{"id":"d533e288-423a-4e96-b566-8bf7f83418df","arxiv_id":"2411.11116","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DBF-Net, a dual-branch network with body and boundary supervision plus feature fusion, reports Dice scores of 81.05%, 76.41%, and 87.75% on BUSI, UNS, and UHES ultrasound datasets.","lead":"This paper introduces DBF-Net, a deep learning model that segments lesions in ultrasound images by training two branches: one for the lesion body and one for its boundary, combined with a feature fusion module. It reports modest accuracy gains over existing methods on three public ultrasound datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim depends on comparing DBF-Net against six 'official' baselines from other papers and a 224x224 transformer table; under an equalized protocol, the reported margins (81.05 vs 80.65 on BUSI) are within the noise.","rationale":"I reviewed the paper in good faith and the architecture is plausible: the dual-branch body/boundary supervision and the FFS module are clearly described, and the ablations show internal improvements. The single most load-bearing weakness is that the experimental evidence for the central claim of state-of-the-art performance rests on a comparison protocol with unequal conditions. Six of the BUSI baselines were not re-run by the authors, and the transformer comparison was done at a lower input resolution with settings the authors admit are not optimal for transformers. The margins are small and no significance tests are reported, so the conclusion 'outperforms existing approaches' is not established by the data as presented. The reader's weakest_assumption identified the same issue, and I agree with the conditional verdict: the paper should be accepted only after all baselines are re-run under identical protocols, the code is released, and the text is cleaned up. A secondary issue worth flagging is Eq. (12), which as written assigns infinite weight to background pixels because Gi=0 gives log(1)=0 in the denominator; the reported training cannot have used the printed formula without an unstated modification, so the released code should be checked as well. This does not change my overall assessment: the concern is about evidence, not about the method's internal logic, and the reader's conditional verdict remains appropriate.","tokens_in":15069,"tokens_out":3959,"duration_ms":40130,"concrete_test":"Re-run the six 'official' baselines in Table 2 (MSSA-Net, V2-CE-CD, HEAT-Net, EHA-Net, DSEU-net, NU-net) on BUSI using the paper's own five-fold splits, 512x512 resizing, and the same augmentation and loss schedule, and re-run the Table 3 transformers at 256x256 or 512x512 instead of 224x224. If, after this equalization, DBF-Net's DSC is not above the best baseline by more than one standard deviation, or if any baseline outperforms it, the state-of-the-art claim in Sections 3.3 and 3.4 and the Conclusion is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2's central evidence for state-of-the-art status on BUSI includes six methods marked with '*' as 'official results' (MSSA-Net, V2-CE-CD, HEAT-Net, EHA-Net, DSEU-net, NU-net), so those numbers come from other papers rather than from this paper's five-fold cross-validation, 512x512 resizing, augmentation, and loss settings. The only in-pipeline comparisons are U-Net, DeepLabV3+, LinkNet, UNeXt, and DBBS-Net, and the largest comfortable margin there is small: 81.05 vs 80.73 against the best ablation, and 81.05 vs 80.65 against the official MSSA-Net. Table 3 compares transformer baselines at 224x224 input resolution after the main experiments use the datasets' standard sizes, and the Discussion itself concedes the transformer training settings may be suboptimal. No statistical significance testing is reported. Because the headline 'outperforms existing approaches' is inferred from these cross-protocol and under-powered comparisons, the claim is load-bearing on an equality of protocols that the paper does not establish.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DBF-Net, a dual-branch encoder-decoder architecture for ultrasound image segmentation that explicitly models lesion body and boundary information. It introduces a Feature Fusion and Supervision (FFS) block that pre-generates body and boundary features, fuses them with a trainable parameter, and supervises both intermediate maps and the final segmentation. The method is evaluated on three public datasets (BUSI, UNS, UHES) with five-fold cross-validation, reporting DSC of 81.05±10.44%, 76.41±5.52%, and 87.75±4.18%, respectively, and is compared with CNN-based and transformer-based methods, with ablations on BUSI. The paper claims state-of-the-art performance and releases code.","tokens_in":15301,"tokens_out":3371,"duration_ms":58408,"significance":"If the empirical claims were established under a uniform evaluation protocol, the contribution would be a modest but useful addition to medical image segmentation: explicit interaction between body and boundary streams, a lightweight architecture (3.2M parameters), and public code. The idea of fusing boundary and body features with learnable weights is reasonable and the ablations show a consistent