{"id":"0c5dfb4a-d1c6-48c6-8b5d-e18fe31028ac","arxiv_id":"2502.00524","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Deep learning ultrasound beamformers that incorporate lesion classification, especially at the channel-data bottleneck, produce higher-contrast images than standard DAS and MV beamforming.","lead":"This paper trains ultrasound beamformers with deep learning to also classify breast lesions, and adds new data tricks to handle noisy, limited patient data. The best system, CDCB, produced cleaner images with better contrast than standard ultrasound methods, which could help doctors see lesions more clearly.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central clinical-relevance claim is untested: no classification accuracy is reported for JBC or CDCB, so 'best in clinical relevance' rests only on contrast metrics that do not measure diagnostic task value.","rationale":"The reader's weakest-assumption pick, MV surrogate labels, is a legitimate concern, but I see a more immediate gap: the paper never evaluates the clinical task that defines the task-based approach. Section II motivates JBC and CDCB by integrating breast lesion classification, and Eqs. 8-10 train classification heads, yet Section IV reports no classification metrics at all. The abstract and conclusion claim 'clinical relevance' superiority, but the only quantitative evidence is contrast metrics, which are not a proxy for diagnostic accuracy. This is load-bearing because the entire novelty of the paper is task-based beamforming; if the integrated classifiers are never validated, the 'clinical relevance' part of the central claim is unsupported. The MV-surrogate issue compounds this: since the targets are MV images, the DL beamformers are trained to imitate MV, and their contrast gains over MV may come from smoothing rather than clinically meaningful feature enhancement. The proposed check, reporting classification performance on held-out patients, directly settles whether the task-based component contributes clinical value. The existing CONDITIONAL verdict remains appropriate because the concern is addressable and does not invalidate the image-quality comparisons, so I do not move the verdict.","tokens_in":13548,"tokens_out":3482,"duration_ms":37916,"concrete_test":"Obtain the trained CDCB and JBC models and the test set of 4 patients (2 benign, 2 malignant). For each test frame, record the classification logit from the integrated classifiers and compute per-patient accuracy, sensitivity, specificity, and ROC-AUC, with patient-level bootstrapped confidence intervals. Compare against a ViT classifier trained on DAS, MV, and UBB images under the same split. In addition, stratify contrast metrics by patient and report paired differences. If CDCB's classification performance is not significantly better than the image-only classifier, or is near chance, the clinical-relevance claim fails; if classification accuracy is high at the patient level, the claim gains direct support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper defines JBC and CDCB as task-based beamformers and trains them with classification losses (Eqs. 9-10), yet Section IV reports no classification accuracy, sensitivity, specificity, or AUC for either model. The abstract's claim that CDCB 'outperforms others in terms of image quality and clinical relevance' is therefore supported only by contrast metrics (CNR, gCNR, CR) in Table II. These metrics are computed after histogram matching to DAS images and measure pixel statistics within manually drawn ROIs, not diagnostic utility. The comparison against MV is also confounded: MV images are the training targets (Section IV-A), so the DL output is a regularized regression toward MV; improved CNR/gCNR over MV can reflect variance reduction from regression and MS-SSIM smoothing rather than task-relevant feature enhancement. Without a task-level evaluation, the central claim that task-based integration improves clinical value is unverified. A related internal inconsistency strengthens the need for direct task metrics: Table I shows augmentations leave UBB CNR unchanged (5.23 vs 5.23) and slightly reduce UBB CR (8.65 vs 8.60 dB), contradicting the text's claim that augmentations 'significantly improve image quality'.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a deep-learning framework for task-based ultrasound beamforming. It introduces two architectures: JBC, which jointly trains a UNet beamformer with a ViT classifier that provides feedback on the generated image, and CDCB, which adds a classification head to the UNet bottleneck so that classification operates on the channel-data representation. The authors also propose several channel-data augmentations (speckle noise, Channel Data SpecAugment, subsampling/masking, and coarse dropout) and train all models using MV beamformed images as surrogate ground truth. On a test set of 4 in-vivo patients (100 frames each), the reported contrast metrics (CNR, gCNR, CR) indicate that DL-based beamformers outperform DAS and MV, and that CDCB achieves the best contrast among all methods.","tokens_in":13703,"tokens_out":4562,"duration_ms":48332,"significance":"If validated, the idea of coupling a clinical classification task directly to the beamforming optimization, and in particular at the channel-data bottleneck, is a useful and timely contribution to DL-based ultrasound imaging. The channel-data augmentation recipes are described in sufficient detail to be reproducible, and the use of real in-vivo breast ultrasound data is a strength. However, the manuscript currently does not provide task-level evidence for the central 'clinical relevance' claim: no classification