Pith. sign in

REVIEW 3 major objections 6 minor 2 references

Incorporating human and learned domain knowledge into training deep neural networks: A differentiable dose volume histogram and adversarial inspired framework for generating Pareto optimal dose distributions in radiation therapy

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a differentiable dose-volume-histogram loss, combines it with adversarial training, and reports that it substantially reduces clinically relevant errors in predicting Pareto-optimal prostate IMRT dose distributions.

desk verdict A solid empirical comparison with a genuinely new differentiable DVH loss, but the head-to-head significance claims rest on treating 1,200 plans per patient as independent. read the letter →

arxiv 1908.05874 v2 pith:RESH6ZGX submitted 2019-08-16 physics.med-ph cs.CVcs.LGeess.IV

Dan Nguyen , Rafe McBeth , Azar Sadeghnejad Barkousaraie , Gyanendra Bohara , Chenyang Shen , Xun Jia , Steve Jiang This is my paper · ORCID
classification physics.med-phcs.CVcs.LGeess.IV
keywords dosepredictionvolumehistogramlossadversarialParetooptimalIMRTprostatecancerdeeplearningdifferentiablefunction
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 tries to show that a neural network can learn to predict Pareto-optimal radiation dose distributions for prostate IMRT in real time, and that the way to make those predictions clinically accurate is to put domain knowledge into the training loss. It introduces a differentiable approximation of the dose-volume histogram (DVH), the standard clinical curve showing what fraction of a structure receives at least a given dose, and uses it alongside an adversarial loss and mean-squared error. Across four trained models, the MSE+DVH+ADV combination had the lowest errors on clinical metrics such as conformation, homogeneity, R50, and D95/D98/D99; the MSE-only model was consistently worse, with conformation error about 3.5 times larger and D95 error about 2.4 times larger. If the claim holds, a physician could interactively adjust tradeoff weights and see a predicted dose in about 0.6 seconds, including display and DVH computation, without recalculating dose influence matrices.

What carries the argument

The load-bearing object is the differentiable DVH surrogate $\widetilde{DVH}_s(D,M_s)$, computed for each structure by summing sigmoid functions of the dose relative to each threshold, weighted by the structure mask, which makes $\partial \widetilde{DVH}/\partial D$ available for backpropagation. The paper sets the steepness $m=1$, choosing smooth gradients over a sharper histogram approximation, and checks on a toy example that the loss landscape keeps the same minima. This term enters the total generator loss $L_{\mathrm{Total}} = L_{\mathrm{MSE}} + \lambda_{\mathrm{DVH}} L_{\mathrm{DVH}} + \lambda_{\mathrm{ADV}} L_{\mathrm{ADV}}^G$ with $\lambda_{\mathrm{DVH}}=0.1$ and $\lambda_{\mathrm{ADV}}=0.001$, alongside an adversarial discriminator that sees the anatomy masks plus either the true or predicted dose. The DVH loss is what directly translates radiation-oncology evaluation metrics into a training signal; the adversarial term supplies learned, unformulated features.

What would settle it

Retrain the four models on ground-truth plans produced by a different dose-calculation engine, a different beam arrangement, or a deliverable MLC-sequenced planning pipeline and compare the same clinical metrics; the central claim is falsified if the error advantage of MSE+DVH+ADV over MSE disappears when the ground-truth proxy changes.

Watch

Extended reading notes

Core claim

The central claim is that a differentiable DVH loss is the largest driver of clinically relevant prediction quality, and that an adversarial loss adds a further gain by capturing features not written into the explicit loss. The DVH loss replaces the non-differentiable volume-above-threshold count with a sigmoid-weighted soft count, so gradients flow from the clinical curve back into the network. Trained for 100,000 iterations on plans from 54 patients and evaluated on 12,000 plans from 10 held-out patients, the MSE+DVH+ADV model achieved a conformation error of 0.038, homogeneity of 0.026, R50 of 0.298, and D95/D98/D99 errors of 1.65%, 2.14%, and 2.43% of prescription dose, compared with 0.134, 0.041, 0.520, 3.91%, 4.33%, and 4.60% for MSE alone. The paper reports that most differences against the combined model were statistically significant, and that raw prediction takes 0.052 seconds per plan.

Load-bearing premise

The load-bearing premise is that the 84,000 pseudo-randomly weighted, optimized plans, computed with a commercial dose engine and a simplified weighted least-squares objective, faithfully represent the clinically relevant Pareto surface for prostate IMRT, so that agreeing with these proxy plans is the right test of a dose prediction.

