Scout-Dose-TCM: Direct and Prospective Scout-Based Estimation of Personalized Organ Doses from Tube Current Modulated CT Exams

Abdullah-Al-Zubaer Imran; Adam S. Wang; Grant Stevens; Jennie Cao; Justin Ruey Tse; Liyan Sun; Maria Jose Medrano; Sen Wang

REVIEW 3 major objections 6 minor 5 references

Scout-Dose-TCM: Direct and Prospective Scout-Based Estimation of Personalized Organ Doses from Tube Current Modulated CT Exams

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A deep-learning model called Scout-Dose-TCM estimates organ-level radiation doses for tube-current-modulated CT scans prospectively from two scout images and the planned scan range, with a mean absolute error of 7% against Monte Carlo…

desk verdict Solid incremental extension of Scout-Net to tube current modulated CT, but the 'any TCM profile' claim outruns the evidence since final testing only uses the vendor's smooth clinical TCM maps. read the letter →

arxiv 2506.24062 v1 pith:OACMWIVL submitted 2025-06-30 physics.med-ph

Maria Jose Medrano , Sen Wang , Liyan Sun , Abdullah-Al-Zubaer Imran , Jennie Cao , Grant Stevens , Justin Ruey Tse , Adam S. Wang This is my paper · ORCID

classification physics.med-ph

keywords radiationdosetubecurrentmodulationorganscoutimagesdeeplearningMonteCarloestimationdiscretecosinetransformpersonalizedCTdosimetry

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that personalized organ-level radiation dose can be computed before a tube-current-modulated CT scan happens, using only the frontal and lateral scout images plus the planned scan range. On 130 contrast-enhanced chest-abdomen-pelvis exams, the model's predicted doses for six organs averaged 7% absolute error against Monte Carlo benchmarks, compared with 13% for the conventional scanner-output method (Global CTDIvol) and 9% for its organ-adjusted variant (Organ CTDIvol), and it needs no organ segmentations at inference. The enabling trick is to write the tube current modulation map as a sum of 27 discrete cosine basis functions and train the network to predict per-basis organ doses, so the dose for any modulation profile is a weighted recombination. If the result holds, radiologists could get a fast, patient-specific estimate of organ dose before the scan and use it to guide scan protocols, instead of relying on a scanner-output number that does not say where dose lands in the body.

What carries the argument

The load-bearing identity is the discrete cosine transform decomposition of the TCM map. Because dose in a CT scan is linear in tube current, an organ dose under TCM map $k$ can be written as $D_{k,l} = \sum_{n=1}^{N} \beta_{k,n} d_{n,l}$, where $\beta_{k,n}$ are the DCT coefficients of the map and $d_{n,l}$ are basis doses for organ $l$; with $N=27$ bases, the TCM map reconstruction error is below 5%. Scout-Dose-TCM is trained to predict the basis doses $d_{n,l}$ from registered frontal and lateral scouts plus a scan-range channel, using a convolutional feature extractor followed by a fully connected dose head, and a joint loss that penalizes both basis-dose error and final TCM-dose error, with the TCM term weighted by $\alpha = 0.2$. Once trained, any feasible TCM profile can be evaluated by a linear recombination of the predicted basis doses.

What would settle it

In an anthropomorphic phantom study, scan the phantom under several DCT-reconstructed TCM maps at 120 kVp, place thermoluminescent or optically stimulated luminescence dosimeters in the six organs, and compare Scout-Dose-TCM's scout-based predictions with those measurements; the claim fails if the mean absolute error is not below the CTDI-based comparators and does not stay near the reported 7%.

Watch

Extended reading notes

Core claim

The paper's central claim is that Scout-Dose-TCM, a convolutional network with a feature-learning module and a dose-learning module, maps scout images and scan range to the 27 DCT-basis organ-dose coefficients, and that recombining those coefficients with the patient's TCM coefficients yields the organ dose under that modulation. In a five-fold cross-validation on 130 adult patients, the mean absolute percentage error over lungs, kidneys, liver, bladder, spleen, and pancreas was 7%, with statistically significant error reductions versus Global CTDIvol for five organs and versus Organ CTDIvol for liver, bladder, and pancreas. The model also produced predictions in about 0.006 seconds per case, orders of magnitude faster than the Monte Carlo simulation used to generate the reference doses, and it required no organ segmentations at inference. Across all six organs, R-squared values were consistently higher for Scout-Dose-TCM than for either CTDI-based method.

