{"id":"a551c63b-7cbb-43fe-a373-53bca472f615","arxiv_id":"2507.03872","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding a hierarchical dual-attention and graph-based prior-reasoning module to existing 3D segmentation models improves liver lesion detection and diagnosis on non-contrast CT by 4 to 6 F1 points.","lead":"PLUS is a plug-in that makes existing CT segmentation models better at telling apart benign and malignant liver lesions on non-contrast CT scans. In 8,651 patients it raises lesion-level F1 by about 5.7 points and malignant patient-level F1 by about 6.3 points over a strong baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training losses L_P and L_S are undefined when a patient has no detected lesions, yet the dataset includes lesion-free patients; the reported PLUS gains rely on unspecified handling of empty S_j.","rationale":"I chose this concern because it attacks the training objective itself, not just the evaluation. The abstract headline improvements are produced by a model trained with L_total; if one term is undefined for a non-negligible part of the dataset, the description of the method is incomplete. The paper is otherwise plausible: code is released, ablations are provided, and the improved F1 over the PLAN baseline is internally consistent with the tables. I agree with the reader that statistical reporting is weak, but the empty-S_j issue is more directly falsifiable: a code inspection settles it. If the code handles empty sets, the concern reduces to a documentation fix and the CONDITIONAL verdict stands; if not, the central claim is in jeopardy. I therefore keep the verdict UNCHANGED relative to the reader's CONDITIONAL, with the explicit condition that the loss handling be clarified.","tokens_in":9070,"tokens_out":11260,"duration_ms":125011,"concrete_test":"Inspect the released repository (github.com/alibaba-damo-academy/plug-and-play-diagnosis) and locate the implementation of L_P and L_S. Check the branch executed when a training sample has zero detected lesions (e.g., a healthy subject). Then run one forward/backward pass on a lesion-free NCCT volume through the full PLUS model using the published loss. If the code defines a fallback (q_j=0 or skipping the sample), the equations in §2.2 must be revised and the impact on Table 1 reported; if the code crashes or leaves the loss undefined, the training procedure is not reproducible as stated and the reported gains are not supported by the described method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 2.2 (Training Paradigm), the patient-level loss L_P = -(1/M) Σ_j [Y_j log(max_{i∈S_j} p^i_malig) + (1-Y_j) log(1 - max_{i∈S_j} p^i_malig)] and screening loss L_S = -(1/M) Σ_j [y_j log q_j + (1-y_j) log(1-q_j)] with q_j = max_{i∈S_j} (p_i,malig ∥ p_i,beni) both take a maximum over the detected lesion set S_j. For any patient with no detected lesion (S_j empty), which the dataset explicitly includes ('patients with or without liver tumors'), the maximum is not defined. L_total = α L_L + β L_P + (1-α-β) L_S is the actual training objective (α=0.5, β=0.3), so every PLUS model depends on this loss. The paper does not state that q_j is set to 0 or that lesion-free patients are excluded from L_P and L_S. This is a concrete gap in the method as written, and the central F1 improvements cannot be reproduced from the equations alone. The released code can resolve whether an undocumented default exists; if it silently omits these patients, the screening objective described in the paper is not what was optimized.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PLUS, a plug-and-play module for enhancing focal liver lesion (FLL) diagnosis on non-contrast CT. Given an arbitrary 3D segmentation model that supplies a liver mask, per-lesion masks, and preliminary classification logits, PLUS refines predictions using three components: a hierarchical dual attention (HDA) mechanism that fuses global liver features with local lesion features, a graph-based prior reasoning (GPR) module that conditionally combines pretrained priors with learned prototypes, and a combined lesion-level, patient-level, and screening loss. Experiments on a curated dataset of 8,651 patients with 28,853 annotated lesions show consistent F1 improvements when PLUS is attached to nnUNet, Mask2Former, and PLAN. The best configuration, PLAN+PLUS, is reported to improve lesion-level F1 by 5.66 percentage points, malignant patient-level F1 by 6.26, and benign patient-level F1 by 4.03 relative to PLAN. Code is made available.","tokens_in":9377,"tokens_out":4853,"duration_ms":54216,"significance":"The strengths of the paper are the large clinical dataset, the consistency of the plug-and-play gains across three structurally different segmentation baselines, and the release of code. If the reported results are reproducible, PLUS is a practically relevant contribution to opportunistic liver lesion screening on non-contrast CT, a setting where contrast-enhanced imaging is often unavailable. However, the manuscript as written has a definitional gap in the training loss for lesion-free patients, and the statistical support for the headline improvements is incomplete (no confidence intervals, garbled p-value text). These issues are load-bearing because the loss definitions are the actual training objective and