{"id":"336d398f-3e0f-457e-8d71-ad82ccb02f99","arxiv_id":"2505.22230","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GradTrack supervises weakly supervised medical segmentation with reverse-truncated gaze track attention maps and track attention modules, improving Dice over GazeMedSeg on Kvasir-SEG and NCI-ISBI.","lead":"GradTrack uses the order in which doctors look at medical images, not just where they look, to train a segmentation model with gaze-based weak supervision. It reports Dice gains over prior gaze methods on two datasets, but the gain specifically from temporal order is small and no code or gaze data is released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 3.21%/2.61% gains over GazeMedSeg are not a controlled comparison: Sec. 2.1 says the backbone is U-Net while Sec. 3.2 says nnU-Net, and the baseline's backbone is never stated, so the gain may come from the architecture rather than the gaze-track modules.","rationale":"The reader's weakest_assumption concerns the identity of the gaze annotations and data split; that is a real problem. I focus instead on the backbone, because it is stated inconsistently inside the paper and is testable without access to the private gaze data. If the nnU-Net sentence is a typo, the test is cheap; if it is not, the comparison to GazeMedSeg and to vanilla U-Net is uncontrolled. This does not prove the method is ineffective, but it means the current evidence does not establish the headline claim. The appropriate disposition remains conditional: require the authors to release code and gaze data and to include an architecture-matched GazeMedSeg baseline before accepting the superiority claim. This is the same verdict as the reader, though for a partly different reason.","tokens_in":7346,"tokens_out":12198,"duration_ms":117350,"concrete_test":"Re-run GazeMedSeg [24] in the GradTrack codebase with the same backbone actually used by GradTrack (U-Net if Sec. 2.1 is authoritative; nnU-Net if Sec. 3.2 is authoritative), using the same gaze tracks, preprocessing, data split, and optimization schedule. If GazeMedSeg's Dice then reaches roughly 80 on Kvasir-SEG and 78-80 on NCI-ISBI, the reported 3.21%/2.61% gain is not attributable to GradTrack's modules. A minimal variant, plain nnU-Net trained only with the under-activated VAM supervision, would also show whether the architecture carries the result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 2.1 says 'GradTrack adopts a U-Net model [16] as the backbone,' but Sec. 3.2 says 'We employ nnU-Net [8] as the backbone architecture,' and the training details (SGD, polynomial LR, 224x224 resizing) are the nnU-Net recipe. These two architectures are not equivalent: Table 1 reports fully supervised Dice of 82.12/80.58 for U-Net versus 88.41/82.43 for nnU-Net on the same tasks, a 6.29/1.85 gap. If GradTrack uses nnU-Net while the inherited GazeMedSeg baselines use vanilla U-Net, the 3.21/2.61 improvements and the '0.33 gap to fully supervised U-Net' could be explained by the backbone alone. The paper gives no backbone specification for GazeMedSeg [24], no baseline re-run under GradTrack's training schedule, and no code or gaze-data release. Because the central claim is an empirical superiority claim, the comparison must be architecture-matched before the GTMG/TA modules can be credited with the gain.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GradTrack, a gaze-based weakly supervised segmentation framework that augments a U-Net-style decoder with gaze track attention. The GTMG module converts fixation points and start timestamps into three reverse-truncated track attention maps (Eqs. 1-2), and TA modules predict and fuse these maps into decoder features (Eqs. 3-4), with total loss Eq. 5. On Kvasir-SEG and NCI-ISBI, GradTrack reports Dice 81.01/80.25, versus 77.80/77.64 for the inherited GazeMedSeg baseline, and 0.33 points below fully supervised U-Net on NCI-ISBI. The paper claims that temporal order of gaze fixations is the key information dividend.","tokens_in":7624,"tokens_out":6061,"duration_ms":55768,"significance":"If the empirical claims are valid, the paper makes a useful contribution by incorporating temporal order into gaze-based WSSS, and the module design is concrete and well-specified. The ablation shows that each component improves over the under-activation baseline, and the multi-dataset evaluation is a strength. However, the headline comparison is not currently architecture-matched, and the gaze annotations are not described or released in sufficient detail; the contribution is therefore conditional on a controlled re-evaluation.","major_comments":[{"comment":"Section 2.1 states that GradTrack adopts a U-Net model as the