{"id":"a0aea7f0-218a-4561-ba34-208873b07784","arxiv_id":"2507.07126","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"DpDNet, a dual-prompt network, achieves top average DSC (74.87%) and IoU (62.56%) on a four-cancer-type whole-body PET-CT segmentation benchmark, and its extracted biomarkers stratify breast cancer survival.","lead":"A new neural network, DpDNet, segments cancer lesions in whole-body PET-CT scans across four cancer types by combining cancer-specific and shared prompts, and reports the highest average Dice score among the compared methods (74.87%). The authors also use its segmentation outputs to compute metabolic tumor volume, total lesion glycolysis, and SUVmax for breast cancer survival risk stratification.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cancer type and data source are confounded for breast vs. AutoPET cancers; the paper does not show that learnable prompts encode biology rather than site artifacts.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: cancer type is confounded with data source, and the paper's attribution of gains to cancer-specific biological features is unsupported. The reader's conditional verdict is appropriate because the architecture is plausible, the code is released, and the ablation shows internally consistent improvements, but the central claim requires either an external multi-center dataset where cancer type and site are decorrelated, or an explicit site-vs-cancer prompt comparison, before the specific-prompt mechanism can be credited. I do not see an internal inconsistency in the method itself; the concern is a correctness risk from confounding and missing statistical detail. The survival results, while interesting, have overlapping confidence intervals and therefore cannot independently carry the claim that segmentation accuracy drives prognostic precision. Given that the reader already conditioned acceptance on these issues, no verdict adjustment is needed beyond what was issued.","tokens_in":8464,"tokens_out":4003,"duration_ms":55014,"concrete_test":"Retrain DpDNet with the same data, architecture, and hyperparameters, but define the prompt condition by acquisition site or scanner instead of cancer type (using the two AutoPET centers and the private breast collection as separate site prompts), with 5 random seeds and reporting mean +/- std DSC. If the site-prompt variant matches or exceeds the cancer-prompt variant's 74.87% average DSC, the reported cancer-prompt advantage is explained by dataset identity; if the cancer-prompt variant remains significantly better with non-overlapping intervals, the confound is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DpDNet's cancer-specific prompts drive its universal PET-CT segmentation gains rests on an untested identification assumption: that the task label 'cancer type' is learnable independently of data source. In Section 3.1, lung cancer, lymphoma, and melanoma come from AutoPET, while breast cancer is a private collection, so cancer type is perfectly confounded with acquisition site, scanner, protocol, and annotation style. The breast cancer column, where the source differs, may be where the prompt has the most room to memorize site-specific features, and the T-SNE separation in Fig. 2 could reflect source identity rather than cancer biology. Because the method requires the cancer type at inference to select the specific prompt, a model that simply encodes scanner or protocol appearance could reproduce the reported results. The claim in Table 1 that all comparisons yield p<0.05 cannot be verified without error bars, repeated-seed variance, or a description of the test; the magnitude of the average gain over the second-best method is 1.32% DSC, which is small relative to the likely run-to-run variability of 3D medical segmentation training. The survival analysis does not resolve this: in Table 3, the proposed method's MTV C-index is 69.33 with 95% CI [62.05, 75.83], while the CLIP-driven model yields 68.83 with CI [61.62, 75.69], so the intervals overlap substantially and the claim that better segmentation improves survival prediction is not statistically established. Thus, the load-bearing vulnerability is that the comparisons do not distinguish a genuine biological task-specific advantage from a dataset-identity shortcut.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DpDNet, a dual-prompt-driven network for universal whole-body PET-CT lesion segmentation across four cancer types. The architecture uses learnable cancer-specific prompts and a common prompt derived by gated fusion of the remaining cancer prompts, both injected into a shared STU-Net-S decoder, followed by prompt-aware heads with channel attention and multi-scale branches. Experiments combine AutoPET lung/lymphoma/melanoma cases with a private breast cancer collection (697 cases, 557/140 train/validation) and report average DSC 74.87% and IoU 62.56%, exceeding a second-best prompt-based model by 1.32% DSC and 1.56% IoU. The authors also compute SUVmax, MTV, and TLG from predicted breast cancer segmentations and use median-derived thresholds to stratify 780 patients, reporting MTV as the strongest prognostic marker (HR=2.88). The code is publicly available.","tokens_in":8809,"tokens_out":5727,"duration_ms":60503,"significance":"If the reported improvements are statistically robust and the cancer-specific prompts generalize beyond acquisition-site identity, the paper would make a genuinely useful contribution: it offers a parameter-efficient universal segmentation architecture that explicitly separates shared and cancer-specific features, with an ablation study isolating each