Editorial extensions

If this is right

  • The DVH loss substantially reduces errors on clinical metrics such as conformation, R50, D95, D98, and D99 relative to MSE-only training, so explicit clinical knowledge can be the main driver of dose-prediction quality.
  • Adding the adversarial loss on top yields the best overall model, implying that learned discriminator features can complement hand-written domain losses rather than replace them.
  • Because prediction takes 0.052 seconds (roughly 0.6 seconds including display and DVH calculation), a physician could interactively explore PTV/OAR tradeoffs and hand the dosimetrist a tangible target dose, potentially shortening plan-tuning feedback loops.
  • At inference the network needs no dose influence matrix, bypassing the roughly 32-minute per-patient dose calculation that dominates the optimization-based Pareto-planning pipeline.
  • The paper's order-of-magnitude weighting scheme for human and learned loss terms is offered as a recipe for applying the same approach to other treatment sites, beam configurations, or optimization formulations.

Reading between the lines

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

  • This suggests the differentiable DVH loss should transfer to any voxelwise prediction task whose evaluation is summarized by a dose-volume curve, such as VMAT dose prediction or adaptive replanning; nothing in the loss is specific to prostate IMRT.
  • The large gap between MSE-only and DVH-trained models indicates that voxel-wise MSE is a weak proxy for clinical acceptability in radiotherapy, so reframing other clinical summaries as differentiable losses could be a general recipe for medical image prediction.
  • A natural stress test would be to retrain the four models on ground-truth plans from a different dose-calculation engine, deliverable MLC-sequenced plans, or physician-approved clinical plans; the reported error advantages might not persist if the weighted-least-squares Pareto proxy differs from routine clinical planning.
  • Because the DVH loss alone improved performance more than the adversarial loss alone, explicit domain knowledge appears more sample-efficient than learned feature discrimination in this setting; varying training set size and measuring the crossover would test that reading.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a differentiable approximation of the dose-volume histogram (DVH) as a training loss for deep neural networks, and combines it with a least-squares adversarial loss to predict Pareto-optimal IMRT dose distributions for prostate cancer. The authors generate 1,200 Pareto plans per patient for 70 patients using a weighted least-squares objective (Eq. 2) with pseudo-random trade-off weights, then train four identical U-net models with different loss combinations: MSE, MSE+ADV, MSE+DVH, and MSE+DVH+ADV. On 10 held-out test patients, the MSE+DVH+ADV model yields the lowest average errors on conformation, homogeneity, R50, and D95/D98/D99, and the authors report that these differences are statistically significant with the largest p-value 0.007. They also report low mean and max dose errors across structures and a prediction time of 0.052 seconds, arguing for real-time physician interaction with the trade-off space.

Significance. If the reported gains are real, the work is a useful contribution to knowledge-based planning: it makes a clinically relevant and previously non-differentiable metric (DVH) usable as a loss, and it provides a systematic head-to-head comparison of domain-knowledge and adversarial losses within an identical architecture and data setup. The strength of the comparison is the controlled design: the same U-net backbone, same training data, same 10 held-out patients, and consistent weight sampling are used across all four models. The paper also provides a constructive demonstration that a smooth sigmoid-based DVH approximation preserves the relevant minima (Section II.1.2, Figures 2 and 3), which is a useful methodological insight. However, the central inferential claim — that MSE+DVH+ADV is statistically significantly better than the alternatives — rests on treating 12,000 test plans from only 10 patients as independent samples. The descriptive ranking of mean errors may well be correct, and the proposed loss is a plausible improvement, but the current statistical evidence is not credible without a clustered analysis.

