{"id":"bf0286ec-3ac7-416a-9749-87b0cad728f5","arxiv_id":"2506.15803","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An unsupervised U-Net trained on spot-count representations preselects energy layers for proton arc therapy, reducing energy switch time and improving nominal dosimetry versus SPArc_ps.","lead":"This paper introduces an unsupervised deep learning model that quickly preselects proton beam energy layers for arc therapy, reducing energy switch time by 37% while improving nominal dose quality on 35 nasopharyngeal cancer cases. The model's plans are less robust to setup and range uncertainties, a limitation the authors acknowledge.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The plan-quality claim rests on the untested assumption that spot-count maximization predicts final dose quality; the paper's own robustness results suggest this proxy is not reliable.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the spot-count representation is used as a proxy for dose coverage, and the network is trained to maximize these counts without ever seeing the final dose distribution. This is indeed the least secure link in the argument. The paper's own robustness analysis provides direct evidence that the proxy does not capture all clinically relevant dose behavior: SPArc_dl plans show wider CTV DVH bands and larger RVH AUC, which the authors attribute to the absence of robustness objectives. That admission does not invalidate the nominal dosimetric results, but it does mean the claimed 'excellent dosimetric performance' is only shown in nominal conditions and is not tied to the training objective in any demonstrated way. The efficiency gain is also partly circular because ELS time is an explicit loss term. However, the paper does present a plausible method, a novel input representation, and a reasonable evaluation pipeline; the deficiencies are about missing evidence and unfair baseline calibration, not about an internal mathematical contradiction. The conditional verdict is therefore appropriate: the paper should be accepted only if the authors release the energy switch time table, the code and data, and provide a direct validation that spot-count correlates with final plan quality or an ablation comparing against a greedy spot-count baseline. My concern reinforces the conditional verdict rather than moving it to acceptance or rejection.","tokens_in":12400,"tokens_out":5585,"duration_ms":66731,"concrete_test":"For each of the 35 cases, record the per-case training objective components (mean target spot-count, mean OAR spot-count, ELS time) and the final DVH metrics (D98, HI, CI, brainstem mean) after the downstream weight optimization, then compute correlations across cases. If the target spot-count component does not positively correlate with D98 or CI, or does not negatively correlate with HI, the proxy is not driving plan quality. Additionally, run a greedy baseline that selects the EL with the maximum M_CTV at each gantry angle and passes it through the same matRad weight optimization; if this baseline matches or exceeds SPArc_dl's nominal dosimetric metrics, the deep learning component is not necessary for the reported dosimetric advantage, and the claim should be substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central dosimetric claims depend on the validity of the spot-count proxy (Sections 2.2--2.3, Eq. 4): the network is trained to maximize the number of proton spots intersecting CTV (and CTV minus brainstem) and to penalize ELS time, but it never evaluates the actual dose distribution produced after the subsequent spot-weight optimization. For the claimed improvements in CI, HI, and brainstem dose to be attributable to SPArc_dl's learned pre-selection, the proxy must be predictive of final plan quality. The paper provides no evidence of this correlation, and its own robustness evaluation (Section 3.3, Fig. 3c) shows SPArc_dl plans have wider CTV DVH bands and larger RVH AUC, which the authors concede reflects reduced robustness. This is direct evidence that the proxy does not capture all clinically relevant dose behavior. Moreover, because the model outputs a single EL per gantry angle via softmax-argmax, the pre-selection drastically constrains the downstream optimization space; the nominal quality gains could be an artifact of the optimizer on this specific reduced set rather than a consequence of the proxy's fidelity. The 37.2% ELS time reduction is largely built into the