{"id":"b1a12a5a-28bb-40e7-b8e1-3f1164cfb7f9","arxiv_id":"2412.14846","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A dual-encoder UNet with CT pre-training and MixUp augmentation achieves 82.38% and 72.53% aggregated Dice on pre- and mid-radiotherapy head and neck tumor segmentation.","lead":"This paper describes an automated method for segmenting head and neck tumors and lymph nodes in MRI scans taken before and during radiotherapy. The method combines CT-based pre-training, MixUp augmentation, and a dual-encoder network, and reports aggregate Dice scores of 82.38% (pre-RT) and 72.53% (mid-RT) on the HNTS-MRG2024 challenge test set.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported challenge scores are credible, but the paper's central claim that the proposed pre-training and DFUNet components drive the result is unsupported by the five-fold ablation due to missing variance estimates and on-average negative DFUNet performance.","rationale":"I read the paper as a challenge-report whose central claim is that a pipeline combining CT pre-training, MixUp, and a dual-encoder DFUNet achieves high aggregated DSC on HNTS-MRG2024. The test scores themselves are likely valid, coming from an external evaluation, and the authors provide code. The soft spot is not the benchmark but the causal attribution: the only controlled comparisons are the five-fold ablations, and these do not show a stable advantage for the proposed components. The reader flagged CT-to-MRI transfer as the weakest assumption, which is one instance of this broader problem; my analysis extends it to MixUp and especially DFUNet, whose average performance is worse than the baseline. This is an internal-evidence issue rather than an external-consensus disagreement. The paper's own Discussion acknowledges the pre-training instability and DFUNet underperformance, which supports a conditional (revise-and-resubmit) rather than reject decision. A paired significance test across folds is a concrete, low-cost check; with only five folds it may not have power, which is itself informative. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":8115,"tokens_out":7477,"duration_ms":66658,"concrete_test":"Run a paired bootstrap (or exact sign test) on the five fold-level aggregated DSC values for (a) Pre-train+MixUp vs Base in Task-1 and (b) DFUNet vs Base+pre-RT in Task-2, reporting 95% confidence intervals for the mean differences and two-sided p-values. If either confidence interval includes zero, the abstract and conclusion should be revised to avoid claiming that the corresponding component improves segmentation, and the final test scores should be attributed to the full ensemble rather than to the individual innovations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The final test scores come from a single submission and are externally credible, so the headline numbers are not in dispute. What is load-bearing is the paper's causal attribution of those scores to the proposed components. The five-fold ablation in Tables 1 and 2 is the only internal evidence, yet it shows (i) for Task-1, Pre-train+MixUp improves GTVp by only 0.47 average points over Base (76.03 vs 75.56), and in Fold 5 the GTVp score drops from 80.39 to 78.82; no variance, confidence intervals, or significance tests are reported, so this difference is within fold noise; (ii) for Task-2, DFUNet is on average worse than the simpler Base+pre-RT in both GTVp (60.37 vs 61.32) and GTVn (86.58 vs 86.65), and the paper's Discussion explicitly concedes that DFUNet underperformed. The final ensemble contains only one DFUNet model out of five, so the 'dual-encoder network' cannot be credited for the 72.53% mid-RT score. Thus the condition required for the central claim - that the described architectural and pre-training choices are responsible for the high test performance - is not established by the paper's own evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes the authors' contribution to the HNTS-MRG2024 challenge for segmenting head and neck gross tumor volumes (GTVp) and metastatic lymph nodes (GTVn) on pre-radiotherapy and mid-radiotherapy T2-weighted MRI. For Task-1 they compare a basic VNet-like segmentation network trained with full supervision, pre-training on the CT-based SegRap2023 dataset with histogram matching and nonlinear intensity transforms, MixUp augmentation, and combinations of these. For Task-2 they introduce a Dual Flow UNet with two encoders and cross-attention to incorporate registered pre-RT images and their labels into the mid-RT segmentation stream. The best model per fold is selected from five-fold cross-validation and ensembled, yielding final test aggregated Dice scores of 82.38% for Task-1 and 72.53% for Task-2. The paper concludes that the proposed pre-training, MixUp, and DFUNet components contributed to the results.","tokens_in":8379,"tokens_out":5720,"duration_ms":36706,"significance":"If the causal claims were established, the paper would provide a useful practical recipe for MRI head and neck tumor segmentation: CT pre-training with intensity alignment, MixUp augmentation, and a dual-encoder network for using registered pre-RT images. The final test scores are measured on the challenge's external held-out test set, which gives objective support for the headline numbers. The authors make their code publicly