{"id":"94df6c10-b8e8-4984-a9a6-a8c2dc0aeb72","arxiv_id":"2502.03781","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Using doctor gaze heatmaps to align features and reweight segmentation losses improves target-domain cardiac ultrasound segmentation over four unsupervised baselines.","lead":"This paper adds doctor eye-gaze heatmaps to a teacher-student domain adaptation pipeline for cardiac ultrasound segmentation. The authors report higher segmentation accuracy than four unsupervised baselines on a CAMUS-to-HMC-QU transfer, but the gaze collection protocol and code are not released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gaze leakage/apples-to-apples: SOTA claim depends on unstated target-side gaze availability; paper never specifies gaze split or protocol.","rationale":"The paper's contribution is a gaze-assisted adaptation framework, so the availability and provenance of gaze is not a reproducibility footnote; it is the condition that makes the central comparison meaningful. The authors compare against methods that use no target-side labels or gaze. If the gaze heatmaps were collected while a doctor viewed the HMC-QU target images used in training, then GAHCDA receives privileged information at adaptation time; the comparison to FDA (a Fourier-based unsupervised method) is no longer a comparison of domain-adaptation algorithms but of problem settings. The paper's one-sentence gaze description (Section III-A) does not rule this out. The internal formula for GBL (Eq. 2) adds a second unverified element: a gaze weight multiplied into the probability inside a log is not a standard loss, so the ablation gain attributed to GBL might be an artifact of an ill-defined objective. Both issues point to the same conclusion: the reported 76.14 DSC is not sufficient evidence for the state-of-the-art claim. A concrete, minimal fix is a controlled gaze-split experiment; if the authors can show that source-only gaze or a disjoint gaze session preserves the DSC gain while matching all baseline training data, the central claim would be materially strengthened. Until then, the verdict should remain rejection.","tokens_in":10270,"tokens_out":4549,"duration_ms":44588,"concrete_test":"Ask the authors for the exact gaze-recording split and run a control: remove target-image gaze from GAA and GBL by training with gaze heatmaps built only from source-domain (CAMUS) gaze recordings, and separately with target gaze maps randomly shuffled across target images. If the shuffled-gaze run reproduces the reported 76.14 DSC, gaze content is not the cause; if the source-gaze run drops toward the 70.14 w/o-DA level or the 71.05 FDA level, the reported gain depends on seeing the target images during gaze collection. Also verify that no HMC-QU test image was included in the gaze session; if it was, report the numbers without that overlap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim 'attains state-of-the-art performance' (Section III-B.1) rests on a comparison of GAHCDA against unsupervised baselines, but the method is trained with doctor gaze heatmaps in both GAA (Eq. 1, via cross-attention on teacher features) and GBL (Eq. 2, via weighted cross-entropy). Section III-A describes the gaze data in one sentence: 'A Tobii Eye-Tracker 5 record the doctor's gaze.' It does not state how many doctors were recorded, how gaze coordinates were converted to heatmaps, whether gaze was recorded on CAMUS, HMC-QU, or both, and, crucially, whether the gaze recordings were made on the same target-domain images used for student training and pseudo-label generation. If gaze was recorded on the HMC-QU training images, GAHCDA is not an unsupervised domain-adaptation method: it injects a target-side privileged signal into feature alignment and loss weighting, so the reported +5.09 DSC over FDA is not an apples-to-apples comparison. If gaze was recorded on the test images, the 76.14±3.2 figure may reflect leakage rather than adaptation. Either way, the current text leaves the core comparison uncontrolled. A secondary internal issue compounds this: Eq. (2) multiplies the predicted probability by a gaze weight w_i inside the logarithm, which is not a valid cross-entropy term when w_i > 1 or w_i = 0; this is the very term whose ablation claims +3.39 DSC.