loss function; the load-bearing part of the claim is that plan quality is maintained or improved, and that part has not been independently established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SPArc_dl, an unsupervised deep learning model for fast energy-layer (EL) pre-selection in proton arc therapy for nasopharyngeal carcinoma. The method introduces a spot-count representation (matrices of proton spots intersecting CTV and CTV-minus-brainstem, structured by gantry angle and energy layer), feeds this representation to a U-Net, and trains with a tri-objective loss: maximizing CTV spot counts, minimizing brainstem spot counts, and penalizing energy-layer-switch (ELS) time. On 35 NPC cases, the method is compared with SPArc_ps, reporting improved conformity index, homogeneity index, D98, and brainstem mean/max dose, a 37.2% reduction in ELS time, and near-instant inference, while also acknowledging inferior plan robustness measured by DVH bands and RVH AUC.","tokens_in":12688,"tokens_out":4188,"duration_ms":52880,"significance":"If the central claims hold, the work would be a useful step toward fast, delivery-efficient proton arc planning: the spot-count representation is an original way to make EL pre-selection amenable to deep learning, the unsupervised formulation avoids supervised ground-truth generation, and the sub-second inference time is practically attractive. The authors also report cross-validation, quantitative metrics, and an explicit robustness limitation, which are strengths. However, the significance is currently limited because the dosimetric superiority claim depends on an unvalidated proxy (spot counts as a surrogate for dose), on a potentially weak baseline, and on a one-energy-layer-per-angle constraint that changes the optimization problem relative to SPArc_ps. The robustness results the authors themselves report suggest the proxy does not fully capture clinically relevant dose behavior.","major_comments":[{"comment":"The network is trained to maximize spot counts intersecting CTV and CTV-minus-brainstem, but the final dosimetric claims (CI, HI, D98, brainstem dose) are measured on plans after spot-weight optimization. No evidence is provided that higher spot counts at the selected energy layers correlate with better final dose distributions. This is load-bearing: if the proxy is not predictive, the reported dosimetric improvements cannot be attributed to the learned pre-selection. The paper's own robustness evaluation (Fig. 3c) shows larger CTV RVH AUC for SPArc_dl, which is consistent with the proxy missing important dose behavior. Please add a direct validation of the spot-count proxy against final dose metrics, or substantially temper the dosimetric claims.","section":"Sections 2.2–2.3, Eq. (4)"},{"comment":"SPArc_dl selects exactly one energy layer per gantry angle via the highest softmax probability. SPArc_ps, by contrast, appears to use multiple energy layers per angle (e.g., the diagonal bands in Fig. 3a). This is not a like-for-like comparison: the one-layer-per-angle constraint drastically reduces the downstream optimization space, and the reported improvements in CI, HI, and brainstem dose could be artifacts of this constraint rather than of the learned pre-selection quality. The authors should report the number of energy layers per gantry angle for both methods and analyze the sensitivity of plan quality to relaxing the one-layer constraint.","section":"Section 2.2, inference stage"},{"comment":"The baseline comparability is not established. The introduction states that SPArc_ps 'may need more objective functions on organs at risk' for head and neck cancer, which suggests the comparator may not be well tuned for NPC. The evaluation section does not state whether the same downstream spot-weight optimizer, objective function, and optimization parameters were used for both SPArc_dl and SPArc_ps. Without this information, the dosimetric comparison may reflect a weak baseline or an inconsistent optimization setup rather than the benefit of SPArc_dl. Please specify the common downstream optimization protocol and, ideally, add a baseline that is optimized with the same OAR objectives.","section":"Sections 1 and 2.4"},{"comment":"The 37.2% ELS time reduction is a direct consequence of the Ltime term in Eq. (4) and of the one-layer-per-angle selection; it is an expected outcome of the loss design, not an independent 'unintentional' finding. Moreover, the statement that unchanged