differentiation was \"impossible\" from noise. The stress-test is right: the fit imposes a Gaussian LSF, and any non-Gaussian tail from backscattered or large-angle electrons changes the high-frequency MTF. The authors' own CASINO simulation in Fig. 2 shows those tails are present. The 100 keV agreement between fitted and unfitted MTF is reassuring but cannot carry the 200 keV case, where clusters are larger and the signal-to-noise is worse. The quoted uncertainties in Table 1 are only sampling-region spreads; they do not include the ESF fit-model uncertainty. So the reported 0.0046 raw MTF at Nyquist and the 3.16x improvement should be treated as fit-model-dependent until the authors demonstrate otherwise.\n\nThere are also smaller issues: the 100 ns grouping, 7x7 window, and 2x2 splitting are hand-set. The authors do justify them, and the centroid uniformity check helps, but they do not quantify sensitivity to these choices. No data or code is released, which makes independent checking harder.\n\nWho gets value: detector developers and TEM users working with hybrid pixel detectors, plus anyone comparing event-based clustering strategies. It is a solid, honest characterization paper with one load-bearing uncertainty in the 200 keV channel. It deserves peer review, not desk rejection. I would send it out and ask for a revision that addresses the 200 keV fit-model dependence, reports sigma and confidence bands, and ideally releases the clustering code and a representative dataset.","headline":"Useful first Timepix4 electron MTF data with a credible clustering gain, but the 200 keV numbers are entirely fit-derived and the quoted errors omit the dominant model uncertainty.","tokens_in":10336,"tokens_out":3129,"would_cite":true,"duration_ms":30818,"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":"The Timepix4 detector's arrival-time and charge data sharpen electron images up to 3.16x at Nyquist.","keywords":["MTF","Transmission Electron Microscopy","Electron Imaging","Timepix4","Hybrid Pixel Detector","ToT-weighted centroiding","Modulation Transfer Function"],"falsifier":"Collect a 200 keV knife-edge dataset with substantially more hits per pixel and differentiate the edge spread function directly, without any error-function fit; if the resulting MTF at Nyquist deviates materially from 0.0046, or the improvement factor from 3.16, the fitted values are artifacts of the assumed Gaussian line spread.","tokens_in":9304,"feed_emoji":"🔬","tokens_out":6964,"duration_ms":59153,"temperature":0.7,"pith_summary":"Timepix4, a hybrid silicon pixel detector read out in data-driven mode, can measure not only where an electron lands but when the charge arrives and how much charge each pixel collects. This paper asks whether that extra temporal and amplitude information can recover spatial resolution lost when an electron scatters sideways inside the 300 µm silicon sensor. Using a slanted knife edge in a transmission electron microscope, the authors measure the modulation transfer function at 100 and 200 keV in two modes: a pseudo-particle-counting mode that mimics a Medipix3, and a clustered mode that replaces each multi-pixel electron hit with the charge-weighted centroid of its cluster, rebinned into quarter-pixel virtual pixels. The clustered mode raises the MTF at Nyquist frequency from 0.16 to 0.34 at 100 keV and from 0.0046 to 0.014 at 200 keV, improvements by factors of 2.12 and 3.16. The result matters because it shows the blur caused by extended electron trajectories inside the sensor can be partially corrected in software rather than by thinning the sensor or shrinking pixels.","feed_headline":"Timepix4 clustering boosts electron-image resolution 2-3x","feed_subtitle":"ToT-weighted centroids lift MTF at Nyquist from 0.16 to 0.34 at 100 keV and 0.0046 to 0.014 at 200 keV.","key_machinery":"The mechanism is ToT-weighted centroid clustering with temporal gating. Each electron event is recovered by sorting hits by Time of Arrival into 100 ns intervals, chosen to match the tens-of-nanosecond drift time of charge in the 300 µm sensor, and then grouping hits whose coordinates fall within a 7x7 pixel window. The cluster is replaced by its Time over Threshold (ToT) weighted centroid, that is, the charge-weighted center of the deposited ionization, and that single sub-pixel coordinate is assigned to one of four virtual pixels per physical pixel. One-by-n clusters, where no centroid can be computed along one axis, are assigned a random uniform position within the pixel for that coordinate. This machinery works because ToT gives a per-pixel measure of deposited charge and because the centroid, while not exactly the entry point for a long track, is systematically closer to it than the whole cluster is.","core_discovery":"On its own terms, the paper establishes that a simple ToT-weighted centroid clustering algorithm, applied offline to Timepix4 event data, substantially improves the spatial resolution of electron images. For 100 keV electrons the MTF at Nyquist rises from 0.16 to 0.34, and for 200 keV it rises from 0.0046 to 0.014, corresponding to improvement factors of 2.12 and 3.16. The improvement is attributed to better estimation of the electron entry point: ToT weights the centroid toward the pixel or pixels where the ionization track deposits most charge, and rebinning into 2x2 virtual pixels samples the edge at a higher effective rate. The authors check that the centroid distribution across the four virtual pixels is even, ruling out an artifact from shrinking the sensitive area, and they show that the improvement is visible in test images as increased contrast and clarity. The raw 200 keV MTF is so low that the edge spread function could not be differentiated directly and was fitted with an error function, so the reported numbers for that energy rest on the fitted line spread being Gaussian.","pith_inferences":["The paper's logic implies that the same ToT-weighted centroid clustering should