{"id":"f7e9ecc0-03cb-4ef0-931b-1a431a782fc6","arxiv_id":"2412.07741","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A self-supervised contrastive model using intra-sweep sampling and probe-location labels retrieves neck ultrasound views with 92.3% accuracy in simulation and shows qualitative feasibility on real patient data.","lead":"This paper trains a contrastive learning model to match live ultrasound images from throat surgery to a pre-recorded scan, so a surgical assistant can be guided back to the correct scanning plane. The method uses the probe's tracked position during training to teach the model which views are similar, and reports 92.3% retrieval accuracy in a simulated test.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 92.30% retrieval accuracy is measured on augmented frames from the same sweep used as the database; the real deployment query is post-retraction US with tissue deformation, and the real-patient evidence is qualitative only, so the feasibility claim is not quantitatively established.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: simulated retrieval accuracy on same-sweep augmented queries does not establish performance under tongue-retraction deformation, and the real-patient evidence is qualitative. My independent reading of Section 4 and the Limitations paragraph confirms this is the least secure link in the argument. The paper does offer real strengths: a self-supervised formulation that avoids manual labels, an ablation (Table 3) showing that probe-location-defined positive pairs and the triplet loss each contribute to the simulated result, and a clinically motivated workflow. However, the central quantitative result is explicitly simulation-only, and the real-patient demonstration (Figures 4 and 5) lacks a measurable success criterion. The proposed concrete test uses the existing held-out patient data with anatomical landmarks to directly measure retrieval under the true deformation; it is feasible because the data were collected before and after retraction with tracking available. No additional concerns about internal consistency or baseline comparisons arose; the main risk is external validity. Because the reader already set the verdict to CONDITIONAL, my stress-test does not change the verdict, but it sharpens the condition: the feasibility claim needs quantitative real-data support before acceptance as a demonstrated clinical capability.","tokens_in":8308,"tokens_out":4463,"duration_ms":46376,"concrete_test":"On the held-out real-patient post-retraction sweeps, construct the retrieval database from pre-retraction frames and use post-retraction frames as queries. Annotate corresponding anatomical landmarks (e.g., carotid bifurcation, laryngeal cartilage, jugular vein) in both sweeps for a subset of frames, and define retrieval success as the retrieved pre-retraction frame sharing the same anatomical structure as the query within a clinically meaningful tolerance. Compute the per-patient success rate with confidence intervals and compare it with the 92.30% simulated rate; if the real success rate is substantially lower, or if the dustbin rejects most real queries, the feasibility claim should be downgraded from demonstrated to preliminary.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the method 'achieves 92.30% retrieval accuracy' (Abstract, Table 2) is load-bearing, but the evaluation protocol in Section 4 (Simulation study) does not exercise the domain shift that defines the deployment scenario. Each test query is generated by 3D affine augmentation of a mini-volume built from frames temporally adjacent to a frame in the same testing sweep, and the retrieval database is effectively drawn from the same sweep. The augmentation can simulate out-of-plane probe orientation changes, but it cannot reproduce the large non-rigid tissue deformation caused by tongue retraction that separates intra-operative post-retraction queries from the pre-retraction database. The paper's own Limitations section concedes that quantitative results are simulation-only and that real-patient evaluation is qualitative 'due to the challenges in directly comparing probe distance before and after tongue retraction.' Consequently, the reported 92.30% is at best an upper bound for the simulated affine domain gap; it does not by itself support the contribution claim of feasibility in neck US localization on real patient data. Figures 4 and 5 show qualitative correspondence, but no quantitative success rate or localization error on post-retraction queries, so a reader cannot tell whether the method would retrieve the correct anatomy under deformation. This is the weakest link in the argument: if the simulation-to-real transfer fails, the headline accuracy is not misleading per se (it is labeled simulated), but the clinically relevant feasibility claim collapses to an anecdote.