{"id":"d881fabc-6f9a-4ecf-b9d0-fcef97f11310","arxiv_id":"2412.02993","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EchoONE, a SAM-based model with prototype-composed dense prompts and a local-feature fusion branch, segments multiple echocardiography planes in one model and reports state-of-the-art Dice scores on internal and external datasets.","lead":"EchoONE is a single deep-learning model that segments the heart's left ventricle, left atrium, and myocardium across multiple ultrasound viewing planes, replacing the usual practice of training one model per plane. If it works as claimed, it could simplify building and deploying echocardiography AI in hospitals and cut development repetition.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The external HMCQU result that supports the generalization claim is computed against algorithmically generated LV cavity masks, so the LV component of the reported Dice may not measure real segmentation.","rationale":"The paper's central claim is that a single SAM-based model with PC-Mask and LFFA is the first uniform multi-plane echo segmenter and is consistently state-of-the-art. The most fragile link is not the architecture but the reference labels used to measure performance on external data. For HMCQU, the only annotation is MYO, and the LV cavity is synthesized by landmark detection and filling. Since Table 4's HMCQU mDice is partly computed against that synthetic reference, the external generalization claim is not yet validly supported. This is load-bearing because external validation is the strongest evidence for generalization, and HMCQU is called out explicitly in Section 4.3.2. The reader's version of this concern overstates its scope: Table 1 lists A, B, and D rows as (2,3) meaning LV plus MYO, not MYO-only, so the synthetic-label problem does not contaminate the internal tables. That is why agreement is partial rather than full. A secondary concern is that the number of PC-Mask clusters K is never reported and it is not stated whether cluster centers are computed only on the training split; if centers include test images, the dense prompt would leak. However, the concrete and textually grounded weakness is the HMCQU LV reference. The proposed check, expert LV traces on a subset of HMCQU and recomputation of LV Dice for EchoONE versus SAMUS, would settle whether the external margin survives. Since the reader already conditioned the verdict on validating synthetic labels, the verdict remains CONDITIONAL, so no adjustment is needed.","tokens_in":13669,"tokens_out":12039,"duration_ms":120076,"concrete_test":"Take a random subset of HMCQU images (e.g., 100 frames) and have an expert echocardiographer trace the true LV endocardial contour. Recompute LV Dice and the HMCQU mDice for EchoONE and for the strongest competitor (SAMUS) using these expert contours, while keeping MYO evaluation unchanged. If EchoONE's HMCQU margin over SAMUS disappears or reverses, the external SOTA claim rests on the synthetic fill; if the margin holds, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 and Figure 5(b) describe generating LV cavity ground truth for images that have only MYO annotations by landmark detection and cavity filling. Table 1 shows this applies to HMCQU, whose only annotation is MYO. Those generated masks are used as the reference for the HMCQU row of Table 4, which the paper explicitly cites as evidence of generalization ('our method still delivers a Dice score of 73.94%'). If the landmark/fill heuristic is biased, e.g., producing smooth convex fills that ignore trabeculations, the valve plane, or the true endocardial boundary, then a model that reproduces the fill gets inflated LV Dice without segmenting the real LV. The paper provides no validation of the generated labels against expert LV contours, and no separate LV-only result on HMCQU. Note that the reader's claim that 'parts of A, B, D provide only MYO annotations' is not supported by Table 1; the A, B, D rows with superscript (2,3) already contain LV annotations. The contamination is therefore concentrated on the external HMCQU result, but that result is load-bearing for the 'consistently state-of-the-art' claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"EchoONE proposes a SAM-based segmentation architecture with two new components: PC-Mask, which generates dense mask prompts by linearly combining cluster prototypes in a ResNet34 latent space without explicit plane labels at inference, and LFFA, a learnable CNN branch with cross-branch attention that fuses local features into the mask decoder. The authors unify heterogeneous annotation protocols, including filling LV cavities from myocardium-only annotations, and train one model on four plane categories from five datasets. They report plane-wise, structure-wise, dataset-wise, and external results showing consistent gains over U-Net, DeepLabV3+, SwinUNet, H2Former, and six SAM-based competitors, with ablations attributing the gains to PC-Mask and LFFA.","tokens_in":13879,"tokens_out":8512,"duration_ms":75223,"significance":"The practical goal is worthwhile, and the evaluation breadth across five internal and two external datasets is a strength. If the results are reliable, EchoONE would be a useful single-model alternative