backbone, while Section 3.2 states that nnU-Net is the backbone architecture. These are not interchangeable: Table 1 reports fully supervised Dice of 82.12/80.58 for U-Net versus 88.41/82.43 for nnU-Net. Because the GazeMedSeg baseline is inherited from [24] and its backbone is not specified, the reported 3.21/2.61 Dice gains and the '0.33 gap to fully supervised U-Net' could stem from the backbone or training recipe rather than from GTMG and TA. Please reconcile the backbone statements and rerun GazeMedSeg under GradTrack's exact training schedule and backbone.","section":"Sec. 2.1 / Sec. 3.2 / Table 1"},{"comment":"Section 3.3 states that the PointSup, AGMM, Boxinst, BoxTeacher, and GazeMedSeg results are inherited from [24]. Section 3.1 says only that training annotations were 'meticulously generated by annotators,' without specifying the gaze-tracking device, fixation/timestamp format, or whether these annotations match the gaze data used by [24]. Without a shared gaze-annotation protocol and data split, the comparison is not controlled; please provide the gaze data or a precise specification and rerun the inherited baseline under the same conditions.","section":"Sec. 3.3 / Sec. 3.1"},{"comment":"The field-of-view threshold tau=0.25 is chosen on Kvasir-SEG by scanning test-set Dice across tau values (0.05-0.5) and selecting the maximum (81.01 at tau=0.25). Since tau appears directly in Eq. (2) and thresholds the attention maps used as supervision, selecting it on the test set can optimistically bias the reported results. Please use a validation split or report the selection procedure with held-out test evaluation.","section":"Fig. 4(b) / Sec. 3.2"},{"comment":"The case for reverse truncation is weak as reported. The full model achieves 81.01 +/- 0.66 while the 'w/o Sequential' variant achieves 80.28 +/- 0.02, a 0.73-point difference with overlapping standard deviations; the central claim that temporal order of fixation points is the 'information dividend' would benefit from statistical testing or a more direct comparison isolating temporal order from the GTMG/TA construction.","section":"Table 2"}],"minor_comments":[{"comment":"Equation (2) contains the typo 'can be is defined as'; replace with 'is defined as'.","section":"Eq. (2)"},{"comment":"The abstract writes 'nnUNet' while Table 1 and Section 3.2 write 'nnU-Net'; use one consistent notation throughout.","section":"Abstract / Table 1 / Sec. 3.2"},{"comment":"The phrase 'basic overly under-activation hard label supervision' is confusing; consider 'basic under-activation hard-label supervision' or define what 'overly' modifies.","section":"Sec. 2.4"},{"comment":"In Eq. (5), both L50 and L75 are weighted by the same lambda_1; if the authors intend different weights at different decoder depths, this should be stated explicitly.","section":"Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The paper's main empirical claim hinges on inherited baselines, and given the explicit contradiction in backbone descriptions, I would not accept without rerunning the baselines under matched conditions. The authors should also clarify whether the 'fully supervised U-Net' comparison in the abstract is with vanilla U-Net, because the abstract says 'such as nnUNet'."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe paper you'll want to know about: GradTrack extends GazeMedSeg by using the temporal order of gaze fixations. The new pieces are a reverse-truncation scheme on the gaze track (50%, 75%, 100%) and a track-attention module that injects these maps into decoder layers. The idea is plausible and the writing is clear. The ablations show each component helps, and the method beats the inherited GazeMedSeg numbers by about 3 points on two datasets.\n\nBut there is a load-bearing inconsistency in the paper's own text. Section 2.1 says the backbone is a U-Net; Section 3.2 says nnU-Net, with SGD, polynomial LR, and 224x224 resizing. Those training details are nnU-Net's recipe. The baselines for GazeMedSeg and others are inherited from [24] and not re-run. If GradTrack uses nnU-Net while GazeMedSeg uses vanilla U-Net, the 3.21%/2.61% gains could be largely backbone, not the gaze-track modules. The paper never states GazeMedSeg's backbone. This is not a minor typo; it changes what the headline numbers mean.\n\nSecond, the field-of-view threshold tau (0.25) appears to have been chosen by scanning the test set (Fig. 4b) with no validation split. That inflates the reported Dice. Third, the gaze annotation data and code are not released, so the 'meticulously generated' annotations cannot be checked, and the temporal contribution is small: 0.73 Dice over the non-sequential variant in Table 2. Fourth, the abstract says the method narrows the gap with fully supervised models 'such as nnUNet', but on Kvasir the gap to nnU-Net is 7.40 Dice, while the gap to U-Net is 1.11. That phrasing is misleading.