major comments (3)
  1. [§III, Figures 8–10, p. 18] The 99% confidence intervals are computed as x̄ ± 2.576·σ/√n over all test plans, and the text states that the large number of test plans makes the comparisons statistically significant (largest p = 0.007). Since the 12,000 test plans come from only 10 patients, with 1,200 plans per patient sharing the same anatomy, beam arrangement, and dose-influence matrix, the plan-level errors are strongly correlated within each patient. Treating each plan as an independent observation inflates the effective sample size by about three orders of magnitude; with an effective sample size near 10, the reported p-values are not credible. The authors should re-analyze the data using patient-level summary statistics (e.g., per-patient mean errors, then a paired test across 10 patients) or a mixed-effects model that accounts for patient-level clustering, and report the resulting confidence intervals and p-values. The descriptive ordering of mean errors may survive, but the current statistical support for the central comparative claim is unestablished without this correction.
  2. [§II.1.2 and §III] The headline evaluation metrics — conformation number, homogeneity, R50, D95, D98, and D99 — are all derived directly from DVHs, and the DVH loss (Eq. 9) is explicitly constructed to minimize DVH differences. Therefore, the improvement of MSE+DVH and MSE+DVH+ADV over MSE on these metrics is partly expected by construction; the paper even acknowledges this on p. 16. To make the claim that the method improves clinically relevant prediction in a way that is not merely circular, the authors should also report a metric that is not directly tied to the DVH loss, such as voxel-wise absolute error, gamma pass rate, or a clinical acceptability score, and should clearly distinguish the by-construction gain from a generalization benefit on unseen patients.
  3. [§II.1 and Table 1] The ground-truth Pareto plans are generated from a simplified weighted least-squares objective (Eq. 2) using Eclipse AAA influence matrices for a fixed 7-beam coplanar setup. As the authors note in the Discussion, the resulting predictions are not guaranteed to be deliverable and the conclusions are limited to this planning paradigm. This limitation is acknowledged, but it also means that the numerical errors reported on test patients are errors relative to this particular computational proxy, not necessarily relative to clinically deliverable IMRT plans. The manuscript should state this more prominently in the abstract and results, so that readers do not interpret the absolute error values as deliverability errors.
minor comments (6)
  1. [Figure 8 caption] There is a typo in the caption: 'Predicion errors' should be 'Prediction errors'.
  2. [p. 16, paragraph after Figure 8] The phrase 'comformation' on p. 16 is a typo and should be 'conformation'.
  3. [Section II, subsections] The subsection numbering is inconsistent: Section II.1.1 and II.1.2 appear after Section II.2, which makes the cross-references in the text confusing. The subsections should be renumbered sequentially.
  4. [Section II.2] The equation numbering in the text is inconsistent: Equation (5) is referred to as 'Equation 3' in the paragraph that introduces the DVH approximation. The equation numbers should be checked and harmonized throughout.
  5. [Section II.4] The description of the adversarial training alternation says the U-net and discriminator 'alternate every 100 iterations,' but it is unclear whether the discriminator updates once per 100 generator updates or whether the alternation is symmetric. This should be clarified for reproducibility.
  6. [Section II.1] The manuscript states that the dose influence arrays were calculated with a 7-beam setup, but later text in Section II.1 says 'on average it takes 32 minutes to use the Eclipse AAA engine to compute beamlet-based dose influence matrices for a 5 beam IMRT plan.' The discrepancy between 7 and 5 beams should be resolved or explained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DVH loss is deliberately optimized against DVH-derived metrics, and the comparative gains are empirical on held-out data, not derived from the loss by construction.

full rationale

The paper's central contribution is an empirical comparison of four training losses (MSE, MSE+ADV, MSE+DVH, MSE+DVH+ADV) on held-out test plans. The fact that the differentiable DVH loss directly penalizes DVH differences while the headline evaluation metrics (conformation, homogeneity, R50, D95, D98, D99) are all DVH-based is a matched objective, not a circular derivation; the paper explicitly acknowledges this mechanism in the Discussion ('the performance of our model improved with respect to our domain relevant metrics, because our domain knowledge-based losses are designed to reduce the error in very specific areas of the model's prediction'). The gains are measured on 12,000 held-out plans from 10 test patients, so the comparison is not fitted to the evaluation set. The only self-citations (refs 48, 55, 56) are used for secondary calibration of loss weights or for time-savings motivation, and are not load-bearing for the main claim. No uniqueness theorem, ansatz, or re-labeling of known results is present. A separate statistical concern is that the 99% confidence intervals and p-values treat the 12,000 test plans as independent samples when they are clustered within 10 patients; this is a pseudo-replication/correctness issue, not circularity, and does not affect the descriptive ordering of mean errors.

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

The central claim rests on hand-set loss weights, the DVH approximation steepness, and the data-generation choices that define what counts as a Pareto-optimal plan. No new physical entities are introduced; the differentiable DVH is a mathematical surrogate and the networks are computational constructs.

