Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Photon-counting CT spectra alone identify 10 phantom materials with 97.45% accuracy.

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 · deepseek-v4-flash

2026-08-02 03:33 UTC pith:QDAT5NZW

load-bearing objection Useful dataset and a genuinely careful slice-level split, but the fixed position of every material in the phantom leaves the central 'material identification' claim underdetermined - a position-control experiment is needed. the 3 major comments →

arxiv 2607.13835 v1 pith:QDAT5NZW submitted 2026-07-15 physics.ins-det

Neural-network-based material identification in photon-counting computed tomography using region-of-interest spectral features

A. Miadzvetski , V. Rozhkov , R. Sotenskii , D. Shashurin , G. Chelkov This is my paper
classification physics.ins-det
keywords photon-counting CTmaterial identificationspectral imagingneural networksmultilayer perceptronregion-of-interest spectrathreshold responsephantom study
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.

Photon-counting computed tomography records how a material's reconstructed signal changes with detector energy threshold. This paper tries to establish whether the full threshold-response curve extracted from a small region of interest is enough for a neural network to tell ten phantom materials apart, including pairs whose spectra are nearly identical, such as water versus PMMA and two different gadolinium-based samples. The authors built a dataset from 216 tomographic slices, split at the slice level to avoid correlated samples inflating performance, and compared preprocessing strategies and two architectures. Their final model, a multilayer perceptron fed with the log-transformed spectrum plus its derivative, reached 0.9745 accuracy, balanced accuracy, and macro F1 on a held-out slice set. A sympathetic reader would care because this suggests machine learning can exploit the full spectral shape, potentially reducing the need for manual selection of energy windows in PCCT material identification.

Core claim

The paper demonstrates that an MLP, a fully connected neural network, operating on the threshold-dependent spectral response of a region of interest can classify ten phantom materials (air, bone, water, PMMA, and La-, Nd-, and Gd-based samples) with accuracy, balanced accuracy, and macro F1 all equal to 0.9745 on a test set whose tomographic slices were completely held out from training. The model distinguishes two Gd samples with nearly equal concentrations (80 vs 78.625 mg/mL Gd) and separates the hardest pair, water and PMMA, based on subtle differences in spectral shape and amplitude. The authors emphasize that this is classification of specific experimental sample classes, not universal

What carries the argument

The central object is the ROI spectral vector: the mean reconstructed tomographic signal inside a region of interest, measured at 12 detector threshold settings (THL 45 to 165) and interpolated onto a uniform grid of 121 points. The best-performing representation concatenates log(1+I) with its first derivative with respect to THL, yielding 242 input features. A multilayer perceptron treats these as a flat feature vector, capturing global relationships across thresholds; the slice-level group split (all spectra from the same slice in the same subset) is the validation device that prevents correlated spectra from inflating performance.

Load-bearing premise

The classification assumes the measured spectral differences between classes come from material properties rather than from systematic spatial artifacts—beam hardening, scatter, detector non-uniformity—that are correlated with each material's fixed position in the phantom; if position-correlated artifacts are the discriminative cue, the 0.9745 accuracy will not transfer even to a repositioned phantom or a different scanner geometry.

What would settle it

Apply the trained model to a second scan in which the same ten materials are placed at different spatial positions, or to the same phantom scanned with a slightly altered geometry or reconstruction; if accuracy drops sharply, or if swapping the ROI positions of two materials makes predictions follow position rather than material, the central claim of material identification from spectra is falsified. A simpler check: retrain on the same data with ROI labels shuffled by spatial position—if accuracy remains around 0.97, the model learned location, not material.

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

If this is right

  • Full threshold-response curves carry enough information to separate spectrally close materials such as H2O/PMMA and two Gd-based samples, without preselected energy windows.
  • An MLP generalizes better than a 1D-CNN on these spectra, indicating that global amplitude relationships matter more than local spectral patterns.
  • Adding the first derivative to the log-transformed spectrum improves classification over raw, log-only, or area-normalized features; area normalization hurts, so absolute signal level is informative.
  • Including an 'empty' class causes air/empty confusion and lowers macro F1 to about 0.93, so it is better treated as a background class for segmentation.
  • The model classifies predefined sample classes; transferring it to new concentrations, matrices, or acquisition conditions requires expanded training data and external validation.

