{"id":"deef802f-7e0f-4e6e-bbfc-2aee36975a43","arxiv_id":"2607.07580","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"CISR-Net achieves SOTA echocardiography segmentation by fusing local transition probability correlations for semantic rectification and frequency-domain denoising pre-training.","lead":"The paper proposes a semi-supervised echocardiography segmentation network (CISR-Net) that uses local transition probability correlations and frequency-aware denoising pre-training to improve robustness against speckle noise. It reports SOTA Dice and HD95 on CAMUS and EchoNet-Dynamic, which matters for automated cardiac function assessment.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"SOTA claim rests on sub-0.5% Dice improvements with no significance testing or multi-seed evaluation; ablation Tables 2(a) and 2(b) also show an unexplained 0.06% discrepancy for the same full model.","rationale":"The reader's weakest_assumption focused on the unvalidated 'transition probability' interpretation (Eq. 6), which is a legitimate conceptual concern but is not the most load-bearing issue for the paper's central SOTA claim. The more load-bearing concern is that the reported improvements — 0.31% Dice over the best baseline, 0.50% from the STLSF ablation, 0.16% from FD pre-training — are within the range of typical run-to-run variance for medical segmentation, yet no significance testing, multi-seed evaluation, or confidence intervals are provided. The model selection criterion ('peak validation Dice') further risks optimistic bias. The reader did identify significance testing as a concern in the rationale, so the verdict of CONDITIONAL is appropriate and I would not change it. I rate agreement as 'partial' because the reader's designated weakest_assumption (transition probability validation) is secondary to the statistical robustness issue, which more directly undermines the SOTA claim. The paper's architectural contribution is legitimate and the ablations are internally consistent in direction, but the magnitude of improvements does not support a confident SOTA assertion without statistical evidence. The unexplained 0.06% discrepancy between Tables 2(a) and 2(b) for the same 'Ours' configuration is a minor but telling indicator that single-run results at this precision level may not be reliable.","tokens_in":9175,"tokens_out":3809,"duration_ms":251600,"concrete_test":"Re-run the full model and the strongest baseline (NCMNet) with at least 5 random seeds each on both CAMUS and EchoNet-Dynamic. Report mean ± std for Dice, HD95, and ASSD. If the standard deviation exceeds ~0.15% Dice (making the 0.31–0.32% margins statistically indistinguishable), the SOTA claim weakens substantially. Additionally, reconcile the 93.25 vs 93.19 discrepancy between Tables 2(a) and 2(b) by confirming both use identical training/evaluation configurations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central SOTA claim depends on improvements of 0.31% Dice over NCMNet on CAMUS (93.87 vs 93.56) and 0.32% on EchoNet-Dynamic (92.62 vs 92.30). The ablation (Table 2a) attributes +0.50% Dice to the full STLSF module (92.75→93.25) and Table 2(b) attributes +0.16% to FD pre-training over the best alternative pre-training task (93.19 vs 93.03). These are single-run results with no reported variance, confidence intervals, or significance tests. §3.2 states the optimal model is selected based on 'peak validation Dice score,' which introduces optimistic selection bias that could easily account for sub-0.5% differences. Furthermore, Table 2(a) 'Ours' = 93.25 and Table 2(b) 'Ours' = 93.19 should represent the same full configuration (STLSF + FD), yet they differ by 0.06% with no explanation — suggesting either different runs or inconsistent evaluation protocols between the two ablation studies. Without multi-seed runs demonstrating non-overlapping confidence intervals, the marginal improvements cannot be distinguished from run-to-run variance, and the SOTA attribution to the proposed modules is not robustly established.