{"id":"b9a0c96c-ce95-41a0-9b60-24d68d3243f2","arxiv_id":"2607.20691","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using lesion masks to ground concept activations in anatomically relevant zones improves both diagnosis and spatial faithfulness of concept bottleneck models on breast ultrasound.","lead":"This paper adds spatial supervision from lesion masks to a concept bottleneck model for breast ultrasound, improving diagnostic accuracy and concept localization. It matters because it shows how weak anatomical labels can make medical AI explanations more trustworthy, and why annotation quality should be tracked.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Spatial-faithfulness claim is largely circular: the grounding metrics are the training objective, so independent validation is needed.","rationale":"The reader's stated weakest assumption is the posterior band assumption, but the rationale explicitly flags the circularity of the grounding metrics. I agree that the posterior band is a secondary risk; the more load-bearing concern is that the spatial-alignment metrics are the training objective itself. This makes the strongest reported effect—the five-fold increase in ROI Energy— an artifact of optimization rather than evidence of anatomical plausibility. The diagnostic and concept AUROC gains are small and lack significance testing, so the central claim's 'simultaneously improves' depends on the circular spatial metric. A mis-specified-zone control or independent pixel-level annotation would settle whether SG-CBM truly improves spatial faithfulness. The paper is otherwise coherent and the method is a reasonable incremental contribution, so a CONDITIONAL verdict (requiring such validation) remains appropriate; my read does not change the reader's verdict.","tokens_in":9841,"tokens_out":2894,"duration_ms":26980,"concrete_test":"Train SG-CBM with intentionally mis-specified zones (e.g., posterior band placed above the lesion, or ROI shifted laterally). Evaluate all grounding metrics on the test set against both the mis-specified zones and the correct anatomical zones. If the model reports high Energy-in-Zone / Hit@1 for the trained (wrong) zones but low for the correct zones, the metrics merely reflect the training supervision, not anatomical faithfulness. Alternatively, obtain independent pixel-level annotations of posterior shadowing/enhancement from a sonographer on a held-out set and compute Energy-in-Zone against those; if SG-CBM aligns no better than a baseline trained with random zones, the spatial faithfulness claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of \"markedly increasing spatial alignment\" rests on metrics that are direct terms of the training loss. Energy-in-Zone (Eq. 8) is exactly the mass concentration term (Eq. 4) evaluated on test images, and Hit@1 / Top-5% Overlap are immediate consequences of the separation loss (Eq. 3). Since SGD explicitly optimizes these quantities, the jump from ROI Energy 0.149 to 0.815 is expected and does not establish anatomical plausibility. The posterior band assumption in Sec. II-A compounds this: if the fixed-size band does not match true posterior acoustic regions, the loss enforces mis-localization and the metrics measure conformity to an imposed prior, not clinical correctness. Diagnostic AUROC improvement (0.869±0.040 → 0.892±0.012) is modest with overlapping standard deviations and no significance testing; concept AUROC gain is small (0.741→0.771). Thus the only strong reported effect is the circular spatial alignment. Without an independent spatial-grounding evaluation, the claim that SG-CBM improves spatial faithfulness is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SG-CBM, a spatially grounded Concept Bottleneck Model for breast ultrasound diagnosis. A lesion mask is used to derive two zones: an in-lesion ROI for morphology-related concepts and a size-adaptive posterior band for posterior acoustic concepts. Concept maps are trained with a grouped grounding loss (separation plus mass concentration) while diagnosis is mediated by a linear classifier over concept logits. Experiments on the BrEaST dataset with 5-fold stratified group cross-validation report that SG-CBM improves diagnostic AUROC and concept macro-AUROC over a vanilla CBM, and markedly increases zone-based grounding metrics (ROI Energy 0.149→0.815). A Train-corrupt/Test-clean stress test is used to study sensitivity to mask quality.","tokens_in":10129,"tokens_out":3838,"duration_ms":37126,"significance":"If the spatial-faithfulness claim