{"id":"88b8884e-83ec-4a2b-b13b-ef5e6ff1ec08","arxiv_id":"2411.11305","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"TP-UNet combines temporal text prompts, contrastive semantic alignment, and cross-attention with UNet, reporting improved Dice on UW-Madison and LiTS datasets, though the organ-name cue is not ablated.","lead":"A new segmentation network, TP-UNet, feeds text prompts that contain the organ name and slice position into a UNet to segment medical scans. It reports higher Dice scores on two public datasets, but the gain may come from telling the model which organ to find rather than from the temporal cue.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The temporal-contribution ablation in §IV-D removes the organ name and the timestamp together; the 2.1% UW-Madison drop may be an organ-name effect rather than temporal, because the 'w/o Temporal Information' prompt becomes 'This is an MRI of the organ'.","rationale":"We focused on the single most load-bearing threat to the central claim: the attribution of performance gains to temporal information. The paper's own ablation is the only direct evidence for this attribution, and it is confounded. The prompt template includes the organ name; the ablation replacing it with 'the organ' changes two variables at once. Because anatomical position and organ identity are correlated in these datasets, the 2.1% drop cannot be read as the timestamp's effect. This is not a dispute with consensus or a stylistic objection; it is an internally incomplete experimental design relative to the mechanism claimed. The proposed test is decisive: it varies the prompt components independently. If the organ-only prompt matches the full prompt, the temporal mechanism is unsubstantiated, and the paper would need to be revised to claim class-conditional prompting rather than temporal prompting. If the timestamp-only prompt retains the gain, the claim survives. We agree with the reader's weakest-assumption analysis and therefore recommend no change to the CONDITIONAL verdict; the condition is exactly this isolation experiment. We also note that the paper's 'SOTA' claim on LiTS is weakened by the absence of strong baselines, but that is secondary to the temporal-attribution issue. The absence of released code also hampers independent verification, but the confound is more directly load-bearing because it targets the proposed mechanism itself.","tokens_in":9492,"tokens_out":9199,"duration_ms":85733,"concrete_test":"Re-run the Table III ablation on UW-Madison with three prompt variants, same architecture and training: (a) full prompt 'This is an MRI of {organ} with a segmentation period of {N_i/N}'; (b) organ-only 'This is an MRI of {organ}'; (c) timestamp-only 'This is an MRI with a segmentation period of {N_i/N}'. If mDice(a) minus mDice(b) is less than 1% while mDice(a) minus mDice(c) is greater than 2%, the Table III 'w/o Temporal Information' drop is predominantly an organ-name artifact, invalidating the temporal attribution. On LiTS, compare 'This is a CT of the liver' versus the full prompt; if the gap is less than 1%, temporal information contributes little when the organ is fixed. Report per-class and average Dice over at least three random seeds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A defines the prompt as 'This is {an MRI / a CT} of the {organ} with a segmentation period of {N_i/N}.' Section IV-D's 'w/o Temporal Information' ablation fixes the template to 'This is an MRI / a CT of the organ' (Table III), simultaneously removing the specific organ name and the timestamp. On UW-Madison, this drops mDice by 2.1% (0.9266 to 0.9048); on LiTS, by 3.9% (0.9125 to 0.8735). Because the prompt contains both a class-conditional organ cue and the temporal fraction, a drop in this ablation does not isolate the temporal contribution. In UW-Madison, the organ named in the prompt is strongly correlated with slice position (stomach earlier, large intestine later), so removing the organ name alone could account for the drop. The baselines in Tables I and II never receive an organ-name textual cue, so the headline SOTA gains may reflect the introduction of a class-conditioning signal rather than temporal information. No ablation removes the timestamp while keeping the filled organ name (e.g., 'This is an MRI of the stomach') or removes the organ while keeping the timestamp. Thus the central claim, that temporal prompts drive the improvement, is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TP-UNet, a UNet-based medical image segmentation framework that injects temporal information through textual prompts of the form \"This is an MRI / a CT of the {organ} with a segmentation period of {N_i/N}.