{"id":"d08005f9-be9c-40ef-8a10-4297f2eeab47","arxiv_id":"2508.18681","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hybrid convolutional-Mamba network segments left ventricular contours in echocardiography videos and reports improved ejection fraction correlation on three benchmarks.","lead":"A new segmentation network for heart ultrasound videos combines single-frame convolutional layers with Mamba video layers, then estimates ejection fraction from the contours. The authors report higher correlation with clinical ejection fraction than eight prior methods on three public datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Oracle-timed input clips leak ED/ES frame identity; reported EF gains do not validate end-to-end EF estimation from unaligned videos.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I identify: evaluation clips are constructed with ground-truth ED/ES timing, leaking the identity of the keyframes. This is not a superficial issue; it changes what the reported EF numbers mean. The paper's stated goal is to improve EF estimation from echocardiography videos, yet the pipeline as described requires external knowledge of which frames are ED and ES. Without a frame-detection mechanism or an evaluation on unaligned videos, the central claim is only validated for a protocol that assumes the most clinically relevant information is already given. I agree with the reader's CONDITIONAL verdict because the segmentation architecture and ablations still support a contribution to frame-wise segmentation, and the comparison to baselines under the same protocol is internally consistent. However, the real-world EF-estimation claim should be conditioned on resolving the timing-leakage issue. The recommended concrete test would settle whether the concern genuinely lands by measuring performance under a realistic unaligned inference setting.","tokens_in":8221,"tokens_out":5786,"duration_ms":70328,"concrete_test":"Run the released HSS-Net inference on the EchoNet-Dynamic test videos in their original, unaligned form: feed full-length videos or uniformly sampled frames without forcing ED as first and ES as last. Compute EF by selecting the frames with predicted minimum and maximum LV volume (or using a simple learned frame detector), then recompute corr, bias, and std. If the correlation drops substantially below the reported 84.50, the headline result is contingent on oracle timing rather than end-to-end EF estimation. Alternatively, retrain the model on randomly shifted clips and evaluate on the same aligned test set; a large drop would confirm reliance on clip position.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that HSS-Net improves ejection-fraction estimation from echocardiography videos is undermined by the evaluation protocol. Section 3 states: 'The video clips are cropped to ensure that the ED frame is the first and the ES frame is the last.' This means every input clip seen in training and testing begins at the ground-truth end-diastolic frame and ends at the ground-truth end-systolic frame. Consequently, the model never has to identify these keyframes; it can learn to segment the first and last frames of the clip, which are precisely the only frames supervised and the only frames used for EF computation. The reported metrics (corr 90.47 on CAMUS, 76.91 on EchoNet-Pediatric, 84.50 on EchoNet-Dynamic) thus measure segmentation quality on oracle-selected keyframes, not the ability to estimate EF from raw video. Because the abstract and title claim EF estimation from videos, and the paper does not report or propose any mechanism for locating ED/ES frames in unaligned sequences, the headline numeric results are likely inflated relative to a realistic end-to-end setting. The comparison with baselines is internally fair (all methods use the same aligned clips), but the absolute SOTA claim overstates real-world utility. This concern is load-bearing because it directly affects whether the primary 'EF estimation' claim transfers beyond the artificial aligned-input protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HSS-Net, a hierarchical spatio-temporal segmentation network for echocardiography videos, combining low-level separable convolution blocks (single-frame detail) with high-level spatio-temporal Mamba blocks (multi-frame global dynamics). A Spatio-temporal Cross Scan (STCS) module with temporal, spatial, diagonal, and anti-diagonal scanning modes is introduced to capture long-range dependencies. EF is computed from predicted ED/ES contours using the Simpson method. The authors report state-of-the-art results on CAMUS, EchoNet-Pediatric, and EchoNet-Dynamic, with correlations of 90.47, 76.91, and 84.50 respectively, along with Dice and HD95 improvements over several segmentation baselines. The code is publicly available.","tokens_in":8544,"tokens_out":4323,"duration_ms":50271,"significance":"If the results hold, HSS-Net offers an efficient (5.6G FLOPs, 31.2M params) hybrid architecture that balances local detail and global temporal modeling, with a credible conceptual motivation for reducing EF bias caused by boundary segmentation errors. The paper ships code and follows a reproducible evaluation protocol, which are strengths. However, the headline EF-estimation claim is weakened by the evaluation protocol: clips are truncated so that the ground-truth ED frame is first and ES frame is