{"id":"56df119d-4fd9-4355-b82c-c54a58fd777d","arxiv_id":"2508.07368","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A K-adaptability clustering heuristic for discrete-uncertainty robust optimization improves worst-case CTV Dmin in proton therapy by up to 4.52 Gy on average versus conventional robust planning, with a proof that the continuous-variable problem is NP-hard.","lead":"Proton therapy plans are usually optimized against the worst-case of many uncertainty scenarios, which can under-dose the tumor. This paper tests a K-adaptability approach that prepares several plans and picks the best one per scenario, reporting an average worst-case tumor-dose improvement of up to 4.52 Gy on five head-and-neck patients.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assignment MIPs (5)/(6) omit the g_k feasibility constraints of Problem (3); reported worst-case CTV Dmin may use infeasible plan-scenario pairs.","rationale":"The reader's identified weakest assumption is external validity: the 57-scenario training set may not capture daily anatomical changes. That is a legitimate limitation, and the paper itself flags it in Section 6. But the more fundamental, internal problem is that the heuristic's assignment MIPs solve a relaxation of Problem (3) by dropping the per-scenario feasibility constraints. This can directly inflate the training-set results that are the paper's main quantitative claim. The issue is concrete and testable from the manuscript's own formulation: Equation (3) has g_k(D x_i) <= 0 inside the inner min, while Equations (5) and (6) have no such conditions and v_ij is defined purely as f(D_j x_i). The paper does not state that infeasible assignments are excluded or penalized. Since the experimental section reports only overdose-type constraints and a max-Dmin objective, a plan that gives high CTV Dmin but violates a brainstem or spinal-cord limit for an assigned scenario would be selected by the MIP, making the reported worst-case Dmin clinically unachievable. This is not a matter of disagreement with clinical consensus; it is an internal consistency gap between the problem formulation and the heuristic. The proposed test would settle it directly. Until then, the strongest claim--that K-adaptability increases worst-case CTV Dmin by up to 4.52 Gy--is not established.","tokens_in":17003,"tokens_out":11071,"duration_ms":130456,"concrete_test":"For Case 4, extract the final X*_K and assignment z from Phase 2. For every assigned pair (i,j), evaluate all Table 2 constraints g_k(D_j x_i) <= 0. If any violation exists, resolve MIPs (5)-(6) with z_ij forbidden whenever g_k(D_j x_i) > 0 and recompute the worst-case CTV Dmin for K = 2,3,4,5,57. Compare to Table 3; if the improvement drops materially (e.g., >0.5 Gy at K=5), the feasibility omission is load-bearing and the headline claim is not supported.","verdict_should_be":"REJECT","load_bearing_attack":"The central dosimetric claim rests on the heuristic's computed worst-case CTV Dmin. Problem (3) explicitly restricts the inner minimization to plans i with g_k(D x_i) <= 0 for all k. However, the clustering MIPs (5) and (6) in Section 3.2 drop this restriction: they use v_ij = f(D_j x_i) and only enforce assignment cardinality and coverage constraints. A solution x_i generated for one cluster is not guaranteed to satisfy the OAR/CTV max-dose constraints of Table 2 for a scenario j outside its cluster, and the MIPs can assign such an infeasible pair if it improves the objective. Section 4.1's statement that adding only overdose constraints ensures feasibility of problem (2) only guarantees some feasible plan exists (e.g., x=0); it does not make every generated plan feasible for every scenario. Consequently, the reported improvements in Table 3, and especially the 4.52 Gy average at K=57, may be optimistic estimates from an infeasible relaxation rather than achievable K-adaptability plans.