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes CISR-Net, a semi-supervised echocardiography segmentation framework with two main contributions: (1) a Spatio-temporal Local Self-Similarity Fusion (STLSF) module that uses local transition probability correlations to rectify noise-induced semantic errors in deep features (TSG) and then guides shallow texture refinement (STG), and (2) a Frequency-aware Denoising (FD) pre-training strategy that injects frequency-domain perturbations to adapt the encoder to ultrasound-specific noise patterns. The backbone is a convolutional network inspired by OverLoCK with a base-overview-focus architecture. The method is evaluated on CAMUS and EchoNet-Dynamic, reporting Dice scores of 93.87% and 92.62% respectively, with ablation studies isolating the contributions of TSG, STG, and FD pre-training.","tokens_in":9354,"tokens_out":1825,"duration_ms":142004,"significance":"The paper addresses a clinically relevant problem and the STLSF module is a reasonable architectural contribution. The frequency-aware denoising pre-training is a sensible domain-specific adaptation. The evaluation covers two standard benchmarks with eight baselines including recent methods (NCMNet, EchoONE, MemSAM), and boundary metrics (HD95, ASSD) are reported alongside Dice. The computational profile (78.62 GFLOPs, 38.5 MB) is practical. However, the magnitude of improvement over the strongest baseline is marginal, and the experimental protocol does not establish that these gains are statistically robust.","major_comments":[{"comment":"§3.2 and Table 1: The SOTA claim rests on improvements of 0.31% Dice over NCMNet on CAMUS (93.87 vs 93.56) and 0.32% on EchoNet-Dynamic (92.62 vs 92.30). No variance, confidence intervals, or significance tests are reported, and §3.2 states the model is selected based on 'peak validation Dice score,' which introduces optimistic selection bias. Without multi-seed runs, sub-0.5% differences cannot be distinguished from run-to-run variance. This is load-bearing because the paper's central claim is SOTA performance. At minimum, the authors should report results over multiple seeds and provide standard deviations or confidence intervals.","section":null},{"comment":"Table 2(a) vs Table 2(b): The 'Ours' row in Table 2(a) reports Dice = 93.25%, while Table 2(b) reports 'Ours' Dice = 93.19%. Both are described as averaged over CAMUS and EchoNet-Dynamic, and both should represent the full model (STLSF + FD pre-training). The 0.06% discrepancy is unexplained and raises questions about whether the two ablation studies used different runs, different evaluation protocols, or different model configurations. The authors should clarify whether these are the same model and, if not, explain the difference.","section":null},{"comment":"§2.3, Eq. (6)–(7): The paper frames the inter-frame affinity matrix computed via local linear projections (Eq. 6) as a 'transition probability distribution' that 'implicitly represents cardiac motion,' and claims this representation is more stable than raw features under speckle noise. This is the conceptual basis for the TSG module, but no empirical evidence is provided that the computed affinities actually correlate with physical cardiac motion or are more stable than alternatives (e.g., raw feature similarity, optical flow). The ablation in Table 2(a) shows TSG contributes +0.26% Dice, which demonstrates practical utility, but does not validate the transition-probability interpretation. A simple analysis (e.g., visualizing the affinity matrices or comparing stability under noise against a baseline) would substantially strengthen the central claim.","section":null}],"minor_comments":[{"comment":"§2.2, Eq. (1): The notation '1/f ultrasound speckle noise map' for N is unclear. Is this a sample from a 1/f power spectral density? The generation procedure for N should be specified.","section":null},{"comment":"§2.3, Eq. (7): The attention formulation uses Q(P_trans(m)), K(P_trans(n)), V(F̃_n). It is unclear whether Q, K, V are shared with or distinct from Linear1, Linear2 in Eq. (6). Clarify the parameter relationship.","section":null},{"comment":"Table 1: The 'Venue/Year' column lists 'UMB 2024' for SimLVSeg — this appears to be an abbreviation for 'Ultrasound in Medicine & Biology'; consider spelling out the journal name for consistency with other entries.","section":null},{"comment":"§3.1: The resolution difference between CAMUS (224×224) and EchoNet-Dynamic (128×128) is noted but its potential impact on cross-dataset comparisons is not discussed. A brief note would help.","section":null},{"comment":"Fig. 1: The figure caption is minimal. Labels for sub-figures (a), (b), (c) referenced in the text (e.g., 'Fig. 1(c)') should be explicitly marked in the figure.","section":null},{"comment":"§2.3: The local window size k is mentioned in the definition of N(m) but its value is not specified in the experimental setup. This is a free parameter that should be reported.","section":null},{"comment":"References [18] (OverLoCK) is cited as the backbone inspiration but the relationship between the proposed architecture and OverLoCK is not detailed enough for reproducibility. Clarify what is borrowed vs. novel.