last, so the model never discovers keyframe timing. The reported correlations are thus conditional on oracle timing, and the advantage over segmentation baselines is shown on this aligned protocol only. The lack of confidence intervals or significance tests further limits the strength of the SOTA claim.","major_comments":[{"comment":"The evaluation protocol crops every input clip so that the ground-truth ED frame is the first frame and the ground-truth ES frame is the last frame, and only these two frames are supervised and used for EF computation. This oracle timing leaks the keyframe identities that a real EF estimator must discover from unaligned video. The reported correlations (90.47/76.91/84.50) therefore measure segmentation quality on physician-timed keyframes, not end-to-end EF estimation. The comparison to baselines is internally fair, but the absolute SOTA claim does not transfer to unaligned clinical videos. Please add an experiment on unaligned clips (e.g., with a separate ED/ES detector or by evaluating on full-length videos) or explicitly re-frame the claim as keyframe-conditional segmentation.","section":"Section 3, Datasets"},{"comment":"All reported numbers are single-run point estimates without confidence intervals, repeated-seed variance, or significance tests. Several Dice gaps are small (e.g., 93.89 vs 93.02 on CAMUS; 92.67 vs 92.48 on EchoNet-Dynamic), so the claimed advantage over baselines may not be statistically reliable. Please report mean±std over at least three runs and apply paired tests (e.g., Wilcoxon or permutation) to the EF correlation and Dice differences.","section":"Tables 1-2, Section 3.1"},{"comment":"The SOTA claim for EF estimation is supported only by comparisons against segmentation baselines (UNet++, TransUNet, VideoMamba, etc.). No direct EF regression or keyframe-detection method is compared, although these methods are the relevant SOTA for EF estimation from videos (e.g., EchoNet). Please include comparisons to such methods under the same aligned protocol and, if possible, on unaligned inputs, or restrict the claim to segmentation-based EF estimation on aligned keyframes.","section":"Section 3.1, Tables 1-2"}],"minor_comments":[{"comment":"The entry '84.500.95±6.75' in the EchoNet-Dynamic row is missing a space; it should read '84.50 0.95±6.75'.","section":"Table 2"},{"comment":"The sentence 'The primary of echocardiography analysis' is ungrammatical; it should be 'The primary goal of echocardiography analysis' or similar.","section":"Section 1"},{"comment":"The text says CAMUS has 'annotations provided for all frames' but later states that only ED and ES annotations are used for training and evaluation. Please clarify which annotations are actually used.","section":"Section 3, Datasets"},{"comment":"The term 'spatio-temporal' is used with inconsistent hyphenation ('spatio-temporal' vs 'spatiotemporal'). Please unify the notation.","section":"Throughout"},{"comment":"Using the same symbol F_i for both input and output of the residual blocks is acceptable but could be clearer with an explicit out-notation (e.g., F_i^out).","section":"Equations (1)-(2)"}],"recommendation":"major_revision","confidential_remarks":"The central concern is the oracle-timed clip construction: the paper's headline EF numbers are obtained under a protocol where ED/ES are given, so the model never performs keyframe detection. This is a load-bearing issue for the paper's main claim of improving EF estimation from videos. The architecture itself is reasonable and the code availability is a plus. I would be willing to accept after the authors add an unaligned evaluation or substantially temper the EF-estimation claim, and after statistical robustness is reported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. This is a genuinely solid architectural paper: a hierarchical encoder where low-level stages are separable convolutions and high-level stages are Mamba blocks, with a spatio-temporal cross-scan module that adds diagonal/anti-diagonal scanning. The ablations show each component helps, and the model is efficient (5.6 GFLOPs on CAMUS). Second, the headline EF numbers should not be read as end-to-end EF estimation from raw video, because Section 3 states plainly that every clip is cropped so the ground-truth ED frame is the first and the ground-truth ES frame is the last. The model therefore never has to find the keyframes; it segments the first and last frames of the given clip. That is a different (and easier) task than estimating EF from an unaligned echocardiography video, which is what the title and abstract promise.\n\nWhat is actually new: the specific combination of convolutional and Mamba stages in this segmentation setup, plus the STCS diagonal modes. This is a legitimate extension of VideoMamba/Vivim, not a breakthrough, but the authors do the work: three public datasets, eight baselines, ablations, and a code link. The relative gains are consistent and, for EF correlation, nontrivial—for instance, CAMUS corr goes from 84.48 for the best baseline to 90.47, and EchoNet-Dynamic from 81.12 to 84.50. Dice gains are small (under a point), but the EF-relevant improvements are larger. Within the aligned-clip protocol, the comparison is fair because all methods see the same clips.