free parameters (5)
  • lambda_DVH = 0.1
    Hand-selected in Table 2 to match the order of magnitude of the converged MSE loss; affects the relative weighting of the DVH loss in Eq. 10.
  • lambda_ADV = 0.001
    Hand-selected to balance the LSGAN loss scale; chosen from estimated converged loss magnitudes rather than optimized on validation data.
  • DVH sigmoid steepness m = 1
    Chosen after the toy example in Figures 2-3 as smoother and better behaved; no formal optimization or proof of optimality.
  • DVH threshold grid and bin widths = not specified
    Equations 5-8 define thresholds d_t and bin widths, but the paper does not report the number of bins n_t or the dose grid used, even though these determine the loss and the predictions.
  • Controlled weight bounds = bladder/rectum rand(0,0.2), fem heads rand(0,0.1), shell rand(0,0.1), skin rand(0,0.3)
    Selected by trial-and-error so generated plans are likely to fall within clinically relevant bounds, as stated in Methods II.1.
assumptions (5)
  • standard math Scalarization with positive weights yields Pareto-optimal solutions of the convex multiobjective problem.
    Invoked in Methods II.1 around Eq. 2, with reference 63 as support.
  • standard math Chambolle-Pock primal-dual algorithm converges to a global optimum for the convex weighted least-squares problem.
    Used to generate all ground-truth Pareto plans; convergence is assumed from reference 64 rather than verified here.
  • domain assumption Eclipse AAA dose influence arrays accurately model deliverable IMRT dose.
    All ground-truth plans depend entirely on these dose arrays; errors in the dose engine would propagate into the training labels.
  • domain assumption Pseudo-random weight sampling categories in Table 1 cover the clinically relevant Pareto surface.
    The paper argues this qualitatively but provides no coverage metric, and the test set is drawn from the same sampling distribution.
  • ad hoc to paper The sigmoid approximation with m=1 preserves the relevant minima and gradient behavior.
    Justified by a two-voxel toy example in Figures 2-3, not by a general derivation; the paper itself notes higher m has steeper and less desirable gradients.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Incorporating human and learned domain knowledge into training deep neural networks: A differentiable dose volume histogram and adversarial inspired framework for generating Pareto optimal dose distributions in radiation therapy." pith.science (2026). https://pith.science/paper/RESH6ZGX

@misc{pith2026190805874,
  author       = {Pith},
  title        = {Pith review of: Incorporating human and learned domain knowledge into training deep neural networks: A differentiable dose volume histogram and adversarial inspired framework for generating Pareto optimal dose distributions in radiation therapy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RESH6ZGX}},
  note         = {Machine review of arXiv:1908.05874}
}
read the original abstract

We propose a novel domain specific loss, which is a differentiable loss function based on the dose volume histogram, and combine it with an adversarial loss for the training of deep neural networks to generate Pareto optimal dose distributions. The mean squared error (MSE) loss, dose volume histogram (DVH) loss, and adversarial (ADV) loss were used to train 4 instances of the neural network model: 1) MSE, 2) MSE+ADV, 3) MSE+DVH, and 4) MSE+DVH+ADV. 70 prostate patients were acquired, and the dose influence arrays were calculated for each patient. 1200 Pareto surface plans per patient were generated by pseudo-randomizing the tradeoff weights (84,000 plans total). We divided the data into 54 training, 6 validation, and 10 testing patients. Each model was trained for 100,000 iterations, with a batch size of 2. The prediction time of each model is 0.052 seconds. Quantitatively, the MSE+DVH+ADV model had the lowest prediction error of 0.038 (conformation), 0.026 (homogeneity), 0.298 (R50), 1.65% (D95), 2.14% (D98), 2.43% (D99). The MSE model had the worst prediction error of 0.134 (conformation), 0.041 (homogeneity), 0.520 (R50), 3.91% (D95), 4.33% (D98), 4.60% (D99). For both the mean dose PTV error and the max dose PTV, Body, Bladder and rectum error, the MSE+DVH+ADV outperformed all other models. All model's predictions have an average mean and max dose error less than 2.8% and 4.2%, respectively. Expert human domain specific knowledge can be the largest driver in the performance improvement, and adversarial learning can be used to further capture nuanced features. The real-time prediction capabilities allow for a physician to quickly navigate the tradeoff space, and produce a dose distribution as a tangible endpoint for the dosimetrist to use for planning. This can considerably reduce the treatment planning time, allowing for clinicians to focus their efforts on challenging cases.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages

  1. [4]

    step and shoot

    MSE+DVH+ADV. We found that the models that included the domain specific DVH loss outperformed the models without the DVH loss in most of the categories, particularly on the evaluations of conformity, heterogeneity, high dose spillage, and planning target volume (PTV) dose coverage. The MSE+DVH+ADV model performed the best in these categories, illustrating...

    arXiv 1988

  2. [56]

    MSE loss is a generalized, domain-agnostic loss function that can be applied to many problems in many domains

    Most of these methods utilize a simple loss function for training the neural network—the mean squared error (MSE) loss. MSE loss is a generalized, domain-agnostic loss function that can be applied to many problems in many domains. It’s large flexibility also means that it is incapable of driving its performance in a domain-specific manner. Mahmood and Bab...

    work page 2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.