\n\nWhat's genuinely good: the reverse-truncation idea is new for gaze-based WSSS, the component ablation is informative, and the method is simple to describe. The paper is not incoherent in its method; the problem is the evaluation.\n\nWho is this for? Researchers working on weakly supervised medical segmentation, especially gaze-based approaches. It deserves a serious referee, but only if the authors resolve the backbone conflict, re-run or at least justify the baselines under the same training setup, report tau selection on a validation set, and release the gaze data. Without that, the 3% gain claim is not supported.\n\nMy recommendation: send it to review but expect heavy revision. If the authors can do a clean same-backbone comparison, the paper could be a solid contribution.\n\nBest.","headline":"A novel reverse-truncation mechanism for gaze-based segmentation, but the evaluation is confounded by a U-Net/nnU-Net backbone inconsistency and test-set-tuned threshold.","tokens_in":8138,"tokens_out":3312,"would_cite":false,"duration_ms":33566,"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":"This paper claims that the order in which a physician's gaze moves across an image is a learnable supervisory signal that brings weakly supervised segmentation within 0.33 Dice of fully supervised U-Net.","keywords":["gaze supervision","weakly supervised segmentation","medical image segmentation","eye tracking","temporal gaze order","gaze track attention","polyp segmentation","prostate MRI"],"falsifier":"Run GradTrack's training protocol on a public gaze dataset with fixation timestamps shuffled but all other signals unchanged; if Dice does not drop appreciably below the reported 81.01, the temporal-order claim is not doing the work attributed to it. Equivalently, retrain the 'w/o Sequential' variant on NCI-ISBI and check whether the gap generalises beyond Kvasir-SEG.","tokens_in":7157,"feed_emoji":"👁","tokens_out":6158,"duration_ms":56832,"temperature":0.7,"pith_summary":"This paper argues that gaze annotations for medical image segmentation carry more information than previous methods use: the order in which a physician looks at regions, not just where and how long, encodes the diagnostic reasoning. To exploit that order, GradTrack builds a gaze track from fixation points and their start timestamps, then truncates the track in reverse to produce several soft attention maps. These maps supervise the decoder at multiple depths, giving the model a progressive, staged view of the diagnostic process. On polyp and prostate MRI datasets, the method reports Dice gains of 3.21 and 2.61 points over the prior gaze-based approach and lands within about a point of fully supervised U-Net.","feed_headline":"Gaze fixation order lifts segmentation close to full supervision","feed_subtitle":"A doctor's scan sequence, not just a heatmap, gives weakly supervised models within 0.33 Dice of U-Net on prostate MRI.","key_machinery":"The load-bearing object is the reversed, truncated gaze track. Given fixation points with start timestamps, GradTrack constructs the gaze track $T$, then produces $T_{50\\%}$, $T_{75\\%}$, $T_{100\\%}$ by removing the earliest fixations in reverse order. For each truncated track it computes a distance map $D^r(\\mathbf{p}) = \\min_{t\\in T^r}\\|\\mathbf{p}-t\\|_2$ and converts it to a soft attention map $G^r = e^{-D^r/\\beta}\\cdot \\mathbb{I}(e^{-D^r/\\beta}>\\tau)$. These maps supervise predictions at decoder blocks 2, 4, and 6 through a weighted cross-entropy loss, providing staged prior information so that later decoder stages are guided by fuller gaze context. The Track Attention module fuses the predicted foreground map into the decoder features with a stop-gradient, letting gaze track knowledge shape segmentation without backpropagating through the fusion.","core_discovery":"GradTrack's central claim is that the temporal order of gaze fixations is a usable supervisory signal for weakly supervised segmentation, and that reversing that order during training stabilises learning. The Gaze Track Map Generation module converts raw gaze data into a continuous track, then forms three track attention maps by keeping the last 50%, 75%, and 100% of the fixation sequence. A distance-based exponential decay turns each truncated track into a soft attention map that supervises the decoder; the Track Attention modules inject the learned maps back into the feature stream. The paper reports that this scheme outperforms GazeMedSeg by 3.21% Dice on Kvasir-SEG and 