also improve the detective quantum efficiency at high spatial frequencies, a quantity not reported here, because it moves signal weight closer to the true entry position.","The 2.12 and 3.16 improvement factors are tied to the threshold setting used in the experiment; a lower threshold would capture more of each ionization track and shift the centroid, so these numbers are thresholds-dependent rather than universal detector properties.","The method should transfer to other Timepix4 applications where charged particles produce multi-pixel clusters, such as low-energy electron microscopy or ion imaging, because the required per-hit ToA and ToT data are detector features rather than TEM-specific.","A natural extension, which the paper lists as future work, is to replace the fixed 2x2 subdivision with a machine-learned or Monte-Carlo-calibrated entry-point estimator; that could push the effective resolution beyond what a simple charge-weighted centroid achieves."],"forward_implications":["At 100 keV the clustering lifts the MTF at Nyquist from 0.16 to 0.34, a 2.12-fold gain; at 200 keV it lifts it from 0.0046 to 0.014, a 3.16-fold gain.","The same Timepix4 hardware, without sensor thinning or smaller pixels, can image electrons with better contrast and clarity, as shown in the clustered Fresnel-disk test images.","Because these gains were achieved before timewalk correction or per-pixel energy calibration, further calibration work is expected to improve the centroid accuracy and push the MTF higher.","The raw Timepix4 MTF curves are consistent with earlier Medipix2 and Medipix3 electron measurements, validating the setup and making the clustered improvement directly comparable to those detectors.","The improvement is a partial correction of the blur created by extended electron trajectories inside the sensing layer, not a removal of the underlying physical spread of charge."],"supporting_citations":[{"why":"Supplies the ISO 12233 slanted-edge MTF method, which the paper adapts with 8-bin oversampling and small slant angles.","marker":"[11]"},{"why":"Provides the Medipix3 baseline MTF at Nyquist for the same 300 µm p-on-n sensor, the comparison that validates the raw Timepix4 curves.","marker":"[20]"},{"why":"Previous electron-detector MTF work that fits the ESF with an error function, the method used for all datasets here.","marker":"[16]"},{"why":"Second precedent for the error-function ESF fit used to derive the LSF and MTF.","marker":"[17]"},{"why":"Establishes sub-pixel resolution by ToT-based centroiding in Timepix3 X-ray imaging, the precedent for the clustering step.","marker":"[22]"},{"why":"Demonstrates centroiding of large clusters for heavy charged particles, supporting use of centroids for particle tracks.","marker":"[23]"},{"why":"Introduces the Timepix4 ASIC with ToA and ToT in each hit packet, the capability the clustering method depends on.","marker":"[2]"}],"fun_headline_variants":["ToT-weighted clustering sharpens electron images 2-3x","Timepix4 centroid trick boosts TEM resolution up to 3x","How Timepix4's arrival-time data rescues 200 keV MTF","Clustering electrons: Timepix4 MTF from 0.16 to 0.34","Timepix4 data processing lifts Nyquist MTF by 2-3x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 200 keV result depends on the assumption that the line spread function is Gaussian: with the limited statistics available, the edge profile had to be fitted with an error function because direct differentiation was impossible, and a non-Gaussian line spread, for example with long tails from backscattered electrons, would change the reported MTF and improvement factor.","fun_headline_variants_meta":{"raw":{"variants":["ToT-weighted clustering sharpens electron images 2-3x","Timepix4 centroid trick boosts TEM resolution up to 3x","How Timepix4's arrival-time data rescues 200 keV MTF","Clustering electrons: Timepix4 MTF from 0.16 to 0.34","Timepix4 data processing lifts Nyquist MTF by 2-3x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000139,"raw_usage":{"total_tokens":1164,"prompt_tokens":962,"completion_tokens":202,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":96}},"tokens_in":578,"tokens_out":202,"duration_ms":13841,"temperature":1.0,"reasoning_tokens":96,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:19:15.298108+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a 200 keV knife-edge dataset with substantially more hits per pixel and differentiate the edge spread function directly, without any error-function fit; if the resulting MTF at Nyquist deviates materially from 0.0046, or the improvement factor from 3.16, the fitted values are artifacts of the assumed Gaussian line spread.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ISO 12233 slanted-edge MTF method, which the paper adapts with 8-bin oversampling and small slant angles."},{"cited_title":"Paton, M.C","cited_arxiv_id":null,"evidence_quote":"Provides the Medipix3 baseline MTF at Nyquist for the same 300 µm p-on-n sensor, the comparison that validates the raw Timepix4 curves."},{"cited_title":"Battaglia, D","cited_arxiv_id":null,"evidence_quote":"Previous electron-detector MTF work that fits the ESF with an error function, the method used for all datasets here."},{"cited_title":"Deptuch, A","cited_arxiv_id":null,"evidence_quote":"Second precedent for the error-function ESF fit used to derive the LSF and MTF."},{"cited_title":"Khalil, E","cited_arxiv_id":null,"evidence_quote":"Establishes sub-pixel resolution by ToT-based centroiding in Timepix3 X-ray imaging, the precedent for the clustering step."},{"cited_title":"Jakubek, A","cited_arxiv_id":null,"evidence_quote":"Demonstrates centroiding of large clusters for heavy charged particles, supporting use of centroids for particle tracks."},{"cited_title":"Llopart, J","cited_arxiv_id":null,"evidence_quote":"Introduces the Timepix4 ASIC with ToA and ToT in each hit packet, the capability the clustering method depends on."}],"review_version":1}