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a self-supervised contrastive learning approach for ultrasound image retrieval to guide neck scanning during transoral robotic surgery. The method samples positive/negative pairs from within a single sweep using tracked probe location distances, adds a learnable dustbin threshold to reject uncertain matches, and uses a symmetric cross-entropy loss plus a triplet term. On a private dataset of 19 TORS patients, the authors report 92.30% retrieval success on simulated queries (augmented frames from the same sweep as the database), outperforming NCC, inter-sweep contrastive learning, and IVPP baselines, with a qualitative demonstration on real post-retraction ultrasound.","tokens_in":8603,"tokens_out":5546,"duration_ms":48603,"significance":"If the reported behavior transfers to the intra-operative setting, the method would provide a tracker-free, self-supervised way to localize a freehand US probe relative to a pre-operative scan, which is clinically useful. The paper contributes a private longitudinal TORS ultrasound dataset, an ablation of the loss components, and comparisons with an intra-video state-of-the-art method. The strengths include the self-supervised formulation that avoids manual labeling and the use of a probe location signal only during training. However, the central quantitative result is measured on a simulation with only affine/out-of-plane augmentation; the real-patient evidence is qualitative, and no statistical significance is reported. The claimed feasibility on real patient data is therefore not yet quantitatively established.","major_comments":[{"comment":"The 92.30% retrieval accuracy in Table 2 is computed on queries generated by 3D affine augmentation of mini-volumes taken from the same sweep as the database. This protocol cannot reproduce the non-rigid tissue deformation caused by tongue retraction that separates the intra-operative query from the pre-operative database. Because the Limitations section concedes that real-patient results are qualitative only, the headline result does not by itself support contribution (2) that the system is feasible on real patient data. Please provide a quantitative post-retraction evaluation (e.g., annotated anatomical correspondences) or explicitly scope the claim to simulated affine perturbations.","section":"4 Results and Discussion, Simulation study; Limitations"},{"comment":"In the triplet loss L_triplet = sum_i sum_j [d_ij * M_ij - (1 - d_ij) * M_ji], for a negative pair (d_ij=0) the term is -M_ji. Since M is described as the dot-product similarity between embeddings, minimizing this term maximizes the similarity of negative pairs, which is the opposite of the intended 'pull and push' behavior. Please clarify whether M denotes a distance in this equation or correct the sign.","section":"2 Methods, Eq. (3)"},{"comment":"All reported success rates are point estimates without error bars, number of random seeds, or significance tests. The improvement over the inter-sweep CL baseline is 1.63 percentage points, and the ablation increments in Table 3 are of similar magnitude, so it is unclear whether these differences are reliable. Please report patient-level or seed-level variance and, if possible, a paired significance test.","section":"Tables 2 and 3"},{"comment":"The retrieval success criterion (probe location difference < 15 mm) is the same physical quantity used to define positive/negative pairs during training. Although test sweeps are held out and queries are augmented, this alignment means the metric primarily measures recovery of probe translation under affine augmentation rather than semantic anatomical matching. Please acknowledge this and add an anatomical-correspondence-based metric, particularly for the patient study.","section":"4 Results and Discussion, metric definition"}],"minor_comments":[{"comment":"The frame rate is reported as '5 .76 ± 0.89 fps'; please fix the spacing and report the unit consistently.","section":"3 Experiments, Dataset"},{"comment":"The sentence 'If there are multiple positive pairs, the image pair with the smallest distance is positive, and the rest will be negative' is ambiguous; clarify whether all pairs below the 1 cm threshold are positives or only the nearest, as this materially changes the loss.","section":"2 Methods, sampling strategy"},{"comment":"The dustbin construction is described only qualitatively; please specify the size of the score matrix after concatenation and how the dustbin value is updated by the symmetric cross-entropy loss.","section":"2 Methods, Figure 2"},{"comment":"The proposed method and NCC both have 0.00% rejected retrieval in Table 2, so the dustbin is never exercised; please report the learned threshold value and add an experiment that triggers rejections.","section":"4 Results and Discussion"},{"comment":"The novelty claim 'first work demonstrating feasibility ... on real patient data' should be reconciled with Yeung et al. [16], who also perform localization with real freehand 2D US; please clarify the precise difference (e.g., retrieval-based vs regression-based).","section":"1 Introduction, contribution 2"},{"comment":"A data or code availability statement is missing; given the private dataset, a code release would aid reproducibility.","section":"Declarations"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable proof-of-concept, but the central quantitative evaluation is on a simulation that does not include the dominant real-world domain shift. The authors should be pushed to either provide quantitative post-retraction results or to reframe the abstract and contributions to separate the simulation claim from the qualitative feasibility claim. Please also verify the correctness of Eq. (3); as written it appears to have the opposite effect from the intended triplet loss."