were supported by an independent evaluation, the paper would make a useful data-centric contribution: it shows how coarse lesion delineations can provide weak spatial supervision for concept localization without pixel-level concept annotations. The Train-corrupt/Test-clean stress test is a sensible methodology for studying annotation-quality effects. However, the central quantitative claim of 'markedly increasing spatial alignment' rests on metrics that are essentially the training objective (Eq. 8 vs. Eq. 4), so the reported gains do not yet demonstrate anatomical plausibility or clinical faithfulness. The diagnostic and concept-quality gains are modest and lack significance testing. The authors partially acknowledge the bias of zone-based evaluation in the Discussion, but the abstract and Section III present the grounding results as definitive evidence.","major_comments":[{"comment":"The primary spatial-grounding metric, Energy-in-Zone (Eq. 8), is exactly 1 - L_mass from Eq. (4), the very term minimized during training. Hit@1 and Top-5% overlap are thresholded variants of the same spatial-concentration objective (Eq. 3). Therefore the improvement from 0.149 to 0.815 in ROI Energy largely reflects the model being optimized for that quantity, not an independent measure of anatomical plausibility. The stress test in Table II uses clean test zones but still evaluates the same objective; it does not break this circularity. The authors should either obtain independent spatial-grounding evidence (e.g., pixel-level concept annotations, radiologist localization judgments, or an out-of-distribution dataset with such annotations) or explicitly re-scope the claims from 'anatomical faithfulness' to 'conformity with the imposed zone prior.'","section":"III-A, Eq. (8) vs. II-B, Eq. (4)"},{"comment":"The diagnostic improvement over vanilla CBM is small and within overlapping standard deviations: AUROC 0.892±0.012 vs. 0.869±0.040; concept macro-AUROC improves only 0.741→0.771. No significance test is reported. Given that the paper's abstract claims SG-CBM 'improves diagnostic AUROC and concept macro-AUROC,' paired tests (e.g., DeLong or a paired permutation test across folds) with confidence intervals should be provided. Without them, the predictive and semantic gains are not statistically supported.","section":"Table I, Section III-B(a)"},{"comment":"The paper states in Section II-A that posterior acoustic features 'are not consistently present or visually salient in all malignant ultrasound scans, but for simplicity, we assume that a posterior band may be informative across cases.' This assumption is load-bearing for the posterior grounding claims: if the band does not contain the true acoustic phenomenon, or if the phenomenon is absent, L_grounding penalizes correct activations and the posterior grounding metrics overstate spatial faithfulness. The authors should report posterior grounding metrics stratified by the presence/absence of posterior enhancement/shadowing (using the BI-RADS labels already available), and ideally analyze failure cases where the posterior band is empty or too tight.","section":"II-A, Posterior Band"}],"minor_comments":[{"comment":"The Discussion correctly notes that predefined zones 'may bias evaluation' and that zone alignment 'should be interpreted alongside, rather than as a substitute for, semantic concept correctness.' This caveat should appear prominently in the abstract and at the first presentation of the grounding results, not only in the final discussion.","section":"Discussion, last paragraph"},{"comment":"The hyperparameters alpha=1.0, H_min=10, H_max=50 are chosen heuristically. A sentence on sensitivity to these parameters (or a small ablation) would improve reproducibility; currently no sensitivity analysis is reported.","section":"II-A, Eq. (1)"},{"comment":"The first row is labeled with a single dash under 'K'; it should be labeled explicitly (e.g., 'none' or 'clean') for readability. Also, K is already used for the number of concepts; using K again for the erosion kernel is confusing.","section":"Table II"},{"comment":"The caption refers to 'the rightmost 2 panels' but the figure appears to show several concept summaries; the panel layout should be described more precisely so the qualitative grounding comparison can be verified.","section":"Fig. 2"},{"comment":"No code availability statement or trained-model release is provided. Given the paper's emphasis on reproducibility and data-centric validation, sharing code and the exact zone-generation implementation would strengthen the contribution.