available, use public datasets for pre-training, and include an honest discussion that acknowledges limitations. However, the internal evidence for the causal contribution of the proposed components is weak: the five-fold ablations are single runs without variance estimates, pre-training is inconsistent and sometimes harmful, and DFUNet is on average worse than the simpler baseline. These issues are load-bearing because the abstract and introduction attribute the final test performance to the proposed components. The manuscript needs sharper statistical evidence and more careful framing before the contribution claims can be accepted.","major_comments":[{"comment":"The five-fold cross-validation results are reported without variance, confidence intervals, or significance tests, with only a single run per condition. The claimed improvements are small relative to fold-to-fold variation: for Task-1, Pre-train+MixUp improves average GTVp by only 0.47 percentage points over Base (76.03 vs 75.56), and for GTVn by 1.30 points (85.12 vs 83.82). Fold-level differences, such as Fold 5 where Pre-train+MixUp reduces GTVp from 80.39 to 78.82, are of comparable or larger magnitude. The Discussion's statement that the strategies led to 'significant improvements' is therefore not supported. Please add repeated experiments or a statistical analysis, or soften the claims to describe observed trends in this specific dataset.","section":"Section 4, Tables 1 and 2"},{"comment":"The claim that pre-training on SegRap2023 CT data transfers useful features to T2-weighted MRI is not consistently supported. With pre-training alone, the average GTVp score is lower than Base (75.23 vs 75.56), and Pre-train+MixUp decreases GTVp relative to Base in Fold 5 (78.82 vs 80.39). The Discussion itself concedes that 'pre-training is not universally effective'. Because the final Task-1 ensemble is a per-fold selection among Base, MixUp, and Pre-train+MixUp, the contribution of pre-training to the 82.38% test score is unclear. Please report results separately for GTVp and GTVn by fold, state explicitly in which folds pre-training was selected, and avoid presenting pre-training as a uniformly beneficial component.","section":"Section 2.2 and Table 1"},{"comment":"DFUNet, presented as a novel contribution in the introduction, underperforms the simpler Base+pre-RT model on average for both GTVp (60.37 vs 61.32) and GTVn (86.58 vs 86.65). The Discussion explicitly states that 'the DFUNet underperformed compared to the basic segmentation model in cross-validation'. The text also notes that DFUNet outperforms Base+pre-RT in folds 1 and 4; if those are the selected folds, the final ensemble contains two DFUNet models, not one. The manuscript should state the selection criterion precisely and quantify the contribution of DFUNet to the final 72.53% score, or it should temper the abstract and conclusion so that DFUNet is presented as a promising but not yet validated architecture.","section":"Section 2.1, Table 2, and Section 5"},{"comment":"The criterion for selecting the best model in each fold is not defined. It is unclear whether selection is based on aggregated DSC over both structures, on GTVp only, on GTVn only, or on a subjective compromise. The bold markers in Tables 1 and 2 are not visible in the text, and the caption says only that bolding indicates the selected models. Because the final ensemble depends on this selection, the criterion must be stated explicitly. Also, selecting the best model per fold on the validation set introduces optimism in the reported cross-validation average of 80.65% and 74.68%; please acknowledge this or report the non-oracle average.","section":"Section 4, 'selected the highest-performing model'"}],"minor_comments":[{"comment":"The phrase 'as HiLab' in the abstract is unclear; please clarify whether it means 'at HiLab' or 'on the HiLab evaluation platform'.","section":"Abstract"},{"comment":"In the second paragraph, 'per-RT images' appears to be a typo for 'pre-RT images'.","section":"Section 1"},{"comment":"Equation (2) sums over d from 0 to l, but the number of deep-supervision resolutions l is not specified in the implementation details. Please state the value used.","section":"Section 2.4"},{"comment":"The preprocessing threshold of intensity greater than 60 for the body mask and the morphological operation parameters are given without rationale or sensitivity analysis; a brief justification would improve reproducibility.","section":"Section 3.2"},{"comment":"The captions say that cases are classified as well-predicted or poor-predicted based on 'whether they meet the method's improvements', which is circular. Please define an objective quantitative criterion for case selection.","section":"Figures 5 and 6"},{"comment":"The nonlinear intensity transformations are visualized in Figure 3, but the reference to 'Bézier Curve [8]' is not explained in the text; please clarify how the Bézier curve is used and why it is cited.","section":"Section 2.2 and Figure 3"},{"comment":"The caption states that 'Pre-train+MixUp' uses the basic segmentation network, but the text in Section 2 for Task-2 lists only DFUNet as an additional training method. Please align the method list with the table entries.