to plane-specific echocardiography segmentation pipelines, and the PC-Mask idea of composing dense prompts from prior prototypes is a transferable contribution. The manuscript does not yet provide code or checkpoints, and the 'first uniform model' claim is stated more strongly than the related-work discussion supports. The central results are plausible, but the external evidence is compromised by the label-generation issue described below, and no uncertainty quantification is provided for the small margins over SAMUS.","major_comments":[{"comment":"Table 1 shows that HMCQU carries only MYO annotations, and §4.1 states that for images with only MYO annotations the LV cavity mask is generated by detecting landmarks and filling the cavity. The HMCQU row of Table 4 is then used in §4.3.2 as evidence of external generalization ('our method still delivers a Dice score of 73.94%'). Because the LV component of that score is computed against algorithmically filled masks rather than expert LV contours, the reported score does not establish segmentation accuracy for the LV in HMCQU. This is load-bearing because HMCQU is one of only two external datasets supporting the generalization claim. The authors should either validate the generated labels against manual LV contours, report HMCQU results restricted to structures with native annotations (MYO), or clearly mark the LV result as a proxy. The description of the landmark-detection and filling procedure is also too brief to assess potential bias.","section":"§4.1, Table 1, Table 4"},{"comment":"The headline claim is 'consistently state-of-the-art', but all reported numbers are single-run point estimates. The margins over SAMUS are, for example, +2.16, +1.73, +1.91, and +2.34 mDice for 2CH, 3CH, 4CH, and PSAX in Table 2, and +0.80 mDice on EchoNet-Dynamic in Table 4. With no error bars, no repeated runs, and no paired significance tests, the reader cannot tell whether these differences are systematic or noise. Reporting per-subject confidence intervals or paired tests across the five internal datasets would materially strengthen the central claim.","section":"Tables 2-4"},{"comment":"The authors emphasize that PC-Mask 'does not know the plane information' and that this property makes EchoONE a uniform model. However, the latent space used for prototypes is obtained from a ResNet34 pretrained for plane classification, so the prompt generation is built on plane-discriminative features, and plane labels are used during that pretraining. The paper should clarify exactly how the classifier is trained, whether plane labels are used, and whether the 'without plane information' claim refers only to inference. If it refers only to inference, the novelty statements in the Abstract and Introduction should be adjusted accordingly.","section":"§3.2, Figure 3"}],"minor_comments":[{"comment":"The unified label definition is inconsistent: the text says '1 for LV, 2 for LV cavity, and 3 for MYO', while Figure 5(a) and the annotation list say '1 LA, 2 LV, 3 MYO'. Please correct the numbering.","section":"§4.1"},{"comment":"The mDice values printed under the four panels do not match Table 2; for example, the line that should correspond to EchoONE shows 25.10, 27.22, 35.30, and 26.24, which are SAM's values in Table 2. The figure and caption need to be re-checked.","section":"Figure 1"},{"comment":"The paper claims six echocardiographic planes, but the experiments merge PSAX-B, PSAX-M, and PSAX-A into a single PSAX category; please state clearly whether the reported results are for four plane groups or six individual planes.","section":"Abstract and §4.1"},{"comment":"The dataset name 'HMC QU' appears with a space; use one consistent name throughout the text and tables.","section":"§4.3.2"},{"comment":"The last sentence reads 'extra benefits four segmentation' and should be 'for segmentation'; the sentence beginning 'it can be drawn that By leveraging' should be rephrased.","section":"§4.4"},{"comment":"The conclusion's claim that the method generalizes to MPS problems in other medical image modalities is speculative; either add supporting evidence or soften the claim.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The strongest issue is the HMCQU label-generation problem, which directly affects the external generalization evidence. If the authors can validate the generated LV labels against expert contours, or restrict the HMCQU claim to MYO, the paper is likely publishable after revision. I would also ask for significance testing, since the margins over SAMUS are small in several comparisons. The 'first uniform model' claim should be checked against prompt-free SAM adaptations such as AutoSAM and ProMISe before final acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"EchoONE is worth reading if you work on SAM adaptation or cardiac ultrasound. The core idea — composing cluster prototype masks into a dense prompt for SAM, without needing plane labels — is genuinely new, and the LFFA branch is a clean way to fuse local CNN features into the mask decoder. The evaluation is extensive: five internal and two external datasets, six planes, ten comparators, and consistent wins across nearly every cell. That is real work, and the gains, while modest over SAMUS (roughly 0.4–2.4 mDice), are consistent.