{"id":"43b41a98-8efc-4544-908e-471393551170","arxiv_id":"2506.19880","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A two-stage deep learning pipeline that first mimics clinical VMAT plans and then uses a frozen learned dose predictor as physics guidance modestly improves dose-metric accuracy on 13 prostate patients, though several organ-at-risk reductions are within noise.","lead":"This paper trains neural networks to plan prostate radiotherapy by first copying existing clinical plans, then adding a second training step that checks whether the generated machine settings would deliver the right radiation dose. The extra dose check improves target coverage metrics and high-dose agreement in a small test set, but the results lack statistical tests and the authors release no code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported plan-quality gains may be circular: evaluation doses appear to come from the same frozen RT Dose Predictor used in the stage-2 loss, so gamma/DVH improvements could reflect surrogate fitting rather than true Pinnacle dose.","rationale":"I agree with the reader that the unvalidated dose surrogate is the weakest link, but I sharpen it: the concern is not only that the surrogate may be biased, but that the evaluation itself likely uses the same surrogate, making the improvement potentially circular. The paper's strongest evidence (gamma pass rate and DVH differences) would then not measure true dosimetric quality. This is testable by recomputing with Pinnacle; no retraining is needed. If the authors already used Pinnacle for evaluation, they should state it explicitly. The rest of the paper, such as the architecture comparison and inference speed, is secondary. I recommend keeping a conditional verdict, with this additional required condition: independent dose calculation for evaluation and, ideally, a validation of the RT Dose Predictor on Pinnacle-recalculated plans.","tokens_in":8446,"tokens_out":5058,"duration_ms":53749,"concrete_test":"Recompute Tables 1 and 2 for the 13 test patients using the actual Pinnacle dose calculation on the predicted MLC and MU sequences, instead of the frozen RT Dose Predictor. If the 3D U-Net's high-dose gamma pass rate and PTV D95 difference move substantially away from 90.5% and 0.42 Gy, or if the stage-1 versus stage-2 ordering reverses, the reported benefit is partly an artifact of evaluating with the training surrogate. As a secondary check, run the RT Dose Predictor on the 13 Pinnacle ground-truth plans and report its 2%/2 mm gamma pass rate against Pinnacle; a large miss would quantify the surrogate bias directly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that stage-2 dose supervision makes predicted plans dosimetrically closer to clinical plans. The only dose information used in stage-2 training is D_pred produced by the frozen RT Dose Predictor (Section 2.3). The paper never states what dose engine generated the dose distributions used in the evaluation tables: Table 1 DVH differences and Table 2 gamma pass rates. Section 3.1 says only that evaluation used 'CUDA-accelerated preprocessing ... and model inference ... in under one second per patient,' which is consistent with running the same learned RT Dose Predictor on the predicted MLC/MU, not with a Pinnacle recalculation. If that is what was done, the evaluation is circular: the planner is trained to minimize MSE against D_true in the surrogate's dose space, and then judged in that same space. A biased surrogate trained on Monaco plans (Section 2.2) and never revalidated on the Pinnacle-recalculated cohort would allow stage 2 to reduce surrogate error while increasing true Pinnacle dose error. The 99.6% gamma figure cited for the dose predictor is from its original Monaco-based validation, not from this cohort, so it does not rule out systematic bias. This makes the reported PTV D95 error 0.42 +/- 1.83 Gy and 90.5 +/- 7.3% high-dose gamma pass unreliable as evidence for the central claim unless the evaluation doses were independently computed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage, physics-guided deep learning pipeline for VMAT prostate radiotherapy planning. In the first stage, a 3D U-Net or UNETR is trained with direct supervision on MLC aperture masks and monitor units from clinical plans. In the second stage, a frozen, differentiable neural dose predictor (the RT Dose Predictor, after Witte and Sonke 2024) maps the predicted plan to a 3D dose distribution, and an additional MSE loss between predicted and ground-truth dose is backpropagated through the cascade. The authors evaluate on 133 prostate cancer patients (104 training, 16 validation, 13 test) treated with a uniform 2-arc VMAT protocol delivering 62 Gy, with all ground-truth plans recalculated in Pinnacle with standardized parameters. They report that the second stage improves PTV/CTV DVH metrics and gamma pass rates for both architectures, with