{"id":"574666ef-543f-426f-adf1-5a228e53556b","arxiv_id":"2509.02710","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Using pretrained MST embeddings and a KAN classifier on DCE-MRI slices, the authors report AUC = 0.80 ± 0.02 for benign vs malignant classification in high-risk women.","lead":"A breast MRI foundation model's features are combined with a KAN classifier to sort benign from malignant lesions in high-risk screening women, reporting AUC 0.80. The base model's score is not reported in text, so the claimed improvement is not fully documented.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Slice-level labels without lesion segmentation allow confound-driven discrimination; AUC may not reflect lesion detection.","rationale":"The reader identified the same fundamental weakness: the slice-level training with patient-level labels and no lesion segmentation. My reading of the paper confirms this is the most load-bearing concern. The central claim is that MST+KAN outperforms the baseline MST and achieves AUC=0.80, but if the model is not actually localizing lesions, then both the performance metric and the interpretability claim are undermined. The lack of a reported baseline AUC and unclear pooling from slice to patient level compound the issue, but the label-mapping problem is the core scientific flaw. The proposed concrete test—comparing AUC on lesion-containing vs. non-lesion slices and measuring attention overlap with expert annotations—would directly determine whether the model's decisions are lesion-driven. Since the reader already assigned CONDITIONAL and my concern aligns with that, I see no reason to change the verdict. The paper could be accepted if such evidence is provided; until then, the claim remains unverified.","tokens_in":4921,"tokens_out":4524,"duration_ms":53262,"concrete_test":"Randomly select 50 malignant and 50 benign studies and have a radiologist annotate 2D bounding boxes around all enhancing lesions on subtraction images. Evaluate the fine-tuned model's attention heatmaps against these boxes using Dice/IOU, and compute slice-level AUC separately for slices containing a lesion vs. slices without any lesion. If AUC on lesion-free slices is substantially above chance or attention Dice is no better than random, the model is exploiting non-lesion confounds rather than performing true lesion detection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 treats each 2D axial slice as an independent input while the label is the breast-level biopsy outcome. The text states that final classification is at the patient level, but Section 2.4 trains a KAN on per-slice embeddings without describing any aggregation to patient-level predictions. Because no lesion segmentation is used, the vast majority of slices in a malignant study do not contain the lesion but are labeled malignant. The model can therefore learn to discriminate benign vs. malignant based on non-lesion correlates such as background parenchymal enhancement, breast density, or biopsy-related artifacts. This would inflate the reported AUC=0.80±0.02 and invalidate the attention heatmaps as evidence of lesion localization. Additionally, the baseline MST AUC is never reported numerically, so the claim of improvement over the baseline is not quantitatively substantiated. The paper even admits in Figure 1B that some attention is incorrectly assigned, underscoring localization limitations. Absent lesion-level annotations or a slice-level localization evaluation, the central claim that this pipeline performs robust lesion detection is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a transfer-learning pipeline for benign/malignant classification of breast DCE-MRI in a high-risk screening cohort. It uses the Medical Slice Transformer (MST), pretrained with DINOv2 on breast MRI, to extract per-slice embeddings from subtraction images. These embeddings are then used to train a Kolmogorov-Arnold Network (KAN) with two hidden layers (128 and 64 neurons) and B-spline activations. Training uses 5-fold stratified cross-validation with Borderline-SMOTE and focal loss, and the authors report AUC = 0.80 ± 0.02 for the MST+KAN pipeline, claiming it outperforms the MST baseline while preserving interpretability through attention heatmaps. The paper includes ROC curves and representative attention maps, but no code or data are provided.","tokens_in":5133,"tokens_out":4600,"duration_ms":54149,"significance":"If the comparative result were properly supported, the contribution would be useful: it demonstrates that frozen embeddings from a medical foundation model can be combined with a compact KAN classifier to handle class imbalance in a challenging high-risk population, and it proposes an attention-based visualization for clinician-facing interpretability. The approach is practical and reproducible in principle. However, the current manuscript does not substantiate the central comparative claim because the baseline MST AUC is never stated numerically, and the slice-level labeling with breast-level labels makes the 'lesion detection' interpretation precarious. Strengths include the explicit preprocessing pipeline, use of a public foundation model, and stratified cross-validation. Weaknesses include the absence of statistical tests, external validation, and code/data availability, as well as the unresolved slice-to-patient aggregation.","major_comments":[{"comment":"The baseline MST AUC is never reported numerically. The only quantitative result is AUC = 0.80 ± 0.02 for the proposed pipeline, so the claims of 'clear improvement' and 'outperforms' cannot be evaluated. Please report the baseline AUC, per-fold AUCs, confidence intervals, and a paired statistical comparison (e.g., DeLong test on patient-level scores, or bootstrap across folds).","section":"§3, Fig. 