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a K-adaptability heuristic for robust proton therapy treatment planning under discrete setup and range uncertainty. The optimization model is a min-max-min problem in which K treatment plans are computed in advance and, after the uncertainty scenario is realized, the best feasible plan among the K is selected. The paper proves that the continuous-variable K-adaptability problem (Problem (3)) is NP-hard, then develops a two-phase heuristic: a solution generation phase that clusters scenarios on the basis of plan-scenario objective values using two MIPs (Problems (5) and (6)), and a re-distribution phase that re-assigns all generated solutions to scenarios. The method is evaluated on five head-and-neck patients with 57 uncertainty scenarios and compared with conventional robust optimization, two variants of the proposed heuristic, and K-medoids. The paper reports an average worst-case CTV Dmin improvement of 4.52 Gy at K=57, and a lower saturation K than the competing clustering methods.","tokens_in":17259,"tokens_out":9015,"duration_ms":88054,"significance":"If the numerical results are valid, the paper would make a useful contribution: it provides the first NP-hardness proof for the continuous-variable K-adaptability problem in this application, and it proposes a general heuristic for min-max-min robust optimization with discrete uncertainty sets that is more principled than generic clustering. The NP-hardness reduction in Section 2.3 is logically sound (up to typos), and the idea of clustering scenarios by plan-scenario objective values is sensible. However, the central dosimetric claim depends on a computationally questionable step: the assignment MIPs in Section 3.2 ignore the feasibility restrictions that appear in Problem (3). In addition, the abstract's claim of time-efficiency superiority is contradicted by Table 5, and the clinical evaluation is limited to the training uncertainty set, as the authors acknowledge in Section 6. No code or data are provided, so independent replication of the 4.52 Gy result is not currently possible.","major_comments":[{"comment":"Problems (5) and (6) drop the feasibility restriction that is explicit in Problem (3). The inner minimization of (3) is over plans i satisfying g_k(D x_i) ≤ 0 for all k, but the MIPs define v_ij = f(D_j x_i) and enforce only assignment and cardinality constraints. A plan generated for one cluster is guaranteed feasible only for scenarios in that cluster (via Problem (2)); it need not satisfy the max/mean-dose constraints of Table 2 for scenarios outside the cluster. The MIPs can therefore assign an infeasible plan-scenario pair if doing so improves the objective, and the re-distribution phase (Section 3.4) performs no feasibility check. Section 4.1's observation that including only overdose constraints makes Problem (2) feasible is insufficient: it ensures some feasible plan exists (e.g., x=0), not that every generated plan is feasible for every scenario. Consequently, the worst-case CTV","section":"Section 3.2, Eq. (5)-(6)"},{"comment":"The abstract and the conclusion state that the developed heuristic is superior in 'time-efficiency' to the competing methods. Table 5 shows average runtimes of 45,256 s for AOSG and 46,592 s for K-medoids, versus 65,314 s for the proposed K-adaptability heuristic; the proposed method is faster only than the LSP variant (83,766 s). Please correct the claim in the abstract and Section 7 to reflect that the proposed heuristic is not the most time-efficient among the tested methods.","section":"Abstract; Section 5.2, Table 5; Section 7"}],"minor_comments":[{"comment":"There are indexing typos in Eq. (4): the summation index in the objective should be over j, not i, and the dose-matrix entries should carry the scenario index l (e.g., D^l_jj). The proof text also writes x_i^i ambiguously; it should be x^i_i or x(i)_i. The reduction is otherwise understandable.","section":"Eq. (4) and proof of Theorem 1"},{"comment":"The text says 'the brainstem should receive at most 57 Gy of dose, D_min^Brainstem ≤ 57 Gy.' This should be D_max (or an upper-bound constraint), not D_min.","section":"Section 2.2"},{"comment":"The sentence 'Figure 1 contains the information on the voxel number...' should refer to Table 1, which is where the patient ROI and beamlet numbers are presented.","section":"Section 4.1"},{"comment":"The while loops in Algorithm B.1 are written as 'while K ≥ 1' with K incremented by 1, which does not terminate. The loops should be 'while K ≤ |U|' in both Phase 1 and Phase 2.","section":"Algorithm B.1"},{"comment":"The caption says 'Worst-case dose distribution of case four,' but the figure is a plot of worst-case CTV Dmin versus K. The caption should be corrected.","section":"Figure 3 caption"},{"comment":"The phrase 'increased the worst-case CTV Dmin dose up to 4.52 Gy on average' is ambiguous. It should specify that this is the average gain at K=57 (and that at K=2-5 the gains are smaller).","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The infeasible-assignment issue in Problems (5)-(6) is the main technical blocker. If the authors add the