{"id":"d3005831-955a-4368-a5c1-f69460b3d00e","arxiv_id":"2411.18767","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A shared multi-task network that predicts contours and dose maps together improves dose prediction relative to sequential contouring-plus-planning models, while a CT-only baseline sometimes does better.","lead":"This paper tests whether one neural network can simultaneously outline organs and predict radiation dose maps, instead of running two separate deep learning models in sequence. On prostate and head-and-neck datasets the integrated model beats the sequential pipeline, but the reported headline improvement does not match the tables.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Capacity confound undermines MTL attribution: Multi-Automated has extra encoder/decoder capacity vs single-decoder baselines, and abstract's 19.82%/16.33% gains use baseline denominator, not Seq-Automated; need parameter-matched ablation before accepting MTL effect.","rationale":"I read the paper in good faith: the empirical comparison between Multi-Automated and Seq-Automated uses reasonable held-out test sets, and the average DVH-MAE values in Tables 2 and 3 are consistent with the qualitative claim that the integrated model improves over the sequential automated pipeline on these two datasets. However, the causal attribution to multi-task learning is not established because the comparison is not capacity-matched. The strongest_claim invokes MTL as the cause, and Sec. 2.2.3 shows Multi-Automated has an additional task decoder and cross-task attention relative to the single-decoder baselines described in Sec. 2.2.1-2.2.2. Without a single-task model with comparable parameters, the improvement could be an architecture/capacity effect. This is highlighted by the OpenKBP result where a simple CT-only baseline already beats Multi-Automated, indicating the comparison to sequential models does not uniquely favor an MTL explanation. The abstract percentages also do not match the correct sequential-comparison denominators. These issues are fixable with an ablation and corrected reporting, so they warrant a conditional acceptance rather than rejection; the reader's conditional verdict is therefore unchanged.","tokens_in":11518,"tokens_out":5739,"duration_ms":48689,"concrete_test":"Train a single-task CT-only dose predictor with the same encoder and dose decoder as Multi-Automated (i.e., remove the contour decoder and contour-specific attention/loss, keeping parameter count approximately equal by adding capacity to the dose decoder if needed). Evaluate on the same prostate and OpenKBP test splits; if its mean DVH-MAE is statistically indistinguishable from Multi-Automated's 3.528 Gy and 10.109 Gy, the reported MTL gains are capacity effects rather than multi-task learning. If instead the single-task matched-capacity model is clearly worse, the MTL attribution is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that multi-task learning—not just a larger network—causes the dose-prediction gain over the sequential pipeline. The paper does not control for model capacity. Section 2.2.3 describes Multi-Automated as a shared encoder plus two task-specific decoders with cross-task attention from [25], while the Seq-Automated dose predictor is a single-decoder U-Net ('same architecture as the baseline dose prediction task,' Sec. 2.2.2). Multi-Automated therefore contains an additional contour decoder and cross-task attention, so its usable capacity for the dose task is not matched to the baseline. The observed average DVH-MAE improvements (4.301 to 3.528 Gy for prostate; 11.650 to 10.109 Gy for OpenKBP) are consistent with a capacity effect: a stronger CT-only dose predictor may simply learn better features from more parameters. This is not a purely semantic point: the OpenKBP baseline, a CT-only single-task U-Net, already achieves 9.436 Gy, beating Multi-Automated's 10.109 Gy, so the integrated model is not universally superior to a strong single-task predictor. Separately, the abstract's 19.82% and 16.33% figures are computed with the baseline model in the denominator, not with Seq-Automated; relative to Seq-Automated the gains are 17.97% and 13.23%. The qualitative direction survives, but the headline numbers overstate the sequential-pipeline comparison. A parameter-matched MTL ablation is therefore the load-bearing missing experiment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-task learning (MTL) framework that simultaneously performs automated contouring and voxel-based dose prediction for radiotherapy, using a shared encoder, two