the 'significant' claim is unverifiable from the paper.","major_comments":[{"comment":"The patient-level loss L_P = -(1/M) Σ_j [Y_j log(max_{i∈S_j} p^j_{i,malig}) + (1-Y_j) log(1 - max_{i∈S_j} p^j_{i,malig})] and the screening loss L_S = -(1/M) Σ_j [y_j log q_j + (1-y_j) log(1-q_j)] with q_j = max_{i∈S_j} (p_{i,malig} ∥ p_{i,beni}) both take a maximum over the detected lesion set S_j. The dataset description explicitly includes 'patients with or without liver tumors', so for a lesion-free patient S_j is empty and both losses are undefined. Since L_total = αL_L + βL_P + (1-α-β)L_S is the stated training objective, the method is not fully specified. Please state exactly how empty S_j is handled (e.g., skipping such patients in L_P/L_S, setting q_j = 0, or using a fallback), and confirm that the released code implements this behavior. This is not a cosmetic issue: every PLUS model is trained with this objective, and the screening and patient-level metrics are directly affected.","section":"Section 2.2, Training Paradigm (loss definitions L_P, L_S, L_total)"},{"comment":"The full PLUS model with combined loss (last row) has a lesion-level F1 of 65.11, while the +GPR variant without the combined loss (third row) has a lesion-level F1 of 65.87. The paper describes the effect of adding the combined loss as 'slight deterioration at the lesion level', but the drop is 0.76 F1 points. Because the combined loss is a core claim of the framework, the authors should explain the trade-off in more detail: why is the full model preferred over the +GPR variant when lesion-level detection is a stated primary goal, and what are the patient-level metrics for the +HDA and +GPR variants? The current presentation leaves the reader uncertain whether the reported 'best' model is actually the best for the intended clinical screening use case.","section":"Table 2, Ablation Study"},{"comment":"No confidence intervals, standard deviations, or per-fold results are reported for any of the F1 or AUC numbers, despite the fact that the segmentation pretraining used 5-fold cross-validation. The text states improvements are 'significant', but the only significance evidence is Fig. 3(a), where the p-value text is unreadable ('p = 4 12 < 0.05' and 'p = 3 10 < 0.05'). Please add bootstrap confidence intervals or per-fold results for the main comparisons, fix the figure text, and describe the statistical test used. This is necessary for the reader to assess whether the headline gains are meaningfully larger than run-to-run variation.","section":"Figures 3(a) and all tables (statistical reporting)"}],"minor_comments":[{"comment":"The axis labels and p-value annotations contain corrupted glyphs ('PLA/glyph1197', 'p = 4 12', 'p = 3 10'); these should be corrected.","section":"Figure 3(a)"},{"comment":"The definition q_j = max_{i∈S_j} (p_{i,malig} ∥ p_{i,beni.}) uses an unexplained parallel symbol; please define it precisely, presumably as an element-wise maximum or a scalar aggregation.","section":"Section 2.2, screening loss"},{"comment":"The evaluation metrics 'lesion-level F1' and 'patient-level F1' are used throughout but never formally defined; please specify how predicted lesions are matched to ground-truth lesions (e.g., IoU threshold) and how per-patient scores are aggregated.","section":"Section 2.2 / Table 1"},{"comment":"The variant rows 'PLAN+PLUS(distill.)', 'PLUS(gated)', and 'PLUS(weighted)' place citations inside the method names; consider describing these fusion baselines in the text or caption instead.","section":"Table 1"},{"comment":"The reported changes (e.g., '5.66%') should be described as percentage-point differences in F1, since F1 is already a percentage-like quantity.","section":"Abstract and Section 1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a clinically relevant problem and the empirical consistency across baselines is encouraging. The empty-S_j issue is the most important technical gap, but it is likely fixable with a clear statement and a code check. The statistical reporting also needs completion. I did not verify the released code, but the authors' willingness to share it is a positive signal. I recommend major revision rather than rejection because the central idea is sound and the identified issues are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a workmanlike plug-and-play module that gives real but modest gains on a large private NCCT dataset, and the central loss equations have a hole for lesion-free patients that needs to be patched before the numbers can be trusted.\n\nWhat's new: the HDA and GPR components are engineering combinations of attention and graph reasoning, but the specific design and the scale of the dataset (8,651 patients, 28,853 lesions) move the needle. The gains are consistent across nnUNet, Mask2Former, and PLAN, which is credible evidence the module is architecture-agnostic. Code is public, which is a plus.\n\nSoft spots: the stress-test concern is on target. L_P and L_S both take a maximum over S_j, and the dataset explicitly includes patients without liver tumors. For empty S_j the max is undefined. The paper does not state the default. This is a methodological gap, not a nitpick. The code could resolve it, but a reader cannot reproduce the reported results from the paper alone.