feasibility restriction and the numerical conclusions change materially, the central claim of the paper would need substantial revision. I would also note that the paper provides no code or data, so the 4.52 Gy result cannot be independently verified. The NP-hardness proof is logically sound despite typos, but it is not the main contribution driving the clinical claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's genuinely new piece is the NP-hardness proof for K-adaptability with continuous decision variables, extending earlier binary-variable results. The reduction from hitting set is valid (modulo minor typos in Eq. (4)). The performance-based clustering heuristic for discrete uncertainty sets is also a reasonable contribution, and the application to proton therapy is well motivated. I want to give credit for those.\n\nThe main problem is in the assignment step. Problems (5) and (6) compute v_ij = f(D_j x_i) but never check whether x_i satisfies the overdose constraints for scenario j. Problem (3) explicitly restricts the inner minimization to plans that are feasible for the realized scenario. As a result, the MIPs can pair a scenario with a plan that violates OAR limits for that scenario and report that as the worst-case objective. The authors say in Section 4.1 that including only overdose constraints ensures feasibility of Problem (2), but that only guarantees some plan (e.g., x=0) is feasible for every scenario, not that every generated cluster plan is feasible off-cluster. This is not a minor technicality: the reported improvements in Table 3, including the 4.52 Gy average at K=57, may be computed from infeasible plan-scenario pairs. A post-hoc feasibility filter or adding the g_k constraints to the assignment would settle it.\n\nOther soft spots: the abstract claims superiority in time-efficiency, but Table 5 shows AOSG and K-medoids are faster than the proposed heuristic. The 4.52 Gy headline is for K=57, which the authors themselves say is not clinically realistic; at K=5 the gain is 2.57 Gy. No code or data are provided, the cohort is five patients, and the evaluation is on the training uncertainty set, with daily-image validation correctly deferred to future work. The saturation-K estimate is a heuristic output without an optimality certificate.\n\nWho is this for? Researchers in K-adaptability or robust optimization in radiotherapy will find the heuristic and complexity result worth knowing. But the quantitative claims should be treated with caution until the feasibility issue is addressed. I'd send it to peer review—the ideas merit scrutiny—but I'd expect referees to require a corrected assignment or re-computed results. If the numbers survive the fix, it's a solid contribution.","headline":"The continuous-variable NP-hardness result and the clustering heuristic are worthwhile, but the reported dosimetric gains may be inflated because the assignment MIPs ignore feasibility constraints.","tokens_in":17741,"tokens_out":5359,"would_cite":false,"duration_ms":50198,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C47","90C11","92C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"Precomputing K treatment plans instead of one robust plan lifts worst-case tumor dose, with an average gain of 4.52 Gy across five head-and-neck proton therapy cases and a proof that the underlying optimization problem is NP-hard.","keywords":["K-adaptability","robust optimization","proton therapy","treatment planning","min-max-min","scenario clustering","setup uncertainty","range uncertainty"],"falsifier":"For the same five head-and-neck patients, acquire daily CT or cone-beam CT images across the treatment course, compute the delivered CTV Dmin for the conventionally robust plan and for the K-adaptability plan set (selecting the best plan per day), and compare the actual worst-case tumor coverage; if K-adaptability does not improve or worsens the coverage relative to the single robust plan on this real anatomical data, the paper's central dosimetric claim fails.","tokens_in":16928,"feed_emoji":"🎯","tokens_out":3483,"duration_ms":32873,"temperature":0.7,"pith_summary":"This paper tries to establish that the standard min-max robust approach to proton therapy planning is unnecessarily conservative, and that a K-adaptability approach—where several plans are computed in advance and the best one is selected after the uncertainty is revealed—can improve worst-case tumor coverage. The authors propose a heuristic that clusters uncertainty scenarios by plan performance, generate