\n\nThe soft spots are proportionate but real. The strongest is the HMCQU external result. Table 1 shows HMCQU annotates only MYO, and Section 4.1 says the LV cavity masks are generated by landmark detection and filling. Those same generated masks are the reference for the HMCQU row in Table 4, which the paper explicitly cites as evidence of generalization. If the fill heuristic is biased — smooth convex fills ignoring trabeculations or the valve plane — the reported LV Dice on that dataset is measuring agreement with the heuristic, not with the true anatomy. The paper provides no validation of these generated labels against expert contours. This does not sink the internal results, since A, B, and D already have LV annotations, but it does undercut the external generalization claim.\n\nTwo other gaps matter: there are no error bars or significance tests, so the small margins over SAMUS could be within run-to-run variance; and the cluster count K in PC-Mask is never reported, so we cannot judge how sensitive the method is to that parameter. The “first uniform MPS model” claim is plausible — none of the cited SAM adapters target multi-plane — but it is asserted rather than systematically verified against prior art. Code is promised but not yet available.\n\nOverall, this is a competent, useful paper with a genuine contribution. The central argument holds up for the internal data; the external claim needs the synthetic-label validation. A serious referee should ask for error bars, expert validation of the generated LV masks on HMCQU, and the value of K. I would send it to review and expect a revised version to be citable.","headline":"EchoONE is a solid engineering contribution to multi-plane echocardiography segmentation, but its headline generalization claim leans on a dataset whose LV ground truth is algorithmically generated.","tokens_in":14407,"tokens_out":1521,"would_cite":true,"duration_ms":16015,"reading_group":"yes","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 a single SAM-based network with two new modules—PC-Mask for dense prompts and LFFA for feature fusion—can segment the left ventricle, left atrium, and myocardium across four echocardiography views in one model…","keywords":["echocardiography segmentation","multi-plane segmentation","Segment Anything Model","dense prompt generation","SAM adaptation","cardiac ultrasound","left ventricle segmentation","medical image segmentation"],"falsifier":"On a dataset with manual LV cavity annotations (e.g., CAMUS, which provides LV labels), recompute EchoONE's LV Dice using the manual masks instead of the filled masks. If EchoONE no longer matches the reported numbers or no longer beats SAMUS on LV, the central state-of-the-art claim for LV segmentation collapses.","tokens_in":13446,"feed_emoji":"🫀","tokens_out":4437,"duration_ms":40022,"temperature":0.7,"pith_summary":"The paper aims to solve the multi-plane segmentation problem for echocardiography: instead of training one network per heart view (2CH, 3CH, 4CH, PSAX), it proposes a single uniform model called EchoONE. The model builds on Segment Anything (SAM), adds a module that generates semantic-aware dense prompts without knowing which plane the input comes from, and a module that fuses local CNN features into SAM's decoder. On internal and external datasets, EchoONE reports consistently higher Dice scores than plane-specific U-Nets and adapted SAM variants across all structures and planes. If true, this would let a clinic deploy one segmentation model for a full echo exam rather than a collection of view-specific models.","feed_headline":"EchoONE segments four heart-ultrasound planes in one model","feed_subtitle":"Combines SAM with learned dense prompts and local feature fusion; reports top scores on internal and external datasets.","key_machinery":"PC-Mask (prior-composable mask learning): a module that clusters images in a pre-trained feature space, stores a center mask per cluster, and for each input computes cosine similarity to cluster prototypes, then builds a dense prompt as a weighted concatenation of the center masks; a light U-Net turns this prior embedding into a mask prompt for SAM. LFFA (local feature fusion and adaption): a set of 1x1 convolutions that concatenate CNN branch features with the transformer block outputs of SAM's mask decoder, injecting local boundary detail and accelerating convergence.","core_discovery":"The central discovery is that the multi-plane segmentation problem in echocardiography can be solved by one model if the segmentation prompt is made plane-aware without using plane labels. EchoONE achieves this with PC-Mask, which groups training images by visual similarity, learns a prototype mask per group, and composes a per-image dense prompt from a weighted combination of these prototypes. A CNN branch with LFFA adapts SAM's encoder and decoder to ultrasound data. The paper reports that EchoONE outperforms both conventional segmentation networks (U-Net, DeepLabV3+, SwinUNet, H2Former) and six SAM-based methods on all four planes, all cardiac structures, and all internal and external datasets tested.","pith_inferences":["If the synthetic LV filling is biased, the reported LV margins over SAMUS may shrink once manual LV