{"id":"f3fd2221-34c5-4c59-b81a-0849c76b5950","arxiv_id":"2509.06618","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FACT measures fibril width distributions from negative-contrast SEM images of branched cellulose nanofibrils by combining ML segmentation with skeletonization; low-branching results match manual measurements, but heavily branched networks show substantial differences.","lead":"A computer-vision pipeline called FACT uses machine-learning segmentation and skeletonization to measure the width of branched cellulose nanofibrils in electron microscope images. It could make routine fiber quality measurements faster and less dependent on the analyst doing the measuring.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"High-branch CNF width comparison cannot support 'comparable': FACT and manual measure different quantities (510 vs 320 nm), and no independent width ground truth exists for branched real CNFs.","rationale":"The paper has genuine independent support: wire measurements are close (89.4 vs 92.4 µm and 118.2 vs 124.0 µm), low-branch CNF means agree within pixel resolution, and synthetic branched phantoms recover branch widths reasonably. The concern is specifically about the headline capability for the high-branch, entangled case, which is the motivating application. There, the manual comparison is non-commensurable, as the authors acknowledge ('FACT is not entirely commensurate with the manual measurement approach'). A 190 nm mean difference (510 vs 320 nm) cannot be dismissed as validation; it could reflect correct sampling weighting or a systematic width bias. The paper also documents segmentation edge defects and an unresolved fraction of unwanted skeleton segments. I therefore do not change the CONDITIONAL verdict, but I would make the condition explicit: demonstrate on a known-width branched phantom, or on equivalently weighted manual segment tracing, that FACT's width estimates are unbiased. The concrete phantom test proposed would settle whether this concern actually lands.","tokens_in":21118,"tokens_out":7408,"duration_ms":69442,"concrete_test":"Build a synthetic NegC-SEM-like phantom of a branched network with known per-pixel widths (e.g., rasterized tapered tubes with diameters drawn from the manual distribution, plus realistic image noise and blur), run the trained U-Net and FACT with the high-branch settings (SST=25%, SSF=25°, Gaussian kernel 2), and compare FACT's width map to the ground-truth width map. If the mean absolute error exceeds 2 pixels (44 nm at 22 nm/pixel) or reproduces the 510-vs-320 nm offset, the central claim of comparable width measurement on branched CNFs is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FACT can measure width distributions of hierarchical, branched CNF structures rests most heavily on the high-level-branching validation. That validation compares FACT against manual measurements, but the paper itself states that the two are not commensurate: FACT's histogram (Fig. 13c) is a pixel/length-weighted distribution (mean 510 nm, 104,891 counts), whereas the manual histogram (Fig. 13d) is a per-fibril number-weighted distribution (mean 320 nm, 978 counts). The 190 nm mean offset is attributed to this sampling difference, which is plausible, but it also means the manual data cannot independently confirm FACT's width accuracy on branched networks. The synthetic validations (Figs. 5 and 6) use clean binary phantoms with known constant-width branches; they do not include the segmentation noise, edge dilation, and ambiguous fibril overlaps that real NegC-SEM images produce, which the paper documents (e.g., 1-to-2-pixel dilation in low-branch images, skeleton mismatch in Fig. 11e, and the statement that a robust way to quantify unwanted skeleton segments remains unclear). Moreover, the U-Net segmentation was trained on Weka segmentations without reporting independent segmentation accuracy metrics (e.g., IoU or Dice) on held-out real images. Therefore, the load-bearing condition—that FACT's width values are accurate on real branched CNFs—is not established; the abstract's phrase 'FACT width results were comparable with manual measurements' overstates the evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents FACT (Fibril Analysis for Cellulose Technology), a semi-automated image analysis framework that measures fibril width distributions from negative-contrast scanning electron microscopy (NegC-SEM) images of cellulose nanofibrils. The pipeline consists of image acquisition, machine-learning segmentation (Weka or U-Net), morphological thinning to produce skeletons, skeleton refinement via segment trimming (SST) and filtering (SSF), and width measurement from the distance transform of the binary segmentation. The authors validate FACT on simulated branched structures, optical micrographs of wires, low-branching CNFs from Beaumont et al., and high-branching CNFs from Ringania et al., comparing FACT results against manual width measurements. The paper claims that FACT can measure width distributions of hierarchical, branched, and entangled CNF structures and that once the U-Net model is trained, each image can be analyzed in under 5 minutes.","tokens_in":21367,"tokens_out":3269,"duration_ms":30593,"significance":"If the central claims hold, FACT would be a useful, open-source tool for objective, high-throughput width characterization of branched and entangled CNF networks, addressing a recognized need in the CNF community. The paper's strengths include systematic validation on idealized binary geometries (Figs. 5, 6 and supporting figures), quantitative validation on constant-diameter wires (Fig. 7), and good agreement with manual measurements for