Where Pith is reading between the lines

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

  • Because each material sat at a fixed position in the phantom, systematic spatial artifacts (beam hardening, scatter, detector non-uniformity) could supply position-correlated cues; a test with repositioned materials or a different phantom geometry would reveal whether the model truly learns material spectra or location.
  • The high performance even with small ROIs (few hundred pixels) suggests the method could serve as a screening or segmentation step in PCCT material maps when averaged spectra are available.
  • The derivative features' contribution hints that the shape of the threshold response (slope changes) is more informative than magnitude alone; running the same pipeline with energy-calibrated thresholds could clarify whether lanthanide K-edge proximity drives the La/Nd confusions.
  • The equality of accuracy, balanced accuracy, and macro F1 (0.9745) reflects a nearly balanced test set; on naturally imbalanced clinical data these metrics would diverge, so the single-number performance should not be read as a universal operating point.

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 paper presents a neural-network-based method to classify ten materials (air, bone, water, PMMA, and several La-, Nd-, and Gd-based samples) from region-of-interest (ROI) spectral features extracted from photon-counting CT slices acquired at twelve detector threshold settings. The authors compare MLP and 1D-CNN architectures, evaluate several preprocessing strategies, and use a slice-level group split to avoid leakage between correlated ROIs from the same tomographic slice. The final MLP with a log1p_deriv feature representation achieves a test-set accuracy, balanced accuracy, and macro F1 of 0.9745 on the 10-class task. The paper also analyzes dominant confusions, the effect of an auxiliary Empty class, and compares with a gradient-boosting baseline.

Significance. If the result holds, the paper makes a useful contribution to PCCT material identification by demonstrating that full ROI threshold-response curves can separate materials with similar attenuation, and by providing a careful group-splitting protocol that avoids a common leakage pitfall. The authors also give a physically interpretable error analysis (H2O/PMMA, Gd_1/Gd_2, etc.) and are explicit about the limited scope. However, the central claim is weakened by a material–position confound inherent to the single-phantom design; without controlling for this, the reported accuracy does not yet establish material identification rather than spatial-pattern recognition. The paper also lacks uncertainty quantification and reproducibility details for the interpolation/derivative preprocessing.

major comments (3)
  1. [Sections 2.1, 2.4, and 4] The phantom places each material at a fixed (x,y) location in every tomographic slice. The slice-level group split removes correlation between slices but does not break the deterministic mapping between material class and ROI coordinates. Because the test slices share the same spatial layout, a classifier trained only on ROI center coordinates would trivially achieve near-perfect accuracy on the same split. Therefore, the 0.9745 accuracy in Table 8 does not by itself establish that the model identifies materials; it may be exploiting position-correlated artifacts (beam hardening, scatter, detector non-uniformity, path-length variations). The Discussion (Section 8) acknowledges the single-phantom limitation but does not address this confound. The authors should provide a control model trained on ROI coordinates alone (with the same group split) and/or perform external validation with a di
  2. [Section 7.3, Table 8] The reported metrics are point estimates from a single group split with one random seed (Section 4). Since all spectra from a slice share the same slice_id, the effective test sample size is the number of test slices (44), not 47,696 individual spectra. The point estimate 0.9745 therefore carries non-negligible uncertainty. The authors should report confidence intervals for the metrics (e.g., bootstrap by slice) or repeat the group-split procedure with multiple seeds to demonstrate stability. Without such uncertainty quantification, it is hard to judge whether the difference between preprocessing strategies (Table 6) is significant.
  3. [Sections 2.3 and 5.4] The preprocessing pipeline that yields the best result (log1p_deriv) relies on interpolating the measured 12-point spectra onto a 121-point uniform grid and then computing the first derivative with respect to THL. The interpolation method (linear, spline, etc.) is not specified, and the derivative is calculated on the interpolated curve rather than on directly measured points. Depending on the interpolation scheme, the derivative may reflect interpolation artifacts rather than physical spectral shape. This is load-bearing because log1p_deriv provides the largest performance gain (Table 6). The authors should specify the interpolation method and the derivative computation, and ideally show the result is robust to the interpolation choice.
minor comments (5)
  1. [Section 2.2] Equation (2.1) uses T for transmission and later p = -ln T; the notation is standard, but the relation between the reconstructed pixel values and the transmission spectra could be clarified.
  2. [Section 5.3] The area normalization in Eq. (5.3) divides by the sum over all THL values; it would be more standard to call this 'sum normalization' or 'integral normalization.'
  3. [Section 6, Table 4] The learning-rate scheduler ReduceLROnPlateau with patience 5 and early stopping patience 12 are both used; it should be stated which patience governs the stopping decision, since the best epoch is reported as 30.
  4. [Section 7.6] The gradient-boosting baseline uses a row-wise split and different features, so the comparison in Table 12 is not direct. The authors correctly state this, but the table may still be misread; adding a clear disclaimer in the caption would help.
  5. [Data and code availability] The statement 'available from the corresponding author upon reasonable request' is not fully reproducible. If possible, deposit the processed data and code in a public repository with a DOI.