ELS is more time-efficient than descended ELS is definitional, since Table 2 assigns zero time to unchanged ELS. The paper should present this result as a designed efficiency feature and focus the novelty claims on plan quality, which is the part that needs independent support.","section":"Section 3.2, Table 2"}],"minor_comments":[{"comment":"The loss uses binarized maximum spot counts as target labels for a cross-entropy term. Calling the method 'unsupervised' is somewhat misleading; 'self-supervised' or 'pseudo-label-based' would be more precise, since the network is trained toward data-derived targets rather than with no labels at all.","section":"Section 2.3, Eq. (1)"},{"comment":"The p-values are reported without specifying the statistical test, whether the comparisons are paired, or whether any multiple-comparison correction was applied. Please add a methods sentence describing the statistical analysis.","section":"Section 3.1, Table 1"},{"comment":"The CT voxel size is reported as '1.36 × 1.36 × 3 cm^3,' which is likely a typo for mm^3. Also, the units for the ELS time table in Fig. 1(b) are not defined in the caption or text.","section":"Section 2.4"},{"comment":"The statement that no suitable repository exists is unusual; if data cannot be shared, at least provide the code or a detailed algorithmic description to support reproducibility of the spot-count representation and loss function.","section":"Data availability statement"}],"recommendation":"major_revision","confidential_remarks":"The paper is a feasibility study with a promising idea, but the central comparison is not yet convincing. The main fixable issue is validation of the spot-count proxy and a fair, well-specified baseline comparison. If the authors can add a proxy-to-dose correlation analysis or an ablation, and document the downstream optimization setup, the work could become publishable. I would also suggest the authors double-check the novelty claim of being 'the first deep learning model' for EL pre-selection, given the ELO-SPAT-related machine learning work cited in the introduction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. The spot-count representation is genuinely new: encoding how many proton spots intersect CTV and CTV-minus-brainstem into matrices over gantry angle and energy, then feeding that to a U-Net, is a clean way around hand-crafted features. The unsupervised framing is also a real step beyond the supervised Ma et al. approach and the heuristic SPArc_ps. But the central claim—better CI, HI, and brainstem dose with 37.2% less ELS time—is plausible rather than proven, because the network is trained on a spot-count proxy and the proxy is never validated against actual dose.\n\nCredit where it is due. The paper identifies a real bottleneck: EL pre-selection in proton arc therapy can take minutes to hours, and this gets inference to 0.05 seconds. The tri-objective loss is simple and interpretable. The robustness evaluation is honest: wider CTV DVH bands and larger RVH AUC for SPArc_dl are shown and acknowledged as a limitation. That self-reporting is worth respecting.\n\nSoft spots, in proportion. First and largest: the proxy-to-dose assumption. The loss maximizes spot counts, not dose, and there is no analysis showing that a higher spot count predicts a better final plan. The gains in Table 1 come after weight optimization, so they could be the optimizer doing well on a constrained EL set rather than evidence the proxy is faithful. The robustness results actually cut against the proxy: nominal quality improves while uncertainty behavior gets worse. Second, the baseline. The paper itself says SPArc_ps \"may need more objective functions on OARs\" for head and neck. If you suspect your comparator is suboptimal for your disease site, the comparison is soft, and the p-values do not fix that. Third, the ELS time reduction is partly by construction—the time penalty is in the loss—so 37.2% is roughly what you would expect; the load-bearing question is whether plan quality holds, and that is where the proxy issue lands. Minor: no code or data released, hyperparameters tuned without sensitivity analysis, and the \"unchanged ELS is more time-efficient\" finding is an observation that follows from the ELS time table itself, which defines unchanged transitions as zero time. The stress-test concern holds up: this is real and it is the load-bearing weakness, not a manufactured one.