Pith. sign in

REVIEW 3 major objections 5 minor 26 references

High-resolution surface topography alone can distinguish carbon ink from carbonized papyrus on Herculaneum fragments.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 16:47 UTC pith:V7OQEXHS

load-bearing objection Topography-only ink signal is real on opened fragments; the closed-scroll resolution target is a useful but still-conditional extrapolation. the 3 major comments →

arxiv 2603.27698 v1 pith:V7OQEXHS submitted 2026-03-29 cs.CV cs.DL

Ink Detection from Surface Topography of the Herculaneum Papyri

Giorgio Angelotti , Federica Nicolardi , Paul Henderson , W. Brent Seales This is my paper
classification cs.CV cs.DL
keywords Herculaneum papyriink detectionsurface topographyoptical profilometrymorphological hypothesismachine learning segmentationlateral resolutionX-ray tomography
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The Herculaneum papyri are carbonized books whose carbon-based ink gives almost no density contrast in ordinary X-ray imaging, so sealed scrolls remain largely unreadable. This paper tests the morphological hypothesis: that ink deposition changes local surface microrelief enough for shape alone to mark written strokes. Using optical profilometry of mechanically opened fragments and segmentation models trained only on heightmaps, the authors obtain held-out Dice scores around 0.88–0.90 at 0.34 µm lateral sampling. Performance falls monotonically as sampling is coarsened, and a high-resolution model loses the signal once effective pixel size exceeds roughly 2 µm. Those resolution curves give concrete spatial targets for any future attempt to read closed scrolls by recovering layer-surface topography from tomography.

Core claim

High-resolution surface topography alone contains a usable signal for ink detection on carbonized Herculaneum papyri. At native 0.34 µm lateral sampling, machine-learning models trained solely on heightmaps separate inked from uninked regions with held-out Dice scores of approximately 0.88–0.90, validating the morphological hypothesis on opened fragments.

What carries the argument

The morphological hypothesis—ink deposition alters local microrelief enough that pure surface-shape information discriminates ink from papyrus—tested by training 2D segmentation models on optical-profilometry heightmaps and measuring how Dice degrades under controlled lateral downsampling.

Load-bearing premise

The ink-related microrelief measured on long-exposed, mechanically opened fragments will still be present and recoverable at similar strength on interior layer surfaces of still-sealed scrolls after virtual unwrapping.

What would settle it

On additional opened fragments or virtually unwrapped sealed layers, train and evaluate heightmap-only ink segmenters at 0.34–1 µm effective sampling with missing-data controls; if Dice falls to chance or near-zero while brightfield ground truth remains clear, the claimed morphological signal does not transfer.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Morphology-only ink detection can succeed without density or composition contrast.
  • Lateral sampling near 1 µm or finer is a practical design target for sealed-scroll tomography that aims to exploit surface microrelief.
  • Vertical quantization matching the lateral pixel size is not the primary bottleneck on this dataset.
  • Cross-papyrus transfer is possible but uneven, so multi-manuscript training data matter for generalization.
  • Opened-fragment profilometry can set resolution requirements before expensive closed-scroll imaging campaigns.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If interior sealed layers preserve comparable microrelief after unwrapping, high-resolution CT protocols should prioritize surface-resolving fidelity over pure attenuation contrast.
  • Any reconstruction step that blurs roughness past roughly 2 µm effective sampling will erase the morphological cue even when nominal voxel size looks adequate.
  • The same shape-based approach may apply to other low-contrast carbonized writing where chemistry-based imaging is difficult.
  • Expanding beyond carefully chosen, still-legible letters would test whether the signal survives typical damage and faint ink.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript tests the morphological hypothesis for carbon ink on carbonized Herculaneum papyri: that ink deposition alters local microrelief enough for topography alone to discriminate inked from uninked regions. Using confocal optical profilometry (0.34 µm lateral sampling, 8 nm vertical sensitivity) on 16 letters from three mechanically opened fragments (PHerc. 248, 250, 500P2), the authors train 2D nnU-Net models on heightmaps with brightfield-derived labels. At native resolution they report held-out Dice ≈0.88–0.90; performance declines monotonically under synthetic block-average coarsening (median Dice 0.890 at 0.34 µm to 0.467 at 10.88 µm; Friedman and Page tests), with a high-resolution model retaining usable signal only up to ~1–2 µm on upsampled coarse inputs. Z-binning (Δz=Δx) changes Dice only modestly; leave-one-papyrus-out yields pooled mean Dice 0.691 with heterogeneous transfer. The authors conclude that high-resolution topography alone supports ink segmentation and that ~1 µm effective lateral sampling is a useful target for morphology-based reading of closed scrolls via X-ray CT.