the 3D U-Net achieving PTV D95% error of 0.42 ± 1.83 Gy and a high-dose (90% threshold) gamma pass rate of 90.5 ± 7.3%.","tokens_in":8730,"tokens_out":4119,"duration_ms":41195,"significance":"If the reported dosimetric gains are confirmed with an independent dose calculation engine, the paper would make a useful contribution to fast adaptive radiotherapy planning: it demonstrates an end-to-end differentiable training scheme that combines direct parameter supervision with a dose-domain objective, and it evaluates two competitive architectures on a clinically standardized dataset. The authors also receive credit for using a uniform treatment protocol, clinical review of the recalculated plans, and a reasonably sized test cohort. However, the central claim rests on the evaluation dose distributions being computed by an independent engine; the manuscript does not state what engine produced the DVH and gamma results for the predicted plans. Because the stage-2 loss is minimized in the space of the frozen learned dose predictor, evaluation in that same space would be circular and would not establish true agreement with Pinnacle dose. The significance of the work is therefore conditional on resolving this issue.","major_comments":[{"comment":"The manuscript never specifies which dose calculation engine produced the dose distributions used to compute the DVH differences in Table 1 and the gamma pass rates in Table 2. Section 3.1 states only that evaluation used CUDA-accelerated preprocessing and model inference on an RTX A6000, which is consistent with running the frozen RT Dose Predictor on the predicted MLC/MU, not with a Pinnacle recalculation. If the evaluation doses come from the same learned RT Dose Predictor used in the stage-2 loss, the comparison is circular: the planner is trained to minimize error in the surrogate's dose space and then judged in that same space. The authors must state explicitly which dose engine was used for the evaluation of predicted plans and, if it was not Pinnacle or another independent engine, must recompute the reported metrics with an independent dose calculation.","section":"Section 3.1, Tables 1 and 2"},{"comment":"The RT Dose Predictor is the load-bearing component of the physics-guided stage, yet its accuracy is only asserted via a 99.6% gamma pass rate (2%/2 mm) from prior validation on 350 Monaco-generated plans. This paper evaluates on 133 Pinnacle-recalculated plans, a different treatment planning system, and does not revalidate the dose predictor on this cohort or report whether any of the 350 training patients overlap with the current 133 patients. A biased surrogate trained on Monaco plans could systematically disagree with Pinnacle dose, and the stage-2 loss would then steer the planner toward the surrogate's errors rather than toward true dose. The authors should report the dose predictor's gamma pass rate and DVH accuracy on the current Pinnacle cohort, including the 13 test patients, and clarify that no overlap exists with the dose predictor's training set.","section":"Section 2.2, RT Dose Predictor"},{"comment":"The paper states that the second training stage 'significantly improved' key dosimetric metrics, but no statistical significance testing is reported. Many OAR metrics in Table 1 have standard deviations that exceed the mean differences (e.g., rectum Dmean for 3D U-Net: -0.82 ± 4.45 Gy; left femoral Dmean for UNETR: -0.72 ± 6.34 Gy), so the claimed OAR sparing is not established without paired tests or confidence intervals. The authors should provide paired statistical tests (e.g., Wilcoxon signed-rank or paired t-tests) across the 13 test patients for the key PTV/CTV and OAR metrics, with appropriate multiple-comparison awareness.","section":"Section 3.1, paragraph 2"},{"comment":"There is a numeric inconsistency: the text reports that the 3D U-Net achieved a mean absolute difference of D95% = 0.59 ± 2.23 Gy, D98% = 0.70 ± 2.14 Gy, and V95% = -0.42 ± 1.12%, but Table 1's PTV rows for the 3D U-Net second stage show D95% = 0.42 ± 1.83, D98% = -0.71 ± 2.12, and V95% = -0.22 ± 1.87. The values in the text match the CTV rows of Table 1, not the PTV rows. The authors should clarify which ROI the text refers to and correct the mismatch, since the abstract quotes PTV values and the reader should be able to trace them.","section":"Section 3.1, paragraph 2 and Table 1"}],"minor_comments":[{"comment":"The abstract says the approach 'consistently produces treatment plans that closely match clinical ground truths,' but Table 1 shows UNETR's second-stage PTV D95% is -1.70 ± 2.21 Gy. Consider adding 'on average' or reporting the range, because 'consistently' overstates the per-patient behavior suggested by the standard deviations.","section":"Abstract"},{"comment":"There are two unit errors: '7 MeV beam energy' should be '7 MV' (megavoltage, not megaelectronvolt), and 'an isotropic resolution of 3.5 mm3' should