{"id":"31ea10b3-68a3-466c-a1d3-1c0745c6ec3c","arxiv_id":"2607.10851","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hinge regularizer on foundation-model anatomical masks forces classifiers to keep more attention energy on foreground than background, improving accuracy and localization without pixel labels.","lead":"Locus steers medical image classifiers toward diagnostic anatomy by regularizing Grad-CAM attention with masks from off-the-shelf segmentation foundation models, using a simple hinge that only penalizes background-dominant focus. The method needs no extra annotations or segmenter training and lifts accuracy plus interpretability across eight imaging modalities.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection beyond the reader's already-identified mask-quality dependence; the hinge design and multi-dataset results keep the central claim intact.","rationale":"The paper's central claim is empirical and modest: a frozen SFM + lightweight hinge regularizer improves classification metrics and attention coherence without dedicated segmentation training. The design choices (aggregate energy comparison rather than Dice, hinge rather than hard constraint, offline pre-computed masks) make the method robust to the very imprecision the reader flags. Multi-modality, multi-backbone, multi-saliency, and perturbation results supply independent support. No equation is internally inconsistent, no baseline is misrepresented, and no result is contradicted by the reported figures. Therefore the reader's CONDITIONAL verdict already correctly balances the useful contribution against the unquantified dependence on SFM mask quality; no further downward adjustment is warranted.","tokens_in":12033,"tokens_out":534,"duration_ms":8540,"concrete_test":"On one held-out dataset (e.g., BUSI or BTXRD), replace the SFM-derived masks M with (a) expert ground-truth segmentations and (b) deliberately corrupted/empty masks; re-train the same backbone under identical λ schedule and report ΔAcc/ΔF1 versus the original Locus run. If performance collapses under (b) but remains comparable under (a), the mask-quality dependence is confirmed as the sole critical assumption; if gains persist even under (b), the hinge is more robust than claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption is already the correct load-bearing point: that un-fine-tuned SFM masks (BiomedParse / Medical SAM3 via patch-wise text prompting, §2.2 and Fig. 3) are sufficiently complete and correctly localized for the diagnostically relevant structures. If those masks systematically miss or misplace true anatomy, R = max(0, A_bg − A_fg) (Eq. 3) supplies incorrect spatial supervision. No stronger internal inconsistency or experimental flaw is present. The hinge formulation itself is deliberately tolerant of coarse/partial masks (it only activates when background attention dominates, not requiring pixel-wise fidelity), the method is plug-and-play across backbones and saliency extractors, and Table 1 + Figs. 4–7 show consistent accuracy/F1 gains plus more anatomically coherent maps on eight datasets. Absence of statistical tests and a few hand-chosen hyperparameters (λ, P, warm-up) are secondary and already noted by the reader.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Locus, an anatomy-guided attention regularization framework for medical image classification. Pretrained segmentation foundation models (BiomedParse, Medical SAM3) supply anatomical masks via a patch-wise text-prompting strategy; a hinge regularizer R = max(0, A_bg − A_fg) (Eq. 3) then penalizes the classifier only when Grad-CAM background attention exceeds foreground attention. The regularizer is added to standard cross-entropy with a warm-up schedule (Eq. 4). The method is claimed to be plug-and-play across backbones and saliency extractors, requiring neither manual pixel annotations nor fine-tuning of the foundation model. Empirical validation spans multiple public datasets (dermoscopy, X-ray, histopathology, ultrasound) plus an in-house cardiac cine MRI set, with reported accuracy/F1 gains over CE-only, masked-input, and MGA baselines, plus qualitative improvements in attention localization and robustness under test-time perturbations.","tokens_in":12326,"tokens_out":1088,"duration_ms":30940,"significance":"If the reported gains hold under stronger statistical scrutiny and more complete reporting, Locus is a practically useful contribution: it converts frozen, off-the-shelf medical SFMs into spatial supervision for classifiers without the annotation cost of mask-guided training or the optimization friction of joint segmentation–classification. The hinge design is deliberately tolerant of coarse masks, the method is architecture- and saliency-agnostic, and the multi-modality evaluation (including temporal cine MRI) is broader than typical single-domain studies. Code is promised (anonymous repository). These are genuine engineering strengths for MICCAI-style venues. The work does not introduce a new theoretical principle; its value is empirical and methodological.","major_comments":[{"comment":"Section 2.2 and Fig. 3: The central claim rests on the assumption that un-fine-tuned SFM masks obtained by patch-wise text prompting are sufficiently complete and correctly localized for the diagnostically relevant anatomy on every dataset. Only qualitative examples are shown; there is no quantitative mask-quality evaluation (e.g., coverage of lesion/organ, false-positive background rate, or