robust plans for each cluster, then re-assign plans to scenarios. On five head-and-neck patients, the heuristic improved worst-case CTV Dmin by on average 4.52 Gy compared to the conventional single robust plan, and it needed only about 15 plans on average to reach the maximum worst-case performance across a 57-scenario uncertainty set. The paper also proves that the continuous-variable K-adaptability problem (3) is NP-hard, so exact solution is not a realistic route for this application.","feed_headline":"Swapping one proton plan for K plans lifts tumor dose 4.5 Gy","feed_subtitle":"A K-adaptability heuristic picks the best of precomputed plans per day, beating conventional robust planning on five head-and-neck cases.","key_machinery":"The key machinery is the scenario-to-plan assignment based on plan-scenario performance. Two MIP problems are used: the first minimizes the worst-case objective by selecting which K solutions to use and how to assign each uncertainty scenario to a selected solution; the second, given the optimal worst-case value, refines the assignment to also optimize the scenario-average objective, achieving Pareto robust optimality. The iterative solution-generation phase starts with one nominal plan per scenario, clusters scenarios by the MIPs, generates robust plans for each cluster, and adds them to a global pool; the re-distribution phase then re-selects the best K plans from the full pool. The cluste","core_discovery":"The central claim is that for proton therapy robust planning with a discrete uncertainty set, the min-max-min K-adaptability problem—though phrased with continuous decision variables and a very specific clinical structure—remains NP-hard (Theorem 1, proven by reduction from the hitting set problem). On the algorithmic side, the paper's scenario-clustering heuristic, which assigns scenarios to solutions based on the objective value each plan achieves on each scenario, produces a set of K plans whose worst-case CTV Dmin is substantially better than that of a conventional single robust plan. Across five head-and-neck patients, the worst-case CTV Dmin improved by 4.52 Gy on average at K = 57 (wi","pith_inferences":["A natural next test the authors leave implicit is evaluating K-adaptability on daily repeated CT or cone-beam CT images, since the current results are computed on the same simulated uncertainty set used for training; if real anatomical changes differ from the 57 setup/range scenarios, the reported 4.52 Gy gain may not translate directly.","The performance-based clustering principle could also be applied to photon therapy or online adaptive radiotherapy, where plan selection per fraction is already a clinical workflow.","The saturation K estimate of about 15 plans suggests that, with automated QA, a manageable number of precomputed plans might cover the practical uncertainty space, which could change the logistics of adaptive proton therapy.","Because the NP-hardness proof is a reduction from hitting set, any exact method would need to exploit further clinical structure—such as the specific dose-influence matrix sparsity—rather than relying on generic continuous optimization."],"forward_implications":["If the K-adaptability results hold in clinical practice, multi-plan adaptive treatment could become a viable alternative to single robust plans, with better worst-case tumor coverage without changing the physics of proton delivery.","The measured saturation K value of about 15 on average indicates that achieving the maximum worst-case performance may require far fewer than the full 57 uncertainty scenarios, which is useful for planning quality-assurance workloads.","The order in which solutions are generated matters: the descending-K strategy outperforms the ascending strategy, suggesting that including less conservative plans from higher K in the pool helps clustering at low K.","The heuristic is presented as a general-purpose method for discrete uncertainty robust optimization, so it could be transferred to other domains such as logistics or knapsack problems that use min-max-min robustness.","The NP-hardness result implies that exact algorithms will not scale to clinically sized problems, making efficient heuristics the practical path for K-adaptability in treatment planning."],"supporting_citations":[{"why":"Introduces the K-adaptability concept of computing K solutions in advance and selecting the best after uncertainty realization.","marker":"Bertsimas and Caramanis (2010)"},{"why":"Provides complexity results for min-max-min