\n\nWho should read this: anyone working on fast PAT planning or on surrogate objectives for ML-driven treatment planning. It is a solid first step, not a final answer. I would send it to review—the approach is novel, the limitations are visible enough for a referee to push on, and the field would benefit from the comparison. I would ask the authors to validate the spot-count proxy against dose, rerun against a stronger baseline like ELO-SPAT, and release code or data before acceptance.","headline":"The spot-count representation is genuinely new, but plan-quality claims depend on a dose proxy the paper never validates and a baseline it itself calls suboptimal.","tokens_in":13225,"tokens_out":6011,"would_cite":true,"duration_ms":57226,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SPArc_dl, an unsupervised deep-learning model, pre-selects proton arc energy layers in under a second and cuts energy-switch time by 37.2% while improving plan quality on 35 nasopharyngeal carcinoma cases.","keywords":["proton arc therapy","energy layer selection","unsupervised deep learning","spot-count representation","nasopharyngeal carcinoma","plan optimization","delivery efficiency","robustness"],"falsifier":"Run SPArc_dl and SPArc_ps on new nasopharyngeal cases with 2-degree angular sampling and full dose optimization; the central claim fails if SPArc_dl does not reduce energy-switch time while keeping CTV D98, CI, HI, and brainstem dose at least as good as SPArc_ps. A more direct test of the proxy: find one case where increasing the spot count $M_{CTV}$ at an angle lowers the actually optimized D98—that would show that the count representation is not measuring the dose effect the model is trained to maximize.","tokens_in":12219,"feed_emoji":"⚡","tokens_out":11170,"duration_ms":118268,"temperature":0.7,"pith_summary":"This paper claims that energy-layer pre-selection for proton arc therapy can be reduced to an unsupervised image-like segmentation problem, with no labeled plans and no dose calculation during the selection step. The model, SPArc_dl, takes spot-count matrices that record how many proton pencil beams hit the target (and how many hit the target while avoiding the brainstem) at each gantry angle and energy layer, and it outputs one energy layer per angle. On 35 nasopharyngeal-carcinoma cases, that recipe delivered improved dose conformity and homogeneity, lower brainstem dose, and 37.2% less energy-switch time than the established SPArc_ps method, with inference in about 0.05 seconds. The paper's own robustness analysis shows the advantage comes with a cost: CTV dose-volume histograms under setup and range uncertainties are broader for SPArc_dl, so the method improves nominal quality and speed more than it improves worst-case stability.","feed_headline":"Unsupervised model cuts proton arc energy-switch time by 37%","feed_subtitle":"On 35 nasopharyngeal cases it also improved conformity, homogeneity, and brainstem sparing.","key_machinery":"The load-bearing object is the spot-count representation: for each structure $S$ a matrix $M \\in \\mathbb{R}^{G \\times E}$ where $G$ is the number of gantry angles, $E$ is the number of energy layers, and entry $(g,e)$ counts the proton spots from angle $g$ at layer $e$ that intersect $S$. Element-wise subtraction gives $M_{CTV-BRS}=M_{CTV}-M_{BRS}$, the count of spots covering the target while missing the brainstem; these two matrices, concatenated with a per-angle WET vector, become a 2-channel image-like tensor for a U-Net. The network outputs a probability $P$ over energy layers for every angle, and the training loss combines two cross-entropy terms that push the selected layer toward the binarized maximum of $M_{CTV}$ and of $M_{CTV-BRS}$ with a third term that extracts the switch time $T(P_{g-1},P_g)$ from an energy-layer time table. This machinery makes clinical priorities directly differentiable, so the model can be trained end-to-end without any dose engine or ground-truth plan.","core_discovery":"The central claim is that the spot-count representation—counting proton spots that intersect a structure, arranged in a matrix over sorted gantry angles and energy layers—captures enough of the dose-coverage information to choose energy layers without supervised ground truth. A U-Net maps the concatenated matrices $M_{CTV}$ and $M_{CTV-BRS}$, together with water-equivalent-thickness values per angle, to a probability tensor over energy layers; a softmax per gantry angle selects the most probable layer. The tri-objective