{"id":"7ef70dae-d539-4624-ad6e-5a9cd9cef4d1","arxiv_id":"2504.13415","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A U-Net variant combining dense blocks, dual channel and spatial attention, edge-based skip connections, and deep supervision reports 98% Dice on left ventricle segmentation in ACDC cardiac MRI.","lead":"DADU is a U-Net based model that adds channel and spatial attention, edge-detection skip connections, and deep supervision to segment the left ventricle, right ventricle, and myocardium in cardiac MRI. It reports slightly higher Dice scores and lower contour distances than several earlier models, which matters for automated cardiac diagnosis if the gains hold under controlled comparison.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 baselines are not shown to be comparable; the outperformance claim rests on unstated protocol and mismatched citations, so it is unsupported as written.","rationale":"The reader's verdict is REJECT based on the unverified comparison protocol in Table 3. My independent reading identifies the same load-bearing weakness, with additional supporting evidence: the citation mismatches in Table 3 (U-Net citing [7], AUN citing [19]) strongly suggest that the baseline numbers were not generated in this paper and may come from incompatible evaluations. The paper's own DADU results also lack error bars and there is an internal inconsistency between the stated fivefold cross-validation in Section 5 and the fixed 100/50 split in Section 5.1. Together these make the central claim unsupported as written. I do not see a need to change the reader's verdict: REJECT remains appropriate because the evidence does not establish the claimed superiority. My concern is about the argument's empirical support, not about any authorial intent. The proposed concrete test would settle whether the outperformance claim survives a controlled comparison.","tokens_in":14122,"tokens_out":3157,"duration_ms":30143,"concrete_test":"Obtain the exact 100/50 split used by the authors (or the official ACDC split if that is what was used), and rerun U-Net, FCN, CNN, MRF, UNet++, AUN, and DADU with identical preprocessing, augmentation, training budget, and evaluation code that computes per-structure DSC and maximum Hausdorff distance at ED and ES. If the reproduced baseline numbers differ from Table 3, or if DADU no longer dominates, the outperformance claim fails. As a cheaper analytical check, verify each Table 3 row against its cited source: if the cited paper used a different dataset/test split or did not report that structure/phase, that row is inadmissible for comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DADU outperforms all compared methods in DSC and HD depends entirely on the validity of Table 3. The paper never establishes that the baseline rows were produced under the same protocol as DADU. Section 5.1 describes a single 100-examination training / 50-examination testing split, while Section 5 also states that fivefold cross-validation was employed; these are different evaluation regimes. Table 3 contains internal citation mismatches: the U-Net row cites reference [7], which is Avendi et al.'s right-ventricle CNN/autoencoder paper, not a U-Net method; the AUN row cites [19], a pediatric echocardiography attention-fusion paper, not the Attention U-Net described in the text. The baseline DSC/HD values therefore appear to be taken from heterogeneous sources with unspecified preprocessing, test sets, and metric definitions (e.g., maximum vs. 95th-percentile Hausdorff distance, per-patient vs. per-slice averaging). If any baseline was evaluated on a different subset or with a different HD convention, the comparison in Table 3 is not apples-to-apples, and the headline claim that DADU outperforms other leading techniques is unsupported. No code, trained models, or evaluation scripts are provided, so the DADU column itself cannot be independently verified either.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DADU, a U-Net variant combining dense blocks, channel and spatial attention (CBAM-style), edge-based skip connections, and deep supervision, for semantic segmentation of the left ventricle (LV), right ventricle (RV), and myocardium in the ACDC cardiac MRI dataset. The authors report DSC values up to 0.98 (LV, end-diastole) and Hausdorff distances lower than those of six compared methods in Table 3, and claim that DADU outperforms other leading techniques in both DSC and HD. The manuscript includes architecture diagrams, pseudo-code for the attention modules, implementation details, and a qualitative comparison of segmentation outputs.","tokens_in":14350,"tokens_out":4972,"duration_ms":42519,"significance":"If