task-specific decoders, and cross-task attention taken from the authors' prior work [25]. The method is evaluated on an in-house prostate dataset (110 patients) and the public OpenKBP head-and-neck dataset (318 patients after exclusions), and is compared against a CT-only baseline, a sequential model using human contours (Seq-Human), and a sequential model using automated contours (Seq-Automated). The authors report average DVH-MAE improvements of Multi-Automated over Seq-Automated on both datasets, together with contouring Dice and Hausdorff distance results, and conclude that MTL can integrate contouring and dose prediction without compromising contouring accuracy.","tokens_in":11841,"tokens_out":5792,"duration_ms":50639,"significance":"If the central attribution to multi-task learning were established, this would be a practically relevant contribution: an integrated model that takes only CT as input and outputs both contours and dose could simplify the radiotherapy planning workflow and reduce sensitivity to contour quality. The use of the public OpenKBP dataset, the inclusion of both human-contour and automated-contour sequential baselines, and the reporting of two anatomically distinct sites are strengths. However, the manuscript as written does not establish that multi-task learning, rather than increased model capacity, drives the observed improvements; the abstract's headline percentages are not supported by the tables; and the statistical reporting is incomplete. The paper is a reasonable empirical study but requires substantial revision before the central claim can be accepted.","major_comments":[{"comment":"The headline percentages do not follow from the reported averages. From Tables 2 and 3, Multi-Automated improves over Seq-Automated by (4.301−3.528)/4.301 = 17.97% on prostate and (11.650−10.109)/11.650 = 13.23% on OpenKBP, not the 19.82% and 16.33% stated in the abstract. The relative differences listed in the tables are computed against the Baseline column, not against Seq-Automated, so the abstract appears to use a different denominator. Because these percentages are the paper's primary quantitative claim, they must be corrected and the denominator explicitly defined.","section":"Abstract; Section 3.1, Tables 2 and 3"},{"comment":"The comparison does not control for model capacity. Multi-Automated uses a shared encoder plus two task-specific decoders and cross-task attention, whereas Seq-Automated and the baseline are single-decoder U-Nets. No parameter-matched single-task dose predictor or MTL ablation is reported, so the observed DVH-MAE gains over Seq-Automated could be due to the larger model rather than to multi-task learning. The OpenKBP result is especially telling: the CT-only single-task Baseline achieves 9.436 Gy, beating Multi-Automated's 10.109 Gy on average, so the integrated model is not universally superior to a strong single-task predictor. A matched-capacity single-task dose model, or an MTL variant without cross-task attention, is required before the central MTL attribution can be accepted.","section":"Section 2.2.2-2.2.3; Tables 2 and 3"},{"comment":"Statistical support is underreported. The tables only mark significance with daggers, without reporting p-values, standard deviations, confidence intervals, number of random seeds, or training epochs. With test sizes of 10 (prostate) and 78 (OpenKBP) and many DVH metrics, the absence of uncertainty measures makes it impossible to assess the stability of the reported average improvements. In addition, the caption of Table 3 states 'p < 0.5' instead of 'p < 0.05', which appears to be a typo but should be corrected.","section":"Section 2.3.3; Tables 2-5"},{"comment":"The claim that contouring accuracy is 'maintained or sometimes even improved' is only partially supported. On the prostate dataset, Multi-Automated's average Hausdorff distance is worse than the baseline (12.049 vs 7.549), and the right femur Hausdorff distance deteriorates significantly (36.441 vs 9.669), while Dice is essentially unchanged. On OpenKBP, Dice improves significantly for several ROIs, but the spinal cord Hausdorff distance worsens. The discussion should report these trade-offs explicitly rather than emphasizing the improvements only.","section":"Section 3.2, Table 4"}],"minor_comments":[{"comment":"The dataset arithmetic is inconsistent: the text says 328 patients were collected and 22 omitted, but the stated split of 200/40/78 sums to 318, whereas 328−22=306. Please reconcile these numbers and state how the omitted cases are distributed across the splits.","section":"Section 2.1"},{"comment":"There