Circularity Check

0 steps flagged

No load-bearing circularity; test accuracy is a genuine out-of-sample measurement, with only non-load-bearing self-citations.

full rationale

The paper's central result (test accuracy/balanced accuracy/macro F1 = 0.9745, Table 8) is not derived from an equation that contains its own conclusion: it is the measured performance of an MLP on a test subset that was not used for training or validation. Section 4 describes a group split by slice_id, with all ROIs from the same slice assigned to a single subset, and states that 'Feature standardization was performed only after the data had been split... estimated exclusively from the training subset.' The 47,696 test spectra therefore constitute an out-of-sample benchmark for the fitted model; no test-label information enters training or model selection. No parameter is fitted to the test labels and then renamed a prediction. The self-citations to [13] (and to [6,7]) are used for phantom provenance and conceptual background ('A closely related phantom configuration and a K-edge-based material-analysis algorithm for PCCT were previously described in [13]'), not as a load-bearing uniqueness or correctness argument; the ML result stands on the measured spectra and holdout evaluation. The main stated limitation is external validity ('the dataset was obtained using a single phantom and a single experimental configuration... it still represents internal validation'), which is a generalization/confound caveat, not a circularity. The fixed material-position correlation in the single phantom is a potential shortcut/validity threat, but it does not make the test-set measurement circular: a coordinate-only control, if run, would be a new experiment testing whether the model uses position, not an algebraic reduction of the reported 0.9745. Accordingly no circular step of the enumerated kinds is present; the only noteworthy item is a minor self-citation that is not load-bearing.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central result is an empirical accuracy measurement, not a derivation. The entries below are the modeling and experimental-design choices on which the reported 0.9745 depends; none are fitted to the test set, but all affect whether the number generalizes.

free parameters (4)
  • MLP hyperparameters (architecture, lr, weight decay, batch size, epochs, dropout) = see Tables 3-4 (e.g., lr=0.001, wd=0.0001, batch=512, max_epochs=30)
    Chosen by the authors; the reported 0.9745 accuracy depends on these choices and on the validation-based selection of the log1p_deriv representation.
  • Preprocessing representation (log1p_deriv) = concatenation of log(1+I) and its derivative with respect to THL (242 features)
    Selected as the best of four preprocessing strategies on the validation set; it is a modeling choice that affects the central result.
  • Interpolation grid (1-THL-unit step from 45 to 165) = 121 interpolated points
    An arbitrary choice for feature-vector dimension; no interpolation scheme (linear/spline) is specified, and no sensitivity analysis is provided.
  • ROI/subsampling design per slice = 16 large-ROI spectra and 90 small-ROI random pixel subsamples per material per slice (Table 1)
    Determines the dataset size, correlation structure, and class balance; these counts are set by hand and not varied.
axioms (4)
  • domain assumption The 216 slice groups are independent replicates of the phantom configuration; slice-level group splitting removes all train/test correlation.
    Section 4 and the Discussion state that the split is by slice_id and that this is internal validation; there is no test of independence across slices (e.g., no spatial gap or scan-order analysis).
  • domain assumption ROI spectral features encode material composition, not systematic spatial artifacts (beam hardening, scatter, detector non-uniformity) tied to fixed ROI locations.
    Each material's ROI sits at a fixed position in the phantom (Figure 2); the model could exploit location-correlated artifacts. The paper does not include a positional control experiment.
  • domain assumption Interpolation from 12 measured THL values to 121 points preserves the discriminative spectral information.
    The interpolation method is not disclosed (Section 2.3); any spectral shape information between measured thresholds is synthesized, and derivative features are computed from the interpolated curve.
  • domain assumption FBP reconstructions at different THL settings are spatially aligned so that the same ROI coordinates correspond to the same physical location.
    Section 2.2 asserts that identical acquisition geometry, reconstruction parameters, and output grid ensure correspondence; this is an untested alignment assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 12892 in / 13021 out tokens · 120610 ms · 2026-08-02T03:33:24.945877+00:00 · methodology

0 comments
read the original abstract

This work investigates the feasibility of neural-network-based material identification based on photon-counting computed tomography (PCCT) data. The input data were spectral vectors extracted from regions of interest (ROIs) in reconstructed tomographic slices of a phantom containing La-, Nd-, and Gd-based samples, as well as air, water, bone, and polymethyl methacrylate (PMMA). The original tomographic slices were acquired at 12 detector threshold settings (THL = 45, 55, 65, 75, 85, 95, 105, 115, 125, 135, 145, 165). For model training, the spectra were interpolated onto a uniform grid from THL 45 to THL 165 with a step of one THL unit. Large ROIs, small ROIs, and their combined dataset were considered. To avoid an overestimated performance caused by correlated ROIs extracted from the same tomographic slice, the data were split into training, validation, and test subsets at the slice level.