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The external test scores are credible and the paper is a reasonable challenge report, but the causal narrative is not supported by the internal ablations. The skeptical review's claim that the final ensemble contains only one DFUNet model appears inconsistent with the manuscript text, which states that DFUNet outperforms the baseline in folds 1 and 4; however, the more important point stands that DFUNet is on average worse and the paper concedes this. A revision that adds variance estimates, defines the selection criterion, and recalibrates the contribution claims would make the paper suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid challenge-participation write-up with externally credible test scores, and the authors are properly cautious in the discussion about which components helped. What it is not is a paper that establishes that the proposed pre-training or the DFUNet architecture produce the headline numbers. The five-fold ablation is single-run, has no variance or significance info, and on average DFUNet is worse than the simpler Base+pre-RT in Task-2. The authors say this in the Discussion, so the paper is not misleading, just under-powered on attribution.\n\nWhat is actually new: results on the HNTS-MRG2024 test set, a pre-training pipeline using SegRap2023 CT with histogram matching and nonlinear intensity transforms, and a dual-encoder DFUNet with cross attention for mid-RT. The strengths: the test set is external, the code is public, and the preprocessing is described well enough to reproduce. The honesty about negative results is refreshing: they explicitly note pre-training can hurt GTVp in some folds and DFUNet underperforms on average.\n\nSoft spots, in proportion: the ablation lacks error bars over folds; differences like +0.47 DSC for GTVp are within noise. DFUNet's improvement appears only in some folds, and the final submission ensemble includes only one of five DFUNet models, so the 72.53% cannot be credited to the architecture. The CT-to-MRI transfer assumption is load-bearing, and the paper's own results show it is inconsistent. These are addressable with more experiments or a more modest claim.\n\nWho benefits: anyone working on the HNTS-MRG2024 benchmark or CT-to-MRI pre-training for segmentation. It is a useful baseline and an honest negative-result data point. It deserves a serious referee because the external scores and reproducible code give it value even if the attribution is weak. My recommendation: engage with it, but ask for variance reporting and a softening of the causal language.","headline":"Credible external challenge scores and an honest write-up, but the ablation is too weak to attribute the gains to any single component.","tokens_in":8936,"tokens_out":1490,"would_cite":true,"duration_ms":13087,"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":"A CT-pretrained dual-flow CNN segments head-and-neck tumors on T2-weighted MRI with 82.38% pre-RT and 72.53% mid-RT aggregated Dice scores.","keywords":["HNTS-MRG2024","head and neck cancer","tumor segmentation","MRI","pre-training","MixUp","dual encoder","Dice similarity coefficient"],"falsifier":"A head-to-head test on the HNTS-MRG2024 training set: train the identical Task-1 pipeline with and without CT pre-training, holding every other component fixed, and compare aggregated DSC for GTVp and GTVn across all five folds. If the no-pre-training model matches or beats the pre-trained model on GTVp in a majority of folds, the claimed benefit of cross-modal pre-training would not hold.","tokens_in":7902,"feed_emoji":"🩻","tokens_out":5721,"duration_ms":35653,"temperature":0.7,"pith_summary":"The paper aims to show that a fully supervised CNN pipeline—CT pre-training with intensity matching, MixUp augmentation, and a dual-encoder network that fuses registered pre-treatment images and labels into the mid-treatment stream—can segment head-and-neck primary tumors (GTVp) and metastatic lymph nodes (GTVn) from T2-weighted MRI with aggregated Dice scores of 82.38% for pre-RT and 72.53% for mid-RT on the HNTS-MRG2024 test set. This matters because pixel-level tumor and node contours are required for adaptive radiotherapy planning, and MRI provides better soft-tissue contrast than the CT images most existing datasets provide. The paper's own ablations show the improvements are concentrated in lymph node segmentation; primary tumor gains are marginal and pre-training can hurt in some training folds.","feed_headline":"CT pre-training lifts MRI tumor contouring to 82% Dice","feed_subtitle":"Lymph-node contours gain most from CT pre-training and MixUp; primary-tumor gains are smaller.","key_machinery":"The Dual Flow UNet (DFUNet) is the paper's central architectural object: two separate encoders, one taking the single-channel mid-RT image and the other taking the registered pre-RT image with its label mask, merged at each resolution stage by a CNN-based cross-attention block that applies both channel and spatial attention (CBAM-style) to the secondary pre-RT features before injecting them into the primary mid-RT stream. For Task-1, the load-bearing mechanism is the training recipe: pre-training a VNet-like encoder-decoder on CT volumes after histogram matching and nonlinear intensity transforms, then fine-tuning on MRI with MixUp-generated mixed patches, using cross-entropy loss for mixed samples and a deep-supervised combination of