masks are used; a re-evaluation with true LV annotations would settle this.","PC-Mask's cluster prototypes act as a soft, unsupervised plane classifier; one could test whether the number of clusters K matters or whether the method degrades when plane-specific appearance shifts (e.g., new ultrasound machines) are not represented in the clusters.","EchoONE could be extended to video (e.g., beat-to-beat assessment) by treating frames as an additional plane-like dimension; the paper does not explore temporal consistency.","For clinical deployment, the model's HD95 on PSAX is the largest gap; boundary quality on short-axis views may be the limiting factor even if Dice is high."],"forward_implications":["A single trained EchoONE model can be deployed for a standard transthoracic echo exam covering 2CH, 3CH, 4CH, and PSAX views, removing the need for a per-view model zoo.","Because the dense prompt is constructed without plane labels, EchoONE can be applied to unseen views whose plane type is unknown at inference time.","The architecture is modally agnostic; the paper argues it should transfer to multi-plane segmentation in other medical imaging modalities.","The LFFA convergence speedup could reduce training cost for SAM-based medical segmentation generally."],"supporting_citations":[{"why":"The Segment Anything Model is the base architecture that EchoONE adapts and fine-tunes.","marker":"[17]"},{"why":"MedSAM is a primary SAM-adapted medical segmentation baseline that EchoONE compares against and improves upon.","marker":"[23]"},{"why":"SAMUS is the strongest SAM-based competitor for ultrasound segmentation and a key comparison point for the central claim.","marker":"[21]"},{"why":"U-Net is the classical segmentation backbone that motivates the multi-plane performance-drop problem EchoONE solves.","marker":"[32]"},{"why":"CAMUS is the public dataset that supplies part of the training and internal evaluation data for EchoONE.","marker":"[18]"},{"why":"EchoNet-Dynamic is one of the external evaluation datasets used to test generalization.","marker":"[31]"},{"why":"HMCQU is the other external evaluation dataset, with noisy low-quality images that stress the model's robustness.","marker":"[16]"},{"why":"SAM-LST provides the ladder fine-tuning idea that the CNN branch and LFFA module build upon.","marker":"[2]"}],"fun_headline_variants":["One model for multiple heart-ultrasound planes: EchoONE","EchoONE: one model to segment any echo plane","Single model beats per-plane specialists in echo segmentation","EchoONE: multi-plane echo segmentation without plane labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper generates synthetic left-ventricle cavity labels from myocardium annotations by detecting landmarks and filling the cavity, and it uses these filled masks as ground truth for both training and evaluation; if this filling is biased, the reported LV Dice improvements do not reflect true segmentation quality.","fun_headline_variants_meta":{"raw":{"variants":["One model for multiple heart-ultrasound planes: EchoONE","EchoONE: one model to segment any echo plane","Single model beats per-plane specialists in echo segmentation","EchoONE: multi-plane echo segmentation without plane labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001572,"raw_usage":{"total_tokens":6244,"prompt_tokens":881,"completion_tokens":5363,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":5296}},"tokens_in":497,"tokens_out":5363,"duration_ms":38126,"temperature":1.0,"reasoning_tokens":5296,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:52:53.259360+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a dataset with manual LV cavity annotations (e.g., CAMUS, which provides LV labels), recompute EchoONE's LV Dice using the manual masks instead of the filled masks. If EchoONE no longer matches the reported numbers or no longer beats SAMUS on LV, the central state-of-the-art claim for LV segmentation collapses.","supporting_citations":[{"cited_title":"Segment any- thing","cited_arxiv_id":null,"evidence_quote":"The Segment Anything Model is the base architecture that EchoONE adapts and fine-tunes."},{"cited_title":"Segment anything in medical images","cited_arxiv_id":null,"evidence_quote":"MedSAM is a primary SAM-adapted medical segmentation baseline that EchoONE compares against and improves upon."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"U-Net is the classical segmentation backbone that motivates the multi-plane performance-drop problem EchoONE solves."},{"cited_title":"Deep learning for segmentation using an open large-scale dataset in 2d echocardiography","cited_arxiv_id":null,"evidence_quote":"CAMUS is the public dataset that supplies part of the training and internal evaluation data for EchoONE."},{"cited_title":"Video-based ai for beat-to-beat assessment of cardiac func- tion","cited_arxiv_id":null,"evidence_quote":"EchoNet-Dynamic is one of the external evaluation datasets used to test generalization."},{"cited_title":"Left ventricular wall motion estimation by active polynomials for acute myocardial infarction detection","cited_arxiv_id":null,"evidence_quote":"HMCQU is the other external evaluation dataset, with noisy low-quality images that stress the model's robustness."}],"review_version":1}