failure-case rate) on the eight evaluation sets. Because R (Eq. 3) uses these masks as the sole spatial prior, systematic mislocalization would supply incorrect supervision. A short quantitative audit of mask quality (or explicit failure analysis) is needed to bound this risk.","section":null},{"comment":"Table 1: Accuracy and F1 gains are reported as single point estimates with no standard deviations across seeds/runs and no statistical significance tests. Several improvements are modest (e.g., PneumoniaMNIST ResNet Acc 89.90 → 92.15; Cardiac DenseNet Acc 78.32 → 82.02). Without multi-seed statistics it is difficult to judge whether the “consistent gains” claim is reliable. At minimum, report mean ± std over ≥3 seeds and a paired test or confidence interval for the main comparisons.","section":null},{"comment":"Abstract and §3 claim validation on eight datasets, yet Table 1 reports only six (HAM10000, PneumoniaMNIST, BTXRD, PanNuke, ChestX-ray8, Cardiac). BUSI and CheXpert appear in figures and text but lack the same quantitative baseline comparison. Either complete the main table for all eight datasets or revise the claim to match the reported numbers.","section":null}],"minor_comments":[{"comment":"Table 2 (λ ablation): performance is non-monotonic (dips at λ = 0.3 and 0.7). A brief discussion of why intermediate strengths can over-constrain the classifier would help readers choose λ.","section":null},{"comment":"§3.1 Implementation: number of patches P is “selected empirically per dataset” with no sensitivity study or default rule. A short ablation or heuristic would improve reproducibility.","section":null},{"comment":"Fig. 1 caption and several figure panels use “SOTA classifiers” without naming which models are shown; label the baselines explicitly.","section":null},{"comment":"Leftover running head “Title Suppressed Due to Excessive Length” appears on multiple pages; remove before camera-ready.","section":null},{"comment":"Medical SAM3 is cited as arXiv:2601.10880 (2026); confirm the citation is stable and publicly available at submission time.","section":null},{"comment":"§3 Datasets: exclusion of “No Finding” / normal classes is reasonable but should be stated once with a clear criterion so readers can judge selection bias.","section":null},{"comment":"Clarify whether Grad-CAM is computed only on the predicted class or on the ground-truth class during training of R; this affects the gradient path.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Fit for a MICCAI-style conference or journal special issue is good; novelty is incremental but useful. The main risk is over-claiming “eight datasets” while only six appear in the primary table, and the unquantified dependence on SFM mask quality. Neither is fatal if the authors complete the table and add a short mask-quality audit plus multi-seed stats. No integrity or circularity concerns. I would not escalate to major_revision unless the missing datasets reverse the trend or mask quality proves systematically poor."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: freeze a medical segmentation foundation model, get coarse anatomical masks via patch-wise text prompts, then add a hinge R = max(0, A_bg − A_fg) that only fires when background attention exceeds foreground. That combination is new relative to the usual CAM-pseudo-mask or joint-seg-and-class pipelines, and it is genuinely plug-and-play.\n\nWhat they do well is the empirics. Table 1 shows accuracy and F1 lifts on six public datasets (plus cardiac cine MRI and BUSI) across ResNet, EfficientNet and DenseNet; the same pattern holds for ConvNeXt and ViT. Grad-CAM maps tighten onto lungs, lesions, myocardium, etc., and the gains survive different saliency extractors and common test-time corruptions. The hinge itself is deliberately tolerant of imperfect masks—it never forces pixel-wise Dice—so the method does not collapse when BiomedParse or Medical SAM3 is only roughly right. Code is promised, masks are pre-computed offline, and the free parameters (λ, patch count, warm-up) are few and ablated.\n\nThe soft spot is exactly the one the reader flagged: everything rides on the quality of those un-fine-tuned SFM masks. If the foundation model systematically misses the true diagnostic structure, the hinge supplies the wrong spatial signal. They show that patch-wise prompting recovers more regions than full-image prompting, but they never quantify mask completeness or run a controlled failure case. Secondary nits: no error bars or significance tests, and a couple of hand-chosen hyperparameters. None of that breaks the central claim.\n\nThis is for people who train medical classifiers and want better localization without buying pixel labels or training a second network. It is solid engineering, not a foundational result. I would send it to peer review; the idea is clean, the evidence is multi-dataset and multi-backbone, and the limitations are already visible. Worth a look if you care about weakly-supervised medical attention.","headline":"Clean plug-and-play hinge regularizer that turns frozen medical SFMs into anatomy priors; consistent gains, real but bounded novelty.","tokens_in":12871,"tokens_out":505,"would_cite":true,"duration_ms":8150,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A hinge regularizer that borrows coarse anatomy masks from pretrained segmentation models steers medical classifiers toward the right tissue and raises accuracy across eight datasets.","keywords":["medical image