robustness with finite uncertainty sets, which the paper extends to the continuous-variable radiotherapy setting.","marker":"Buchheim and Kurtz (2018)"},{"why":"Derives complexity results and exact mixed-integer programming approaches for K-adaptability in two-stage robust binary programming.","marker":"Hanasusanto et al. (2015)"},{"why":"Establishes the minimax optimization formulation for handling range and setup uncertainties in proton therapy, the baseline this work compares against.","marker":"Fredriksson et al. (2011)"},{"why":"Supplies the clinical rationale and magnitude of range uncertainty (3%) used to construct the uncertainty scenario set.","marker":"Lomax (2019)"},{"why":"Provides the MOQUI Monte Carlo dose calculation engine used for dose-influence matrix computation and uncertainty simulation.","marker":"Lee et al. (2022)"},{"why":"Provides the Nymph optimizer used to solve the robust optimization problems for each scenario cluster.","marker":"Gorissen (2022)"},{"why":"Supplies the QUANTEC normal tissue constraints used as planning constraints in the treatment planning optimization.","marker":"Bentzen et al. (2010)"}],"fun_headline_variants":["K plans beat one: +4.5 Gy in proton therapy","Proton therapy: K-adaptability lifts dose 4.5 Gy","Multiple proton plans improve worst-case dose","K-adaptability boosts proton robust planning","Cluster heuristic ups tumor dose in proton therapy"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The 57-scenario uncertainty set (3 mm setup shifts in 19 directions plus ±3% range errors) is treated as a valid surrogate for the anatomical changes that actually occur during a course of proton therapy, and the evaluation is performed on that same training set rather than on daily temporal images.","fun_headline_variants_meta":{"raw":{"variants":["K plans beat one: +4.5 Gy in proton therapy","Proton therapy: K-adaptability lifts dose 4.5 Gy","Multiple proton plans improve worst-case dose","K-adaptability boosts proton robust planning","Cluster heuristic ups tumor dose in proton therapy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000141,"raw_usage":{"total_tokens":994,"prompt_tokens":730,"completion_tokens":264,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":187}},"tokens_in":474,"tokens_out":264,"duration_ms":3068,"temperature":1.0,"reasoning_tokens":187,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:11:09.661031+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the same five head-and-neck patients, acquire daily CT or cone-beam CT images across the treatment course, compute the delivered CTV Dmin for the conventionally robust plan and for the K-adaptability plan set (selecting the best plan per day), and compare the actual worst-case tumor coverage; if K-adaptability does not improve or worsens the coverage relative to the single robust plan on this real anatomical data, the paper's central dosimetric claim fails.","supporting_citations":[{"cited_title":"and Caramanis, C","cited_arxiv_id":null,"evidence_quote":"Introduces the K-adaptability concept of computing K solutions in advance and selecting the best after uncertainty realization."},{"cited_title":"and Kurtz, J","cited_arxiv_id":null,"evidence_quote":"Provides complexity results for min-max-min robustness with finite uncertainty sets, which the paper extends to the continuous-variable radiotherapy setting."},{"cited_title":"A., Kuhn, D., and Wiesemann, W","cited_arxiv_id":null,"evidence_quote":"Derives complexity results and exact mixed-integer programming approaches for K-adaptability in two-stage robust binary programming."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the minimax optimization formulation for handling range and setup uncertainties in proton therapy, the baseline this work compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the clinical rationale and magnitude of range uncertainty (3%) used to construct the uncertainty scenario set."},{"cited_title":"M., Bobić, M., Winey, B., Schuemann, J., and Paganetti, H","cited_arxiv_id":null,"evidence_quote":"Provides the MOQUI Monte Carlo dose calculation engine used for dose-influence matrix computation and uncertainty simulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Nymph optimizer used to solve the robust optimization problems for each scenario cluster."},{"cited_title":"M., Constine, L","cited_arxiv_id":null,"evidence_quote":"Supplies the QUANTEC normal tissue constraints used as planning constraints in the treatment planning optimization."}],"review_version":1}