accuracy, sensitivity, specificity, or AUC is reported for JBC or CDCB. The comparison against MV is also confounded because MV images serve as the training target, and the statistical analysis pools correlated frames from only four patients. These are load-bearing gaps, but they are addressable in revision, so the contribution is worth further consideration rather than rejection.","major_comments":[{"comment":"The claim in Section IV-C that 'the integration of channel data augmentations led to a notable improvement in performance' is not supported by Table I for the baseline UBB: CNR is unchanged (5.23 vs. 5.23 dB) and CR decreases slightly (8.65 vs. 8.60 dB); JBC CR also decreases (8.58 vs. 8.31 dB). Only CDCB improves on all three metrics. Because the augmentation contribution is one of the two headline contributions, the authors should either report a statistical test that justifies 'significantly improves' or limit the claim to the specific configurations where the improvement is consistent.","section":"IV-C, Table I"},{"comment":"The central claim that task-based beamformers improve 'clinical relevance' is unverified: no classification accuracy, sensitivity, specificity, or AUC is reported for JBC or CDCB, even though Eqs. (8)-(10) define classification losses and the clinical task is breast-lesion classification. The reported CNR/gCNR/CR are pixel-statistics contrast measures computed on manually drawn ROIs, not task-performance measures. The authors should report held-out classification metrics for the ViT and the CDCB classifier head, and ideally task-level performance on the beamformed images, before claiming clinical relevance.","section":"Abstract, IV-C, V"},{"comment":"The comparison against MV is confounded because MV beamformed images are used as the training target (Section IV-A), yet MV appears as a baseline in Table II. Any DL model trained with an MSE+MS-SSIM loss toward MV can show lower variance and higher contrast than MV itself, so the reported gains over MV do not by themselves establish that task integration is the cause. The authors should address this circularity, for example by reporting (i) the distance between DL outputs and MV in terms of the training loss, (ii) performance of the UBB baseline trained on DAS targets, or (iii) task-level metrics that are not defined by similarity to MV.","section":"IV-A, Table II"},{"comment":"The statistics are not adequate for the claims. The test set consists of 4 patients and 100 frames per patient, and the metrics appear to pool all 400 frames as independent samples. Frames from the same patient are highly correlated, so the standard deviations and point estimates in Tables I-II cannot support conclusions about methods without patient-level aggregation or paired tests. The differences among DL methods (e.g., CDCB CNR 5.65 vs. UBB 5.23 dB) are smaller than the reported standard deviations, and no significance testing is reported. Please report per-patient results, confidence intervals, and a paired test across patients, and state explicitly that the 4 test patients are disjoint from the training patients.","section":"IV-C"}],"minor_comments":[{"comment":"The contrast metrics are computed 'after histogram matching to DAS images'; this normalization should be justified, since it can reduce or remove absolute contrast differences between methods and may affect the reported CR/CNR values in favor of methods whose histograms are closer to DAS.","section":"IV-B"},{"comment":"The scalar γ is described as balancing the beamforming loss and the classification feedback loss, but no sensitivity analysis is provided; given the very different scales of MSE and cross-entropy, a brief discussion or ablation of γ would improve confidence in the reported results.","section":"II-C, Eq. (9)"},{"comment":"The indexing in Algorithm 1 is confusing: the line '˜Xi,˜j,k,l ← Xi,j,k,l' with the condition 'j ∈ {p | p ∈ [C] ∧ p|S}' does not clearly define which channels are sliced; please rewrite this step with explicit index sets.","section":"III, Algorithm 1"},{"comment":"The augmentation probabilities are specified in prose (e.g., speckle noise with p=0.5, Gaussian noise with p=1/3, coarse dropout with p=1/3), but a precise schedule of which augmentations are applied per batch, and whether they are mutually exclusive or composed, is needed for reproducibility.","section":"III"},{"comment":"The text in Section IV-D refers to 'lesion detection' although the clinical task throughout the paper is lesion classification; please use consistent terminology.","section":"IV-D"},{"comment":"There is a typo in the introduction: 'Weiner Beamforming' should be 'Wiener Beamforming.' Also, reference [31] appears to be a duplicate of the arXiv preprint cited elsewhere; please unify the citation.","section":"I"}],"recommendation":"major_revision","confidential_remarks":"The paper has a potentially useful contribution, but the evidence for its central claims is currently incomplete: no classification results are reported, the MV-target circularity is unaddressed, and the statistical analysis is based on a small number of correlated frames. These are fixable with additional experiments and revised wording, so I recommend major revision rather than rejection. If a revision adds patient-level task metrics and addresses the augmentation inconsistencies, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible extension of deep-learning beamforming with two genuinely new pieces—the CDCB bottleneck classifier and the Channel Data SpecAugment—and real in-vivo breast data. The contrast metrics do show the DL beamformers beating DAS and MV on their test set. But the headline claim about clinical relevance is not supported, because the paper never reports classification accuracy for either task-based model. The stress-test note is right: no sensitivity, specificity, or AUC anywhere.