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GAHCDA, a gaze-assisted human-centric domain adaptation framework for cardiac ultrasound image segmentation. It contains two modules: Gaze Augment Alignment (GAA), which uses cross-attention between teacher features and gaze heatmaps to align student features via MSE loss, and Gaze Balance Loss (GBL), which weights segmentation outputs by a gaze-derived mask before computing cross-entropy against pseudo-labels. The method is evaluated by transferring from CAMUS to HMC-QU, reporting DSC of 76.14±3.2% and ASSD of 6.976±1.4 mm, claimed as state-of-the-art. Ablations attribute +1.53 DSC to GAA and +3.39 DSC to GBL relative to no adaptation.","tokens_in":10561,"tokens_out":3659,"duration_ms":36761,"significance":"The idea of injecting clinician gaze as human cognitive guidance into medical-image domain adaptation is clinically motivated and potentially novel. If the method were shown to work without target-test leakage and with a fair comparison to unsupervised baselines, the work would be of interest to the medical imaging community. However, the current manuscript does not provide this evidence: the gaze recording protocol is almost entirely unspecified, the target-side gaze signal may be a privileged input that makes the comparison to unsupervised baselines unfair, Eq. (2) defines a mathematically invalid loss, and the ablation study lacks error bars or statistical tests. The paper ships no code, no machine-checked derivations, and no reproducible protocol, so the quantitative claims rest entirely on the incomplete experiment description.","major_comments":[{"comment":"The gaze collection protocol is unspecified to a degree that undermines the central comparison. Section III-A says only 'A Tobii Eye-Tracker 5 record the doctor's gaze' with 5 Hz sampling. It does not state how many doctors participated, whether calibration was performed, how raw gaze coordinates were converted to the heatmaps used in Fig. 2 and Fig. 3, which dataset images were viewed, or--critically--whether the gaze recordings were made on the same target-domain images used for student training and pseudo-label generation. Since GAA (Eq. 1) and GBL (Eq. 2) both use these heatmaps on target-domain data, if gaze was recorded on the HMC-QU training set, the comparison against unsupervised baselines such as FDA is not apples-to-apples: GAHCDA uses a privileged target-side signal. If gaze was recorded on the test images, the reported 76.14±3.2 DSC may reflect leakage. Either way, the current text does not establish a controlled experimental setting, and the claim 'attains state-of-the-art performance' in Section III-B.1 is unsupported. The authors must disclose the full recording and heatmap generation protocol, state the train/test split of the gaze data, and, if gaze is available only on target data, reframe the method as privileged-information domain adaptation and compare against appropriate baselines.","section":"Section III-A and II-B/II-C"},{"comment":"Eq. (2) is not a valid cross-entropy loss. Writing L_GB = -(1/N) Σ [y_i log(w_i ŷ_i) + (1 - y_i) log(1 - w_i ŷ_i)] places the gaze weight w_i inside the logarithm. For w_i = 0, the term y_i log(0) is undefined when y_i = 1, and for w_i > 1 the argument w_i ŷ_i can exceed 1, giving a non-positive 'log-probability' and producing a per-sample loss with incorrect sign and curvature. Since the ablation in Section III-C.1 attributes +3.39 DSC to GBL on the basis of this term, the loss formulation is load-bearing for the central claim. The gaze weight should instead multiply the loss outside the logarithm, e.g., -w_i [y_i log ŷ_i + (1 - y_i) log(1 - ŷ_i)], or be incorporated through a different mechanism such as a spatial constraint. The authors should correct the formula and rerun the affected experiments.","section":"Section II-C, Eq. (2)"},{"comment":"The ablation study is reported only as bar charts without numerical values, error bars, or significance tests. Given that the reported DSC standard deviations for the full method are 3.2 percentage points, the claimed improvements of +1.53 (GAA) and +3.39 (GBL) could be within the run-to-run noise; the paper provides no repeated-run statistics for the individual ablations. The hyperparameter analysis in Fig. 5(b) and 5(c) states that performance 'continuously improves' as λ_gaa and λ_gb increase, but no exact numbers or variance are given, and the range tested is unclear. These results directly support the claim that both modules are effective, so the manuscript needs error bars, the exact table of ablation values, and a statement of the number of independent runs.","section":"Section III-C.1 and III-C.2"},{"comment":"The claim of state-of-the-art performance is not supported by the baseline selection. Table I compares GAHCDA with only four baselines: SIFA, CycleGAN, VAE, and FDA. Modern unsupervised domain adaptation methods for medical image segmentation, such as DaFormer (cited as [29] but not evaluated) and other recent self-training/feature-alignment methods, are missing. Without these comparisons, the phrase 'state-of-the-art' in Section III-B.1 overstates what the experimental evidence demonstrates. The authors should include recent UDA baselines and ideally also a gaze-free variant of their own architecture to isolate the effect of the gaze signal from the effect of the added attention module.","section":"Section III-B.1, Table I"}],"minor_comments":[{"comment":"The text says the method achieves '30% DSC improvements' over SIFA, but the actual difference is 30 percentage points (76.14 vs. 46.14), which is an approximately 65% relative improvement; please correct the phrasing.","section":"Section