Load-bearing premise

The reference doses that define ground truth come from a Monte Carlo simulation of the scanner that is validated against CTDI phantom measurements, not against measured organ doses in patients, and it treats injected contrast agent as dense water; if those simulated doses are biased, the reported 7% error inherits that bias.

Editorial extensions

If this is right

A clinician could obtain a personalized six-organ dose estimate before the scan begins, in about 6 ms, from data already acquired for positioning (scouts) plus the planned scan range.
Scout-Dose-TCM's error advantage over Global CTDIvol is statistically significant for lungs, liver, bladder, spleen, and pancreas, and over Organ CTDIvol for liver, bladder, and pancreas, so it is a candidate replacement for those prospective estimates.
Because the network outputs DCT basis doses rather than a single dose, the same prediction can be recombined with any feasible TCM profile, making organ dose available as an objective in real-time TCM optimization.
The model requires no organ segmentations during inference, removing the main workflow barrier that kept prospective organ-dose methods from being used before scanning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The linearity that makes DCT recombination possible also suggests the model could be transferred to other scanner models or TCM algorithms by recalibrating the DCT coefficients, without retraining the whole network, as long as dose remains linear in tube current.
The paper does not validate against measured patient organ doses, so the true clinical accuracy could differ from 7%; a dosimeter-based phantom study is the obvious next test.
Because contrast is modeled as dense water and the paper cites evidence that contrast can raise organ dose by up to 50%, the absolute doses for these contrast-enhanced exams may be systematically low even if the relative ranking of organs is correct.
Extending the same architecture to more organs or to direct effective-dose prediction seems straightforward, since each organ only adds another output head over the same learned features.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

The bottom line: this is a competent, incremental extension of the authors' earlier Scout-Net idea to tube-current-modulated CT. The DCT basis trick is the right way to make the model TCM-aware, and the reported 7% organ dose error beats both CTDIvol baselines on their cohort. But before you rely on the 'any TCM profile' claim, note that the held-out test set uses each patient's original vendor-generated TCM map—which is smooth. The random TCM maps appear only in the alpha ablation on validation, not on the test fold. So the model's ability to handle extreme or unusual TCM patterns is asserted, not yet demonstrated.

What's new and good: representing TCM maps as DCT basis functions and predicting basis doses from scouts is a clean, physics-motivated idea. The loss with a TCM dose term is sensible. The 5-fold cross-validation is done properly, the Wilcoxon signed-rank tests are appropriate for paired non-normal errors, and the comparison against Global and Organ CTDIvol is useful. Not needing organ segmentations at inference is a real practical advantage.

Soft spots, in rough order:

1. The 'any TCM' claim. The final evaluation only tests the exact vendor TCM map of each patient. That is a single smooth profile per patient. To support the generalization claim, they should hold out random or perturbed TCM maps on the test fold and show the DCT combination still works. The validation-based alpha selection does not substitute.

2. MC reference organ doses. MC-GPU is validated only against CTDI phantom measurements (within 1.5%). Organ doses in patients are not validated against measurements. The contrast-as-water limitation is acknowledged and could bias doses for contrast-enhanced scans.

3. Truncation error. The <5% TCM representation error at N=27 is stated but not propagated into the dose error budget. For organs near steep modulation gradients (e.g., bladder), a 5% mA error could translate to a larger dose error, though it likely does not explain the whole 7%.

4. No code or data released. Given the reliance on proprietary scanner parameters and TotalSegmentator, reproducibility is limited.

None of these are load-bearing enough to reject the central idea. The method is sound within its stated scope. I would send it to peer review and ask for a test-fold random-TCM evaluation and better uncertainty quantification.