are several typographical errors: 'able able' in the abstract, 'respectivley' in the Figure 1 caption, 'radiotherpay' in Table 1, 'coeffcient' in Tables 4 and 5, and the 'p < 0.5' captions in Tables 3 and 5. These should be corrected.","section":"Throughout"},{"comment":"Training details are incomplete: the number of epochs, learning-rate schedule, dynamic weight average hyperparameters, and whether results come from a single run or multiple seeds are not reported. A code or data availability statement would also improve reproducibility.","section":"Section 2.2.4"},{"comment":"The paper says the framework is novel 'to the best of our knowledge,' but the architecture is from the authors' prior paper [25] and another multi-task dose-prediction/contouring study [24] is cited. Please state explicitly which components are new in this work, e.g., the systematic sequential-model comparison and dual-site validation, to avoid overclaiming novelty.","section":"Section 1"},{"comment":"The description of D0.1cc is awkward: it is written as 'Maximum dose received by the smallest volume of 0.1 cc.' Standard D0.1cc is the dose received by 0.1 cc of the structure, not a maximum-dose definition; please rephrase for clarity.","section":"Section 2.3.1, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper leans heavily on the authors' own MICCAI 2023 paper [25] for the core architecture, and the novelty framing should be softened to reflect that this is an empirical evaluation of a previously proposed MTL model for integrated contouring and dose prediction rather than a new architectural contribution. The abstract's numerical claims are inconsistent with the tables, and the capacity confound is serious enough that I cannot recommend acceptance without a matched-capacity ablation. The topic is within the journal's scope, and the empirical comparison to sequential pipelines is useful, but the central claim needs stronger support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe useful thing here is a head-to-head comparison: an integrated multi-task network that segments and predicts dose in one pass versus the standard sequential pipeline (automated contouring, then dose prediction) on prostate and OpenKBP head-and-neck data. The integrated model beats the sequential automated pipeline on average DVH-MAE at both sites, and the authors are honest enough to report that a CT-only single-task baseline beats them on OpenKBP. That is the kind of result worth discussing.\n\nWhat is new: prior work from this group introduced the cross-task attention architecture [25], and Jiao et al. did MTL for dose prediction. This paper's contribution is the explicit comparison to sequential human and automated pipelines, plus the observation that sequential automated contouring can degrade dose prediction relative to a CT-only baseline. The contouring results are also presented cleanly, with dice and Hausdorff per structure.\n\nThe soft spots are real. The MTL model has a shared encoder, two task-specific decoders, and cross-task attention, while the sequential and baseline models are single-decoder U-Nets. No parameter-matched ablation is run, so the improvement could be a capacity effect, not multi-task learning. On OpenKBP the simple CT-only baseline (9.436 Gy) actually beats the MTL model (10.109 Gy), which alone undercuts the claim that MTL is the best way to predict dose.\n\nThe abstract's headline numbers are also misleading: 19.82% and 16.33% are improvements relative to the baseline model, not against the Seq-Automated pipeline. Against Seq-Automated the gains are 17.97% and 13.23%—still positive but smaller. Reporting lacks error bars, seeds, and full p-values, and no code or data are released. The prostate D99 is worse with MTL, so the average gain is carried by other metrics.\n\nNone of this sinks the paper's value as an empirical comparison. The fix is straightforward: add a parameter-matched single-task dose predictor (or a capacity-matched MTL) and report the Seq-Automated-relative numbers correctly. I would send this to peer review, but with the expectation of major revisions. The question is important and the comparison is useful, but the attribution to multi-task learning is not yet established.\n\nIt is a maybe for our reading group—useful for a discussion on confounds in MTL papers, but not a must-read.