low-branching CNFs at multiple magnifications (Figs. 9, 10), where mean differences are within pixel resolution. The public availability of the code on Zenodo and the clear description of the skeleton-refinement steps are also positive features. However, the evidence for the high-branching CNF case, which is the most consequential claim, is weaker, as detailed in the major comments.","major_comments":[{"comment":"The comparison in Fig. 13 does not support the abstract's claim that \"FACT width results were comparable with manual measurements\" for high-branching CNFs. The FACT histogram (Fig. 13c, mean 510 nm, n = 104,891 pixels) is pixel/length-weighted, whereas the manual histogram (Fig. 13d, mean 320 nm, n = 978 fibrils) is per-fibril weighted, and the authors themselves state that \"FACT is not entirely commensurate with the manual measurement approach.\" The 190 nm mean offset is attributed to this sampling difference, which is plausible, but the manual data therefore cannot independently confirm the accuracy of FACT's width values on branched networks. I recommend one of the following: (a) reweight FACT's pixel-based distribution to a per-fibril basis using skeleton-segment lengths, which would allow a direct statistical comparison; or (b) validate on realistic synthetic branched phantoms that include segmentation noise, edge dilation, and overlap ambiguity with known ground-truth widths; or (c) include internal calibration features of known diameter in the SEM images. Without such a test, the high-branching validation remains inconclusive.","section":"High-Level Branching CNFs, Fig. 13"},{"comment":"The U-Net segmentation accuracy is not quantified. Section \"Image Segmentation\" states that Weka segmentation on eight images was used as \"ground truth images\" to train the U-Net, but the Weka segmentations are themselves described as \"too noisy,\" producing jagged edges. No IoU, Dice, or pixel-wise accuracy is reported on held-out real images or against independently labeled regions. Since the width values in FACT are directly derived from the segmented binary image, errors in segmentation propagate directly into width bias and skeleton artifacts. I recommend reporting segmentation metrics on a held-out subset of real NegC-SEM images, at minimum, and ideally comparing the U-Net output against manually corrected segmentations rather than against the same noisy Weka output used for training.","section":"Image Segmentation (U-Net training)"},{"comment":"The width measurement assumes that, after thinning and refinement, the skeleton coincides with the medial axis of each fibril so that twice the distance-transform value equals the local fibril width. The manuscript documents multiple violations of this assumption: skeleton encroachment at branch junctions (Fig. S9), spurious segments from edge defects (Fig. 4), and skeleton mismatch in regions of extensive bundling (Fig. 11e). The paper also states that \"a robust approach to quantifying the fraction of unwanted skeleton segments remains unclear.\" The simulated validations (Figs. 5, 6) use clean binary phantoms without segmentation noise or edge dilation, so they do not bound the magnitude of these errors on real images. I recommend adding a sensitivity analysis that quantifies how many skeleton pixels are affected by encroachment and edge defects in the real CNF images (e.g., by manual labeling or by deliberately introducing controlled edge noise in phantoms) and reporting the resulting effect on the width distribution.","section":"Measurement & Analysis (width definition)"},{"comment":"The SST and SSF parameters are set by visual inspection and trial and error, with values that differ between image sets and even between figures (e.g., Fig. S11 caption states SST=10% and SSF=20°, while Fig. 13 and the text for high-branching CNFs state SST=25% and SSF=25°). This introduces analyst-dependent choices into the pipeline, which partially undermines the claim of removing analyst bias. The paper should either provide a principled, automated criterion for selecting these parameters or explicitly characterize how sensitive the reported width means and distributions are to the chosen SST/SSF values across a plausible range.","section":"Skeleton Refining (SST and SSF parameters)"}],"minor_comments":[{"comment":"There is a typo in the abstract: \"mfeasurement\" should be \"measurement.\"","section":"Abstract"},{"comment":"The SST/SSF values in the Fig. S11 caption (SST=10%, SSF=20°) are inconsistent with the values reported for the same high-branching images in the main text and Fig. 13 (SST=25%, SSF=25°); please correct the inconsistency.","section":"Fig. S11 caption"},{"comment":"The Fig. 13 caption contains a typo: \"ixels along these lines\" should be \"Pixels along these lines.\"","section":"Fig. 13 caption"},{"comment":"Several reference entries contain typos: \"disertation\" should be \"dissertation\" in Ringania (2023), and \"Instutue\" should be \"Institute.