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper is technically competent but the SOTA claim is fragile given the marginal improvements and lack of statistical rigor. The 'transition probability' framing is conceptually borrowed from [25, 21] and the paper would benefit from either validating this interpretation empirically or toning down the theoretical claims. The table discrepancy (93.25 vs 93.19) is small but suggests possible carelessness in reporting that an editor should flag. If the authors can provide multi-seed results and clarify the ablation inconsistency, this could become a solid contribution."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee raises three major points: (1) the absence of multi-seed runs and significance testing for sub-0.5% Dice improvements, (2) an unexplained 0.06% discrepancy between the 'Ours' rows in Tables 2(a) and 2(b), and (3) the lack of empirical validation for the transition-probability interpretation underlying the TSG module. We agree that all three points require action and will address each in the revised manuscript.","responses":[{"response":"The referee is correct. We will conduct multi-seed experiments (at least 5 seeds with different random initializations) for both CISR-Net and the strongest baselines (NCMNet, EchoONE, MemSAM) on both CAMUS and EchoNet-Dynamic. We will report mean Dice, standard deviation, and paired t-tests (or Wilcoxon signed-rank tests where appropriate). We will also revise the model selection description in §3.2 to clarify that selection is based on validation set performance only, and we will report test-set results corresponding to the best validation epoch. If the significance tests do not support the SOTA claim at conventional thresholds, we will revise the language from 'SOTA performance' to 'competitive performance with improvements in boundary metrics,' since our HD95 and ASSD improvements are more substantial (e.g., 3.29mm vs 4.08mm on CAMUS, 2.73mm vs 4.04mm on EchoNet-Dynamic relative to NCMNet). We acknowledge that the Dice improvements alone are marginal and that the boundary metric improvements, while larger in relative terms, also require statistical validation.","revision_made":"yes","referee_comment":"§3.2 and Table 1: The SOTA claim rests on improvements of 0.31% Dice over NCMNet on CAMUS and 0.32% on EchoNet-Dynamic. No variance, confidence intervals, or significance tests are reported, and the model is selected based on 'peak validation Dice score,' introducing optimistic selection bias. Without multi-seed runs, sub-0.5% differences cannot be distinguished from run-to-run variance."},{"response":"The referee has identified a genuine inconsistency. The two ablation tables were conducted at different stages of our experiments: Table 2(a) was run during the STLSF module development phase, while Table 2(b) was run later when evaluating pre-training strategies. Although both use the same model configuration, minor differences in random seeds and the order in which experiments were conducted led to the 0.06% discrepancy. This is not acceptable for a paper claiming precise ablation results. In the revision, we will re-run both ablation studies under identical conditions (same seeds, same data splits, same evaluation protocol) and ensure the full-model rows match exactly. We will also report standard deviations across seeds for all ablation entries, which will contextualize whether 0.06% is within run-to-run variance (which we expect it is).","revision_made":"yes","referee_comment":"Table 2(a) vs Table 2(b): The 'Ours' row in Table 2(a) reports Dice = 93.25%, while Table 2(b) reports 'Ours' Dice = 93.19%. Both are described as averaged over CAMUS and EchoNet-Dynamic and both should represent the full model. The 0.06% discrepancy is unexplained."},{"response":"This is a fair criticism. The conceptual claim that the affinity matrix represents transition probabilities correlated with cardiac motion is currently asserted but not empirically validated. We will add the following analyses to the revised manuscript: (1) Visualization of the learned affinity matrices P_trans at different cardiac phases (ED, mid-systole, ES, mid-diastole) to show that the attention patterns shift in a manner consistent with known cardiac motion patterns. (2) A quantitative stability comparison: we will add controlled speckle noise at varying levels to the same input frames and measure the variance of the affinity matrices versus raw feature similarity matrices, demonstrating that the transition-probability formulation is more stable under noise perturbation. (3) We will compare TSG against a variant using raw feature similarity (without the linear projection + softmax normalization) to directly test whether the transition-probability formulation provides benefits beyond simple feature matching. If these analyses do not support the stability claim, we will revise the framing to describe TSG as a learned attention-based semantic rectification module without the transition-probability interpretation.","revision_made":"yes","referee_comment":"§2.3, Eq. (6)–(7): The paper frames the inter-frame affinity matrix as a 'transition probability distribution' that 'implicitly represents cardiac motion' and claims this representation is more stable than raw features under speckle noise. No empirical evidence is provided that the computed affinities actually correlate with physical cardiac motion or are more stable than alternatives. The ablation shows TSG contributes +0.26% Dice but does not validate the transition-probability interpretation."