For you: worth reading if you work on CT dosimetry or scout-based methods; I would bring it to a reading group. I would cite it as the TCM-aware successor to Scout-Net, with the caveat about generalization.

Referee Report

3 major / 6 minor

Summary. This paper proposes Scout-Dose-TCM, a deep-learning method that takes frontal and lateral scout images plus the scan range as input and predicts organ-level doses for six organs under tube current modulation (TCM). Rather than predicting a dose for a single TCM map, the model predicts doses for N=27 discrete cosine transform (DCT) basis functions, and the final dose for a given TCM map is computed as a linear combination weighted by that map's DCT coefficients (Eq. 2). Reference doses were obtained from MC-GPU simulations of 130 contrast-enhanced chest-abdomen-pelvis CT exams, with organ segmentation by TotalSegmentator. The method is compared with AAPM TG 204 Global CTDIvol and a TCM-adapted Organ CTDIvol using 5-fold cross-validation. The authors report mean absolute percentage errors of 13%, 9%, and 7% for Global CTDIvol, Organ CTDIvol, and Scout-Dose-TCM, respectively, with statistical significance for most organ-level comparisons, and a runtime of 0.006 s per case.

Significance. If the reported performance is borne out, Scout-Dose-TCM is a clinically relevant advance: it provides prospective, real-time, patient-specific organ dose estimates without requiring organ segmentations or full CT reconstruction at inference, and it outperforms standard CTDI-based estimates. The DCT linear-superposition formulation is an elegant and physically motivated way to decouple patient anatomy from TCM variability, and the held-out-patient cross-validation design is appropriate. The paper also makes good use of open-source components (MC-GPU, TotalSegmentator) and reports statistical tests rather than only means. The main reservations are that the empirical support for 'any TCM profile' is currently incomplete, and the reference doses themselves are simulation-based without organ-level validation; these issues are addressable and should be fixed before publication.

major comments (3)

[Section 3.3, Table II and Figure 6] The central claim that Scout-Dose-TCM estimates organ doses for 'any TCM profile' is not tested on held-out test data. The final evaluation uses only each patient's original vendor-generated TCM map; the random TCM maps used in the α ablation (Figure 5) are confined to the validation split, and no test-fold results with diverse or random TCM profiles are reported. Because Eq. 2 is the mechanism by which arbitrary TCM maps are handled, the paper should include a test-fold evaluation with a set of TCM maps not seen during training (e.g., random feasible DCT coefficients or profiles from other vendors/protocols) and report organ-level errors for those maps. Without this, the reported 7% error may reflect the smoothness of the single vendor's clinical TCM maps rather than the model's dose-prediction skill.
[Section 2.2, Eq. 2] The <5% DCT representation error is stated in tube-current (mA) space and is never propagated to organ dose. Because dose is linear in tube current, the tail DCT basis functions omitted by N=27 can contribute to organ dose with different weights than the mA-space residual suggests; an mA-space residual of 5% does not by itself bound the organ-dose truncation error to 5%. The paper should report the DCT truncation error in dose space for each organ (e.g., by comparing full and truncated MC-GPU organ doses), and should include this truncation error in the error budget when interpreting the 7% mean absolute error.
[Section 2.1 and Discussion (Limitations)] The MC-GPU reference doses are validated only against CTDI phantom measurements (reported within 1.5%), not against measured organ doses in patients, and the contrast agent is modeled as dense water even though the paper acknowledges that contrast can increase organ dose by up to 50%. Since these reference doses are both the training targets and the evaluation benchmark, the reported accuracy is an agreement with a simulation that may carry systematic organ-level bias. The authors should provide an uncertainty or sensitivity analysis for the MC organ doses (e.g., contrast modeling, material density assignment, and the empirical CTDI scaling factor), or at least explicitly frame the 7% figure as consistency with the MC-based benchmark rather than absolute accuracy.

minor comments (6)