\n\nThe soft spots. The oracle timing is the load-bearing one. The paper gives no mechanism for locating ED/ES frames in unaligned sequences, and no experiment that removes this assumption. So the absolute correlations and biases are likely inflated relative to a realistic automatic pipeline. This does not invalidate the relative ranking of methods, but it does invalidate the overstatement that HSS-Net 'estimates EF' from videos. Second, there are no confidence intervals or repeated-run variance; some baseline gaps are small enough that significance is unclear. Third, the training loss supervises only the first and last frames of each clip, but the paper does not say whether the unlabeled middle frames are masked out or somehow used; that needs clarification.\n\nWho should read it: people working on echo segmentation and medical video Mamba architectures, and anyone thinking about evaluation protocols for cardiac function estimation. I'd bring it to a reading group and I'd cite it, but with the protocol caveat. It deserves a real referee. Send it out, but the authors should be pushed to either report results on unaligned clips (e.g., using a standard keyframe detector) or reframe the claims as 'segmentation given ED/ES timing,' and to add uncertainty quantification.\n\nBest,\n[Your name]","headline":"Solid segmentation paper with a real protocol flaw: EF numbers come from oracle-cropped clips (ED first, ES last), so the headline 'EF estimation from videos' overstates what is actually evaluated.","tokens_in":8986,"tokens_out":3775,"would_cite":true,"duration_ms":36406,"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":"HSS-Net: a hierarchical CNN-Mamba video segmentation network that estimates ejection fraction from echocardiography videos more accurately than eight baselines on three datasets.","keywords":["Echocardiography","Ejection fraction estimation","Left ventricular segmentation","Video segmentation","Mamba","State space models","Spatio-temporal scan","Medical image analysis"],"falsifier":"Retrain and test HSS-Net on full-length, unaligned echo clips without physician-provided ED/ES alignment, then compare the EF Pearson correlation with the reported aligned numbers (90.47 on CAMUS, 84.50 on EchoNet-Dynamic). If the correlation falls toward the image-only baseline or the bias grows beyond the reported values, the advantage is partly an artifact of the alignment protocol; testing on randomly cropped non-aligned clips would also reveal timing leakage.","tokens_in":8143,"feed_emoji":"🫀","tokens_out":7020,"duration_ms":72685,"temperature":0.7,"pith_summary":"This paper tries to show that ejection fraction (EF) estimation from echocardiography videos gets more accurate when a network is explicitly split into a local-detail pathway and a global-motion pathway. HSS-Net uses cheap separable convolutions on single frames at low resolution levels, then switches to Mamba-based spatio-temporal blocks that scan across frames and positions to track how the left ventricle deforms over the heartbeat. The paper reports that this division of labour avoids the two failure modes of previous work: image-only methods miss the continuity of cardiac motion, and video-only methods lose fine boundaries that determine volume. On three public video datasets—CAMUS, EchoNet-Pediatric, and EchoNet-Dynamic—HSS-Net reports higher EF correlation and lower bias than eight compared baselines, with per-frame computation of 5.6 GFLOPs at 256×256. A reader should care because EF is the routine clinical measure of systolic function, and small segmentation errors at the base and apex translate into large EF errors.","feed_headline":"Hybrid CNN-Mamba network improves ejection-fraction estimates in echo","feed_subtitle":"By pairing single-frame detail with cross-frame motion, HSS-Net beats eight baselines on three cardiac datasets.","key_machinery":"The central machinery is the hierarchical encoder-decoder combined with the STCS scanning module. Low-level stages use separable convolution blocks (MobileNetV2-style inverted residuals) on single frames to retain fine local detail. High-level stages use spatio-temporal Mamba blocks that reshape the video into a long one-dimensional sequence and pass it through STCS, which scans along bidirectional temporal, spatial, diagonal, and anti-diagonal routes. This lets a selective state-space model build long-range cross-frame and cross-position dependencies, capturing structures such as apex motion and lateral wall contraction that a single-frame segmentation would miss, while avoiding the full co","core_discovery":"The paper's central claim is that deliberately separating single-frame detail from multi-frame motion modeling improves ejection-fraction estimation. The proposed HSS-Net keeps convolutional blocks in the low-level encoder stages to preserve edges, textures, and boundary detail, and switches to spatio-temporal Mamba blocks in the high-level stages to capture how the ventricle moves across frames. To keep the multi-frame modeling from staying local, the Spatio-temporal Cross Scan (STCS) module flattens video patches into one-dimensional sequences and scans them along temporal, spatial, diagonal, and anti-diagonal directions in forward and backward passes. On CAMUS, HSS-Net reports a Pearson c","pith_inferences":["The reported numbers depend on a protocol where each clip is cropped so the physician-annotated end-diastolic frame is first and the end-systolic frame is last; whether the advantage survives without