component, and it goes beyond segmentation accuracy by linking predicted volumes to survival stratification. The main segmentation evaluation is on a held-out validation split, so the central claim is not circular. The empirical evidence for the segmentation claim is encouraging but currently under-supported; the manuscript would be strengthened substantially by error bars, repeated-seed variability, per-cancer sample sizes, a source-confounding check, and a statistically defensible survival comparison.","major_comments":[{"comment":"The central interpretation of the cancer-specific prompts is threatened by a confound between cancer type and data source: lung cancer, lymphoma, and melanoma are taken from AutoPET, while breast cancer is a private collection, and the model requires the cancer type as input at inference time. Consequently, the specific prompts, and the T-SNE separation in Fig. 2, could encode scanner, protocol, or annotation-style differences rather than cancer biology. To support the claim that the prompts capture cancer-specific metastatic patterns, the authors should provide a control experiment that disentangles source from cancer type — for example, adding acquisition site as an additional prompt or label, training on subsets matched by acquisition protocol, or validating on an external multi-site dataset where the same cancer type comes from different scanners and different cancer types share a scanner. Without such a test, the mechanism attributed to cancer specificity remains an identification assumption.","section":"Section 3.1, Table 1, Fig. 2"},{"comment":"The headline performance claim is not statistically supported as reported. The caption states that 'all comparison results yielded p <0.05' but gives no test procedure, variance estimates, confidence intervals, or number of repeated runs, and the average gain over the second-best method is only 1.32% DSC. Given the well-known run-to-run variability of 3D medical segmentation training, the authors should report mean ± standard deviation over at least three random seeds for each method, specify the paired statistical test and adjustment for multiple comparisons, and give per-cancer validation sample sizes. This is essential before readers can assess whether the reported ranking is trustworthy.","section":"Table 1, Section 3.3"},{"comment":"The survival analysis does not establish that better segmentation leads to better survival prediction. The C-indices for MTV are 69.33 [62.05, 75.83] for DpDNet, 68.83 [61.62, 75.69] for CLIP-driven, and 68.51 [61.79, 75.30] for DoDNet, so the 95% confidence intervals overlap almost completely; the same holds for TLG and SUVmax. Moreover, the median thresholds are computed on 480 patients and applied to a disjoint 780-patient set, but no details are given for how the C-index and hazard ratio are computed (censoring, covariates, whether thresholds are re-estimated, or whether the 780 patients are used for the C-index). The authors should report a formal comparison of the survival models (e.g., difference in C-index with confidence intervals, or a test of equivalence) and validate the threshold derivation, for example by internal cross-validation, before claiming clinical utility.","section":"Table 3, Section 3.3"},{"comment":"The composition of the training and validation sets is under-specified. The text reports 697 total cases with 557 training and 140 validation cases, but it does not state the number of cases per cancer type, whether the split is at the patient level (e.g., multiple scans per patient), or the acquisition protocols and annotation guidelines for the private breast cancer collection. This information is necessary to interpret the averaged metrics, to assess class balance, and to reproduce the experiments. It also bears directly on the source-confounding concern raised above.","section":"Section 3.1"}],"minor_comments":[{"comment":"The phrase 'An Dual-Prompt-Driven Network' should read 'A Dual-Prompt-Driven Network'; similar grammatical issues appear elsewhere in the main text.","section":"Abstract and Title"},{"comment":"The sentence introducing evaluation metrics says 'four metrics' but then lists only DSC and IoU; the C-index and HR are described in the next sentence, so the wording should be corrected to avoid confusion.","section":"Section 3.1"},{"comment":"The notation `Fcat(...)` is undefined; if it denotes concatenation, it should be written `cat(...)` consistently with Eq. (2).","section":"Equation (3)"},{"comment":"The caption says 'Weighted Averages' but the weighting scheme is not described anywhere; either define the weights or remove the term.","section":"Table 1 caption"},{"comment":"The table labels the first row 'Baseline' but does not specify what the baseline includes; the text should state explicitly that it is STU-Net-S without any prompt module.","section":"Table 2 and Section 3.3"},{"comment":"The T-SNE visualization is split into four separate plots, making it difficult to compare the prompt geometries across cancer types; a single joint plot or a fixed embedding space would be more informative.