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe honest summary: this paper has a sensible idea — use probe-location-defined positive/negative pairs within a single US sweep to learn a retrieval embedding, plus a learnable rejection threshold — and it tests it with a clean ablation. The headline 92.30% retrieval accuracy is real for the simulated protocol, but the protocol is easier than the clinical use case, and the paper itself says real-patient evaluation is qualitative. The feasibility claim in the abstract and contributions goes beyond what the evidence supports.\n\nWhat's new: rather than treating frames as independent or all frames in a video as positive, they define positives by physical distance of the tracked probe. That is a good use of supervision that is only needed during training. The dustbin threshold is a reasonable idea for clinical rejection. The ablation (Table 3) shows probe-location labeling and the triplet term each help over plain symmetric cross-entropy. The writing is clear and the limitations section is candid.\n\nThe main soft spot is the evaluation. In the simulation study, queries are augmented frames drawn from the same sweep that forms the database. That tests robustness to affine/out-of-plane variation but not the actual deployment gap: pre-retraction database vs post-retraction intraoperative queries. The patient study shows only qualitative correspondence, with no success rate or localization error. So the 92.30% number is not an upper bound for the real task in any rigorous sense. The difference over inter-sweep CL is 1.63 percentage points with no error bars, significance testing, or per-sweep breakdown; given 5 test patients, that could be noise. Also, the final method rejects 0.00% of queries, so the dustbin threshold is not actually exercised in the reported results. That makes the rejection mechanism's value undemonstrated, even if ablated variants show some rejection.\n\nMinor: no code or data release, which makes reproducing the comparison hard. The 'first feasibility on real patient data' claim is overreach for qualitative figures.\n\nWho should read it: people working on contrastive learning for US retrieval or intra-sweep representation learning will find the sampling scheme worth considering. It deserves a serious peer review — the method is coherent, clearly described, and the clinical motivation is real — but the revision needs to either quantify the real-patient evaluation or temper the feasibility claim. If I were the editor, I'd send it out rather than desk-reject.","headline":"A sensible intra-sweep contrastive idea with an honest but limited evaluation; the 92.30% is simulation-only and the real-patient feasibility claim outruns the evidence.","tokens_in":9146,"tokens_out":2709,"would_cite":false,"duration_ms":26304,"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 self-supervised retrieval method matches neck ultrasound views to guide surgical scanning, reaching 92.3% accuracy in simulation.","keywords":["image retrieval","contrastive learning","ultrasound guidance","transoral robotic surgery","self-supervised learning","neck ultrasound","probe localization","intra-sweep representation learning"],"falsifier":"Track ultrasound probe positions in a prospective transoral robotic surgery case, build the database from the pre-retraction sweep, issue post-retraction frames as queries, and compare retrieval success (probe distance below 15 mm) against the inter-sweep baseline; if the gap narrows substantially or success falls toward the 68-90% baseline range, the simulation result does not transfer.","tokens_in":8130,"feed_emoji":"🩺","tokens_out":3488,"duration_ms":28908,"temperature":0.7,"pith_summary":"The paper aims to make freehand neck ultrasound scanning guidance practical during transoral robotic surgery by retrieving the most similar frame from a patient's pre-operative sweep rather than estimating probe pose directly. It proposes a self-supervised contrastive learning method that uses intra-sweep similarity and probe translation data recorded only during training to learn a frame representation for retrieval. On simulated queries, the method reports 92.30% retrieval accuracy and lower probe-distance error than temporal contrastive baselines, and it shows qualitative retrieval on real post-retraction ultrasound. If the approach holds, an assistant could receive probe-movement hints from 2D images alone, without an external tracking system during surgery.","feed_headline":"Retrieval hits 92.3% for guiding neck ultrasound scans","feed_subtitle":"Self-supervised matching localizes the probe from 2D images alone, with no tracker needed during surgery.","key_machinery":"Intra-sweep contrastive learning with probe-location supervision. The model samples frame batches from one sweep, labels pairs positive when the tracked probe translation distance is below 1 cm, computes a dot-product similarity matrix without normalizing embeddings, appends