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

22 extracted references · 14 canonical work pages

  1. [1]

    Taguchi and J.S

    K. Taguchi and J.S. Iwanczyk,Vision 20/20: Single photon counting x-ray detectors in medical imaging, Med. Phys.40(2013) 100901, doi:10.1118/1.4820371. – 22 –

    work page doi:10.1118/1.4820371 2013

  2. [2]

    Llopart, R

    X. Llopart, R. Ballabriga, M. Campbell, L. Tlustos and W. Wong,Timepix, a 65k programmable pixel readout chip for arrival time, energy and/or photon counting measurements, Nucl. Instrum. Meth. A 581(2007) 485–494, doi:10.1016/j.nima.2007.08.079

    work page doi:10.1016/j.nima.2007.08.079 2007

  3. [3]

    Ballabriga, J

    R. Ballabriga, J. Alozy, G. Blaj, M. Campbell, M. Fiederle, E. Frojdh, E.H.M. Heijne, X. Llopart, M. Pichotka, S. Procz, L. Tlustos and W. Wong,The Medipix3RX: a high resolution, zero dead-time pixel detector readout chip allowing spectroscopic imaging, JINST8(2013) C02016, doi:10.1088/1748-0221/8/02/C02016

    work page doi:10.1088/1748-0221/8/02/c02016 2013

  4. [4]

    Dudak,High-resolution X-ray imaging applications of hybrid-pixel photon counting detectors Timepix, Radiat

    J. Dudak,High-resolution X-ray imaging applications of hybrid-pixel photon counting detectors Timepix, Radiat. Meas.137(2020) 106409, doi:10.1016/j.radmeas.2020.106409

    arXiv 2020

  5. [5]

    Procz, K.-A

    S. Procz, K.-A. Wartig, A. Fauler, A. Zwerger and M. Fiederle,Medipix3 CT for material sciences, JINST8(2013) C01025, doi:10.1088/1748-0221/8/01/C01025

    work page doi:10.1088/1748-0221/8/01/c01025 2013

  6. [6]

    Suslova, A.P

    E.V. Suslova, A.P. Kozlov, D.A. Shashurin, V.A. Rozhkov, R.V. Sotenskii, S.V. Maximov, S.V. Savilov, O.S. Medvedev and G.A. Chelkov,New Composite Contrast Agents Based on Ln and Graphene Matrix for Multi-Energy Computed Tomography, Nanomaterials12(2022) 4110, doi:10.3390/nano12234110

    work page doi:10.3390/nano12234110 2022

  7. [7]

    Suslova, D

    E. Suslova, D. Shashurin, A. Kozlov, S. Maximov, V. Rozhkov, R. Sotenskii, S. Savilov, O. Medvedev and G. Chelkov,Development of La–graphene composite contrasting agents for photon-counting computed tomography, Functional Materials Letters15(2022) 2250029, doi:10.1142/S1793604722500291

    work page doi:10.1142/s1793604722500291 2022

  8. [8]

    Schlomka, E

    J.-P. Schlomka, E. Roessl, R. Dorscheid, S. Dill, G. Martens, T. Istel, C. Bäumer, C. Herrmann, R. Steadman, G. Zeitler, A. Livne and R. Proksa,Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical computed tomography, Phys. Med. Biol.53(2008) 4031–4047, doi:10.1088/0031-9155/53/15/002

    work page doi:10.1088/0031-9155/53/15/002 2008

  9. [9]

    S.A.Si-Mohamed, M.Sigovan, J.C.Hsu, V.Tatard-Leitman, L.Chalabreysse, P.C.Nahaetal.,Invivo molecular K-edge imaging of atherosclerotic plaque using photon-counting CT, Radiology300 (2021) 98–107, doi:10.1148/radiol.2021203968

    work page doi:10.1148/radiol.2021203968 2021

  10. [10]

    Boccalini, R

    S. Boccalini, R. Dessouky, P.-A. Rodesch, H. Lacombe, Y. Yagil, E. Lahoud et al.,Gadolinium K-edge angiography with a spectral photon-counting CT in atherosclerotic rabbits, Diagn. Interv. Imaging104(2023) 490–499, doi:10.1016/j.diii.2023.05.002

    work page doi:10.1016/j.diii.2023.05.002 2023

  11. [11]