\n\nBest,\n[Your name]","headline":"A useful sequential-vs-MTL comparison in automated radiotherapy planning, but capacity is unmatched and the headline gains are overstated.","tokens_in":12406,"tokens_out":3367,"would_cite":false,"duration_ms":28949,"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":"This paper argues that automated contouring and dose prediction in radiotherapy should be trained jointly, reporting DVH error reductions of 19.82% (prostate) and 16.33% (head and neck) versus a sequential automated pipeline.","keywords":["multi-task learning","automated contouring","dose prediction","radiotherapy treatment planning","dose volume histogram","deep learning","prostate cancer","head and neck cancer"],"falsifier":"Train a single-task dose-prediction model with the same encoder capacity, pre-trained backbone, and attention machinery as Multi-Automated but without the contouring output branch; if its mean absolute error of the dose-volume histogram (DVH) metrics matches the reported 3.528 Gy and 10.109 Gy values, the improvement is due to model capacity rather than to joint contouring and dose training.","tokens_in":11276,"feed_emoji":"🎯","tokens_out":6351,"duration_ms":54932,"temperature":0.7,"pith_summary":"Automated radiotherapy planning today runs two deep-learning jobs in series: first a model contours the organs, then a second model uses those contours to predict the radiation dose distribution. This paper argues that the two jobs should be trained together in one multi-task network that takes only a CT scan and outputs contours and a dose map at the same time. On an in-house prostate dataset and the public OpenKBP head-and-neck dataset, the integrated model reduced the mean absolute error of dose-volume histogram metrics by 19.82% and 16.33%, respectively, compared with the sequential automated pipeline, while matching or improving contouring accuracy. If the finding holds, integrated training could make fully automated treatment planning faster and less sensitive to imperfect input contours.","feed_headline":"Contours and dose from one network beat a two-step pipeline","feed_subtitle":"Multi-task model cuts DVH errors by 19.8% (prostate) and 16.3% (head and neck) over sequential AI planning.","key_machinery":"The load-bearing mechanism is a cross-task attention network: one shared encoder extracts common CT features, while separate output branches predict contours and dose, and specialized attention modules within the encoder let each task weigh the shared features differently. Training balances a mean-absolute-error loss for dose with a combo loss (dice plus cross-entropy) for contours, and a dynamic weight average adjusts the relative loss weights during training. This mechanism is what lets the model output contours and dose simultaneously from CT alone, and it is the component the paper credits for the reported DVH-MAE improvements.","core_discovery":"The paper's central claim is that multi-task learning, rather than a sequential contour-then-plan workflow, is the right way to structure automated radiotherapy planning. In the proposed Multi-Automated model, a single shared encoder processes the CT image and feeds two task-specific output branches, one for organ contours and one for voxel-wise dose, linked by cross-task attention. The authors report that this integrated model outperforms both sequential baselines: it achieves a DVH-MAE of 3.528 Gy versus 4.301 Gy for sequential automated planning on prostate and 10.109 Gy versus 11.650 Gy on head and neck, and it improves average contouring Dice coefficients from 0.818 to 0.824 on prostate and from 0.674 to 0.716 on head and neck. The authors interpret this as evidence that sharing features between contouring and dose prediction lets the model learn anatomical relationships that a dose-only or contour-then-dose model cannot exploit.","pith_inferences":["The paper does not compare Multi-Automated against a single-task dose predictor with the same parameter count or the same pre-trained encoder; until such a comparison is run, part of the reported gain could be attributed to model capacity rather than to multi-task learning.","Because the OpenKBP CT-only baseline already beat both contour-based sequential models, the benefit of the integrated model there may partly come from its ability to ignore noisy contour labels; corrupting input contours in a controlled experiment would test this.","The method is evaluated on 2D slices and a limited set of organs; translating it to full 3D plans with more anatomical structures is the natural next test of clinical value."],"forward_implications":["A fully automated planning pipeline no longer needs a separate contouring model or pre-computed contours at inference time: CT goes in, contours and dose come out.","Dose prediction becomes less sensitive to contour quality, because the