be '3.5 mm' (voxel spacing, not volume).","section":"Section 2.1"},{"comment":"The description of the five input channels is unclear: the first channel is the CT, and 'the remaining four channels contain the rotation and projection at each control point for the CT, PTV, CTV, and OARs.' Please clarify whether the four channels are rotated versions of the CT and three masks for each control point, and how the control point index is provided to the network.","section":"Section 2.2, first paragraph"},{"comment":"The gamma evaluation uses 3%/3 mm criteria but does not state the software or implementation used to compute the gamma index, nor whether the comparison is performed in the 3D dose grid after resampling. Adding this detail would improve reproducibility.","section":"Section 3.2"},{"comment":"The Discussion does not mention the reliance on a learned dose predictor as a limitation. Given that the stage-2 objective uses a surrogate dose engine, a sentence acknowledging that surrogate errors could propagate and that independent dose recalculation is needed for clinical validation would strengthen the paper.","section":"Discussion"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the engineering is solid, but the evaluation-engine ambiguity is the kind of issue that, if unresolved, would undermine the paper's central claim. The authors should be given an opportunity to specify the evaluation dose engine and, if necessary, rerun the evaluation with an independent dose calculation. I also note that the RT Dose Predictor comes from prior work by co-authors; the manuscript should report its accuracy on the current dataset rather than citing the original validation, to avoid any appearance of evaluating a surrogate in its own training space."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: this paper has a genuinely useful idea—train a direct MLC/MU predictor with a second-stage loss from a frozen differentiable dose predictor—and the 3D U-Net numbers look impressive. But the manuscript never states how the predicted doses in the evaluation tables were obtained. Section 3.1 mentions CUDA-accelerated preprocessing and model inference in under one second per patient, which strongly suggests they ran the same frozen RT Dose Predictor on the predicted plan parameters rather than recalculating with Pinnacle. If that's the case, the reported gamma pass rates and DVH differences are measured in the surrogate's dose space, and the stage-2 loss is being optimized against the same function used for evaluation. That is a circularity problem, and it makes the central claim—that physics-guided training produces dosimetrically closer plans—unreliable as presented. The 99.6% gamma pass rate cited for the dose predictor comes from its prior Monaco-based validation, not from this Pinnacle-recalculated cohort, so it doesn't resolve the bias question.\n\nWhat is actually new and good here: cascading a frozen, differentiable dose predictor as a second supervision signal for MLC/MU prediction is a sensible and non-obvious approach. The authors use two architectures, report full DVH and gamma tables, and the 3D U-Net improvements are large enough to be clinically interesting even with small n. The motivation—direct MLC/MU supervision is ambiguous because many apertures give equivalent dose—is well articulated. The prior dose engine (Witte and Sonke) is a serious piece of work, so the building blocks are solid.\n\nSoft spots, in order of severity. First is the evaluation engine ambiguity: the authors must state explicitly whether predicted doses came from the RT Dose Predictor or from Pinnacle, and if the former, they need to redo the evaluation with independent dose calculation. This is a load-bearing omission. Second, there are no statistical tests; with 13 test patients and standard deviations as large as the reported differences, some of the claims (e.g., OAR reductions) may not be robust. Third, the abstract says the approach 'consistently' matches clinical plans, but the UNETR variant still has a PTV D95 error of -1.70 ± 2.21 Gy, which is not a great match; the abstract reads as if the 3D U-Net results apply to both architectures. Fourth, no code or model weights are provided, and the hyperparameters λ1 and λ2 are chosen without sensitivity analysis—minor but worth noting.