Significance. If the open-fragment result holds, the paper provides the first quantitative, machine-learning demonstration that surface morphology alone can segment carbon ink on carbonized Herculaneum papyrus at high lateral sampling, with careful missingness controls, non-parametric resolution trends, and isotropic-voxel emulation. That is a genuine contribution to the morphological hypothesis and to heritage imaging. The resolution–performance curves and the ~1 µm guidance are useful design inputs for future CT campaigns, provided they are read as dataset-level observations rather than universal bounds. The impact claim for sealed scrolls remains conditional on open-to-closed transfer of microrelief and SNR after virtual unwrapping—an assumption the Conclusions state carefully but that the abstract and framing still lean on. Strengths include transparent statistics, artifact controls, and planned data release.

major comments (3)
  1. Abstract and Conclusions: the central open-fragment claim (topography-only Dice ~0.88–0.90 at 0.34 µm) is well supported, but the impact framing—that these findings “inform spatial resolution targets for morphology-based reading of closed scrolls through X-ray tomography”—rests on an untested transfer premise. Letters were selected for visible legibility on long-exposed scorze; leave-one-papyrus-out already drops to mean Dice 0.691 (PHerc. 500P2 median 0.475; Table 2); synthetic block averaging does not reproduce CT PSF, partial-volume, or reconstruction effects. The Conclusions correctly call the sealed-scroll case conditional; the abstract and introduction should state the same caveat with equal prominence so the resolution targets are not over-read as CT requirements.
  2. Methods (Labelling) and Discussion: ground-truth masks are drawn on co-registered brightfield photographs. At feathered stroke edges, where height relief approaches background micro-roughness (Figure 2), registration error or optical visibility bias can inflate Dice and couple the “topography-only” evaluation to a modality that already sees ink. A short sensitivity analysis (e.g., morphological erosion/dilation of labels, or independent re-annotation of a subset) would strengthen the claim that the reported scores are not partly label-driven.
  3. Dataset size and generalization (Results, Table 2): n=14 samples / 16 letters from three papyri is small for the cross-papyrus and coarsest-resolution claims. Heterogeneous leave-one-out performance (especially the drop on PHerc. 500P2) shows that transfer is not yet established. The paper should either expand the held-out set or more explicitly bound the resolution targets and generalization statements as preliminary and dataset-specific, as already hinted in the Discussion.
minor comments (5)
  1. Figure 1 caption and Methods: state explicitly that no plane subtraction or tilt correction was applied (mentioned only once in Methods); this affects how readers interpret the heightmaps dominated by inclination and fiber structure.
  2. Table 1 / Figure 3: the Dice=0.70 reference line is useful but arbitrary; a brief note that it is a conventional segmentation threshold (ref. 24) rather than a domain-specific readability criterion would help.
  3. Methods (Preprocessing): Telea inpainting radius = 3 px and the 0.5–99.5 percentile scaling are free parameters; a one-sentence robustness check or statement that finite pixels are restored exactly is already present—consider moving the latter earlier for clarity.
  4. Cross-resolution experiment: “upsampled back to the original grid using bilinear interpolation” introduces a second filter; a short remark that the steeper blue-curve drop partly reflects this pipeline (as the Discussion already suggests) would tighten the interpretation.
  5. Typographical / formatting: occasional spacing issues around units (e.g., “0.34 µm”) and the garbled figure-panel text in the source PDF should be cleaned for production.