loss maximizes spot counts on CTV, maximizes spot counts on CTV while avoiding the brainstem, and penalizes energy-layer switch time read from a machine-specific time table. Across five-fold cross-validation on 35 patients, the resulting plans beat SPArc_ps on conformity index (0.70 vs 0.60), homogeneity index (0.26 vs 0.97), brainstem mean dose (0.54 vs 0.79 Gy) and energy-switch time (51.0 vs 81.1 s, a 37.2% reduction), all with $p<0.01$. The paper also reports that the selected sequences use many unchanged energy layers between consecutive gantry angles, and that unchanged transitions are more time-efficient than descended ones, contradicting the usual preference for descending energy sequences. The paper acknowledges that robustness is weaker: SPArc_dl shows broader CTV DVH bands and larger CTV RVH AUC than SPArc_ps under uncertainty.","pith_inferences":["A testable extension: replace the raw spot-count matrices with a differentiable approximate dose objective; the difference in final plan quality would quantify how much information the spot-count proxy discards.","Adding a robustness term to the same unsupervised loss—for example, computing spot counts under several setup shifts and range uncertainties and penalizing their variance—would directly target the paper's reported robustness limitation.","The unchanged-energy-layer efficiency result depends on the shape of the machine's ELS time table; on systems where descending energy switches are cheap, the optimal balance of unchanged, descended, and ascended transitions could shift.","The 2 × 72 × 103 input is architecture-agnostic, so the same spot-count representation could be used with sequence models or multi-energy-layer-per-angle selection to gain dosimetric flexibility beyond the single-layer-per-angle choice."],"forward_implications":["Energy-layer selection becomes a sub-second forward pass: SPArc_dl infers a full arc's energy sequence in about 0.05 seconds on a CPU, compared with 13.5 seconds for SPArc_ps, enabling interactive replanning or rapid what-if exploration.","Treatment delivery time drops 37.2% on average (from 81.1 to 51.0 seconds of energy-switch time per arc) because the model chooses many unchanged energy layers between consecutive gantry angles, and unchanged transitions cost zero switch time.","Nominal plan quality improves across all reported metrics on the 35 nasopharyngeal cases: higher conformity index, lower homogeneity index, higher D98, lower D2, and lower brainstem mean and max dose, all with $p<0.01$.","The unsupervised formulation leaves the objective open-ended: additional OARs, beam-delivery constraints, or robustness scenarios can be added as extra loss terms without generating labels or running a full optimizer to create training targets.","The robustness results are a direct corollary of the paper's claim: SPArc_dl's CTV DVH bands and RVH AUC are worse than SPArc_ps, so the speed and nominal-quality gains do not extend to uncertainty robustness."],"supporting_citations":[{"why":"Introduces SPArc, the spot-scanning proton arc therapy modality whose energy-layer selection problem this paper addresses.","marker":"[6]"},{"why":"SPArc_ps is the baseline planning method: the paper compares plan quality, energy-switch time, and robustness directly against it.","marker":"[17]"},{"why":"SPArc_seq is an earlier energy-sequence optimizer whose preference for descended energy switches is the assumption the present results challenge.","marker":"[12]"},{"why":"matRad supplies the initial full set of energy layers and performs dose calculation for evaluation, making the experimental pipeline possible.","marker":"[26]"},{"why":"U-Net is the convolutional architecture adopted for the SPArc_dl network's encoder-decoder with skip connections.","marker":"[18]"},{"why":"Prior supervised machine-learning prediction of energy layers provides the contrast that motivates the unsupervised approach.","marker":"[15]"},{"why":"Frames proton arc optimization as two steps with energy-layer pre-selection and identifies energy-switch time as a major delivery bottleneck.","marker":"[28]"},{"why":"ELO-SPAT is a multi-objective energy-sequencing formulation whose long optimization times the fast inference of SPArc_dl is meant to overcome.","marker":"[29]"},{"why":"Supplies the homogeneity-index formula used to evaluate CTV dose uniformity.","marker":"[27]"},{"why":"Supplies the conformity-index formula used to evaluate how well the high-dose region matches