III-B.1"},{"comment":"The sentence 'A Tobii Eye-Tracker 5 record the doctor's gaze' is grammatically incomplete and imprecise; specify the exact device model and the subject population.","section":"Section III-A"},{"comment":"The heading 'GAHCDA Framwork' contains a typo; it should be 'Framework'.","section":"Section II-A"},{"comment":"The embedded text in the figures appears as garbled glyphs in the PDF rendering, making the architecture diagrams difficult to interpret; please provide clean vector figures with readable labels.","section":"Figures 2 and 3"},{"comment":"The construction of the 'regularized gaze heatmap' is not described anywhere; at minimum define the regularization (e.g., Gaussian smoothing, normalization) used to produce the weight mask w.","section":"Section II-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a very early draft: the gaze protocol, the loss formula, and the experimental statistics all need substantial revision before the central claim can be evaluated. I would not rule out the contribution on novelty grounds, but the current framing as unsupervised domain adaptation is untenable given the target-side gaze signal; the authors should reposition the work as privileged-information or gaze-assisted domain adaptation and redo the comparisons. If the authors can provide a controlled protocol and correct the loss, I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper proposes gaze-assisted domain adaptation for cardiac ultrasound segmentation. The idea is genuinely new: using doctor gaze heatmaps both as a feature-alignment target and as a loss-weighting mask. That's a plausible way to inject human prior into UDA, and the reported gains over the four baselines are consistent with the ablation, so I don't think the numbers are fabricated.\n\nThe central evaluation, though, is uncontrolled. The paper never says whether the gaze was recorded on the target training frames, the target test frames, or only on source-domain images. If it was recorded on the target training set, the method is not unsupervised: it has access to a per-image signal from the target domain that the baselines don't, so the +5 DSC over FDA is not apples-to-apples. If it was recorded on test frames, that's leakage. Either way, the SOTA claim in Section III-B.1 is not supported as stated.\n\nThe baselines are old (SIFA, CycleGAN, FDA, and a shape-modeling VAE); DAFormer is cited but not compared. There are no significance tests and no error bars on the ablation. The GBL loss in Eq. (2) is odd — multiplying the predicted probability by a weight inside the log is not standard weighted cross-entropy, and the paper doesn't justify it. The gaze protocol is one sentence: no number of doctors, no heatmap construction, no train/test split. Given the central claim depends on that information, this is a load-bearing gap.\n\nWhat the paper does well: the ablation is internally consistent, the qualitative feature-map analysis is informative, and the idea is concretely different from the cited UDA literature. If the authors could clarify the gaze collection and re-benchmark against modern UDA with a proper protocol, the contribution could be real.\n\nAs it stands, the paper isn't ready to be accepted, but it deserves a serious referee. I'd send it to review with a request for major revision, and I'd want the reviewer to push hard on the gaze split.","headline":"Interesting gaze-guided UDA idea, but the evaluation leaves the source of the gain uncontrolled and the reported SOTA claim doesn't hold as written.","tokens_in":11102,"tokens_out":2951,"would_cite":false,"duration_ms":27196,"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":"The paper claims that gaze heatmaps from doctors, injected through a cross-attention feature alignment and a gaze-weighted loss, lift cardiac ultrasound segmentation from 70.14% to 76.14% Dice under domain adaptation, beating GAN- and…","keywords":["cardiac ultrasound","domain adaptation","image segmentation","gaze heatmap","human-centric learning","teacher-student self-training","cross-attention","pseudo-labels"],"falsifier":"Repeat the CAMUS→HMC-QU experiment with target-image gaze heatmaps replaced by smooth random fixation maps matched in spatial extent, or by saliency maps from a generic model, keeping all other settings identical: if Dice stays near 76.14%, the reported gain comes from any spatially concentrated attention prior rather than from doctors' domain knowledge. Also rerun the comparison while holding gaze out of the exact test images and reporting train/test separation: if the margin over FDA shrinks, the baseline comparison was not fully unsupervised.","tokens_in":10069,"feed_emoji":"👁️","tokens_out":7015,"duration_ms":63503,"temperature":0.7,"pith_summary":"Cardiac ultrasound segmentation models lose accuracy when they move between hospitals or machines, and existing fixes either generate distorted synthetic images (GAN-based adaptation) or trust unreliable