\" The text prompt is encoded with CLIP (LoRA-tuned) or Electra (SFT-tuned), aligned to image features via an unsupervised contrastive loss, and fused through cross-attention with UNet encoder features before decoding. Experiments on the UW-Madison gastrointestinal MRI dataset and the LiTS 2017 liver CT dataset report improved Dice/Jaccard scores over several UNet variants, with the best configuration reaching 0.9286 average Dice on UW-Madison and 0.9125 liver Dice on LiTS. Ablations remove the temporal information, the whole prompt, the semantic alignment, and the modality fusion. The authors claim state-of-the-art performance and plan to release the code after acceptance.","tokens_in":9797,"tokens_out":2770,"duration_ms":28179,"significance":"If the reported gains are attributable to temporal information, the work would be a useful and simple contribution: it formulates slice position as a continuous prompt, aligns text and image modalities with contrastive learning, and fuses the prompt through cross-attention. The automatic prompt generation (under 1 ms per slice) and the use of parameter-efficient text encoders are practical strengths, and the paper is commendably explicit about its template. However, the central claim that temporal information drives the improvement is not established by the current experiments because the temporal ablation is confounded with removal of the organ name, and the baseline comparisons omit sequence-aware and prompt-based methods. The LiTS preprocessing is also under-specified. These issues are fixable with additional experiments, so the manuscript has potential, but the current evidence is not sufficient to support the headline conclusion.","major_comments":[{"comment":"The LiTS preprocessing is under-specified: the paper states that 58,638 2D slices were divided but \"a large number of 2D slices are also redundant\" and finally 10,967 slices were selected, without giving the selection criterion. If the retained slices are chosen based on organ presence, image quality, or another property correlated with difficulty, the comparison against baselines trained on the same subset could be biased. Please state the exact filtering rule and any exclusion statistics, and confirm that all compared methods use the same filtered slices with the same train/validation/test split.","section":"IV-A.1"},{"comment":"The comparison omits the most relevant sequence-aware and prompt-based methods. The paper cites SIA-UNet [3] as a sequence-aware UNet variant but does not include it in Tables I or II, and it cites prompt-based segmentation works [5], [6] but does not compare against them. Without these baselines, the claim of \"state-of-the-art performance\" is not supported. Please add at least one sequence-aware baseline (SIA-UNet or a temporal-unaware variant) and one prompt-based segmentation baseline, or justify their exclusion with a concrete explanation.","section":"II-A and IV-B"}],"minor_comments":[{"comment":"The LiTS paragraph states \"the most significant improvement of 9.47% in the Small Intestine category,\" but LiTS in Table II only reports liver segmentation; this sentence appears to be carried over from the UW-Madison discussion and should be corrected.","section":"IV-B, last paragraph"},{"comment":"The caption says \"a lower score indicates a greater contribution of the module to the TP-UNet model,\" but the table reports Dice and Jaccard where higher is better; please rephrase the caption to state that removing a module decreases the score.","section":"IV-D, Table III caption"},{"comment":"The notation {N_i/N} is used without defining N_i; if N is the total number of slices in a scan, N_i should be explicitly defined as the slice index (or normalized position) so that the timestamp lies in [0,1].","section":"III-A"},{"comment":"The phrase \"with a segmentation period of {N_i/N}\" is awkward; consider \"at normalized slice position {N_i/N}\" or \"at timestamp {N_i/N}\" for clarity. Also, the implementation details mention initial temperature 25 and maximum 96.875 but do not explain how the temperature schedule in Eq. (1) is updated; please clarify.","section":"III-A and IV-A"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the idea of encoding slice position as a textual prompt — \"This is an MRI of the {organ} with a segmentation period of {N_i/N}\" — and feeding it into a UNet through contrastive alignment and cross-attention. That specific combination is not in the cited prompt-based or sequence-aware baselines, and the paper does show a consistent improvement over Swin UNet on two datasets. The components are standard but sensibly assembled, and the writing is clear enough to follow.\n\nThe main soft spot is exactly what the stress-test note flags. Section IV-D's \"w/o Temporal Information\" ablation fixes the prompt to \"This is an MRI / a CT of the organ,\" which drops the organ name together with the timestamp. On UW-Madison, the organ name is strongly correlated with slice position (stomach early, large intestine late), so the 2.1% drop could just as easily come from losing the class-conditional cue. The baselines never receive an organ-name prompt at all, so the headline gains over Swin UNet may reflect adding a localization/class signal rather than temporal reasoning. There is no ablation that removes only the timestamp while keeping the filled organ name, or vice versa. That is a real design flaw, and it lands on the central claim.