\"","section":"References"},{"comment":"The Zenodo link is given as a placeholder in the Conclusion and Declarations; please include the actual DOI or URL before publication so that the reproducibility claim can be verified.","section":"Code availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically sound in its low-branching CNF and wire validations, and the simulated phantoms support the methodological machinery. The main concern is that the high-branching CNF validation, which is the paper's most distinctive claim, lacks an independent width ground truth; the current comparison against manual measurements is between non-commensurate distributions. This is fixable with additional analysis or a clearly worded revision of the claim, so I recommend major revision rather than rejection. The authors should also add quantitative segmentation metrics for the U-Net and either automate or sensitivity-test the SST/SSF parameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a practical methods paper that does something real: it assembles Weka or U-Net segmentation, morphological thinning, skeleton refinement, and distance-transform width measurement into a single pipeline aimed at branched and entangled CNF networks. That integration is genuinely missing from earlier tools, and the validation on synthetic hierarchical branched phantoms and wire micrographs is a reasonable check. The low-branching CNF comparison agrees with manual measurements within pixel resolution, and the wire tests show the width extraction itself works. The authors also deserve credit for documenting junction artifacts, branch encroachment, and edge defects carefully rather than sweeping them under the rug.\n\nThe soft spots are real but concentrated. First, the high-branching validation does not do what the abstract says. FACT's 510 nm mean comes from a pixel/length-weighted distribution with 104,891 counts; the manual 320 nm mean comes from a per-fibril number-weighted distribution with 978 counts. The paper itself says the two are not commensurate. So the manual data cannot independently confirm FACT's width accuracy on branched networks; it only demonstrates that the two sampling schemes differ. The claim that FACT results were 'comparable' with manual measurements is overstatement for this case. Second, the U-Net segmentation was trained on Weka-generated ground truth with no independent segmentation accuracy metric on held-out real images. Given that width values inherit every segmentation edge error, this is the main missing evidence. Third, the code link is a Zenodo placeholder, so reproducibility cannot currently be checked; that is a concrete problem, not a nit. The SST and SSF thresholds are tuned per image by visual trial and error, which limits turnkey use, but this is disclosed and the limitations section is unusually candid.\n\nThe citation pattern is fine, covering prior tools and the relevant standards discussions. The authors are not overselling the underlying machinery; the oversell is limited to the abstract and conclusion phrasing around the high-branching comparison.\n\nThis paper deserves peer review. A serious referee should ask for the code, an independent segmentation accuracy metric, and either a softened claim about high-branching agreement or an independent width benchmark for real branched CNFs. As is, it is a credible contribution for CNF quality control and image analysis researchers, but it needs those fixes before publication.","headline":"A genuinely useful, honest methods paper for semi-automated CNF width measurement, but the high-branching validation does not support the 'comparable to manual' claim as stated.","tokens_in":21943,"tokens_out":1811,"would_cite":true,"duration_ms":19388,"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":"FACT, a semi-automated image-analysis pipeline, measures the width distribution of branched and entangled cellulose nanofibrils from negative-contrast SEM images, producing thousands of width measurements per image with results comparable…","keywords":["cellulose nanofibrils","negative-contrast SEM","image segmentation","machine learning","morphological thinning","skeletonization","width distribution","semi-automated image analysis"],"falsifier":"Take nanofibers or wires whose widths are already known from transmission electron microscopy or lithographic dimensions, image them with the same NegC-SEM protocol, run FACT, and compare each reported width distribution with the known values. If the FACT means are off by more than the two-pixel resolution limit, or if widths systematically inflate near branch points and junctions, then the central comparability claim fails.","tokens_in":20871,"feed_emoji":"🔬","tokens_out":7005,"duration_ms":56391,"temperature":0.7,"pith_summary":"FACT is a semi-automated image-analysis pipeline that measures the widths of cellulose nanofibrils from negative-contrast SEM images, including hierarchical, branched, and entangled networks that existing tools and manual measurement handle poorly. The paper claims that machine-learning segmentation plus morphological thinning produces a one-pixel-wide skeleton of the fibril network, and reading the local distance to the fibril edge along that skeleton yields width distributions comparable to skilled manual measurements. If correct, the method removes analyst bias while measuring every fibril many times along its length, and once the segmentation model is trained, each image can be analyzed in under five minutes. Validation uses idealized branched geometries, standard wire micrographs, and two published CNF data sets with contrasting branching levels.","feed_headline":"Machine learning reads nanofibril widths across whole SEM images","feed_subtitle":"The pipeline segments branched nanofibril networks, measures width along every fibril, and matches manual results without analyst bias.","key_machinery":"The load-bearing mechanism is the skeleton-distance width identity: after thinning, local fibril width at a skeleton pixel is twice the distance-transform value at that pixel, so $\\text{width} = 2d$ where $d$ is the Euclidean distance to the nearest background pixel. The thinning operator erodes a binary object until a one-pixel-wide skeleton remains, and the distance transform records, for each foreground pixel, its distance to the nearest background pixel. Two refinement filters protect this measurement: skeleton