}],"tokens_in":9135,"tokens_out":1057,"duration_ms":139480,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Two things to know upfront: this is a competent architectural contribution to echocardiography segmentation, and the SOTA claim rests on sub-0.5% Dice improvements with no multi-seed evaluation or significance testing. The stress-test concern about this is correct and is the main issue a referee should press on.","headline":"Legitimate engineering contribution to echo segmentation, but SOTA claim is not robustly supported without significance testing.","tokens_in":9949,"tokens_out":127,"would_cite":false,"duration_ms":92221,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Heart Ultrasound Segmentation Hits 93.87% Dice via Transition Probabilities","keywords":["echocardiography segmentation","transition probability","spatiotemporal features","semi-supervised learning","frequency-domain pre-training","speckle noise","cardiac motion","local self-similarity"],"falsifier":"If one replaced the transition probability computation (Eq. 6-7) with a standard local self-attention using raw features as queries, keys, and values—without any motion-probability interpretation—and achieved equivalent segmentation performance, the central claim that transition probability correlations specifically drive the improvement would be undermined.","tokens_in":9277,"feed_emoji":"💓","tokens_out":1291,"duration_ms":198673,"temperature":0.7,"pith_summary":"The paper tries to prove that deep learning segmentation of echocardiography videos can be substantially improved by exploiting a specific kind of local motion cue: inter-frame affinity matrices interpreted as pixel-level transition probability distributions of cardiac motion. The authors argue that these transition probabilities are more stable under speckle noise than raw texture features, and that using them to guide semantic correction produces smoother, more anatomically consistent segmentations. They build a semi-supervised network called CISR-Net around a Spatio-Temporal Local Self-Similarity Fusion (STLSF) module, which first rectifies deep semantic features using local transition probability correlations and then uses those refined semantics to clean up noisy shallow texture features. A frequency-domain denoising pre-training strategy adapts the encoder to ultrasound-specific noise patterns before fine-tuning. On two public benchmarks, CAMUS and EchoNet-Dynamic, the method achieves Dice scores of 93.87% and 92.62% with boundary errors (HD95) of 3.29mm and 2.73mm, surpassing eight competing methods including both general video models and echocardiography-specific networks. The paper's central claim is that transition probability correlations, computed via local linear projections and softmax normalization within spatiotemporal windows, capture structural motion patterns that resist imaging artifacts better than direct feature matching, and that this stability propagates through the network to yield better boundaries.","feed_headline":"Heart Ultrasound Segmentation Hits 93.87% Dice via Transition Probabilities","feed_subtitle":"A new network treats inter-frame pixel affinities as cardiac motion probabilities, using them to correct noise-induced segmentation errors.","key_machinery":"Spatio-Temporal Local Self-Similarity Fusion (STLSF) module with two components: (1) Transition-to-Semantics Guidance (TSG), which computes local inter-frame transition probability distributions via linear projections and softmax, then uses them as structural queries/keys in an attention operation to rectify deep semantic features; (2) Semantics-to-Texture Guidance (STG), which uses the refined deep semantics to guide local cross-attention over shallow texture features for boundary-consistent output. A Frequency-aware Denoising (FD) pre-training strategy injects speckle-like noise in the DCT frequency domain and trains the encoder to recover clean frequency spectra.","core_discovery":"The central mechanism is the Transition-to-Semantics Guidance (TSG) component