\n\nTwo smaller issues: the LiTS preprocessing filters 58,638 slices down to 10,967 with no stated criterion, which matters because slice selection can bias the temporal distribution; and the tables report no error bars or significance tests, so the 1.3% UW-Madison gain over Swin UNet is not clearly beyond noise. The text also contains a stray line about \"Small Intestine\" improvement on LiTS, where only liver is evaluated — minor, but symptomatic of sloppy reporting in Section IV-B.\n\nTo be fair, the paper does not fit a parameter to a known answer; the temporal prompt is an input, and the normal-distribution observation is a motivating prior. The circularity concern is not the real problem here. The real problem is that the experimental design cannot distinguish temporal information from organ-name leakage.\n\nBottom line: this is a reasonable incremental idea, worth a serious referee, but the authors need to rerun the ablations with the organ name held fixed before the temporal claim can be taken at face value. I would send it to review, but with a clear request for that experiment and for code or detailed preprocessing.\n\nRecommendation: engage with it, require the missing ablation and error bars.","headline":"A plausible temporal-prompt segmentation idea with strong reported numbers, but the key ablation removes the organ name along with the timestamp, so the central claim is not yet isolated.","tokens_in":10287,"tokens_out":1090,"would_cite":false,"duration_ms":12434,"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":"TP-UNet claims that feeding a UNet a text prompt naming the imaging modality, the organ, and its normalized position in the slice stack yields state-of-the-art segmentation on two medical benchmarks.","keywords":["medical image segmentation","temporal prompts","UNet","cross-attention","contrastive learning","semantic alignment","prompt learning"],"falsifier":"Train the full TP-UNet with the same prompts but replace each timestamp $N_i/N$ with a random value in $[0,1]$, keeping organ names and modalities unchanged. If average Dice on UW-Madison and liver Dice on LiTS stay near 0.9286 and 0.9125, the temporal fraction carries no learnable signal and the paper's central claim would be refuted.","tokens_in":9341,"feed_emoji":"🩻","tokens_out":7699,"duration_ms":66750,"temperature":0.7,"pith_summary":"This paper claims that the ordering of organs across a stack of MRI or CT slices, a temporal cue that standard UNet segmenters ignore, can be injected as a short text prompt and improve segmentation accuracy. On the UW-Madison gastrointestinal MRI dataset and the LiTS 2017 liver CT dataset, the proposed TP-UNet reports the highest average Dice among compared models: 0.9286 average Dice on UW-Madison with the CLIP encoder and 0.9125 liver Dice on LiTS with Electra. The prompt names the imaging modality, the organ, and the normalized slice position $N_i/N$; a text encoder embeds it, a contrastive step aligns text and image features, and a cross-attention block carries the fused signal into the UNet decoder. A sympathetic reader would take this as evidence that cheap, automatically generated textual prompts can give segmentation models a useful positional prior without redesigning the vision backbone.","feed_headline":"Time-stamped prompts lift medical segmentation past Swin UNet","feed_subtitle":"TP-UNet tells a UNet where in an MRI/CT slice stack each organ appears, reporting a 0.9286 Dice on UW-Madison.","key_machinery":"The load-bearing mechanism is the temporal prompt template, where the normalized slice index $N_i/N$ turns the ordering of a scan into a token the text encoder can represent. Around that prompt, two learned components carry the argument: the semantic align module, a bidirectional contrastive loss that pulls matching prompt-image pairs together and pushes non-matching pairs apart, and the modality fusion block, a cross-attention over concatenated projected text and image features whose output is concatenated with the UNet's first-level skip connection before decoding. Together they implement the prior that each organ appears with roughly a normal probability over the scan's temporal interval.","core_discovery":"The paper's central claim is that temporal prompts—short texts of the form \"This is {an MRI / a CT} of the {organ} with a segmentation period of {N_i/N}\"—guide a standard UNet to state-of-the-art segmentation. The temporal signal has real content because organ occurrence over the normalized scan position roughly follows a normal distribution: stomach peaks early, small intestine in the middle, large intestine late, and liver near 0.78 in CT. TP-UNet encodes