trend. However, the central 'state-of-the-art' claim currently rests on comparisons that are partly cross-protocol, partly under-powered, and not statistically tested; the transformer comparison is conducted at a different input resolution and even the paper's own discussion concedes suboptimal training settings. The evidence is therefore not yet sufficient to support the headline claim.","major_comments":[{"comment":"The state-of-the-art claim relies on six methods marked with '*' whose results are quoted as 'official results' from other papers, not re-run under this paper's five-fold cross-validation, 512x512 resizing, augmentation, and loss settings. The only fully in-pipeline competitors are U-Net, DeepLabV3+, LinkNet, UNeXt, and DBBS-Net. The margin over the best quoted official result (81.05 vs 80.65 on BUSI for MSSA-Net) is smaller than the reported standard deviations, and no statistical significance test is provided. Because the central claim is that DBF-Net 'outperforms existing approaches', this is a load-bearing protocol gap: the claim should be re-supported by re-running the baselines under identical conditions, or by restricting the claim to in-pipeline comparisons and adding significance tests.","section":"3.3, Table 2"},{"comment":"The transformer comparison is performed at 224x224 input resolution, whereas the main experiments use 512x512 (BUSI), 580x420 (UNS), and 448x256 (UHES) resolutions; the Discussion in Section 4 itself concedes that the transformer training settings may be suboptimal. Moreover, the DBF-Net results in Table 3 differ drastically from those in Table 2 for the same dataset (e.g., 55.49 vs 81.05 DSC on BUSI), confirming that the comparison is not protocol-equivalent. Table 3 therefore does not support the stated claim that DBF-Net outperforms transformer-based methods; it should be framed as a low-resolution auxiliary experiment, or all methods should be run at the same resolution.","section":"3.4, Table 3"},{"comment":"The ablation conclusions are based on small performance differences relative to the reported standard deviations. For example, the full configuration with two FFS blocks and feature fusion achieves 81.05±10.44, while the one-FFS-with-fusion configuration achieves 80.73±7.68; the improvement is well within one standard deviation. No paired significance tests or confidence intervals are reported, so the claim that the two-FFS configuration is 'optimal' is not statistically supported. This matters because the ablation is the primary evidence for the contribution of the FFS module.","section":"3.6, Tables 5 and 6"}],"minor_comments":[{"comment":"The model name is inconsistent: the title, methods, and experiments use 'DBF-Net', while the abstract and conclusion use 'UBBS-Net'. A conference-paper baseline named 'DBBS-Net' also appears in Table 2, making the naming confusion particularly problematic. Please unify the terminology.","section":"Title, Abstract, Conclusion"},{"comment":"The text 'In accordance with the configuration specified in the UNeXt [ ? ]' contains an unresolved citation placeholder '[ ? ]'; this should be fixed before publication.","section":"3.2, BUSI description"},{"comment":"Equation (9) states that L* is given by Lwbce and 'Dice loss', but the Dice term is not explicitly named in the sentence; please introduce Ldice explicitly before the equation.","section":"2.5, Equation (9)"},{"comment":"The P-R and ROC curves are rendered at low resolution with garbled axis labels and legend text, making them difficult to read. Higher-resolution figures with legible annotations are needed.","section":"Figures 5 and 7"},{"comment":"The learned fusion parameter λ is a post-hoc observation from five training runs, not a prediction, and the claim that FFM 'harmonizes' the weights is based on small numerical differences without variance or statistical support. Please either soften this interpretation or provide additional evidence.","section":"Table 7"},{"comment":"The manuscript header reads 'Preprint submitted to Nuclear Physics B', which appears to be a template artifact and is inconsistent with the medical imaging content; this should be corrected.","section":"Header"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a journal extension of the authors' BIBM 2022 work (DBBS-Net), and the novelty with respect to that prior publication is not sharply delineated. The header indicating submission to Nuclear Physics B and the unresolved citation placeholder suggest the manuscript was assembled from a template and needs careful editorial cleanup. The central empirical claim depends on protocol-equivalent re-evaluation of baselines, which is feasible within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plain take: this is an incremental but sensible extension of the authors' own BIBM 2022 DBBS-Net, with a simple feature-fusion module, tested on three public ultrasound datasets. The main results are probably real, but the paper's 'state-of-the-art' claim is not supported by the evidence as presented.