the target shape.","marker":"[24]"}],"fun_headline_variants":["Unsupervised model cuts proton arc energy-switch time by 37%","Spot-count deep learning trims proton arc switch time 37%","Proton arc planning: unsupervised model cuts energy switches 37%","Model pre-picks energy layers to cut proton arc switch time 37%","Deep learning speeds proton arc therapy: 37% less energy switching"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the number of proton spots intersecting a structure is a faithful proxy for radiation dose coverage, so the layer with the most intersecting spots will also be the layer that best covers the target after weight optimization; if that link breaks, the model's dosimetric gains could disappear.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised model cuts proton arc energy-switch time by 37%","Spot-count deep learning trims proton arc switch time 37%","Proton arc planning: unsupervised model cuts energy switches 37%","Model pre-picks energy layers to cut proton arc switch time 37%","Deep learning speeds proton arc therapy: 37% less energy switching"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000448,"raw_usage":{"total_tokens":2403,"prompt_tokens":1230,"completion_tokens":1173,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":846,"completion_tokens_details":{"reasoning_tokens":1078}},"tokens_in":846,"tokens_out":1173,"duration_ms":11046,"temperature":1.0,"reasoning_tokens":1078,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:50:41.909526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SPArc_dl and SPArc_ps on new nasopharyngeal cases with 2-degree angular sampling and full dose optimization; the central claim fails if SPArc_dl does not reduce energy-switch time while keeping CTV D98, CI, HI, and brainstem dose at least as good as SPArc_ps. A more direct test of the proxy: find one case where increasing the spot count $M_{CTV}$ at an angle lowers the actually optimized D98—that would show that the count representation is not measuring the dose effect the model is trained to maximize.","supporting_citations":[{"cited_title":"International Journal of Radiation Oncology* Biology* Physics 96(5), 1107–1116 (2016)","cited_arxiv_id":null,"evidence_quote":"Introduces SPArc, the spot-scanning proton arc therapy modality whose energy-layer selection problem this paper addresses."},{"cited_title":"Physics in Medicine & Biology 69(1), 015004 (2023)","cited_arxiv_id":null,"evidence_quote":"SPArc_ps is the baseline planning method: the paper compares plan quality, energy-switch time, and robustness directly against it."},{"cited_title":"Acta Oncologica 59(10), 1178–1185 (2020)","cited_arxiv_id":null,"evidence_quote":"SPArc_seq is an earlier energy-sequence optimizer whose preference for descended energy switches is the assumption the present results challenge."},{"cited_title":"Medical physics 44(6), 2556–2568 (2017)","cited_arxiv_id":null,"evidence_quote":"matRad supplies the initial full set of energy layers and performs dose calculation for evaluation, making the experimental pipeline possible."},{"cited_title":"In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18","cited_arxiv_id":null,"evidence_quote":"U-Net is the convolutional architecture adopted for the SPArc_dl network's encoder-decoder with skip connections."},{"cited_title":"Medical Physics 51(7), 4970–4981 (2024)","cited_arxiv_id":null,"evidence_quote":"Prior supervised machine-learning prediction of energy layers provides the contrast that motivates the unsupervised approach."},{"cited_title":"Physics in Medicine and Biology (2025)","cited_arxiv_id":null,"evidence_quote":"Frames proton arc optimization as two steps with energy-layer pre-selection and identifies energy-switch time as a major delivery bottleneck."},{"cited_title":"Computers in biology and medicine 148, Preprint 17 105609 (2022)","cited_arxiv_id":null,"evidence_quote":"ELO-SPAT is a multi-objective energy-sequencing formulation whose long optimization times the fast inference of SPArc_dl is meant to overcome."},{"cited_title":"i: dosimetric results","cited_arxiv_id":null,"evidence_quote":"Supplies the homogeneity-index formula used to evaluate CTV dose uniformity."},{"cited_title":"International Journal of Radiation Oncology* Biology* Physics 37(3), 731–736 (1997)","cited_arxiv_id":null,"evidence_quote":"Supplies the conformity-index formula used to evaluate how well the high-dose region matches the target shape."}],"review_version":1}