\n\nWhat's good: the CDCB idea is clean—classifying the channel-data representation at the bottleneck as a regularizer is more direct than feeding back through a full image classifier. The augmentation set is thoughtful, especially the speckle-noise adaptation to channel data and the SpecAugment extension to spatial/temporal dimensions. The paper is honest about using MV as surrogate labels, which is standard in this literature. And the qualitative figures show real differences.\n\nThe soft spots are proportionate. The missing classification results are the big one: the entire 'task-based' motivation is about the clinical task, and contrast metrics after histogram matching do not measure that. The test set is four patients with 100 frames each, all pooled; no per-patient or per-lesion statistics, so the standard deviations are over frames, not patients. That is a real statistical issue. The MV-as-target-and-baseline comparison is partly circular: the network regresses to MV, so beating MV on contrast may just reflect regression-to-mean smoothing. Also, Table I contradicts the text's claim that augmentations 'significantly improve image quality': UBB CNR is unchanged at 5.23 and CR actually drops from 8.65 to 8.60 dB. The JBC CR also drops. So that claim needs softening.\n\nThese are all fixable. The architecture and augmentations deserve a serious look, but the paper as written overclaims. I'd send it to review, but it needs major revisions: report classification performance, do patient-level analysis, and temper the augmentation claim. It's worth engaging with rather than desk-rejecting.","headline":"Plausible task-based beamforming extension with a genuinely new CDCB bottleneck classifier and Channel Data SpecAugment, but the clinical-relevance claim is untested because no classification results are reported.","tokens_in":14341,"tokens_out":2547,"would_cite":false,"duration_ms":24516,"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":"Task-based deep learning beamformers—especially one that classifies at the channel-data bottleneck—produce breast ultrasound images with better lesion contrast and clinical relevance than DAS, MV, or task-free deep beamforming.","keywords":["ultrasound beamforming","task-based beamforming","deep learning","channel data augmentation","breast lesion classification","minimum variance beamforming","UNet","speckle noise"],"falsifier":"Train CDCB, JBC, and UBB on simulated or ex vivo channel data for which a true scattering ground truth exists, and compare them against MV on that ground truth; if CDCB does not beat MV when the target is not an MV image, the reported superiority depends on the surrogate label rather than on task-based learning.","tokens_in":13270,"feed_emoji":"🩺","tokens_out":5889,"duration_ms":53004,"temperature":0.7,"pith_summary":"This paper tries to show that a deep-learning ultrasound beamformer can be improved by making it do a clinical task while forming the image. Using breast-lesion classification as the task, the authors train two variants: one that classifies the beamformed image and feeds that back, and one that classifies a compressed channel-data representation inside the network's bottleneck. They report that both beat conventional delay-and-sum and minimum-variance beamforming, and that the channel-data classifier (CDCB) is the best. The reason to care is that clinical ultrasound is judged by whether lesions are visible and diagnosable, not just by pixel accuracy against a reference. If the claim holds, embedding task information at the raw-data level is a way to get more diagnostically useful images without a higher-quality physical reference.","feed_headline":"Task-aware deep beamformer beats standard ultrasound imaging","feed_subtitle":"Adding a lesion-classification head inside the network lifts contrast and diagnostic relevance on real breast scans.","key_machinery":"The load-bearing object is the UNet-based beamformer with a classification task attached at two possible points. In CDCB, a ChannelAttention module at the bottleneck reweights the compressed channel-data tensor, and convolutional layers plus a fully connected layer classify it as benign or malignant while the decoder still produces the beamformed image; this classification loss acts as a regularizer that highlights clinically relevant features. In JBC, a ViT image classifier is trained alongside the beamformer and its cross-entropy feedback is added to the beamforming loss. Both are trained to regress to MV beamformed images using MSE plus MS-SSIM loss, and the channel-data augmentations are applied to the raw input during training.","core_discovery":"The central claim, stated in the paper's own terms, is that task-based beamforming—training the beamformer jointly with a lesion-classification objective—yields ultrasound images with higher contrast between lesion and surrounding tissue, and greater clinical relevance, than conventional DAS, MV, or a UNet beamformer trained only to regress to MV images. The CDCB variant, which places the classification head at the UNet bottleneck and thus constrains the channel-data representation, outperforms the JBC variant, which classifies the output image. The paper also claims that its new channel-data augmentations (speckle noise, Gaussian noise, Channel Data SpecAugment, subsampling and masking, and coarse dropout) improve image quality, especially for CDCB, as measured by CNR, gCNR, and CR after histogram matching. Across all comparisons, the