that timing alignment is not tested in the paper and is the most direct open question.","Because STCS's diagonal scanning is a generic mechanism, the same hierarchical CNN-Mamba recipe may transfer to other sparsely annotated video segmentation tasks, such as fetal ultrasound or cardiac MRI, where only two key frames are labelled.","The paper still follows the segmentation-then-EF pipeline; a direct end-to-end regression head on the same spatio-temporal features could potentially capture motion cues that the two-frame volume formula discards.","The reported bias and standard deviation on CAMUS are close to the range of inter-observer variability typically reported in echocardiography, so part of the remaining error may be label noise rather than architecture; an observer-agreement comparison would clarify the practical ceiling."],"forward_implications":["Low-level single-frame detail and high-level video modeling are complementary: the paper's ablations show that removing either (image-only or video-only variants) lowers EF correlation on both CAMUS and EchoNet-Dynamic.","Every STCS scan direction contributes: dropping temporal, spatial, diagonal, or anti-diagonal scanning degrades at least one EF metric, which supports the design of multi-perspective scanning.","The model's per-frame cost (5.6 GFLOPs, 31.2M parameters) is in the range of much weaker baselines, suggesting the accuracy gain does not require a large computational budget.","On the two larger datasets the reported EF bias is small (1.29 ± 8.68 on EchoNet-Pediatric, 0.95 ± 6.75 on EchoNet-Dynamic), which is the clinically relevant quantity for screening and follow-up."],"supporting_citations":[{"why":"Supplies the S6 selective state-space model that the Spatio-temporal Cross Scan module uses to model long spatio-temporal sequences.","marker":"[6]"},{"why":"Supplies the inverted separable convolution module used in the low-level single-frame stages of the encoder.","marker":"[15]"},{"why":"Provides the CAMUS dataset and the Simpson's single-plane method-of-disks EF calculation used for the EchoNet datasets.","marker":"[9]"},{"why":"Provides the pediatric echocardiography dataset with ED/ES annotations used as a second benchmark.","marker":"[14]"},{"why":"Provides the EchoNet-Dynamic dataset splits and the Simpson's biplane method-of-disks EF calculation used for CAMUS.","marker":"[24]"},{"why":"Defines the 10-frame uniform sampling protocol and the EF metric conventions the experiments follow.","marker":"[5]"},{"why":"Shares the sampling and evaluation protocol and serves as a strong semi-supervised baseline that HSS-Net must beat.","marker":"[21]"}],"fun_headline_variants":["Hybrid CNN-Mamba net sharpens echo EF estimates","Two-tier network boosts echo ejection fraction accuracy","STCS module bridges frames to refine echo EF scores","Split-level design lifts EF prediction in echo videos","HSS-Net balances detail and motion for cardiac EF"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that each input clip is pre-aligned so the end-diastolic frame is the first frame and the end-systolic frame is the last frame, using physician-provided timing; real workflows must discover those two frames from an unaligned video, and the reported EF correlations may depend on having been given that timing.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid CNN-Mamba net sharpens echo EF estimates","Two-tier network boosts echo ejection fraction accuracy","STCS module bridges frames to refine echo EF scores","Split-level design lifts EF prediction in echo videos","HSS-Net balances detail and motion for cardiac EF"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000319,"raw_usage":{"total_tokens":1626,"prompt_tokens":724,"completion_tokens":902,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":827}},"tokens_in":468,"tokens_out":902,"duration_ms":10107,"temperature":1.0,"reasoning_tokens":827,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:17:26.874631+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain and test HSS-Net on full-length, unaligned echo clips without physician-provided ED/ES alignment, then compare the EF Pearson correlation with the reported aligned numbers (90.47 on CAMUS, 84.50 on EchoNet-Dynamic). If the correlation falls toward the image-only baseline or the bias grows beyond the reported values, the advantage is partly an artifact of the alignment protocol; testing on randomly cropped non-aligned clips would also reveal timing leakage.","supporting_citations":[{"cited_title":"In: Computer Vision and Pattern Recog- nition","cited_arxiv_id":null,"evidence_quote":"Supplies the inverted separable convolution module used in the low-level single-frame stages of the encoder."},{"cited_title":"IEEE Transactions on Medical Imaging38(9), 2198–2210 (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the CAMUS dataset and the Simpson's single-plane method-of-disks EF calculation used for the EchoNet datasets."},{"cited_title":"In: International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Provides the EchoNet-Dynamic dataset splits and the Simpson's biplane method-of-disks EF calculation used for CAMUS."},{"cited_title":"In: Computer Vision and Pattern Recogni- tion","cited_arxiv_id":null,"evidence_quote":"Defines the 10-frame uniform sampling protocol and the EF metric conventions the experiments follow."}],"review_version":1}