[Throughout] The method is referred to inconsistently as 'Scout-Dose-TCM' in the abstract and introduction, 'Scout-Net-TCM' in Table II and Figure 6, and 'final Scout-Net-TCM model' in Section 2.3; please use a single name throughout.
[Section 2.2 and Eq. 5] The equations contain OCR artifacts (e.g., missing subscripts and placeholder characters in Eq. 5), and the definition of the TCM dose loss term does not clarify whether the K random TCM maps are regenerated per batch or per epoch; please correct the typography and specify the sampling procedure for reproducibility.
[Section 3.1, Figure 5] The α ablation reports validation performance but does not state whether the same set of 50 random TCM maps per patient is used across all α values; please clarify whether the random maps are fixed or resampled, as this affects the comparability of the ablation curves.
[Section 2.1, Eq. 1] The notation '𝛩𝛩 = 4' for the number of start angles is confusing; please define Θ as the number of start angles and δ(i) as the specific angle values (0°, 90°, 180°, 270°) to avoid ambiguity.
[Section 2.4] The organ-specific constants A and B for the Global and Organ CTDIvol methods are apparently fit on the training folds, but this is not stated explicitly; please confirm that no test-fold data are used in fitting the comparators, and state this in the evaluation section.
[Discussion] The evaluation is limited to 130 adult patients, a single institution, one scanner vendor, and 120 kVp contrast-enhanced chest-abdomen-pelvis scans; the authors acknowledge this in the limitations, but the 'clinically useful tool' claim would be strengthened by a brief discussion of how the DCT-basis approach is expected to transfer to other scanners and protocols.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: held-out training/evaluation and a physics-based DCT composition keep the central claim independent of its inputs.

full rationale

The central performance claim is not circular. Reference organ doses come from MC-GPU simulations whose scanner model was validated against CTDI phantom measurements, and the same reference doses serve as supervised targets for training and as held-out test targets under 5-fold cross-validation. Equation 2 (D_{k,l} = sum_n beta_{k,n} d_{n,l}) is a linear-superposition statement justified by the proportionality of dose to tube current, not a fitted target; the network learns the basis doses d_{n,l} from scout images, and final doses are composed via the DCT coefficients of the TCM map. The customized loss (Eq. 5) includes a TCM-dose term whose random-TCM targets are generated from the same linear model, so it enforces internal consistency rather than encoding the test answer. Self-citations to prior Scout-Net work and to prior MC-GPU validation (Wang et al. 2021, 2025; Imran et al. 2021-2023) are supporting references, not uniqueness constraints or an ansatz smuggled in; the DCT representation is standard mathematics. The absence of a held-out random-TCM test is a generalization/correctness limitation, not a circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the validity of the Monte Carlo reference doses, the DCT linear superposition, and the chosen basis count and loss weighting. No new physical entities are introduced. The main free parameters are modeling choices rather than fitted physics constants.

free parameters (4)

Number of DCT basis functions N = 27
Chosen by hand to keep TCM representation residual under 5%; affects the expressiveness of the dose model.
Loss weighting alpha = 0.2
Selected via ablation on validation with 50 random TCM maps; balances basis-dose and TCM-dose loss terms.
Number of random TCM maps K per patient = 50
Chosen for the TCM-dose loss during training; no sensitivity analysis is reported.
CTDIvol method organ-specific constants A and B
Fitted by logarithmic regression per organ and per cross-validation fold for the Global and Organ CTDIvol baselines; the paper reports fitted curves but not the values.

assumptions (5)