\n\nAlso missing: confidence intervals, per-fold variance, and a clean significance test (the p-value text in Fig. 3(a) is garbled). The full model's lesion-level F1 is lower than the HDA+GPR variant (65.11 vs 65.87), so the combined loss is a trade-off, not a free lunch; that's fine, but it should be framed as such.\n\nThe reader's worry about reliance on pre-segmentation masks is reasonable but proportionate: any plug-and-play method depends on the upstream model. The circularity concern is a non-issue.\n\nNet: the empirical claim is plausible and the architecture is sensible. The loss undefinedness is the one load-bearing flaw I would want fixed before publication, and the statistical reporting needs a second pass.\n\nRecommendation: send to peer review. The paper deserves referee time. If the authors clarify the empty-lesion handling, add variance, and fix the garbled text, this becomes a solid contribution for anyone working on NCCT-based liver screening.","headline":"Useful plug-and-play for NCCT liver lesion screening, but the loss equations are undefined when no lesions are detected and the statistics need a second pass.","tokens_in":9895,"tokens_out":2903,"would_cite":true,"duration_ms":31168,"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":"PLUS wraps any pre-trained 3D segmentation model and raises liver-lesion F1 by 5.66% on non-contrast CT.","keywords":["focal liver lesions","non-contrast CT","plug-and-play","hierarchical dual attention","graph-based prior reasoning","lesion classification","patient-level diagnosis","3D segmentation"],"falsifier":"Run PLUS on a held-out set of NCCT cases where the base segmenter's lesion recall is low, or corrupt the input lesion masks by randomly adding or removing voxels; if lesion-level F1 does not drop or the gains disappear, the proposed mask-conditioned reasoning is not the cause of the improvement.","tokens_in":8880,"feed_emoji":"🩻","tokens_out":6054,"duration_ms":58022,"temperature":0.7,"pith_summary":"PLUS is a plug-and-play module that upgrades any pre-trained 3D liver lesion segmentation model to diagnose focal liver lesions directly on non-contrast CT scans. It takes the segmentation model's liver mask, lesion masks, and preliminary class predictions as inputs, then re-reads the original image through two encoders and fuses the two scales before making final lesion and patient-level decisions. On a dataset of 8,651 patients and 28,853 lesions, adding PLUS to the PLAN baseline raises lesion-level F1 from 59.45% to 65.11%, malignant patient-level F1 from 83.86% to 90.12%, and benign patient-level F1 from 69.07% to 73.10%. The same module also improves nnUNet and Mask2Former, suggesting the gains come from the plug-in reasoning rather than a particular backbone. If correct, this offers a route to opportunistic liver cancer screening on the most widely available CT protocol.","feed_headline":"PLUS: plug-and-play module lifts liver-lesion F1 by 5.66%","feed_subtitle":"A wrapper around any 3D segmenter raises malignant patient-level F1 to 90.12% on plain non-contrast CT.","key_machinery":"The framework rests on three components. Hierarchical dual attention (HDA) performs bidirectional cross-attention between global liver features and local lesion features at multiple pooled scales, mimicking a radiologist zooming in and out and using liver context such as cirrhosis or steatosis to interpret a lesion. Graph-based prior reasoning (GPR) builds a graph whose nodes are the enhanced lesion features and learnable class prototypes, and uses the base model's prior class probabilities as conditional messages so reliable priors are amplified and unreliable ones suppressed. The combined loss has three terms — lesion-level focal loss, patient-level malignancy loss, and screening loss — that reproduce the clinical cascade from screening to diagnosis.","core_discovery":"The paper's central claim is that a classification-focused wrapper, rather than a new segmentation architecture, is what closes the gap between finding lesions and telling malignant from benign on NCCT. PLUS does this by extracting features only inside the liver mask and each lesion mask, exchanging information between the two with hierarchical dual cross-attention at multiple scales, and then refining the base model's uncertain prior probabilities with a graph that connects lesion features to class prototypes. A three-part loss mirrors clinical workflow: lesion-level focal loss, patient-level malignancy loss over the worst lesion, and a screening loss for any tumor. The reported result is consistent gains across three different segmentation baselines, with the largest on PLAN. The authors attribute the improvement to explicit modeling of the lesion–liver relationship that detection-only models lack.","pith_inferences":["PLUS's global-local recipe is organ-agnostic: the same HDA plus GPR wrapper could be pointed at other NCCT screening targets such as the pancreas or esophagus, though the paper only demonstrates liver lesions.","Because PLUS only touches features inside an existing model's masks, it should in principle compose with future backbones without modification; a cheap test is to plug it into a Mamba or pure-transformer segmenter not in the paper.","The reported gains are on a single test split of 1,200 patients; external multi-center validation would clarify