2.61% on NCI-ISBI, and comes within 1.11% and 0.33% of a fully supervised U-Net on the two datasets.","pith_inferences":["Beyond the paper, the same ordering signal might transfer to other sequential human annotations, such as mouse-trace or touch-screen scribbles with timestamps, turning any time-stamped annotation trace into this style of staged supervision.","The 0.73 Dice gain of the full design over the non-sequential variant (81.01 vs 80.28) suggests that temporal order contributes only modestly on this benchmark; a larger gain would need datasets where diagnostic scanning order varies more between images.","A testable extension would replace the hand-set truncation ratios of 50/75/100% with learned or per-image ratios, letting the model decide how much of a physician's early scanning to trust.","If the gaze annotations were released, re-running the exact protocol on a second institution's data would clarify whether the 3.21% and 2.61% gains are tied to the specific annotators or generalise across gaze patterns."],"forward_implications":["Gaze-based weakly supervised segmentation can close most of the gap to full supervision: on NCI-ISBI the gap to U-Net is 0.33 Dice, so gaze annotation may serve as a practical substitute for pixel-level labels.","Multi-level supervision from truncated tracks gives increasing benefit at deeper decoder stages; placing guidance in the encoder hurts, so decoder-only integration is the design to keep.","Reverse truncation is intentionally conservative: it biases predictions toward under-activation, trading recall of uncertain regions for precision against noise.","The component ablation on Kvasir-SEG shows a cumulative climb from 63.63 Dice with only the under-activated VAM to 81.01 for the full design, so each module adds measurable value.","The field-of-view threshold is a sensitive knob: 0.25 works best, small values inject redundant noise and large values starve the model, so the method needs calibration per dataset."],"supporting_citations":[{"why":"The gaze-based WSSS method GradTrack is compared against; its reported numbers are inherited as the baseline.","marker":"[24]"},{"why":"Supplies the nnU-Net backbone and the training recipe (SGD, polynomial decay, preprocessing) used by GradTrack.","marker":"[8]"},{"why":"Kvasir-SEG polyp segmentation dataset used for evaluation.","marker":"[9]"},{"why":"NCI-ISBI prostate MRI dataset used for evaluation.","marker":"[2]"},{"why":"D-CRF refinement used to produce the under-activation hard masks that provide the VAM supervision.","marker":"[12]"},{"why":"Fully supervised U-Net baseline whose performance on the two datasets defines the gap GradTrack narrows.","marker":"[16]"}],"fun_headline_variants":["Gaze track order lifts weak supervision close to full models","Doctor's gaze sequence narrows gap to full supervision","Gaze track attention boosts weak segmentation by 3.2% Dice","Temporal gaze order gives near-full supervision without labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method presumes that every training image has gaze annotations with reliable start timestamps, and that the comparison numbers inherited from GazeMedSeg were produced on the same annotations and data split.","fun_headline_variants_meta":{"raw":{"variants":["Gaze track order lifts weak supervision close to full models","Doctor's gaze sequence narrows gap to full supervision","Gaze track attention boosts weak segmentation by 3.2% Dice","Temporal gaze order gives near-full supervision without labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000531,"raw_usage":{"total_tokens":2535,"prompt_tokens":898,"completion_tokens":1637,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":1569}},"tokens_in":514,"tokens_out":1637,"duration_ms":13889,"temperature":1.0,"reasoning_tokens":1569,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:11:48.956174+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GradTrack's training protocol on a public gaze dataset with fixation timestamps shuffled but all other signals unchanged; if Dice does not drop appreciably below the reported 81.01, the temporal-order claim is not doing the work attributed to it. Equivalently, retrain the 'w/o Sequential' variant on NCI-ISBI and check whether the gap generalises beyond Kvasir-SEG.","supporting_citations":[{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"The gaze-based WSSS method GradTrack is compared against; its reported numbers are inherited as the baseline."},{"cited_title":"Advances in neural information processing systems24(2011)","cited_arxiv_id":null,"evidence_quote":"D-CRF refinement used to produce the under-activation hard masks that provide the VAM supervision."}],"review_version":1}