cross-entropy and Dice loss for raw samples.","core_discovery":"On the paper's own terms, the central discovery is that combining (1) pre-training on the SegRap2023 CT dataset after histogram matching and nonlinear intensity transformations, (2) MixUp augmentation of training patches, and (3) a Dual Flow UNet whose second encoder ingests the registered pre-RT image and its mask through CNN-based cross-attention blocks, yields a competitive segmentation pipeline for the HNTS-MRG2024 challenge, with 82.38% aggregated DSC on pre-RT and 72.53% on mid-RT test data. The ablation tables show that adding the pre-RT image and label to the mid-RT input is the single largest factor for Task-2, lifting average aggregated DSC from 37.60% to 61.32% for GTVp, while the DFUNet architecture itself only surpasses the basic network in two of five folds. For Task-1, the combination of pre-training and MixUp improves GTVn by 1.30% on average while leaving GTVp essentially unchanged.","pith_inferences":["A natural extension the authors leave implicit is a standalone validation of the cross-modal pre-training: if the same fine-tuning is repeated with a modality-invariant pretraining objective (e.g., style augmentation or contrastive alignment), the GTVp gains might become consistent where the current intensity-matching transfer fails.","The dual-encoder design could transfer to other settings where registered baseline and follow-up scans exist, such as treatment response assessment or longitudinal tumor tracking, not just mid-RT radiotherapy.","The reversed difficulty ordering between modalities (GTVp easier than GTVn on CT, harder on MRI) suggests the intensity and contrast characteristics of each structure are modality-specific; a testable consequence is that reweighting the loss per structure or per fold should improve the weaker class's Dice without hurting the stronger one."],"forward_implications":["If the reported test scores hold, this pipeline establishes a practical baseline for automated GTVp/GTVn contouring in adaptive radiotherapy workflows using T2-weighted MRI alone.","The large Task-2 gain from adding registered pre-RT images and labels implies that longitudinal imaging is a high-value input for mid-treatment segmentation, worth exploiting beyond simple input concatenation.","The observation that GTVn benefits more than GTVp from pre-training and MixUp suggests class- and structure-specific augmentation or loss strategies may be needed for hard classes.","The DFUNet's inconsistent gains indicate that two-stream attention fusion is not uniformly better than a single-stream network, and further architecture refinement is needed before it is reliable."],"supporting_citations":[{"why":"Supplies the SegRap2023 CT dataset used for pre-training the Task-1 model.","marker":"[9]"},{"why":"Defines the MixUp augmentation that creates mixed training patches and is reported to improve GTVn segmentation.","marker":"[18]"},{"why":"Provides the nnU-Net v2 framework that automates training configuration for all experiments.","marker":"[6]"},{"why":"Gives the VNet-like encoder-decoder architecture that both the basic network and DFUNet build on.","marker":"[10]"},{"why":"Supplies the CBAM channel-and-spatial attention mechanism used inside the CNN-based cross-attention block.","marker":"[16]"},{"why":"Inspires the CNN-based cross-attention fusion that integrates the pre-RT features into the mid-RT stream.","marker":"[13]"}],"fun_headline_variants":["CT pretraining and MixUp push MRI tumor Dice to 82%","Dual-flow UNet with CT priors scores 82% on head-neck MRI","Pre-RT mask is key for mid-RT: 72% Dice from HN MRI","82% pre-RT Dice via CT pretraining and dual encoders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pre-training on the SegRap2023 CT dataset, after histogram matching and nonlinear intensity transforms, transfers features that help T2-weighted MRI segmentation; the paper's own fold-level results show this transfer is inconsistent, lowering GTVp scores in some folds.","fun_headline_variants_meta":{"raw":{"variants":["CT pretraining and MixUp push MRI tumor Dice to 82%","Dual-flow UNet with CT priors scores 82% on head-neck MRI","Pre-RT mask is key for mid-RT: 72% Dice from HN MRI","82% pre-RT Dice via CT pretraining and dual encoders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001483,"raw_usage":{"total_tokens":5988,"prompt_tokens":1007,"completion_tokens":4981,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":4894}},"tokens_in":623,"tokens_out":4981,"duration_ms":23127,"temperature":1.0,"reasoning_tokens":4894,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:51:06.341478+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A head-to-head test on the HNTS-MRG2024 training set: train the identical Task-1 pipeline with and without CT pre-training, holding every other component fixed, and compare aggregated DSC for GTVp and GTVn across all five folds. If the no-pre-training model matches or beats the pre-trained model on GTVp in a majority of folds, the claimed benefit of cross-modal pre-training would not hold.","supporting_citations":[{"cited_title":"In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)","cited_arxiv_id":null,"evidence_quote":"Inspires the CNN-based cross-attention fusion that integrates the pre-RT features into the mid-RT stream."}],"review_version":1}