segment trimming removes pixels near junction points where branch encroachment distorts width values, and skeleton segment filtering discards short spurious segments that arise from edge defects and have a steep slope in distance-transform value along the segment.","core_discovery":"The central claim is that a fully specified image-analysis pipeline, FACT, can measure the width distribution of hierarchical, branched, and entangled cellulose nanofibrils from NegC-SEM images, and that the resulting width statistics are comparable to manual measurements. After machine-learning segmentation converts each micrograph to a binary foreground/background image, morphological thinning produces a one-pixel-wide skeleton; the distance from each skeleton pixel to the nearest background pixel, doubled, gives the local fibril width. The authors demonstrate the claim on simulated branch hierarchies, on optical micrographs of uniform wires, and on two published CNF data sets, one lightly branched and one heavily branched and networked. They report that FACT resolved the distinct widths of simulated branch levels and gave mean widths close to manual measurements on real CNF images, while producing thousands of width measurements per image instead of one per fibril.","pith_inferences":["A natural next test is applying FACT to the same CNF suspension imaged at several magnifications and checking whether the width distribution becomes magnification-independent after weighting by pixel counts; if not, the remaining differences quantify residual segmentation bias.","The skeleton-as-medial-axis assumption could be checked quantitatively by comparing FACT widths against TEM widths of the same fibrils or against synthetic images with known ground-truth widths, especially near branch points.","The orientation correction suggests a further step: converting FACT's length-weighted width histogram into a number-weighted fibril width distribution, which would make it directly comparable to manual measurements.","Because FACT separates segmentation from measurement, the framework could be extended to other fibrous materials and imaging modalities by retraining the segmentation model while keeping the thinning and distance-transform measurement intact."],"forward_implications":["Width statistics can be collected from entire branched CNF networks, with every fibril measured many times along its length, eliminating analyst bias in fibril selection and placement.","Once a U-Net segmentation model is trained, each NegC-SEM image can be analyzed in under five minutes, making batch comparisons across CNF grades practical.","Width resolution is limited to two pixels, so reliable measurement requires fibrils to span at least five pixels across; digital zoom cannot substitute for higher magnification or higher pixel density.","Skeleton segment lengths and junction counts should not be used as direct measures of fibril length or branching density on real CNFs, because segmentation edge defects create artificial junctions."],"supporting_citations":[{"why":"Supplies the U-Net convolutional architecture used for segmentation of the high-branching CNF images.","marker":"Ronneberger et al. 2015"},{"why":"Provides the low-level branching NegC-SEM imaging protocol and images that FACT analyzes.","marker":"Mattos et al. 2019"},{"why":"Supplies the low-branching CNF material, its NegC-SEM images, and the manual width measurements used for comparison.","marker":"Beaumont et al. 2021"},{"why":"Supplies the high-branching, networked CNF material, its NegC-SEM images, and the manual width measurements used for comparison.","marker":"Ringania et al. 2022"},{"why":"Provides the SEM imaging and reporting recommendations that the study follows when acquiring and analyzing images.","marker":"Moon et al. 2025"},{"why":"Defines the morphological thinning operation that produces the one-pixel-wide skeleton from the binary image.","marker":"Wolfram Research 2010a"}],"fun_headline_variants":["AI thins SEM, maps every fibril width","ML fibril width mapping for tangled nanofibers","FACT: no analyst bias in fibril width measurement","ML segments, thins, and sizes branched CNF networks","Whole-network fibril widths in under five minutes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"FACT assumes that after machine-learning segmentation and thinning, the one-pixel line left through each fibril is exactly its centerline, so that twice the distance from that line to the nearest background pixel equals the true local width.","fun_headline_variants_meta":{"raw":{"variants":["AI thins SEM, maps every fibril width","ML fibril width mapping for tangled nanofibers","FACT: no analyst bias in fibril width measurement","ML segments, thins, and sizes branched CNF networks","Whole-network fibril widths in under five minutes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00129,"raw_usage":{"total_tokens":5316,"prompt_tokens":1040,"completion_tokens":4276,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":4198}},"tokens_in":656,"tokens_out":4276,"duration_ms":25590,"temperature":1.0,"reasoning_tokens":4198,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:14:53.985366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take nanofibers or wires whose widths are already known from transmission electron microscopy or lithographic dimensions, image them with the same NegC-SEM protocol, run FACT, and compare each reported width distribution with the known values. If the FACT means are off by more than the two-pixel resolution limit, or if widths systematically inflate near branch points and junctions, then the central comparability claim fails.","supporting_citations":[],"review_version":2}