Circularity Check

0 steps flagged

No significant circularity: held-out Dice on independent brightfield labels is an empirical external metric, not a result forced by definition or self-citation.

full rationale

The paper’s load-bearing claim is empirical segmentation performance: nnU-Net models trained only on optical-profilometry heightmaps achieve held-out Dice ≈0.88–0.90 against manually drawn ink masks from co-registered brightfield photographs. Labels are independent of the topographic inputs; training loss (Dice+CE) and evaluation Dice are standard external metrics, not quantities defined from the same fit. The morphological hypothesis is prior literature (Parsons 2023; SEM observations) used as motivation and then tested, not defined by the reported scores. Resolution–performance trends, cross-resolution inference, z-binning, missingness controls, and leave-one-papyrus-out are likewise measured outcomes on held-out samples, not predictions that reduce to fitted constants by construction. Self-citations to Vesuvius Challenge / EduceLab virtual-unwrapping work and the morphological hypothesis are background and framing; they do not force the numerical Dice results or the resolution trend. No uniqueness theorem, ansatz smuggled via self-citation, or renaming of a known result appears in the derivation chain. The open-to-sealed transfer premise is a conditional impact claim, not a circular derivation. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper is an empirical CV study. Load-bearing premises are domain assumptions about material and labeling, plus the modeling choice that synthetic block averaging and z-binning stand in for coarser volumetric sampling. No new physical entities are postulated; free parameters are ordinary ML/preprocessing knobs that do not define the scientific claim.

free parameters (3)
  • nnU-Net learning rate and default hyperparameters
    Adam LR 1e-3 and nnU-Net defaults are chosen by convention; they affect absolute Dice but are not fitted to invent the existence of a morphological signal.
  • Telea inpainting radius = 3 px
    Hand-chosen fill radius for non-finite confocal pixels; authors restore original finite values and show missingness is not ink-aligned, so the claim is not driven by this choice.
  • Dice = 0.70 reference threshold
    Used only as a visual/reporting guide from prior segmentation literature, not as a fitted decision boundary for the scientific claim.
axioms (4)
  • domain assumption Ink deposition on carbonized papyrus alters local microrelief enough that a purely morphological signal can discriminate inked from uninked regions (morphological hypothesis).
    Stated as the hypothesis under test; supported by prior SEM observations the paper cites, then measured here.
  • domain assumption Manually drawn brightfield ink masks are sufficiently accurate ground truth for evaluating topography-only segmentation, including at feathered edges.
    Labels come from co-registered photographs (Methods, Labelling); edge registration bias is acknowledged as a limitation.
  • ad hoc to paper Block averaging of profilometry heightmaps (then optional bilinear upsampling) adequately approximates coarser effective lateral sampling for resolution-sensitivity experiments.
    Synthetic degradation used throughout Results; authors note it does not replicate real instrument PSF/MTF.
  • ad hoc to paper Discretizing height into bins of width Δz = Δx emulates isotropic-voxel volumetric sampling well enough to test vertical quantization effects.
    Isotropic-voxel emulation experiment in Methods/Results; modest Dice change is reported as evidence vertical quantization is not primary bottleneck.

pith-pipeline@v1.1.0-grok45 · 15561 in / 3055 out tokens · 38355 ms · 2026-07-13T16:47:56.591750+00:00 · methodology

0 comments
read the original abstract

Reading the Herculaneum papyri is challenging because both the scrolls and the ink, which is carbon-based, are carbonized. In X-ray radiography and tomography, ink detection typically relies on density- or composition-driven contrast, but carbon ink on carbonized papyrus provides little attenuation contrast. Building on the morphological hypothesis, we show that the surface morphology of written regions contains enough signal to distinguish ink from papyrus. To this end, we train machine learning models on three-dimensional optical profilometry from mechanically opened Herculaneum papyri to separate inked and uninked areas. We further quantify how lateral sampling governs learnability and how a native-resolution model behaves on coarsened inputs. We show that high-resolution topography alone contains a usable signal for ink detection. Diminishing segmentation performance with decreasing lateral resolution provides insight into the characteristic spatial scales that must be resolved on our dataset to exploit the morphological signal. These findings inform spatial resolution targets for morphology-based reading of closed scrolls through X-ray tomography.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 16 canonical work pages

  1. [1]

    (Carocci, Roma 2020)

    Longo Auricchio, F., Indelli, G., Leone, G., Del Mastro, G.La Villa dei Papiri: una residenza antica e la sua biblioteca. (Carocci, Roma 2020)

    2020

  2. [2]

    Seales, W. B. and Delattre, D. Virtual unrolling of carbonized Herculaneum scrolls: research status (2007–2012).Cronache Ercolanesi43, 191–208 (2013)

    2007

  3. [3]