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the code release is a plus. My main concern is not the architecture itself but the evidential gap between the reported numbers and the strong mechanistic claims. The source confounding issue is particularly important because the method's name and design imply biology-specific prompts; a reviewer or reader versed in medical imaging will immediately ask whether the private breast cohort is the actual driver. I would encourage the editor to solicit the additional control experiments and statistical reporting described in the report before further consideration. There is no indication of misconduct; the issues are standard rigor problems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a sensible architectural idea and a full benchmark, but the central claim is weaker than it looks. Cancer type and data source are perfectly entangled: lung, lymphoma, and melanoma come from AutoPET, while breast cancer comes from a private collection. So the reported gains could just be the model separating scanner or protocol artifacts, not cancer biology. The T-SNE in Fig. 2 could be showing site identity, not metastatic patterns. That's a load-bearing flaw, and the paper does not address it.\n\nWhat is genuinely new: the dual-prompt design—splitting a UniSeg-style prompt into cancer-specific prompts plus a gated common prompt, and adding prompt-aware heads to each task. The ablation is clean and shows each component contributes. The paper is clearly written, the code is public, and using the segmentation output for MTV/TLG/SUVmax-based survival analysis is a practical direction.\n\nThe soft spots beyond the confound: no error bars anywhere. Table 1 reports single runs, no standard deviations or repeated-seed variance, and \"p<0.05\" appears without any test description. A 1.3% DSC gain over the second-best method could easily be training noise in 3D segmentation. The survival analysis does not solidify anything: the CIs for MTV overlap substantially across DoD-Net, CLIP-driven, and the proposed method (69.33 with CI [62.05, 75.83] vs. 68.83 with CI [61.62, 75.69]). So the claim that better segmentation accuracy improves survival prediction is not statistically supported.\n\nAll that said, this is a well-engineered paper with a testable hypothesis and a useful benchmark. It deserves a serious referee, but the authors need to either remove the confound (e.g., multi-source validation, site-adversarial checks) or openly discuss it, report repeated-run variance, and temper the survival claims. A revision could make it a solid contribution.\n\nMy take: send it to review, but expect major revisions.","headline":"Clean architecture and honest ablation, but the main benchmark is confounded by data source and the statistics are too thin to support the headline claims.","tokens_in":9345,"tokens_out":1931,"would_cite":false,"duration_ms":22863,"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":"DpDNet splits cancer-specific and shared prompts to lift PET-CT segmentation across four cancer types.","keywords":["PET-CT segmentation","universal segmentation","dual prompts","cancer-specific prompts","prompt-aware heads","survival analysis","metabolic tumor volume"],"falsifier":"Retrain DpDNet with the cancer-type prompts randomly shuffled across datasets, or evaluate it on breast-cancer scans from a public multicenter cohort; if the reported 1.32-point DSC advantage vanishes or fails to transfer, the claimed benefit is attributable to site separation rather than cancer-specific knowledge.","tokens_in":8303,"feed_emoji":"🩻","tokens_out":4621,"duration_ms":46806,"temperature":0.7,"pith_summary":"PET-CT lesion segmentation usually treats all cancers as one task, but different cancers spread and light up differently. DpDNet instead learns a separate prompt for each cancer plus a shared common prompt, feeding both into the decoder early, with prompt-aware heads that adapt the final prediction per cancer. On a four-cancer whole-body dataset, the paper reports the best average Dice of 74.87% and IoU of 62.56%, beating the second-best method by 1.32 and 1.56 points. It then shows the segmentations can drive breast-cancer survival stratification, with metabolic tumor volume giving a hazard ratio of 2.88. The claim is that explicitly decoupling shared and cancer-specific knowledge improves universal segmentation and downstream prognosis.","feed_headline":"Dual-prompt network beats baselines on four-cancer PET-CT segmentation","feed_subtitle":"Separate cancer prompts plus prompt-aware heads reach 74.87% average Dice and sharpen breast-cancer survival risk scores.","key_machinery":"The engine is the dual-prompt block plus prompt-aware heads. A universal prompt $F_{\\text{uni}}\\in\\mathbb{R}^{N\\times D/16\\times H/32\\times W/32}$ is split along the channel dimension into cancer-specific prompts $\\{F_{\\text{can}}^1,\\ldots,F_{\\text{can}}^N\\}$; the target prompt is selected according to the current task, while the remaining prompts are gated-fused into a common prompt $F_{\\text{com}}^i$ that encodes shared metastatic behavior. These are concatenated with the bottleneck features and fed into the shared decoder, so task awareness arrives early. Prompt-aware heads then replace one shared head with per-cancer heads that add channel attention and multi-scale branches, preventing the early prompts from being forgotten and improving small-target extraction. The T-SNE visualizations show the common prompt sitting centrally among separated cancer-specific prompts, supporting the claim that the two prompt families capture different information.","core_discovery":"The paper's central discovery is that a dual-prompt design can outperform both single-task and flat multi-task PET-CT segmentation. DpDNet computes cancer-specific prompts from the bottleneck features plus a common prompt formed by gated fusion of the other cancers' prompts, then concatenates both with the bottleneck as decoder input. After the decoder, a prompt-aware head selected by cancer type uses channel attention and 1x1x1, 3x3x3, and 5x5x5 convolution branches to capture small lesions. Across lung cancer, lymphoma, melanoma, and