    G. Jost, M. McDermott, R. Gutjahr, T. Nowak, B. Schmidt and H. Pietsch,New contrast media for K-edge imaging with photon-counting detector CT, Invest. Radiol.58(2023) 515–522, doi:10.1097/RLI.0000000000000978

    work page doi:10.1097/rli.0000000000000978 2023

  12. [12]

    Clark and C.T

    D.P. Clark and C.T. Badea,Spectral diffusion: an algorithm for robust material decomposition of spectral CT data, Phys. Med. Biol.59(2014) 6445–6466, doi:10.1088/0031-9155/59/21/6445

    work page doi:10.1088/0031-9155/59/21/6445 2014

  13. [13]

    Sotenskii, V.A

    R.V. Sotenskii, V.A. Rozhkov, D.A. Shashurin, E.V. Suslova and G.A. Chelkov,Novel algorithm for qualitative and quantitative material analysis by the K-edges for photon-counting computed tomography, JINST19(2024) P04009, doi:10.1088/1748-0221/19/04/P04009

    work page doi:10.1088/1748-0221/19/04/p04009 2024

  14. [14]

    Abascal, N

    J.F.P.J. Abascal, N. Ducros, V. Pronina, S. Rit, P.-A. Rodesch, T. Broussaud, S. Bussod, P. Douek, A. Hauptmann, S. Arridge and F. Peyrin,Material decomposition in spectral CT using deep learning: a Sim2Real transfer approach, IEEE Access9(2021) 25632–25647, doi:10.1109/ACCESS.2021.3056150

    arXiv 2021

  15. [15]

    Bousse, V.S.S

    A. Bousse, V.S.S. Kandarpa, S. Rit, A. Perelli, M. Li, G. Wang, J. Zhou and G. Wang,Systematic – 23 – review on learning-based spectral CT, IEEE Trans. Radiat. Plasma Med. Sci.8(2024) 113–137, doi:10.1109/TRPMS.2023.3314131

    arXiv 2024

  16. [16]

    Rajagopal, S

    J.R. Rajagopal, S. Rapaka, F. Farhadi et al.,Development of a deep learning based approach for multi-material decomposition in spectral CT: a proof of principle in silico study, Sci. Rep.15(2025) 28814, doi:10.1038/s41598-025-09739-9

    work page doi:10.1038/s41598-025-09739-9 2025

  17. [17]

    Source-Ray, Inc.,SB-120 X-ray generators: SB-120-350, Source-Ray product information, accessed 14 July 2026

    2026

  18. [18]

    Jakubek, M

    J. Jakubek, M. Jakubek, M. Platkevic, P. Soukup, D. Turecek, V. Sykora and D. Vavrik,Large area pixel detector WIDEPIX with full area sensitivity composed of 100 Timepix assemblies with edgeless sensors, JINST9(2014) C04018, doi:10.1088/1748-0221/9/04/C04018

    work page doi:10.1088/1748-0221/9/04/c04018 2014

  19. [19]

    van Aarle, W.J

    W. van Aarle, W.J. Palenstijn, J. Cant, E. Janssens, F. Bleichrodt, A. Dabravolski, J. De Beenhouwer, K.J. Batenburg and J. Sijbers,Fast and flexible X-ray tomography using the ASTRA Toolbox, Opt. Express24(2016) 25129–25147, doi:10.1364/OE.24.025129

    doi:10.1364/oe.24.025129 2016

  20. [20]

    Paszke et al.,PyTorch: An imperative style, high-performance deep learning library, inAdvances in Neural Information Processing Systems 32 (NeurIPS 2019), article 721 (2019)

    A. Paszke et al.,PyTorch: An imperative style, high-performance deep learning library, inAdvances in Neural Information Processing Systems 32 (NeurIPS 2019), article 721 (2019)

    2019

  21. [21]

    Pedregosa et al.,Scikit-learn: Machine Learning in Python, J

    F. Pedregosa et al.,Scikit-learn: Machine Learning in Python, J. Mach. Learn. Res.12(2011) 2825–2830

    2011

  22. [22]

    Loshchilov and F

    I. Loshchilov and F. Hutter,Decoupled Weight Decay Regularization, inProc. International Conference on Learning Representations(ICLR 2019), arXiv:1711.05101, doi:10.48550/arXiv.1711.05101. – 24 –

    Pith/arXiv arXiv doi:10.48550/arxiv.1711.05101 2019