classification","attention regularization","anatomical shapes","segmentation foundation models","Grad-CAM","hinge loss","interpretability"],"falsifier":"On a held-out set where the foundation model systematically omits the true lesion or organ, measure whether Locus accuracy drops below plain cross-entropy while Grad-CAM still concentrates on the incorrect masked region; that outcome would refute the claim that the free masks supply useful supervision.","tokens_in":12957,"feed_emoji":"🩻","tokens_out":849,"duration_ms":22838,"temperature":0.7,"pith_summary":"Medical classifiers usually train on image-level labels alone, so they freely attend to background artifacts and scanner cues that do not generalize. This paper argues that you can supply the missing spatial common sense without any manual pixel masks or a second trained segmentation network. Frozen segmentation foundation models, prompted patch-wise with simple text, yield usable anatomical priors; a hinge loss then penalizes the classifier only when its attention energy on background exceeds that on the anatomy. The resulting framework, Locus, plugs into ordinary CNN and transformer backbones, needs no foundation-model fine-tuning, and produces both higher accuracy/F1 and more coherent Grad-CAM maps on dermoscopy, X-ray, histopathology, ultrasound and cardiac cine MRI. A sympathetic reader cares because it turns an expensive annotation problem into a free, off-the-shelf regularizer that also improves robustness to common imaging corruptions.","feed_headline":"Anatomy hinge lifts medical classifiers without new labels","feed_subtitle":"Pretrained masks plus a simple background-attention penalty raise accuracy and focus across eight datasets.","key_machinery":"The anatomy-centric hinge regularizer R(M,A)=max(0,A_bg−A_fg), which activates only when background attention dominates foreground attention; masks M are obtained by union of patch-wise text prompts to a frozen segmentation foundation model.","core_discovery":"Locus establishes that an adaptive hinge comparing aggregate attention inside versus outside a foundation-model anatomical mask is sufficient spatial supervision for medical image classification: it yields consistent gains in accuracy and F1-score together with more anatomically grounded attention maps across eight diverse datasets, without dedicated segmentation training or expert pixel annotations.","pith_inferences":["As segmentation foundation models improve, free anatomical priors of this kind could become default training practice for any anatomy-dependent medical task.","The hinge idea may transfer outside medicine whenever a cheap off-the-shelf detector can supply a soft region prior.","Learning the patch grid end-to-end, as the authors note for future work, would likely close remaining cases of missed small or sparse structures.","Uncertainty-aware down-weighting of unreliable foundation masks would let the regularizer back off automatically on out-of-distribution anatomy."],"forward_implications":["Medical classifiers can obtain anatomical focus and higher scores with zero extra pixel annotation cost.","The same plug-in regularizer works across CNN and transformer backbones and multiple saliency extractors.","Coarse foundation-model masks suffice when the loss only demands anatomy-dominant rather than pixel-perfect attention.","Robustness to brightness, contrast, blur and noise improves because background-correlated shortcuts are discouraged.","Temporal sequences such as cardiac cine MRI can be regularized with the same spatial hinge."],"fun_headline_variants":["Anatomy hinge from foundation masks focuses medical classifiers","Free anatomy priors plus background penalty improve classifiers","Locus adaptive hinge steers attention without pixel labels","Foundation segments regularize attention for medical classification","Simple hinge on anatomy masks lifts accuracy across eight datasets"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method assumes that the coarse, un-fine-tuned masks from patch-wise prompting of the foundation model correctly cover the diagnostically relevant anatomy on every dataset; if those masks systematically miss or misplace the true structures, the hinge supplies wrong spatial guidance.","fun_headline_variants_meta":{"raw":{"variants":["Anatomy hinge from foundation masks focuses medical classifiers","Free anatomy priors plus background penalty improve classifiers","Locus adaptive hinge steers attention without pixel labels","Foundation segments regularize attention for medical classification","Simple hinge on anatomy masks lifts accuracy across eight datasets"]},"model":"grok-4.5","effort":"low","cost_usd":0.006722,"raw_usage":{"total_tokens":1677,"prompt_tokens":739,"num_sources_used":0,"completion_tokens":72,"cost_in_usd_ticks":67220000,"prompt_tokens_details":{"text_tokens":739,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":866,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":739,"tokens_out":72,"duration_ms":12489,"temperature":1.0,"reasoning_tokens":866,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T08:46:32.937568+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out set where the foundation model systematically omits the true lesion or organ, measure whether Locus accuracy drops below plain cross-entropy while Grad-CAM still concentrates on the incorrect masked region; that outcome would refute the claim that the free masks supply useful supervision.","supporting_citations":[],"review_version":1}