guessed labels (self-training). This paper's proposal is to record where a doctor looks while reading the target-domain images and feed those gaze heatmaps into the adaptation loop. Gaze guides a cross-attention feature-alignment step (Gaze Augment Alignment, GAA) and a loss-reweighting step (Gaze Balance Loss, GBL) that concentrates training on structures the doctor actually scans. On the CAMUS-to-HMC-QU transfer, the method reports 76.14% Dice similarity coefficient versus 70.14% with no adaptation and 71.05% for the Fourier-domain FDA baseline, against a 90.23% fully supervised upper bound. The claim is that human gaze supplies label-free cross-domain guidance that narrows the gap to supervised training by roughly six Dice points.","feed_headline":"Doctor gaze maps lift heart ultrasound segmentation by 6 DSC points","feed_subtitle":"Gaze heatmaps replace labels: cardiac ultrasound segmentation jumps from 70.14 to 76.14 Dice without target annotations.","key_machinery":"The load-bearing object is the doctor gaze heatmap: a 2D map of where the eye tracker sampled during reading, treated as a human-cognition prior that transfers across domains. An extractor converts it to $f_G$; GAA then uses cross-attention (query from the gaze feature, keys and values from the teacher's feature map) to assemble $f_{GA}$, a feature set shared between source and target, and pulls the student features toward it with an MSE loss. GBL regularizes the same heatmap into a weight mask $w$ and multiplies it into the student output before cross-entropy against pseudo-labels, so under-segmented gazed-at regions receive larger gradients and over-segmented distractors are suppressed. The teacher trained on the source domain initializes the student, and the pseudo-labels, the gaze heatmaps, and the two gaze losses together drive adaptation.","core_discovery":"The central claim is that a doctor's gaze heatmap, extracted from eye-tracking on target-domain ultrasound images, carries the cross-domain “human cognition general features” that standard unsupervised domain adaptation lacks, and that two gaze-driven modules turn that signal into better segmentation. GAA uses the gaze heatmap to produce a gaze feature $f_G$, then applies cross-attention with the teacher encoder's features $f_T$ as key and value and $f_G$ as query, forming $f_{GA} = \\mathrm{Concat}(f_T, \\mathrm{Softmax}(f_G f_T^T / \\sqrt{d}) f_T)$; the student's features are aligned to $f_{GA}$ with an MSE loss. GBL turns the gaze heatmap into a weight mask $w$ and computes cross-entropy between the student's gaze-weighted output and the teacher's pseudo-labels, so the loss favors complete segmentation in areas the doctor actually inspected. With both modules, teacher–student training on CAMUS as source and HMC-QU as target reaches 76.14±3.2% DSC and 6.976±1.4 ASSD, which the paper reports as state-of-the-art for this task and as a 5.09-point improvement over FDA and a 30-point improvement over SIFA.","pith_inferences":["The reported comparison is only apples-to-apples if gaze heatmaps were recorded on target-domain training images and withheld from the test set; if gaze was collected on test images, part of the 6-point gain could be a form of target-side supervision that the unsupervised baselines do not get.","Because the eye tracker samples only five times per second, the construction of gaze heatmaps from sparse fixation coordinates (interpolation, smoothing, choice of doctor) is a design choice that can dominate results; ablating gaze quality, number of doctors, and calibration would clarify how much of the gain comes from gaze content versus its implementation.","A testable extension is to replace doctor gaze with saliency maps or smooth random heatmaps of similar shape: if DSC remains near 76.14%, the method is using a generic attention prior, whereas if performance drops, the domain-specific human guidance is doing the work.","If the mechanism transfers, the same GAA/GBL recipe could be applied to endoscopy, fundus imaging, or other expert-reviewed modalities where doctors already look at images during diagnosis, without requiring new annotations."],"forward_implications":["With gaze heatmaps available, adaptation on CAMUS→HMC-QU reaches 76.14% DSC, a 5.28-point gain over no adaptation and a 5.09-point gain over FDA.","Adding GAA alone adds 1.53 points and GBL alone adds 3.39 points, so the two gaze modules are complementary rather than redundant.","The gaze-weighted loss is the larger single contributor, supporting the paper's claim that it counteracts over- and under-segmentation caused by incomplete pseudo-labels.","If replicated, the result suggests that a low-cost eye tracker can substitute for part of the annotation effort in clinical domain adaptation, since gaze is recorded passively while a doctor reads images.","The same teacher–student recipe with gaze guidance could apply to other cardiac views or to any imaging domain where expert reading behavior is