\n\nWhat's new: the FFS module (residual addition of boundary features into body stream and vice versa) plus a trainable λ for fusion, combined with body/boundary deep supervision. The design is simple and the ablation study shows each piece adds something, though the gains are small. The paper deserves credit for running five-fold cross-validation on three public datasets (BUSI, UNS, UHES) and for being honest in the Discussion: it concedes that the transformer comparisons used suboptimal training settings and that its own ROC/AUC is worse than transformers. That honesty is genuine evidence of good faith.\n\nSoft spots: the central SOTA claim rests on Table 2, which mixes six 'official results' from other papers (MSSA-Net, V2-CE-CD, HEAT-Net, EHA-Net, DSEU-net, NU-net) with the authors' own pipeline. Those numbers are not produced under the same pre-processing, augmentation, five-fold splits, or loss settings, so they are not directly comparable. The only in-pipeline comparisons show DBF-Net at 81.05 vs 80.73 for the best ablation, and vs 80.65 for official MSSA-Net; those margins are well within the reported standard deviations. No statistical significance tests are provided. Table 3 compares transformers at 224x224 input resolution while the main experiments use 512x512 (and 448x256 for UHES), and the Discussion admits the settings may be suboptimal—so that table is weak evidence at best. The trainable λ in Table 7 is a post-hoc observation, not a predictive experiment; that is a minor issue. There is also a naming inconsistency: the abstract says UBBS-Net while the title and rest of the paper say DBF-Net, which suggests a sloppy revision.\n\nWho this is for: a reader working on ultrasound segmentation or boundary-aware medical segmentation will find this a reasonable incremental baseline and a useful comparison point. It is not a breakthrough and does not open a new direction. The code link is provided (though I haven't verified it).\n\nRecommendation: send to peer review with a request for major revision. The method is sensible and the experiments are mostly reproducible in spirit, but the SOTA claim needs to be either backed by equalized baseline runs and significance testing, or softened. If the authors cannot re-run the official baselines, they should present their in-pipeline comparisons as the primary evidence.","headline":"A modest, plausible dual-branch ultrasound segmentation net whose SOTA claim is undercut by cross-protocol baseline comparisons and missing significance tests.","tokens_in":15883,"tokens_out":2615,"would_cite":false,"duration_ms":22177,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dual-branch network that explicitly models how lesion boundaries relate to lesion interiors achieves higher Dice scores than prior methods on three public ultrasound datasets.","keywords":["ultrasound image segmentation","dual-branch network","feature fusion","boundary segmentation","body-boundary relationship","Dice similarity coefficient","deep learning"],"falsifier":"Re-run every baseline listed in Table 2 on the same five-fold splits, input resolutions, and training schedule as DBF-Net; if the best alternative matches or exceeds DBF-Net's Dice score within one standard deviation on BUSI (81.05±10.44%), the paper's central claim of superiority is not supported.","tokens_in":14832,"feed_emoji":"🩺","tokens_out":6028,"duration_ms":54220,"temperature":0.7,"pith_summary":"Ultrasound images make lesion boundaries hard to distinguish, and the paper claims that segmentation improves when a network explicitly learns how a lesion's boundary relates to its interior (the 'body'). To test this, the authors build DBF-Net, a dual-branch encoder-decoder where one branch segments the lesion body and another segments the boundary, with a feature fusion module that lets the two branches exchange information. On three public datasets—breast cancer (BUSI), brachial plexus nerves (UNS), and infantile hemangioma (UHES)—the network reports Dice coefficients of 81.05%, 76.41%, and 87.75%, outperforming the CNN-based and transformer-based methods compared. A careful reader would care because boundary quality matters clinically, and the paper offers a lightweight way to improve it without a large parameter increase.","feed_headline":"Fusing body and boundary maps lifts ultrasound segmentation","feed_subtitle":"Adding a body-plus-boundary branch with feature fusion pushes Dice scores to 81–88% on breast, nerve, and hemangioma scans.","key_machinery":"The load-bearing mechanism is the Feature Fusion and Supervision (FFS) block, placed at two decoder stages. Each FFS block takes concatenated encoder and decoder features, splits them into two parallel convolutional streams that pre-generate body features and boundary features, then exchanges information through a feature fusion module. The fused body and boundary maps are each supervised with a combined weighted cross-entropy and Dice loss, and the final output is a trainable-weighted sum $\\hat{F} = \\lambda F^*_{\\text{body}} + F^*_{\\text{bound}}$ with $\\lambda$ initialized to 1 and learned during training. The body and boundary ground truths are generated from the original mask by a distance transform with a threshold $\\alpha = 1$.","core_discovery":"The paper asserts that the key to better ultrasound segmentation is to treat boundary and body as