domain assumption Organ dose is a linear function of tube current, allowing DCT basis doses to be summed with beta coefficients (Eq. 2).
Used throughout Section 2.2 and in the loss function; ignores nonlinear effects such as beam hardening and detector response saturation.
ad hoc to paper N=27 DCT basis functions represent clinically relevant TCM maps with less than 5% residual error.
The basis count is a modeling choice; the residual error is not propagated into the dose error budget.
domain assumption MC-GPU simulation with vendor-specific bowtie, anode heel, and TCM parameters produces accurate reference organ doses.
Validated only against CTDI phantom measurements within 1.5%; organ-level accuracy is assumed.
domain assumption Contrast agent can be modeled as dense water in the Monte Carlo simulation.
The paper states organ dose can increase up to 50% with contrast, so this assumption affects absolute dose values.
domain assumption TotalSegmentator-derived organ masks are accurate enough to extract organ doses from dose maps.
Segmentation errors contribute to uncertainty; the paper cites 7% errors from segmentation in the discussion but does not quantify it in this cohort.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Scout-Dose-TCM: Direct and Prospective Scout-Based Estimation of Personalized Organ Doses from Tube Current Modulated CT Exams." pith.science (2026). https://pith.science/paper/OACMWIVL

@misc{pith2026250624062,
  author       = {Pith},
  title        = {Pith review of: Scout-Dose-TCM: Direct and Prospective Scout-Based Estimation of Personalized Organ Doses from Tube Current Modulated CT Exams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OACMWIVL}},
  note         = {Machine review of arXiv:2506.24062}
}

read the original abstract

This study proposes Scout-Dose-TCM for direct, prospective estimation of organ-level doses under tube current modulation (TCM) and compares its performance to two established methods. We analyzed contrast-enhanced chest-abdomen-pelvis CT scans from 130 adults (120 kVp, TCM). Reference doses for six organs (lungs, kidneys, liver, pancreas, bladder, spleen) were calculated using MC-GPU and TotalSegmentator. Based on these, we trained Scout-Dose-TCM, a deep learning model that predicts organ doses corresponding to discrete cosine transform (DCT) basis functions, enabling real-time estimates for any TCM profile. The model combines a feature learning module that extracts contextual information from lateral and frontal scouts and scan range with a dose learning module that output DCT-based dose estimates. A customized loss function incorporated the DCT formulation during training. For comparison, we implemented size-specific dose estimation per AAPM TG 204 (Global CTDIvol) and its organ-level TCM-adapted version (Organ CTDIvol). A 5-fold cross-validation assessed generalizability by comparing mean absolute percentage dose errors and r-squared correlations with benchmark doses. Average absolute percentage errors were 13% (Global CTDIvol), 9% (Organ CTDIvol), and 7% (Scout-Dose-TCM), with bladder showing the largest discrepancies (15%, 13%, and 9%). Statistical tests confirmed Scout-Dose-TCM significantly reduced errors vs. Global CTDIvol across most organs and improved over Organ CTDIvol for the liver, bladder, and pancreas. It also achieved higher r-squared values, indicating stronger agreement with Monte Carlo benchmarks. Scout-Dose-TCM outperformed Global CTDIvol and was comparable to or better than Organ CTDIvol, without requiring organ segmentations at inference, demonstrating its promise as a tool for prospective organ-level dose estimation in CT.

Figures

Figures reproduced from arXiv: 2506.24062 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

[1]

and Vandermeulen, D

Almeida, D.F., Astudillo, P. and Vandermeulen, D. (2021) ‘Three -dimensional image volumes from two-dimensional digitally reconstructed radiographs: A deep learning approach in lower limb CT scans’, Medical Physics, 48(5), pp. 2448–2457. Badal, A. and Badano, A. (2009) ‘Accelerating Monte Carlo simulations of photon transport in a voxelized geometry using...

work page 2021
[87]

Scout-Net: Prospective Personalized Estimation of CT Organ Doses from Scout Views

Fillon, M., Si-Mohamed, S., Coulon, P., Vuillod, A., Klahr, P. and Boussel, L. (2018) ‘Reduction of patient radiation dose with a new organ based dose modulation technique for thoraco- abdominopelvic computed tomography (CT) (Liver dose right index)’, Diagnostic and Interventional Imaging, 99(7–8), pp. 483–492. Gandhi, D., Crotty, D.J., Stevens, G.M. and ...