model is not forced to trust imperfect input contours.","The reported gains on prostate and head-and-neck datasets suggest the benefit is not site-specific and may transfer to other treatment sites.","Integrated training can slightly improve contouring as well as dose prediction, so automation need not trade one task against the other."],"supporting_citations":[{"why":"Supplies the cross-task attention network architecture used for the Multi-Automated model, including the shared encoder and two task-specific output branches.","marker":"[25]"},{"why":"Provides the OpenKBP head-and-neck dataset, the train/validation/test split, and the DVH evaluation framework used for the head-and-neck experiments.","marker":"[7]"},{"why":"Shows that dose models can be trained with limited contour information, motivating the integrated approach that avoids relying on input contour quality.","marker":"[15]"},{"why":"Earlier multi-task learning work for dose prediction and contouring with tumor masks; the paper positions its sequential comparison against this line of work.","marker":"[24]"},{"why":"Foundational definition of multi-task learning and its expected benefit of shared training signals.","marker":"[16]"},{"why":"Provides the combo loss used to train the contouring branch of the multi-task model.","marker":"[26]"},{"why":"Supplies the dynamic weight average technique used to balance the contouring and dose losses during training.","marker":"[27]"}],"fun_headline_variants":["Multi-task AI cuts DVH errors by 19.8% and 16.3%","Shared network for contouring and dose outperforms sequential models","Integrated multi-task model improves both contours and dose accuracy","One-pass AI planning: contouring and dose prediction together","Multi-task learning boosts radiotherapy planning accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central comparison assumes that the dose-prediction gain is caused by multi-task learning itself, but the multi-task model is a larger, more complex network than the single-task baselines, and no capacity-matched single-task model is tested.","fun_headline_variants_meta":{"raw":{"variants":["Multi-task AI cuts DVH errors by 19.8% and 16.3%","Shared network for contouring and dose outperforms sequential models","Integrated multi-task model improves both contours and dose accuracy","One-pass AI planning: contouring and dose prediction together","Multi-task learning boosts radiotherapy planning accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001228,"raw_usage":{"total_tokens":5091,"prompt_tokens":1037,"completion_tokens":4054,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":3970}},"tokens_in":653,"tokens_out":4054,"duration_ms":26237,"temperature":1.0,"reasoning_tokens":3970,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:53:49.698851+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a single-task dose-prediction model with the same encoder capacity, pre-trained backbone, and attention machinery as Multi-Automated but without the contouring output branch; if its mean absolute error of the dose-volume histogram (DVH) metrics matches the reported 3.528 Gy and 10.109 Gy values, the improvement is due to model capacity rather than to joint contouring and dose training.","supporting_citations":[{"cited_title":"Cross-task attention network: Improving multi-task learning for medical imaging applications","cited_arxiv_id":null,"evidence_quote":"Supplies the cross-task attention network architecture used for the Multi-Automated model, including the shared encoder and two task-specific output branches."},{"cited_title":"Moore, Thomas G","cited_arxiv_id":null,"evidence_quote":"Provides the OpenKBP head-and-neck dataset, the train/validation/test split, and the DVH evaluation framework used for the head-and-neck experiments."},{"cited_title":"Dose distribution prediction for head-and-neck cancer radiotherapy using a generative adversarial network: influence of input data","cited_arxiv_id":null,"evidence_quote":"Shows that dose models can be trained with limited contour information, motivating the integrated approach that avoids relying on input contour quality."},{"cited_title":"Mask-free radiotherapy dose prediction via multi-task learning","cited_arxiv_id":null,"evidence_quote":"Earlier multi-task learning work for dose prediction and contouring with tumor masks; the paper positions its sequential comparison against this line of work."},{"cited_title":"Loss odyssey in medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the combo loss used to train the contouring branch of the multi-task model."}],"review_version":1}