    Qualche considerazione sui rotoli ercolanesi ancora non svolti, in vista di un futuro svolgimento virtuale

    Nicolardi, F. Qualche considerazione sui rotoli ercolanesi ancora non svolti, in vista di un futuro svolgimento virtuale. Atene e Roma18, 1-4, 121–130 (2024). doi:10.7347/AR-2024-p121

    work page doi:10.7347/ar-2024-p121 2024

  4. [4]

    and Delattre, D

    Mocella, V ., Brun, E., Ferrero, C. and Delattre, D. Revealing letters in rolled Herculaneum papyri by X-ray phase-contrast imaging.Nature Communications6, 5895 (2015). doi:10.1038/ncomms6895. 7/9

    work page doi:10.1038/ncomms6895 2015

  5. [5]

    doi:10.1038/srep20763

    Tack, P.et al.Tracking ink composition on Herculaneum papyrus scrolls: quantification and speciation of lead by X-ray-based techniques and Monte Carlo simulations.Scientific Reports6, 20763 (2016). doi:10.1038/srep20763

    work page doi:10.1038/srep20763 2016

  6. [6]

    B.et al.From damage to discovery via virtual unwrapping: reading the scroll from En-Gedi.Science Advances 2, e1601247 (2016)

    Seales, W. B.et al.From damage to discovery via virtual unwrapping: reading the scroll from En-Gedi.Science Advances 2, e1601247 (2016). doi:10.1126/sciadv.1601247

    work page doi:10.1126/sciadv.1601247 2016

  7. [7]

    doi:10.1038/srep27227

    Bukreeva, I.et al.Virtual unrolling and deciphering of Herculaneum papyri by X-ray phase-contrast tomography.Scientific Reports6, 27227 (2016). doi:10.1038/srep27227

    work page doi:10.1038/srep27227 2016

  8. [8]

    4).Zeitschrift für Papyrologie und Epigraphik229, 1–13 (2024)

    Nicolardi, F., Parsons, S.et al.Revealing Text from a Still-rolled Herculaneum Papyrus Scroll (PHerc.Paris. 4).Zeitschrift für Papyrologie und Epigraphik229, 1–13 (2024)

    2024

  9. [9]

    4 Revealed through Virtual Unwrapping,Cronache Ercolanesi54, 9–27 (2024)

    Nicolardi, F., Delattre, D.et al.The Final Columns of PHerc.Paris. 4 Revealed through Virtual Unwrapping,Cronache Ercolanesi54, 9–27 (2024)

    2024

  10. [10]

    Virtually Unrolling the Herculaneum Papyri by Diffeomorphic Spiral Fitting

    Henderson, P. Virtually Unrolling the Herculaneum Papyri by Diffeomorphic Spiral Fitting. InProceedings of the IEEE/CVF Winter Conference on Computer Vision (WACV)(2026)

    2026

  11. [11]

    doi:10.1073/pnas.1519958113

    Brun, E.et al.Revealing metallic ink in Herculaneum papyri.Proceedings of the National Academy of Sciences113, 3751–3754 (2016). doi:10.1073/pnas.1519958113

    work page doi:10.1073/pnas.1519958113 2016

  12. [12]

    S.et al.From invisibility to readability: recovering the ink of Herculaneum.PLOS ONE14, e0215775 (2019)

    Parker, C. S.et al.From invisibility to readability: recovering the ink of Herculaneum.PLOS ONE14, e0215775 (2019). doi:10.1371/journal.pone.0215775

    work page doi:10.1371/journal.pone.0215775 2019

  13. [13]

    Theses and Dissertations–Computer Science

    Parsons, Stephen Hard-Hearted Scrolls: A Noninvasive Method for Reading the Herculaneum Papyri. Theses and Dissertations–Computer Science. 138. (2023) doi:10.13023/etd.2023.372

    work page doi:10.13023/etd.2023.372 2023

  14. [14]

    XRF ink analysis of some Herculaneum papyri

    Bonnerot, O., Del Mastro, G., Hammerstaedt, J., Mocella, V ., Rabin, I. XRF ink analysis of some Herculaneum papyri. Zeitschrift für Papyrologie und Epigraphik21650–52. 16.Booras, S.W., Seely, D.R. Multispectral Imaging of the Herculaneum Papyri,Cronache Ercolanesi29, 95–100 (1999)