the reported results are reproducible and the comparisons in Table 3 are apples-to-apples, DADU would be a useful incremental contribution to cardiac MRI segmentation, particularly for the improved Hausdorff distance. The paper evaluates on the external ACDC benchmark and reports quantitative metrics for end-diastolic and end-systolic phases, which is a strength. However, the paper provides no code, trained models, or evaluation scripts, and the central comparative claim rests on baseline numbers whose provenance and protocol are not established. The claimed edge-based skip connection, one of the two stated contributions, is not actually described in the manuscript. As written, the significance of the reported results is therefore not yet established.","major_comments":[{"comment":"The claim that DADU outperforms all other methods in DSC and HD depends entirely on Table 3, but the manuscript does not state whether the U-Net, FCN, CNN, MRF, UNET++, and AUN rows were recomputed under the same protocol or quoted from the cited papers. Section 5 states that fivefold cross-validation was used, while Section 5.1 describes a fixed 100-examination training / 50-examination testing split; the baseline rows are not tied to either regime. Moreover, the U-Net row cites [7], which is Avendi et al.'s right-ventricle CNN/autoencoder paper, and the AUN row cites [19], a pediatric echocardiography segmentation paper, rather than the Attention U-Net reference [17]. Without a controlled re-run under the same training, preprocessing, and evaluation protocol, the outperformance claim is unsupported.","section":"Section 6, Table 3"},{"comment":"Equations (7) and (8) are identical: both define M_{c in C}^{SpE} as f^{7x7}([F_{c in C}^{ChE-avg} • F_{c in C}^{ChE-1x1} • F_{c in C}^{ChE-max}]). One of these should presumably define the decoder map M_{c in C}^{SpD}. As written, the equations do not provide the decoder branch, and Eq. (9) then sums M_{c in C}^{SpE} with itself, so the spatial attention module is not well defined and cannot be implemented from the text.","section":"Section 4.3, Eqs. (7)-(8)"},{"comment":"The abstract and contributions list 'edge-detection based skip-connection' and 'real edge extraction' as central components, and Section 4 says skip connections include derived edge information. However, no equation, algorithm, or textual description specifies how edges are extracted (e.g., Sobel or Canny), how edge maps are fused with the DAB feature maps, or whether the edge branch is trained with a separate loss. This missing technical core makes the method unreproducible and prevents assessment of one of the two stated contributions.","section":"Sections 4 and 4.1"},{"comment":"The evaluation protocol is internally inconsistent: Section 5 states 'We employed a fivefold cross-validation approach,' while Section 5.1 states 'We divided the dataset into a training set comprising 100 examinations and a testing set comprising 50 examinations.' If the reported DSC/HD values are from the 50-examination test set, then the cross-validation sentence is misleading; if they are averaged over folds, the fixed split is misleading. Either way, Tables 1 and 3 cannot be interpreted without clarification of which regime produced the numbers.","section":"Sections 5 and 5.1"},{"comment":"All reported metrics are single point estimates without standard deviations or confidence intervals. For a dataset with 150 examinations and two phases per patient, per-structure performance varies considerably across patients and slices; without variance or a paired significance test, the small HD improvements claimed in Table 3 (e.g., LV ED HD of 1.9 mm vs. 2.9 mm for AUN) cannot be distinguished from noise. The paper also does not specify whether HD is the maximum or 95th-percentile Hausdorff distance used in the ACDC challenge.","section":"Section 6, Tables 1-3"}],"minor_comments":[{"comment":"The Hausdorff distance is defined as a max-min distance, but the ACDC challenge typically uses the 95th-percentile Hausdorff distance; please state explicitly which definition was used.","section":"Section 5.3, Eq. (13)"},{"comment":"The abstract refers to 'myocardium scar tissue,' but the ACDC dataset provides myocardium labels, not scar labels, and the paper does not segment scar tissue; please correct the terminology.","section":"Abstract"},{"comment":"The LMyo end-diastolic HD value is 8.3 mm in Table 1 but 8.2 mm in Table 3 for DADU; please reconcile the rounding or identify which table is correct.","section":"Tables 1 and 3"},{"comment":"In the spatial attention loop, F_i^{SpD-max} = maxPool(F_i^{ChE}) uses the encoder feature for the decoder branch; this should presumably read F_i^{ChD}.","section":"Algorithm 3"},{"comment":"References [15] and [32] are the same CMS-UNet paper and should be consolidated; also several in-text citations in Section 3 and Table 3 do not match the cited works.","section":"References"},{"comment":"Figure 1 is low-resolution and the components described in the text (ED, DAB, deep supervision) are difficult to identify; Figures 7 and 8 would benefit from more detailed captions specifying what each column shows.","section":"Figures 1 and 7-8"},{"comment":"The irregular behavior of HD around epoch 40 is mentioned but not explained; please provide an explanation or remove the anomaly from the discussion.