whether the mid-probability calibration improvements shown in the probability correlation analysis generalize.","If mask dependence is the bottleneck, ensembling segmentations from multiple base models into the ROI inputs could push the F1 gains further than any single segmenter."],"forward_implications":["Existing NCCT lesion screening pipelines can be upgraded without retraining the segmentation backbone, since PLUS only consumes the backbone's masks and class logits.","Lesion-level detection and patient-level malignant versus benign diagnosis improve together, which is the combination needed for opportunistic cancer screening.","The architecture-agnostic design means the same wrapper should transfer to newer 3D segmentation models as they appear.","The graph-based prior reasoning path provides a template for correcting unreliable base-model predictions in other medical imaging tasks.","The hierarchical loss structure ties model training to the clinical cascade of screening, lesion analysis, and diagnosis, making the outputs directly relevant to care decisions."],"supporting_citations":[{"why":"PLAN is the main base model; PLUS's largest reported gains (5.66/6.26/4.03) are measured by adding the plugin to PLAN.","marker":"Yan et al. (2023)"},{"why":"nnUNet supplies one of the three segmentation backbones and the preprocessing or resampling pipeline used in all experiments.","marker":"Isensee et al. (2021)"},{"why":"Mask2Former is the second architecture-agnostic benchmark and the underlying segmentation design for PLAN.","marker":"Cheng et al. (2022)"},{"why":"The graph neural network message-passing formulation that GPR uses to connect lesion features to class prototypes.","marker":"Kipf and Welling, 2016"},{"why":"The self-attention operation used inside both the dual cross-attention module and the GPR graph reasoning.","marker":"Vaswani et al., 2017"},{"why":"Cited as evidence that prior segmentation predictions can be unreliable, motivating GPR's conditional rather than direct use of priors.","marker":"Meng et al., 2021"},{"why":"Knowledge distillation is one of the alternative fusion strategies that PLUS is compared against in Table 1.","marker":"Gou et al. (2021)"},{"why":"Gated fusion is another alternative fusion strategy compared against GPR in Table 1.","marker":"Arevalo et al. (2017)"},{"why":"Weighted fusion is the third alternative fusion strategy compared against GPR in Table 1.","marker":"Yin (2018)"}],"fun_headline_variants":["PLUS: plug-and-play lifts liver-lesion F1 by 5.66% on plain CT","Wrap any 3D segmenter with PLUS and gain 5.66% liver-lesion F1","No contrast: PLUS tells benign from malignant on routine CT","PLUS adapts any segmenter to screen liver cancer without contrast","Plug-in PLUS improves liver-lesion F1 by 5.66% on non-contrast CT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gains presume the pre-trained segmentation model already finds the liver and lesions well; if the base model misses lesions or produces false positive masks on NCCT, the feature extraction and prior predictions degrade and the reported F1 improvements shrink or vanish.","fun_headline_variants_meta":{"raw":{"variants":["PLUS: plug-and-play lifts liver-lesion F1 by 5.66% on plain CT","Wrap any 3D segmenter with PLUS and gain 5.66% liver-lesion F1","No contrast: PLUS tells benign from malignant on routine CT","PLUS adapts any segmenter to screen liver cancer without contrast","Plug-in PLUS improves liver-lesion F1 by 5.66% on non-contrast CT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1692,"prompt_tokens":909,"completion_tokens":783,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":670}},"tokens_in":525,"tokens_out":783,"duration_ms":8744,"temperature":1.0,"reasoning_tokens":670,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:59:53.875725+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PLUS on a held-out set of NCCT cases where the base segmenter's lesion recall is low, or corrupt the input lesion masks by randomly adding or removing voxels; if lesion-level F1 does not drop or the gains disappear, the proposed mask-conditioned reasoning is not the cause of the improvement.","supporting_citations":[{"cited_title":"nnu-net: a self-configuring method for deep learning-based biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"nnUNet supplies one of the three segmentation backbones and the preprocessing or resampling pipeline used in all experiments."},{"cited_title":"Schwing, Alexander Kirillov, and Rohit Girdhar","cited_arxiv_id":null,"evidence_quote":"Mask2Former is the second architecture-agnostic benchmark and the underlying segmentation design for PLAN."},{"cited_title":"Graph-based region and boundary aggregation for biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"Cited as evidence that prior segmentation predictions can be unreliable, motivating GPR's conditional rather than direct use of priors."},{"cited_title":"González","cited_arxiv_id":null,"evidence_quote":"Gated fusion is another alternative fusion strategy compared against GPR in Table 1."},{"cited_title":"Tensor sparse representation for 3-d medical image fusion using weighted average rule","cited_arxiv_id":null,"evidence_quote":"Weighted fusion is the third alternative fusion strategy compared against GPR in Table 1."}],"review_version":1}