    1999

  15. [15]

    doi:10.1126/sciadv.aav8936

    Tourni’e, A.et al.Ancient Greek text concealed on the back of unrolled papyrus revealed through shortwave-infrared hyperspectral imaging.Science Advances5, eaav8936 (2019). doi:10.1126/sciadv.aav8936

    work page doi:10.1126/sciadv.aav8936 2019

  16. [16]

    The digital restoration of the Herculaneum Papyri.Cronache Ercolanesi 52, 201–211 (2023)

    Seales, B., Chapman, C., Nicolardi, F., Parker, C.S. The digital restoration of the Herculaneum Papyri.Cronache Ercolanesi 52, 201–211 (2023)

    2023

  17. [17]

    doi:10.1038/s41598-025-19911-w

    Ceccarelli, S.et al.Pulsed thermographic analysis of Herculaneum papyri.Scientific Reports15, 34466 (2025). doi:10.1038/s41598-025-19911-w

    work page doi:10.1038/s41598-025-19911-w 2025

  18. [18]

    and Daffara, C

    Mazzocato, S., Cimino, D. and Daffara, C. Integrated microprofilometry and multispectral imaging for full field analysis of ancient manuscripts.Journal of Cultural Heritage66, 110–116 (2024). doi:10.1016/j.culher.2023.11.014

    work page doi:10.1016/j.culher.2023.11.014 2024

  19. [19]

    B., Mark, R

    Borch, J., Lyne, M. B., Mark, R. E. and Habeger, C. C. Characterization of paper surfaces using optical profilometry. In Handbook of Physical Testing of Paper, 2nd edn (eds Borch, J., Lyne, M. B., Mark, R. E. and Habeger, C. C.) 429–450 (CRC Press, Boca Raton, 2002). doi:10.1201/9780203910498-25

    work page doi:10.1201/9780203910498-25 2002

  20. [20]

    Zienkiewicz, E.et al.Optical profilometry as a method for detecting indented writing.Kryminalistyka i S ˛ adowa Ekspertyza 66(2020)

    2020

  21. [21]

    F., Kohl, S

    Isensee, F., Jaeger, P. F., Kohl, S. A. A., Petersen, J. and Maier-Hein, K. H. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation.Nature Methods18, 203–211 (2021). doi:10.1038/s41592-020-01008-z

    doi:10.1038/s41592-020-01008-z 2021

  22. [22]

    Morphometric analysis of white matter lesions in MR images: method and validation

    Zijdenbos AP, Dawant BM, Margolin RA, Palmer AC. Morphometric analysis of white matter lesions in MR images: method and validation. IEEE Trans Med Imaging. 1994;13(4):716-24. doi: 10.1109/42.363096. PMID: 18218550

    work page doi:10.1109/42.363096 1994

  23. [23]

    An Image Inpainting Technique Based on the Fast Marching Method.Journal of Graphics Tools9(1), 25–36 (2004)

    Telea, A. An Image Inpainting Technique Based on the Fast Marching Method.Journal of Graphics Tools9(1), 25–36 (2004). doi:10.1080/10867651.2004.10487596

    work page doi:10.1080/10867651.2004.10487596 2004

  24. [24]

    and Boykov, Y

    Marin, D., He, Z., Vajda, P., Chatterjee, P., Tsai, S., Yang, F. and Boykov, Y . Efficient segmentation: learning downsampling near semantic boundaries. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)2131– 2141 (2019). doi:10.1109/ICCV .2019.00222

    doi:10.1109/iccv 2019

  25. [25]

    and Pizurica, A

    Benˇcevi’c, M., Qiu, Y ., Gali’c, I. and Pizurica, A. Segment-then-segment: context-preserving crop-based segmentation for large biomedical images.Sensors23, 633 (2023). doi:10.3390/s23020633

    work page doi:10.3390/s23020633 2023

  26. [26]

    Vittorio Emanuele III

    Li, L., Gao, S., Wu, F. and An, X. MBAN: multi-branch attention network for small object detection.PeerJ Computer Science10, e1965 (2024). doi:10.7717/peerj-cs.1965. 8/9 Acknowledgements We thank the curators and conservators responsible for the PHerc. material at the National Library of Naples "Vittorio Emanuele III" for access and guidance, in particula...

    work page doi:10.7717/peerj-cs.1965 2024