\n\nOverall: the core mechanism is plausible and worthy of serious refereeing, but the current manuscript needs major revision to prove the central claim. If the authors can show Pinnacle-recalculated doses for the predicted plans and pass the surrogate-fit test, this could be a solid contribution to the ART planning literature. I would send it to peer review, but flag the dose-evaluation question as a requirement before acceptance.","headline":"A worthwhile two-stage training idea with strong 3D U-Net numbers, but the paper never says what dose engine produced the reported evaluation doses—if it is the same frozen learned dose predictor used in training, the headline results are not yet evidence of true dosimetric improvement.","tokens_in":9300,"tokens_out":3368,"would_cite":false,"duration_ms":36782,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dose-supervised second training stage brings deep-learning VMAT plans to a 90.5 percent high-dose gamma pass rate.","keywords":["radiotherapy treatment planning","volumetric modulated arc therapy","physics-guided deep learning","MLC aperture prediction","3D dose prediction","prostate cancer","adaptive radiotherapy","gamma pass rate"],"falsifier":"Compute the frozen dose predictor's 2%/2 mm gamma pass rate on the 13 test patients by comparing its predicted dose with the recalculated clinical ground-truth dose; if the pass rate is far below the 99.6 percent cited from its original validation, the reported stage-2 gains could be an artifact of fitting the surrogate. A complementary experiment is to train stage 2 with a classical differentiable dose engine in place of the learned predictor and check whether the PTV $D_{95\\%}$ error and high-dose gamma pass-rate gains persist.","tokens_in":8204,"feed_emoji":"🎯","tokens_out":10472,"duration_ms":97113,"temperature":0.7,"pith_summary":"Radiotherapy planning currently requires solving a high-dimensional, non-convex optimization problem, and adaptive radiotherapy makes the time pressure worse. This paper tries to show that a two-stage deep-learning pipeline can produce clinically usable VMAT plans in under a second. Stage one trains a network to predict treatment plan parameters, the multileaf collimator apertures and monitor units, directly from the CT and structure masks; stage two adds dose-domain supervision by sending the predicted plan through a frozen, differentiable dose predictor and penalizing the resulting 3D dose against the clinical dose. The authors claim, based on 133 prostate patients treated with a uniform 62 Gy two-arc protocol, that this physics-guided second stage substantially improves dosimetric agreement, with the best model reaching a PTV $D_{95\\%}$ error of $0.42 \\pm 1.83$ Gy and a high-dose gamma pass rate of $90.5 \\pm 7.3$ percent while keeping organ-at-risk doses lower. If true, the method gives adaptive radiotherapy an automatic replanning step scored on dose rather than on one arbitrary parameter solution.","feed_headline":"Dose-guided AI planning hits 90.5 percent gamma pass","feed_subtitle":"Adding a frozen dose predictor to MLC and MU supervision cuts PTV D95 error to 0.42 Gy and lowers OAR dose.","key_machinery":"The load-bearing object is the RT Dose Predictor: a fully differentiable convolutional gated recurrent unit, pretrained on 350 patients with plans generated by one commercial planning system, that maps a CT and treatment plan to a 3D dose distribution. During stage two it is frozen and cascaded after the Deep RT Planner, so the $L_2$ distance between predicted and ground-truth dose backpropagates through the dose predictor into the MLC and MU decoders. This is what makes the training physics-guided: the planner is optimized for the dose its parameters imply rather than for matching a single MLC/MU ground truth, and the gradient flows through a model of dose transport instead of through the plan parameters alone.","core_discovery":"Direct supervision on MLC apertures and monitor units is hampered by non-uniqueness: many different parameter sets deliver clinically equivalent dose. The central discovery is that adding a second-stage dose-supervision term, computed by pushing the predicted plan through a pretrained, frozen RT Dose Predictor and comparing its 3D dose to the ground-truth dose with an L2 loss, systematically brings predictions closer to clinical plans. For the 3D U-Net, the PTV $D_{95\\%}$ error falls from $1.75 \\pm 3.55$ Gy to $0.42 \\pm 1.83$ Gy, and the gamma pass rate above 90 percent of maximum dose rises from $56.46 \\pm 32.68$ percent to $90.50 \\pm 7.28$ percent; OAR mean and maximum doses shift lower or stay comparable. The same dose-guided stage also improves the UNETR variant, though less strongly, which supports the interpretation that the mechanism, not a single architecture, drives the gain. The paper concludes that training in the clinically relevant dose domain mitigates multi-arc redundancy and makes deep-learning plans feasible for adaptive radiotherapy.","pith_inferences":["The gains depend on the frozen dose predictor transferring to the target planning system; an unstated check would be to re-evaluate that predictor's 2%/2 mm gamma pass rate directly on the recalculated doses of the 13 test patients and to exclude patient overlap between its 350 training cases and this cohort.","A clean way to locate the source of the physics guidance is to replace the learned dose predictor with a classical differentiable dose calculator, for example a collapsed-cone or Monte Carlo-based engine, in the