a learnable dustbin threshold to form a score matrix, and trains with a symmetric cross-entropy loss (CLIP-style) plus a triplet loss weighted by probe distance.","core_discovery":"The central claim is that treating frames from the same ultrasound sweep with probe-location proximity as positives, and using a cross-entropy matching loss plus a triplet loss, yields an embedding where the closest database frame to a query accurately identifies the ultrasound probe location. The paper further claims that this is the first demonstration of image-retrieval-based neck ultrasound localization on real patient data after tongue retraction, and that a learned dustbin threshold can reject uncertain retrievals so that guidance is not given when the match is unreliable.","pith_inferences":["The same intra-sweep contrastive recipe could transfer to other freehand ultrasound tasks where standard planes are not well defined, such as thyroid or musculoskeletal scanning, since it only needs a tracked training sweep.","Because the quantitative evaluation is simulation-only, a natural next test is a prospective patient study with tracked post-retraction probes to see whether the 92% accuracy survives real tissue deformation; the paper's qualitative results suggest it may degrade gracefully but do not yet quantify this.","The probe-location signal at training time could be replaced by an inertial measurement unit or image-based motion estimation, making data collection easier outside operating rooms with optical trackers."],"forward_implications":["At test time, only the 2D ultrasound image is needed; the probe location is not required, so no external tracker is needed during surgery.","The retrieved frame gives the current probe location in the pre-operative coordinate system, and the difference to the surgeon-selected target view can be turned into a probe-motion hint.","The dustbin threshold can reject low-confidence matches, which matters for clinical safety.","The method outperforms inter-sweep and temporal-weighting contrastive baselines on the simulation study, indicating that probe-location supervision captures semantic similarity better than temporal proximity alone."],"supporting_citations":[{"why":"Supplies the symmetric cross-entropy loss and the view-retrieval framing adapted for image-image alignment.","marker":"[17]"},{"why":"Provides the base contrastive optimization goal (SimCLRv2) that the intra-sweep sampling strategy modifies.","marker":"[18]"},{"why":"Serves as the state-of-the-art temporal intra-video contrastive baseline that our method is compared against.","marker":"[21]"},{"why":"Defines the retrieval success metric (probe distance below 15 mm) and the prior simulation-based landmark retrieval context.","marker":"[14]"},{"why":"Motivates the choice not to normalize embeddings before computing similarity, because feature magnitude may encode confidence.","marker":"[19]"}],"fun_headline_variants":["92.3% match accuracy for neck ultrasound guidance via intra-sweep learning","Self-supervised retrieval localizes neck ultrasound without a tracker","Intra-sweep similarity boosts ultrasound retrieval to 92.3% accuracy","First real-patient ultrasound retrieval after tongue retraction","Uncertain ultrasound matches rejected for safer neck guidance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative evaluation is performed on simulated queries created by augmenting frames from the same sweep that makes up the retrieval database, so the measured accuracy assumes that such augmented same-sweep views predict performance on real intra-operative images with tongue-retraction deformation.","fun_headline_variants_meta":{"raw":{"variants":["92.3% match accuracy for neck ultrasound guidance via intra-sweep learning","Self-supervised retrieval localizes neck ultrasound without a tracker","Intra-sweep similarity boosts ultrasound retrieval to 92.3% accuracy","First real-patient ultrasound retrieval after tongue retraction","Uncertain ultrasound matches rejected for safer neck guidance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1696,"prompt_tokens":909,"completion_tokens":787,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":700}},"tokens_in":525,"tokens_out":787,"duration_ms":6838,"temperature":1.0,"reasoning_tokens":700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:32:35.578191+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track ultrasound probe positions in a prospective transoral robotic surgery case, build the database from the pre-retraction sweep, issue post-retraction frames as queries, and compare retrieval success (probe distance below 15 mm) against the inter-sweep baseline; if the gap narrows substantially or success falls toward the 68-90% baseline range, the simulation result does not transfer.","supporting_citations":[{"cited_title":"Frontiers in Imaging3 (2024) https://doi.org/ 10.3389/fimag.2024.1416114 12","cited_arxiv_id":null,"evidence_quote":"Serves as the state-of-the-art temporal intra-video contrastive baseline that our method is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the retrieval success metric (probe distance below 15 mm) and the prior simulation-based landmark retrieval context."}],"review_version":1}