work page Pith review arXiv 2018
[169]

and Schmidt, T.G

Principi, S., Lu, Y., Liu, Y., Wang, A., Maslowski, A., Wareing, T., Van Heteren, J. and Schmidt, T.G. (2021) ‘Validation of a deterministic linear Boltzmann transport equation solver for rapid CT dose computation using physical dose measurements in pediatric phantoms’, Medical Physics, 48(12), pp. 8075–8088. Principi, S., Wang, A., Maslowski, A., Wareing...

work page 2021
[220]

and Kofler, J.M

McCollough, C.H., Bruesewitz, M.R. and Kofler, J.M. (2006) ‘CT dose reduction and dose management tools: Overview of available options’, Radiographics, pp. 503–512. Medrano, M.J., Wang, S., Imran, A.-A.-Z., Pal, D., Stevens, G.M., Tse, J. and Wang, A.S. (2023) ‘Organ-Aware Scout -Based Approach for Scout Segmentation and Prospective, Personalized Organ CT...

work page 2006
[710]

X2CT-GAN: Reconstructing CT from Biplanar X-Rays with Generative Adversarial Networks

Summerlin, A.L., Lockhart, M.E., Strang, A.M., Kolettis, P.N., Fineberg, N.S. and Smith, J.K. (2008) ‘Determination of split renal function by 3D reconstruction of CT angiograms: A comparison with gamma camera renography’, American Journal of Roentgenology , 191(5), pp. 1552–1558. Tanriover, B., Fernandez, S., Campenot, E.S., Newhouse, J.H., Oyfe, I., Moh...

work page Pith review arXiv 2008

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

and Vandermeulen, D

Almeida, D.F., Astudillo, P. and Vandermeulen, D. (2021) ‘Three -dimensional image volumes from two-dimensional digitally reconstructed radiographs: A deep learning approach in lower limb CT scans’, Medical Physics, 48(5), pp. 2448–2457. Badal, A. and Badano, A. (2009) ‘Accelerating Monte Carlo simulations of photon transport in a voxelized geometry using...

work page 2021

[2] [87]

Scout-Net: Prospective Personalized Estimation of CT Organ Doses from Scout Views

Fillon, M., Si-Mohamed, S., Coulon, P., Vuillod, A., Klahr, P. and Boussel, L. (2018) ‘Reduction of patient radiation dose with a new organ based dose modulation technique for thoraco- abdominopelvic computed tomography (CT) (Liver dose right index)’, Diagnostic and Interventional Imaging, 99(7–8), pp. 483–492. Gandhi, D., Crotty, D.J., Stevens, G.M. and ...

work page Pith review arXiv 2018

[3] [169]

and Schmidt, T.G

Principi, S., Lu, Y., Liu, Y., Wang, A., Maslowski, A., Wareing, T., Van Heteren, J. and Schmidt, T.G. (2021) ‘Validation of a deterministic linear Boltzmann transport equation solver for rapid CT dose computation using physical dose measurements in pediatric phantoms’, Medical Physics, 48(12), pp. 8075–8088. Principi, S., Wang, A., Maslowski, A., Wareing...

work page 2021

[4] [220]

and Kofler, J.M

McCollough, C.H., Bruesewitz, M.R. and Kofler, J.M. (2006) ‘CT dose reduction and dose management tools: Overview of available options’, Radiographics, pp. 503–512. Medrano, M.J., Wang, S., Imran, A.-A.-Z., Pal, D., Stevens, G.M., Tse, J. and Wang, A.S. (2023) ‘Organ-Aware Scout -Based Approach for Scout Segmentation and Prospective, Personalized Organ CT...

work page 2006

[5] [710]

X2CT-GAN: Reconstructing CT from Biplanar X-Rays with Generative Adversarial Networks

Summerlin, A.L., Lockhart, M.E., Strang, A.M., Kolettis, P.N., Fineberg, N.S. and Smith, J.K. (2008) ‘Determination of split renal function by 3D reconstruction of CT angiograms: A comparison with gamma camera renography’, American Journal of Roentgenology , 191(5), pp. 1552–1558. Tanriover, B., Fernandez, S., Campenot, E.S., Newhouse, J.H., Oyfe, I., Moh...

work page Pith review arXiv 2008