the prompt with a fine-tuned text encoder (CLIP with LoRA or Electra with SFT), aligns the text and image embedding spaces with an unsupervised contrastive loss, and fuses them through a cross-attention block whose output joins the UNet decoder's first skip connection. Reported numbers beat the strongest compared baseline, Swin UNet, by 1.3% average Dice on UW-Madison and 9.21% liver Dice on LiTS; ablations attribute 2.1% on UW-Madison to the timestamp and 5.36% on LiTS to the full prompt.","pith_inferences":["Beyond the paper: a direct ablation that swaps the timestamp for a random value while keeping organ names fixed would settle whether the temporal fraction or the organ-localization cue drives the gains.","Beyond the paper: the same prompt mechanism could be tested on other ordered imaging, such as cardiac cine series or pathology z-stacks, where structure occurrence follows a known sequence.","Beyond the paper: because the contrastive alignment is unsupervised, the framework could be pretrained on unlabeled paired scans and prompts, then adapted to new organs with fewer annotations."],"forward_implications":["On UW-Madison, replacing the prompt's timestamp with no timestamp lowers average Dice by 2.1%, so the temporal fraction itself contributes to the reported result.","On LiTS 2017, removing the temporal prompt entirely lowers liver Dice by 5.36%, showing the prompt as a whole is a sizable part of the gain.","Removing the semantic-alignment contrastive loss costs 1.01% average Dice on UW-Madison, so aligning modalities before fusing them matters.","Both text encoders work: CLIP with LoRA reaches 0.9286 average Dice on UW-Madison, and Electra with SFT reaches 0.9125 liver Dice on LiTS."],"supporting_citations":[{"why":"Supplies the UW-Madison GI-tract MRI dataset on which the main comparison is run.","marker":"[24]"},{"why":"Supplies the LiTS 2017 liver CT benchmark and its 2D slices used for the second comparison.","marker":"[25]"},{"why":"Provides the UNet architecture that TP-UNet extends with prompt fusion.","marker":"[14]"},{"why":"CLIP is the first text encoder option, fine-tuned with LoRA.","marker":"[9]"},{"why":"Electra is the second text encoder option, fine-tuned with SFT.","marker":"[12]"},{"why":"Gives the LoRA parameter-efficient fine-tuning method for CLIP.","marker":"[11]"},{"why":"Supplies the align-before-fuse principle implemented by the semantic alignment module.","marker":"[4]"},{"why":"Swin UNet is the strongest compared baseline, beaten by 1.3% on UW-Madison and 9.21% on LiTS.","marker":"[22]"}],"fun_headline_variants":["Temporal prompts guide UNet to sharper organ segmentation","Position-aware prompts boost UNet on medical scans","TP-UNet uses organ order to beat Swin UNet","Prompting UNet with scan position improves Dice scores","Time-stamped prompts: a new lever for medical segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's claimed advantage rests on the premise that the normalized slice position in the prompt, rather than the organ name, is the cue responsible for the reported gains; the ablations vary the timestamp or remove the whole prompt but never vary the organ name alone while holding the timestamp fixed.","fun_headline_variants_meta":{"raw":{"variants":["Temporal prompts guide UNet to sharper organ segmentation","Position-aware prompts boost UNet on medical scans","TP-UNet uses organ order to beat Swin UNet","Prompting UNet with scan position improves Dice scores","Time-stamped prompts: a new lever for medical segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1534,"prompt_tokens":920,"completion_tokens":614,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":535}},"tokens_in":536,"tokens_out":614,"duration_ms":7225,"temperature":1.0,"reasoning_tokens":535,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:39:46.437996+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full TP-UNet with the same prompts but replace each timestamp $N_i/N$ with a random value in $[0,1]$, keeping organ names and modalities unchanged. If average Dice on UW-Madison and liver Dice on LiTS stay near 0.9286 and 0.9125, the temporal fraction carries no learnable signal and the paper's central claim would be refuted.","supporting_citations":[{"cited_title":"Uw-madison gi tract image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the LiTS 2017 liver CT benchmark and its 2D slices used for the second comparison."},{"cited_title":"U-net: Convol utional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the UNet architecture that TP-UNet extends with prompt fusion."},{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"CLIP is the first text encoder option, fine-tuned with LoRA."},{"cited_title":"Align before fuse: Vision and language representatio n learning with momentum distillation,","cited_arxiv_id":null,"evidence_quote":"Supplies the align-before-fuse principle implemented by the semantic alignment module."}],"review_version":1}