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a conference-style paper with a plausible architecture, but the evaluation and method description are incomplete. I recommend major revision rather than rejection because the identified problems—baseline re-evaluation, the missing edge-detection description, and the equation/algorithm corrections—are addressable in a revision. I would also ask the editor to verify the reference list for citation integrity, as several cited baselines do not match the methods they are claimed to support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an engineering combination of known components—dense blocks, CBAM-style channel/spatial attention, edge-based skip connections, deep supervision—applied to ACDC CMR segmentation. That combination is not in the literature, so there is a small novelty claim, and the model's reported training cost (2.2 hours, 4GB GPU) is useful practical info. The qualitative slices in Figures 6-8 look like a working segmentation.\n\nThe problems start at the central comparison. Table 3 lists baselines U-Net, FCN, CNN, MRF, UNet++, AUN, but the citations don't match the methods: the U-Net row cites an RV-segmentation paper that isn't U-Net, and the AUN row cites a pediatric echocardiography paper rather than Attention U-Net. The paper never says whether baseline numbers were rerun on the same 100/50 split or quoted from papers with different preprocessing, test sets, and Hausdorff definitions. Section 5 says fivefold cross-validation, then Section 5.1 says a 100/50 split. Without a controlled protocol, 'outperforms all other models' is unsupported. On top of that, there are no error bars or significance tests, so the 0.01–0.02 DSC gains over AUN could easily be noise.\n\nTwo more technical soft spots: the edge-detection mechanism, which is listed as a contribution, is never defined or shown anywhere; and equations 7 and 8 are literally identical, which suggests an editing error. The deep supervision loss equation has garbled summation notation. The abstract's '98% DSC' applies only to LV at end-diastole; across classes and phases it's 0.92–0.98, so the headline overstates things.\n\nWhat does the paper do well? It is clearly written at the architectural level, and the idea of adding dual attention and edge guidance to a U-Net is reasonable. For a reader wanting a compact overview of a plausible lightweight cardiac segmentation model, it has some value. But the evaluation that would establish superiority isn't there.\n\nIf this lands on my desk, I would not send it to review in its current form. The authors need to rerun baselines under the same protocol, report variance, specify the edge extraction, and fix the equations. After that it might be a modest but honest application paper. As written, the central claim fails the evidence.","headline":"An incremental U-Net variant whose architecture is plausible but whose headline comparison to prior work is not supported by the experiments as reported.","tokens_in":14893,"tokens_out":3009,"would_cite":false,"duration_ms":27553,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that its DADU model, a U-Net with dual attention, edge-based skip connections, and deep supervision, segments cardiac MRI structures with 98% Dice and lower Hausdorff distances than leading methods.","keywords":["cardiac MRI segmentation","U-Net","dual attention","channel attention","spatial attention","deep supervision","Dice similarity coefficient","Hausdorff distance"],"falsifier":"Re-run U-Net, UNet++, and Attention U-Net on the same 100-exam training and 50-exam test split, the same PNG/resizing preprocessing, and the same DSC and Hausdorff-distance computation used for DADU, and check whether DADU's Table 3 margins persist. If any baseline reaches or exceeds 0.98 DSC and matches or beats DADU's Hausdorff distances under identical conditions, the claimed superiority would not survive.","tokens_in":13905,"feed_emoji":"❤️","tokens_out":9270,"duration_ms":80489,"temperature":0.7,"pith_summary":"The paper sets out