two complementary supervised tasks under a shared encoder, then fuse their features at multiple scales. This design, implemented through the Feature Fusion and Supervision (FFS) block, consistently beats the considered baselines on all three datasets, and the ablation study shows that both the auxiliary body/boundary supervision and the trainable-parameter fusion contribute to the gain. The paper further claims that this is achieved with only 3.2 million parameters, about a third of U-Net's parameter count.","pith_inferences":["The body-boundary fusion idea could be extended to 3D ultrasound volumes by replacing the 2D convolution branches with 3D convolutions, a step the paper does not discuss.","The distance-transform threshold $\\alpha$ that defines the boundary is fixed at 1 pixel; tuning $\\alpha$ per dataset or lesion size might change results, which the paper does not explore.","The learned $\\lambda$ values staying near 1 when fusion is enabled suggests the fusion module acts as a regularizer that stabilizes the relative weight of the boundary branch; removing fusion leads to $\\lambda$ growing above 1, implying the body branch dominates."],"forward_implications":["If the claim holds, explicitly coupling boundary and body segmentation is a reliable way to improve accuracy on noisy, low-contrast imaging modalities like ultrasound.","The same dual-branch supervision and fusion design can be transferred to other boundary-ambiguous segmentation tasks, including CT, MRI, or optical coherence tomography.","The trainable fusion parameter $\\lambda$ offers a principled way to balance boundary and interior cues, and could be studied as a per-image or per-patient variable.","Because DBF-Net uses only 3.2 million parameters, it is plausible to deploy it in real-time ultrasound systems without specialized hardware."],"supporting_citations":[{"why":"U-Net encoder-decoder baseline that DBF-Net extends and compares against.","marker":"[7]"},{"why":"OC-Net supplies the ASP OC multi-scale context module used in the encoder.","marker":"[24]"},{"why":"DeepLabV3+ baseline for comparison.","marker":"[25]"},{"why":"DBBS-Net, the authors' prior dual-branch network, which DBF-Net improves with feature fusion.","marker":"[26]"},{"why":"UNeXt baseline for lightweight segmentation.","marker":"[27]"},{"why":"BUSI breast ultrasound dataset, one of the three evaluation sets.","marker":"[28]"},{"why":"UNS brachial plexus nerve ultrasound dataset, one of the three evaluation sets.","marker":"[29]"}],"fun_headline_variants":["Dual-branch fusion network sharpens ultrasound lesion boundaries","Body-boundary fusion boosts ultrasound segmentation accuracy","Feature-fused dual branch lifts ultrasound lesion Dice scores","Two-branch feature fusion improves ultrasound lesion boundary detection","Ultrasound segmentation gains from body-boundary feature fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that DBF-Net outperforms existing methods assumes that the comparison numbers for the other methods were obtained under the same training and evaluation conditions, yet several baseline results were taken from other papers without rerunning, and the transformer comparison was run at a lower input resolution.","fun_headline_variants_meta":{"raw":{"variants":["Dual-branch fusion network sharpens ultrasound lesion boundaries","Body-boundary fusion boosts ultrasound segmentation accuracy","Feature-fused dual branch lifts ultrasound lesion Dice scores","Two-branch feature fusion improves ultrasound lesion boundary detection","Ultrasound segmentation gains from body-boundary feature fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1329,"prompt_tokens":816,"completion_tokens":513,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":437}},"tokens_in":432,"tokens_out":513,"duration_ms":5602,"temperature":1.0,"reasoning_tokens":437,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:53:11.399828+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run every baseline listed in Table 2 on the same five-fold splits, input resolutions, and training schedule as DBF-Net; if the best alternative matches or exceeds DBF-Net's Dice score within one standard deviation on BUSI (81.05±10.44%), the paper's central claim of superiority is not supported.","supporting_citations":[{"cited_title":"Ronneberger, P","cited_arxiv_id":null,"evidence_quote":"U-Net encoder-decoder baseline that DBF-Net extends and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"OC-Net supplies the ASP OC multi-scale context module used in the encoder."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DeepLabV3+ baseline for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DBBS-Net, the authors' prior dual-branch network, which DBF-Net improves with feature fusion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"UNeXt baseline for lightweight segmentation."},{"cited_title":"Al-Dhabyani, M","cited_arxiv_id":null,"evidence_quote":"BUSI breast ultrasound dataset, one of the three evaluation sets."},{"cited_title":"Montoya, Hasnin, shirzad, W","cited_arxiv_id":null,"evidence_quote":"UNS brachial plexus nerve ultrasound dataset, one of the three evaluation sets."}],"review_version":1}