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea is timely and the stress-test design is a nice addition, but the main spatial-faithfulness result is circular as measured. I would not reject the paper—the circularity can be addressed by re-framing claims and adding an independent grounding probe or a clear caveat—but as written the central claim is overstated. The diagnostic and concept AUROC gains also need significance testing. This is a major revision rather than a minor one because the abstract and Table I currently present the grounding improvement as strong evidence of anatomical plausibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: if you read this as a method for injecting lesion-mask spatial priors into CBMs, it's a sensible, clearly written piece. If you read the abstract's \"markedly increasing spatial alignment\" as evidence of anatomical plausibility, that part is not supported by the metrics as reported.\n\nWhat's new: combining lesion-mask-derived ROI/posterior zones with a grouped separation/mass-concentration loss for CBMs is a natural extension that I haven't seen in the CBM literature. The BrEaST dataset with BI-RADS concepts is a reasonable testbed, and the 5-fold group CV is proper. The train-corrupt/test-clean erosion stress test is a genuinely useful idea, and the discussion is honest about needing human validation and physics-aware posterior modeling.\n\nWhere it goes soft: the primary grounding metric, Energy-in-Zone (Eq. 8), is exactly the quantity minimized by L_mass (Eq. 4), up to epsilon. Hit@1 and Top-5% overlap are close correlates of the separation loss. So the 0.149 to 0.815 jump mostly says the optimizer minimized its own objective—not that the model localizes evidence the way a radiologist would. That's a real circularity problem, and it's compounded by the posterior band assumption (Sec. II-A) that a fixed band under the lesion is informative across all cases; the paper itself acknowledges this is a simplification. Diagnostic gains are modest: AUROC 0.869 to 0.892 with overlapping stds and no significance test; concept AUROC 0.741 to 0.771. These are real but small.\n\nTo be fair, the authors do flag in the Discussion that zone alignment should not be equated with clinical validity and that a radiologist study is needed. That softens the blow. The method is also a form of controllable regularization, and the erosion results show the framework is sensitive to supervision quality in a non-obvious way.\n\nBottom line: this is a workshop-grade contribution with a promising core idea and an evaluation that needs an independent spatial metric. I'd send it to peer review, but I'd ask the authors to add either a human-annotation study, a comparison against an alternative spatial-grounding method, or at least an evaluation where the grounding metric is computed from independently drawn zones. Without that, the claim of trustworthiness is under-supported.","headline":"Useful methods paper, but the spatial-faithfulness claim is mostly self-measuring: the headline metric is the training loss.","tokens_in":10623,"tokens_out":3075,"would_cite":true,"duration_ms":26299,"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":"Coarse lesion masks can spatially ground concept-bottleneck explanations in breast ultrasound, improving both diagnosis and the anatomical plausibility of concept evidence.","keywords":["concept bottleneck models","spatial grounding","breast ultrasound","weak supervision","BI-RADS","explainable AI","data quality","trustworthiness"],"falsifier":"Measure posterior-concept AUROC on a held-out set of malignant cases where radiologists confirm posterior shadowing appears outside the defined band or is absent; if SG-CBM's posterior grounding loss degrades those concepts relative to vanilla CBM, the band assumption is wrong.","tokens_in":9738,"feed_emoji":"🩺","tokens_out":4540,"duration_ms":37238,"temperature":0.7,"pith_summary":"The paper aims to show that concept-bottleneck models — which explain a diagnosis through human-readable clinical concepts — can be made spatially trustworthy without expensive pixel-level concept annotations. It does this by deriving two clinically motivated zones from coarse lesion masks (the lesion region itself and a band immediately beneath it for posterior acoustic effects) and training the concept maps to concentrate their evidence inside their assigned zone. On breast ultrasound, this spatially grounded CBM reportedly improves diagnostic AUROC from 0.869 to 0.892, concept