to show that a U-Net architecture augmented with DenseNet-style dense blocks, a dual attention block combining channel and spatial attention, edge-detection-based skip connections, and deep supervision can automatically segment the left ventricle, right ventricle, and myocardium from cardiac cine MRI more accurately than existing methods. It reports a mean Dice similarity score up to 0.98 and lower Hausdorff distances than the compared techniques, with the clearest gains on boundary-sensitive contour error. A sympathetic reader would care because reliable automated contours of these heart structures are the first step toward clinical measurements such as ventricular volumes and ejection fraction. The paper presents DADU as both more accurate and light enough to run quickly on modest hardware, which would matter for deployment in clinical settings.","feed_headline":"Heart MRI contours hit 98% Dice with dual-attention U-Net","feed_subtitle":"Edge-aware skip connections and deep supervision shrink contour errors, improving automated cardiac chamber delineation.","key_machinery":"The load-bearing object is the Dual Attention Block (DAB), installed at each skip connection between encoder and decoder feature maps. DAB runs a Channel Attention Module first: it average-pools and max-pools each channel of the encoder map and the decoder map separately, adds the pooled vectors within each side, adds encoder and decoder summaries, and applies a sigmoid to produce channel weights. Those weights multiply the input feature map. Then a Spatial Attention Module average-pools, max-pools, and 1×1-convolves the refined map, concatenates the three spatial descriptors, passes them through a 7×7 convolution, and applies a sigmoid to produce a spatial weight map. The DAB output is the spatially reweighted feature map, and real extracted edges are also fused into the skip connection. Around this core, dense blocks in the encoder reuse features and auxiliary decoder paths contribute to the total loss, implementing deep supervision.","core_discovery":"On its own terms, the paper's central discovery is that combining four mechanisms in one network—dense feature reuse in the encoder, channel and spatial attention at every skip connection, real edge information carried into the decoder, and deep supervision from auxiliary decoder paths—produces segmentation that the authors report as 0.98 Dice for the left ventricle at end-diastole, 0.97 for the right ventricle, and 0.93 for the myocardium, with Hausdorff distances lower than all compared models, especially at end-systole. The paper attributes the edge-quality improvement to the edge-based skip connections and attention-guided refinement, and the training stability to deep supervision. It concludes that DADU outperforms U-Net, FCN, CNN-based methods, MRF, UNet++, and Attention U-Net in both Dice score and Hausdorff distance.","pith_inferences":["Inference: Table 3's baseline numbers appear to be quoted from earlier papers rather than re-run under DADU's 100/50 split and preprocessing; until the comparison is repeated under identical conditions, the claimed margin over Attention U-Net remains unverified.","Inference: No ablation is reported, so the individual contributions of edge skip connections, dual attention, and deep supervision to the Hausdorff-distance gain are untested; an ablation study would isolate which mechanism carries the improvement.","Inference: Because DADU processes 2D slices independently, enforcing through-plane or cardiac-phase consistency could further sharpen end-systolic contours and is a natural next test.","Inference: The same recipe—encoder-decoder with channel-spatial attention and edge-aware skip connections—could transfer to other boundary-sensitive segmentation problems, such as echocardiographic chamber tracing or vessel-wall segmentation."],"forward_implications":["At the reported accuracy, automated LV, RV, and myocardium contours from cine MRI reach near-expert overlap—around 0.98 Dice for the LV at end-diastole—with boundary errors low enough to support downstream volume and ejection-fraction estimates.","The edge-based skip connection recovers boundary detail lost in downsampling, so other U-Net-style segmentation models could adopt this mechanism without changing imaging hardware or acquisition protocols.","Deep supervision plus dense blocks allow the model to train in about 2.2 hours on 4 GB GPU memory and to segment a full phase in 0.2 seconds at test time, which makes clinical deployment plausible.","Because end-diastolic segmentation is consistently more accurate than end-systolic in the reported epochs, the main remaining risk to ejection-fraction accuracy is systolic contour