of the STLSF module. For each pixel in a deep feature map, TSG computes a local transition probability distribution by projecting the pixel and its neighbors through separate linear layers, taking their dot product, and applying softmax. This distribution is treated as an implicit representation of cardiac motion. An attention operation then uses these transition probability distributions as queries and keys while the original features serve as values, so that semantic alignment is driven by structural motion coherence rather than by noise-corrupted texture. The paper demonstrates that this two-step process—compute","pith_inferences":["The paper does not empirically validate that the computed transition probabilities actually correspond to physical cardiac motion or are more stable than alternatives under noise. If they are simply re-encoded attention affinities relabeled as 'transition probabilities,' the improvement may come from the local windowing and dual-projection design rather than from any motion-modeling property, in w","The ablation shows TSG and STG each contribute roughly 0.25% Dice improvement and combine for 0.50%, suggesting the gains are additive but modest. The frequency-domain pre-training contributes a larger share of the total improvement (roughly 0.3-0.4% Dice over alternatives), raising the question of whether the STLSF module's complexity is justified relative to simply investing more in pre-training","The method processes 10-frame clips at 150 frames/s, which is clinically practical, but the paper does not report inference latency on variable-length or streaming inputs, leaving open whether the local-window approach scales to continuous acquisition scenarios."],"forward_implications":["If transition probability correlations are genuinely more stable than raw features under noise, the same mechanism could improve segmentation in other imaging modalities with heavy speckle or artifact corruption, such as optical coherence tomography or ultrasound of other organs.","The frequency-domain pre-training strategy could be extended to incorporate modality-specific spectral priors beyond speckle noise, potentially improving encoder initialization for any imaging system with known noise spectral characteristics.","The local-window restriction for transition probability computation keeps computational cost manageable, which could make the approach viable for real-time clinical deployment where 3D or global temporal models are too expensive.","The semi-supervised design—temporal constraints from intermediate frames with labels only at end-diastolic and end-systolic frames—suggests a practical path for clinical settings where exhaustive frame annotation is infeasible."],"fun_headline_variants":["Echo Segmentation Corrects Noise via Cardiac Motion Probabilities","Transition Probabilities Guide Echo Segmentation to 93.87% Dice","Pixel Affinities as Motion Priors Improve Heart Ultrasound Segmentation","Cardiac Motion Probabilities Enhance Echocardiography Segmentation","Treating Pixel Affinities as Cardiac Motion Guides Echo Segmentation"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that inter-frame affinity matrices computed via local linear projections and softmax meaningfully represent transition probability distributions of cardiac motion and are more stable than raw features under speckle noise. This is stated conceptually but never empirically validated—no analysis shows the computed transition probabilities correlate with actual cardiac motion or are demonstrably more stable than alternatives.","fun_headline_variants_meta":{"raw":{"variants":["Echo Segmentation Corrects Noise via Cardiac Motion Probabilities","Transition Probabilities Guide Echo Segmentation to 93.87% Dice","Pixel Affinities as Motion Priors Improve Heart Ultrasound Segmentation","Cardiac Motion Probabilities Enhance Echocardiography Segmentation","Treating Pixel Affinities as Cardiac Motion Guides Echo Segmentation"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1247,"prompt_tokens":525,"completion_tokens":722,"prompt_tokens_details":null},"tokens_in":525,"tokens_out":722,"duration_ms":51757,"temperature":1.0,"reasoning_tokens":653,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T06:07:24.233654+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If one replaced the transition probability computation (Eq. 6-7) with a standard local self-attention using raw features as queries, keys, and values—without any motion-probability interpretation—and achieved equivalent segmentation performance, the central claim that transition probability correlations specifically drive the improvement would be undermined.","supporting_citations":[],"review_version":1}