macro-AUROC from 0.741 to 0.771, and the fraction of concept evidence inside the lesion region from 0.149 to 0.815. A sympathetic reader would care because it offers a data-centric, low-cost mechanism for making medical AI explanations align with how radiologists actually reason.","feed_headline":"Zone-guided loss lifts in-lesion concept evidence from 15% to 82%","feed_subtitle":"Coarse lesion masks alone make concept-bottleneck explanations anatomically faithful and lift diagnostic AUROC.","key_machinery":"The grounding objective is built from two zone masks derived from each lesion mask: an in-lesion ROI (morphology concepts) and a size-adaptive posterior band (posterior acoustic concepts). Concept activations are pushed into their assigned zones via a separation loss (in-zone vs out-of-zone mean activation, with a 10% margin) and a mass-concentration loss (fraction of activation mass inside the zone), combined with group weights so each concept contributes comparably.","core_discovery":"The paper claims that SG-CBM, a concept-bottleneck model with a grouped weak spatial supervision objective, achieves simultaneous gains in diagnosis, concept prediction, and spatial faithfulness on breast ultrasound. Compared to an ungrounded CBM, SG-CBM raises diagnostic AUROC (0.869 → 0.892), concept macro-AUROC (0.741 → 0.771), and the Energy-in-Zone grounding metric (ROI: 0.149 → 0.815; Hit@1: 0.915). These gains come from coarse lesion masks alone, requiring only two zone definitions per image, and hold across five-fold stratified group cross-validation.","pith_inferences":["The same zone-based mechanism could be transferred to other imaging tasks with available coarse anatomical priors, such as thyroid nodules or lung lesions, by replacing BI-RADS zones with the relevant anatomy.","The mild-erosion improvement hints that zone tightness itself is a tunable regularizer; a systematic sweep over dilation sizes could reveal an optimal supervision tightness independent of mask correctness.","Treating the grounding metrics as an online quality signal could let a deployment pipeline flag cases where automatic segmentation quality is low and route them to human review."],"forward_implications":["Spatial grounding can serve as a replacement for expensive pixel-level concept annotation in practice.","Evidence-alignment constraints regularize concept representations, improving rather than sacrificing diagnostic performance.","Grounding metrics make concept explanations auditable both semantically and spatially.","The Train-corrupt/Test-clean stress test provides a methodology for quantifying how annotation quality affects trustworthiness."],"fun_headline_variants":["Zone-guided loss boosts concept grounding from 15% to 82% in breast ultrasound","Zone-guided loss lifts diagnostic AUROC and concept grounding with just masks","Zone-guided CBMs: masks boost faithfulness and accuracy in breast US","Grounding concepts with lesion masks lifts AUROC and map fidelity","From 15% to 82% grounding: zone-based CBMs for trustworthy US"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The framework assumes both that reliable lesion masks are available at training time and that posterior acoustic evidence reliably lives in a fixed-size band immediately beneath the lesion; if either fails, the grounding loss can penalize correct concept activations and the grounding metrics overstate faithfulness.","fun_headline_variants_meta":{"raw":{"variants":["Zone-guided loss boosts concept grounding from 15% to 82% in breast ultrasound","Zone-guided loss lifts diagnostic AUROC and concept grounding with just masks","Zone-guided CBMs: masks boost faithfulness and accuracy in breast US","Grounding concepts with lesion masks lifts AUROC and map fidelity","From 15% to 82% grounding: zone-based CBMs for trustworthy US"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000635,"raw_usage":{"total_tokens":2765,"prompt_tokens":740,"completion_tokens":2025,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":1937}},"tokens_in":484,"tokens_out":2025,"duration_ms":12272,"temperature":1.0,"reasoning_tokens":1937,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:37:46.893023+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure posterior-concept AUROC on a held-out set of malignant cases where radiologists confirm posterior shadowing appears outside the defined band or is absent; if SG-CBM's posterior grounding loss degrades those concepts relative to vanilla CBM, the band assumption is wrong.","supporting_citations":[],"review_version":1}