quality, which the low systolic Hausdorff distances are meant to address."],"supporting_citations":[{"why":"Supplies the U-Net encoder-decoder architecture that DADU extends with dense blocks, attention, and edge-based skip connections.","marker":"[9]"},{"why":"Provide the cardiac MRI benchmark dataset with expert masks for left ventricle, right ventricle, and myocardium used for training and testing.","marker":"[30, 31]"},{"why":"Supplies the channel-then-spatial attention design that the Dual Attention Block builds on.","marker":"[24]"},{"why":"Supplies the deep-supervision strategy of auxiliary losses that DADU uses to reduce vanishing gradients.","marker":"[29]"},{"why":"The Attention U-Net baseline whose Dice and Hausdorff-distance results DADU claims to beat in Table 3.","marker":"[19]"},{"why":"The UNet++ baseline with redesigned skip connections that DADU is compared against.","marker":"[18]"},{"why":"The FCN baseline in Table 3 whose performance DADU claims to exceed.","marker":"[11]"},{"why":"The MRF and active-contour baseline in Table 3 whose performance DADU claims to exceed.","marker":"[14]"}],"fun_headline_variants":["Dual-attention U-Net with edge skip-connections hits 98% Dice","Edge-guided attention deep-supervision sharpens heart MRI contours","Attention and edge features refine cardiac MRI segmentation to 98%","Deep-supervised U-Net plus edges and attention: 98% Dice on hearts","Cardiac MRI: Edge-aware attention U-Net reaches 0.98 Dice score"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison numbers in Table 3 for U-Net, FCN, CNN, MRF, UNet++, and Attention U-Net were obtained under the same training split, preprocessing, and evaluation protocol as DADU; the paper does not state this.","fun_headline_variants_meta":{"raw":{"variants":["Dual-attention U-Net with edge skip-connections hits 98% Dice","Edge-guided attention deep-supervision sharpens heart MRI contours","Attention and edge features refine cardiac MRI segmentation to 98%","Deep-supervised U-Net plus edges and attention: 98% Dice on hearts","Cardiac MRI: Edge-aware attention U-Net reaches 0.98 Dice score"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000731,"raw_usage":{"total_tokens":3243,"prompt_tokens":887,"completion_tokens":2356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":2255}},"tokens_in":503,"tokens_out":2356,"duration_ms":17268,"temperature":1.0,"reasoning_tokens":2255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:08:29.287383+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run U-Net, UNet++, and Attention U-Net on the same 100-exam training and 50-exam test split, the same PNG/resizing preprocessing, and the same DSC and Hausdorff-distance computation used for DADU, and check whether DADU's Table 3 margins persist. If any baseline reaches or exceeds 0.98 DSC and matches or beats DADU's Hausdorff distances under identical conditions, the claimed superiority would not survive.","supporting_citations":[{"cited_title":"In: Nawab, N., Hornegger, J., Wells, W","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net encoder-decoder architecture that DADU extends with dense blocks, attention, and edge-based skip connections."},{"cited_title":"S.: CBAM: convolutional block attention module","cited_arxiv_id":null,"evidence_quote":"Supplies the channel-then-spatial attention design that the Dual Attention Block builds on."},{"cited_title":"Y., Xie, S., Gallagher, P., Zhang, Z., & Tu, Z.: Deeply Supervised Nets","cited_arxiv_id":null,"evidence_quote":"Supplies the deep-supervision strategy of auxiliary losses that DADU uses to reduce vanishing gradients."},{"cited_title":"F., Qin, J., et el.: Dual Attention Enhancement Feature Fusion Network for Segmentation and Quantitative Analysis of Pediatric Echocardiography","cited_arxiv_id":null,"evidence_quote":"The Attention U-Net baseline whose Dice and Hausdorff-distance results DADU claims to beat in Table 3."},{"cited_title":"-M.-R., Tajbakhsh, N., Liang, J.: UNET++ : Redesigning Skip Connections to Exploit Multiscale Features in Image Segmentation","cited_arxiv_id":null,"evidence_quote":"The UNet++ baseline with redesigned skip connections that DADU is compared against."},{"cited_title":"A., Krishnamurthi, G.: Fully Convolutional Multi-Scale Residual DenseNets for Cardiac Segmentation and Automated Cardiac Diagnosis using Ensemble of Classifiers","cited_arxiv_id":null,"evidence_quote":"The FCN baseline in Table 3 whose performance DADU claims to exceed."},{"cited_title":"In: Pop, M., Sermesant M., Jodoin, P.-M., Lalande, A., et al","cited_arxiv_id":null,"evidence_quote":"The MRF and active-contour baseline in Table 3 whose performance DADU claims to exceed."}],"review_version":1}