available."],"supporting_citations":[{"why":"supplies the CAMUS source-domain dataset of 9964 cardiac ultrasound images used to train the teacher.","marker":"[32]"},{"why":"supplies the HMC-QU target-domain dataset of 2349 images used for adaptation and evaluation.","marker":"[33]"},{"why":"provides the cross-attention mechanism that GAA uses to fuse gaze features with teacher features.","marker":"[30]"},{"why":"is the prior feature-alignment module that GAA modifies by adding gaze-guided human cognition features.","marker":"[31]"},{"why":"serves as a GAN-based domain-adaptation baseline whose comparison result (46.14% DSC) supports the claimed improvement.","marker":"[7]"},{"why":"supplies the CycleGAN baseline used to illustrate mode collapse in GAN-based adaptation.","marker":"[9]"},{"why":"provides the Fourier-domain adaptation baseline (FDA, 71.05% DSC) that GAHCDA claims to beat by 5.09 points.","marker":"[21]"},{"why":"provides the shape-modeling self-training baseline (VAE, 54.16% DSC) contrasted with the gaze-guided self-training approach.","marker":"[34]"}],"fun_headline_variants":["Gaze heatmaps lift cardiac ultrasound segmentation to 76.14 Dice","Gaze-guided domain adaptation boosts cardiac ultrasound segmentation","Doctor gaze replaces labels for cardiac ultrasound domain adaptation","Eye-tracking heatmaps improve heart ultrasound segmentation by 6 DSC","Gaze-assisted DA hits 76.14 Dice on cardiac ultrasound without labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that doctor gaze heatmaps are available on the target-domain images used for adaptation and that those eye movements reliably point at the structures to segment; the paper does not report the number of doctors, calibration procedure, gaze-to-heatmap construction, or train/test separation at gaze-collection time.","fun_headline_variants_meta":{"raw":{"variants":["Gaze heatmaps lift cardiac ultrasound segmentation to 76.14 Dice","Gaze-guided domain adaptation boosts cardiac ultrasound segmentation","Doctor gaze replaces labels for cardiac ultrasound domain adaptation","Eye-tracking heatmaps improve heart ultrasound segmentation by 6 DSC","Gaze-assisted DA hits 76.14 Dice on cardiac ultrasound without labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000597,"raw_usage":{"total_tokens":2829,"prompt_tokens":1020,"completion_tokens":1809,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":1737}},"tokens_in":636,"tokens_out":1809,"duration_ms":14276,"temperature":1.0,"reasoning_tokens":1737,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:45:29.733480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the CAMUS→HMC-QU experiment with target-image gaze heatmaps replaced by smooth random fixation maps matched in spatial extent, or by saliency maps from a generic model, keeping all other settings identical: if Dice stays near 76.14%, the reported gain comes from any spatially concentrated attention prior rather than from doctors' domain knowledge. Also rerun the comparison while holding gaze out of the exact test images and reporting train/test separation: if the margin over FDA shrinks, the baseline comparison was not fully unsupervised.","supporting_citations":[{"cited_title":"Deep learning for segmentation using an open large-scale d ataset in 2d echocardiography,","cited_arxiv_id":null,"evidence_quote":"supplies the CAMUS source-domain dataset of 9964 cardiac ultrasound images used to train the teacher."},{"cited_title":"Early myocardial infarction detection over multi-view ec hocardiogra- phy,","cited_arxiv_id":null,"evidence_quote":"supplies the HMC-QU target-domain dataset of 2349 images used for adaptation and evaluation."},{"cited_title":"Unsup ervised domain adaptation for medical image segmentation by select ive entropy constraints and adaptive semantic alignment,","cited_arxiv_id":null,"evidence_quote":"is the prior feature-alignment module that GAA modifies by adding gaze-guided human cognition features."},{"cited_title":"Synergis tic im- age and feature adaptation: Towards cross-modality domain adaptation for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"serves as a GAN-based domain-adaptation baseline whose comparison result (46.14% DSC) supports the claimed improvement."},{"cited_title":"Unpaired im age-to-image translation using cycle-consistent adversarial networks ,","cited_arxiv_id":null,"evidence_quote":"supplies the CycleGAN baseline used to illustrate mode collapse in GAN-based adaptation."},{"cited_title":"Fda: Fourier domain adaptation f or semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"provides the Fourier-domain adaptation baseline (FDA, 71.05% DSC) that GAHCDA claims to beat by 5.09 points."},{"cited_title":"Gaze-Assisted Human-Centric Domain Adaptation for Cardiac Ultrasound Image Segmentation","cited_arxiv_id":"2502.03781","evidence_quote":"provides the shape-modeling self-training baseline (VAE, 54.16% DSC) contrasted with the gaze-guided self-training approach."}],"review_version":1}