CDCB model is reported as the best performer in both image quality and clinical relevance.","pith_inferences":["Editorial inference: the CDCB gains could also come partly from added parameters and regularization at the bottleneck rather than from the clinical task itself; an ablation with a randomly labeled classification head would separate the two.","Editorial inference: if the breast-lesion classifier at the bottleneck is actually learning meaningful features, the bottleneck representation could double as a feature extractor for downstream tasks such as lesion segmentation or BI-RADS scoring, which the paper does not test.","Editorial inference: because MV images are the training target throughout, the method inherits MV's bias; the clinical-relevance claim would be stronger if validated against biopsy-confirmed outcomes or an independent radiologist rating rather than contrast metrics alone."],"forward_implications":["If the central claim is right, clinical task feedback at the channel-data level (CDCB) is a stronger regularizer than feedback at the image level (JBC), so future beamformers should attach task heads deep in the network rather than only at the output.","If CDCB truly beats MV while remaining a single forward pass, it offers a route to real-time beamforming with higher contrast than the computationally heavy MV method.","If the channel-data augmentations are what make the difference, similar corruptions (speckle, masking, and spectral masking) can be used to train other channel-data-based ultrasound networks when in-vivo data are scarce.","If the framework generalizes as the paper suggests, the same dual-beamformer-plus-task design can be applied to segmentation, detection, or regression tasks beyond breast-lesion classification."],"supporting_citations":[{"why":"Establishes the practice of using MV beamformed images as surrogate training targets for deep-learning beamformers.","marker":"[8]"},{"why":"Also cited for using MV images as surrogate labels and for end-to-end reconstruction, reinforcing the training target choice.","marker":"[31]"},{"why":"Supplies the speckle noise simulation algorithm that the paper adapts to channel data.","marker":"[33]"},{"why":"Provides the SpecAugment time-frequency masking idea that Channel Data SpecAugment extends to ultrasound channel data.","marker":"[34]"},{"why":"Contributes the subsampling of channel data in time and element dimensions that the paper adopts as augmentation.","marker":"[25]"},{"why":"Defines SSIM and MS-SSIM used in the beamforming loss to preserve luminance, contrast, and structure.","marker":"[32]"},{"why":"Defines the minimum-variance beamformer used as the training target and as one of the baselines to beat.","marker":"[3]"}],"fun_headline_variants":["Task-based deep beamformer improves ultrasound image contrast","Channel-level classifier in beamformer yields best ultrasound images","Ultrasound beamforming with lesion task beats conventional methods","Data augmentation boosts deep beamformer for ultrasound imaging","Joint classifier and beamformer enhances clinical relevance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole training pipeline regresses to MV beamformed images as ground truth, so if MV images are not a reliable target for breast-lesion appearance, any apparent improvement over MV could be partly learned from MV's own imperfections.","fun_headline_variants_meta":{"raw":{"variants":["Task-based deep beamformer improves ultrasound image contrast","Channel-level classifier in beamformer yields best ultrasound images","Ultrasound beamforming with lesion task beats conventional methods","Data augmentation boosts deep beamformer for ultrasound imaging","Joint classifier and beamformer enhances clinical relevance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000368,"raw_usage":{"total_tokens":1977,"prompt_tokens":952,"completion_tokens":1025,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":966}},"tokens_in":568,"tokens_out":1025,"duration_ms":10371,"temperature":1.0,"reasoning_tokens":966,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T18:39:37.573516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CDCB, JBC, and UBB on simulated or ex vivo channel data for which a true scattering ground truth exists, and compare them against MV on that ground truth; if CDCB does not beat MV when the target is not an MV image, the reported superiority depends on the surrogate label rather than on task-based learning.","supporting_citations":[{"cited_title":"Adaptive ultrasound beamforming using deep learning,","cited_arxiv_id":null,"evidence_quote":"Establishes the practice of using MV beamformed images as surrogate training targets for deep-learning beamformers."},{"cited_title":"End-to-end learning-based ultrasound reconstruction,","cited_arxiv_id":null,"evidence_quote":"Also cited for using MV images as surrogate labels and for end-to-end reconstruction, reinforcing the training target choice."},{"cited_title":"Speckle simulation based on b-mode echographic image acquisition model,","cited_arxiv_id":null,"evidence_quote":"Supplies the speckle noise simulation algorithm that the paper adapts to channel data."},{"cited_title":"Deep-learning based adaptive ultrasound imaging from sub-Nyquist channel data,","cited_arxiv_id":null,"evidence_quote":"Contributes the subsampling of channel data in time and element dimensions that the paper adopts as augmentation."},{"cited_title":"Image quality assessment: from error visibility to structural similarity,","cited_arxiv_id":null,"evidence_quote":"Defines SSIM and MS-SSIM used in the beamforming loss to preserve luminance, contrast, and structure."}],"review_version":1}