breast cancer, the model reaches average DSC 74.87% and IoU 62.56%, with the best scores in lung, lymphoma, and breast cancer; in survival analysis on 1,210 breast-cancer patients, segmentation-derived MTV stratifies overall survival with HR 2.88. The authors conclude that separating shared from cancer-specific features is the mechanism that produces both better masks and more clinically usable biomarkers.","pith_inferences":["The central comparison is confounded: lung, lymphoma, and melanoma scans come from a public dataset while breast scans come from a private collection, so the task identity is entangled with scanner and protocol differences; a cleaner test would use each cancer from multiple sites.","The survival analysis derives median thresholds from 480 patients and applies them to 780 others, and no external validation cohort is reported, so the HR of 2.88 for MTV should be read as an exploratory result until replicated.","The dual-prompt idea is modality-agnostic and could be tested on CT or MRI organ segmentation or on other multi-cancer imaging tasks where shared and specific anatomical patterns matter.","One testable extension is to condition prompts on patient-level variables such as stage, subtype, or lesion site instead of only cancer type, which could improve stratification further."],"forward_implications":["Because the architecture is prompt-based, adding a new cancer type requires a new learnable prompt and head rather than retraining the whole network from scratch.","Decoupling cancer-specific and shared prompts should help when each cancer has few annotated cases, since the common prompt can transfer knowledge across tasks.","Prompt-aware heads allow task adaptation with only about 15 million parameters and roughly 139 GFLOPs, so the approach remains practical for whole-body volumes.","Segmentation quality directly affects downstream clinical quantities: the better segmentor yields higher C-indices for MTV, TLG, and SUVmax than DoDNet and CLIP-driven baselines.","The reported hazard ratios imply that accurate automated PET-CT segmentation could be used for noninvasive risk stratification in breast cancer without manual lesion delineation."],"supporting_citations":[{"why":"Supplies the STU-Net-S encoder-decoder backbone that DpDNet builds on.","marker":"[11]"},{"why":"Provides the UniSeg method whose universal-prompt formulation DpDNet extends by separating specific and common prompts.","marker":"[23]"},{"why":"Provides DoDNet's dynamic convolution head, the main baseline for the prompt-aware head design.","marker":"[25]"},{"why":"Provides the CLIP-driven universal model baseline and the idea of embedding label descriptions as external prompts.","marker":"[16]"},{"why":"Supplies the public whole-body FDG-PET/CT data for lung cancer, lymphoma, and melanoma used in training and validation.","marker":"[6]"},{"why":"Supplies the nnU-Net baseline and the general segmentation approach that the backbone inherits.","marker":"[12]"},{"why":"Grounds the biological premise that different cancers have distinct yet overlapping metastatic progression patterns.","marker":"[2]"}],"fun_headline_variants":["Dual-prompt PET-CT segmentation tops four-cancer benchmarks","Cancer-specific prompts boost PET-CT lesion segmentation","Prompt-aware heads sharpen four-cancer PET-CT masks","Dual-prompt model improves PET-CT segmentation and survival risk","Personalized PET-CT segmentation with dual-prompt design"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the learnable cancer-specific prompts capture biological differences between cancers rather than differences in where the scans came from, since three cancer types come from one public dataset and breast cancer comes from a private collection.","fun_headline_variants_meta":{"raw":{"variants":["Dual-prompt PET-CT segmentation tops four-cancer benchmarks","Cancer-specific prompts boost PET-CT lesion segmentation","Prompt-aware heads sharpen four-cancer PET-CT masks","Dual-prompt model improves PET-CT segmentation and survival risk","Personalized PET-CT segmentation with dual-prompt design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1293,"prompt_tokens":965,"completion_tokens":328,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":252}},"tokens_in":581,"tokens_out":328,"duration_ms":4332,"temperature":1.0,"reasoning_tokens":252,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:07:08.346408+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain DpDNet with the cancer-type prompts randomly shuffled across datasets, or evaluate it on breast-cancer scans from a public multicenter cohort; if the reported 1.32-point DSC advantage vanishes or fails to transfer, the claimed benefit is attributable to site separation rather than cancer-specific knowledge.","supporting_citations":[{"cited_title":"In: International Title Suppressed Due to Excessive Length 11 Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Provides the UniSeg method whose universal-prompt formulation DpDNet extends by separating specific and common prompts."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Provides the CLIP-driven universal model baseline and the idea of embedding label descriptions as external prompts."},{"cited_title":"Data9(1), 601 (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the public whole-body FDG-PET/CT data for lung cancer, lymphoma, and melanoma used in training and validation."},{"cited_title":"Oncotarget 6(1), 570 (2014)","cited_arxiv_id":null,"evidence_quote":"Grounds the biological premise that different cancers have distinct yet overlapping metastatic progression patterns."}],"review_version":1}