same two-stage loop; if the dosimetric gains vanish, they come from the learned surrogate rather than from dose-level supervision itself.","Because the non-uniqueness of plan parameters is generic, dose-level supervision should transfer across planning systems better than parameter-level supervision; a natural extension is to train the planner on plans recalculated with several different systems and test on a held-out system."],"forward_implications":["A single forward pass of the two-stage pipeline produces a full 144-control-point VMAT plan in under one second per patient, contrasted with minutes for conventional GPU-based planning systems.","Dose-domain supervision improves PTV and CTV coverage metrics such as $D_{95\\%}$, $D_{98\\%}$, and $V_{95\\%}$ and high-dose gamma pass rates for both a convolutional and a transformer-based planner, so the benefit does not depend on the specific network.","After the physics-guided stage, rectum and femoral-head doses are lower than or clinically comparable to the clinical ground truth, suggesting the method could reduce organ-at-risk toxicity in replanning.","The largest gamma improvement occurs in the highest-dose region, where the 3D U-Net's pass rate roughly doubles, which is the region most relevant for target coverage."],"supporting_citations":[{"why":"Supplies the frozen differentiable RT Dose Predictor (convolutional GRU) that computes the 3D dose used as stage-2 supervision.","marker":"[33]"},{"why":"Defines the direct MLC-aperture and MU-prediction strategy that stage 1 follows as its baseline.","marker":"[28]"},{"why":"Provides the 3D U-Net architecture used as one Deep RT Planner variant.","marker":"[34]"},{"why":"Provides the UNETR architecture used as the second Deep RT Planner variant, showing the stage-2 benefit is architecture-independent.","marker":"[13]"},{"why":"Defines the gamma pass-rate criterion used both to validate the dose predictor and to score predicted dose distributions.","marker":"[31]"},{"why":"Supplies the planning system used to generate the 350 training plans for the RT Dose Predictor.","marker":"[10]"},{"why":"Supplies the planning system used to recalculate all cohort plans, giving the ground-truth dose distributions and DVH metrics.","marker":"[24]"}],"fun_headline_variants":["Two-stage physics-guided AI sharpens prostate RT plans","Dose supervision in training boosts gamma pass to 90.5%","Frozen dose predictor reduces PTV D95 error to 0.42 Gy","AI planning gains from dose-domain supervision for VMAT","Physics guidance in deep learning improves VMAT plan quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the frozen dose calculator, trained on plans from other patients and a different planning system, is accurate enough on the current cohort's recalculated doses, so the second-stage loss steers the planner toward true dose instead of toward the calculator's own errors.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage physics-guided AI sharpens prostate RT plans","Dose supervision in training boosts gamma pass to 90.5%","Frozen dose predictor reduces PTV D95 error to 0.42 Gy","AI planning gains from dose-domain supervision for VMAT","Physics guidance in deep learning improves VMAT plan quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000635,"raw_usage":{"total_tokens":2985,"prompt_tokens":1060,"completion_tokens":1925,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":1839}},"tokens_in":676,"tokens_out":1925,"duration_ms":14278,"temperature":1.0,"reasoning_tokens":1839,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:37:43.193406+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the frozen dose predictor's 2%/2 mm gamma pass rate on the 13 test patients by comparing its predicted dose with the recalculated clinical ground-truth dose; if the pass rate is far below the 99.6 percent cited from its original validation, the reported stage-2 gains could be an artifact of fitting the surrogate. A complementary experiment is to train stage 2 with a classical differentiable dose engine in place of the learned predictor and check whether the PTV $D_{95\\%}$ error and high-dose gamma pass-rate gains persist.","supporting_citations":[{"cited_title":"Physics in Medicine & Biology (2023)","cited_arxiv_id":null,"evidence_quote":"Defines the direct MLC-aperture and MU-prediction strategy that stage 1 follows as its baseline."},{"cited_title":"Medical Physics34(5), 1647–1654 (2007)","cited_arxiv_id":null,"evidence_quote":"Defines the gamma pass-rate criterion used both to validate the dose predictor and to score predicted dose distributions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the planning system used to generate the 350 training plans for the RT Dose Predictor."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the planning system used to recalculate all cohort plans, giving the ground-truth dose distributions and DVH metrics."}],"review_version":2}