2"},{"comment":"The model is trained on per-slice inputs labeled by breast-level biopsy outcome, and no aggregation from slice-level to patient-level predictions is described, despite §2.3 stating that final classification is at the patient level. Since no lesion segmentation is used, most slices in malignant studies likely contain no lesion but are labeled malignant. This permits the classifier to exploit non-lesion confounds (e.g., background parenchymal enhancement, breast density, biopsy-related artifacts) and would inflate the reported AUC. The attention heatmaps would then not reflect lesion detection. The authors should either train/evaluate with lesion-level annotations or provide a slice-level localization evaluation and explicitly describe how per-slice predictions are aggregated to patient level.","section":"§2.3–2.4"},{"comment":"The '±0.02' is reported without stating how the uncertainty was computed, and the text uses 'significant performance improvement' without any statistical test. Additionally, 'non-overlapping groups' is not defined, so it is unclear whether the same patient can appear in both training and validation folds (e.g., through multiple studies, laterality, or repeated exams). Address this grouping explicitly, report per-fold scores, and provide confidence intervals or tests. Without these, the robustness/generalizability claims are not supported.","section":"§2.4, §3"},{"comment":"The interpretability argument rests on attention heatmaps, but the paper's own Figure 1B states that 'some attention is incorrectly assigned,' and there is no quantitative localization evaluation (e.g., Dice score against lesion masks, or distance to annotated lesion centers). The heatmaps are anecdotal. To support the claim that attention identifies lesion-relevant regions, the authors need a quantitative evaluation or should temper the interpretability claim.","section":"§3, Fig. 1"}],"minor_comments":[{"comment":"Reference [10] is cited for the Medical Slice Transformer, but the MST model used is from Müller-Franzes et al. [13]. Please correct the citation in the introduction.","section":"§1, References"},{"comment":"Typo: 'We excluded studies malignant studies where cancer laterality was not provided' should read 'We excluded malignant studies where cancer laterality was not provided.'","section":"§2.1"},{"comment":"The text uses 'studies' and 'patients' interchangeably; clarify whether 4,900 benign and 1,214 malignant entries are unique patients or imaging studies, and whether a patient may contribute both a benign and a malignant study.","section":"§2.1, §2.4"},{"comment":"The phrase 'we randomly sample a volume from the cohort of benign patients such that each sample consists of one breast' is unclear. Specify the unit of analysis and how left/right breast volumes are constructed and split across folds.","section":"§2.4"},{"comment":"Figure 2 would benefit from adding AUC values to the legend or panel labels, and from clearly distinguishing the two ROC curves; the current text does not give the baseline AUC.","section":"§3, Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a preliminary report. The central idea is plausible and within the journal's scope, but the missing baseline AUC and the slice-label/patient-aggregation issues are load-bearing. These are fixable within the manuscript's scope if the authors provide a quantitative baseline comparison, a patient-level aggregation protocol, and an assessment of label-to-slice noise. I would not recommend rejection, but the current evidence is insufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest transfer-learning paper with a real empirical result, but you cannot yet verify the main claim because the baseline AUC is never reported and the slice-to-patient label mapping is underspecified.\n\nThe new bit is narrow: taking the MST embeddings from Müller-Franzes et al. and feeding them into a KAN classifier on their UPHS high-risk cohort. That combination is not in the literature as far as I can tell, so on that narrow point the result is new. The AUC of 0.80 on 6,114 biopsy-confirmed studies is plausible and could mean something clinically if it holds up. What the paper does well: it follows a sensible transfer-learning recipe, uses group-stratified 5-fold CV, addresses imbalance with focal loss and Borderline-SMOTE, and it is honest about several limitations (small cohort, slice-wise loses 3D context, expert-label subjectivity).\n\nThe problems are in the evidence for the headline claim. First, the baseline MST AUC is never given numerically. The text says 'clear improvement' and Figure 2 shows two ROC curves, but without the baseline number the comparison is not assessable. Second, there is no statistical test, so the ±0.02 is just fold spread. Third, the label question is potentially load-bearing. Each 2D axial slice is treated as independent while the label is the biopsy outcome for the breast. The text says final classification is at patient level, but no aggregation step is described. If slices are labeled benign/malignant by whole-breast outcome, most slices in a malignant study contain no lesion, so the model can learn non-lesion correlates like background parenchymal enhancement or density. That could inflate AUC and make the attention heatmaps misleading. The paper admits in Figure 1B that some attention is wrongly assigned, which is consistent with this worry. Absent lesion-level annotations or a slice-level localization evaluation, the 'lesion detection' framing is not established.\n\nNone of this makes the paper worthless. The KAN-on-frozen-embeddings trick is cheap and easy to reproduce if the embeddings are available, and the dataset size is a plus. But as written, it supports a 'we tried this and it seems promising' claim, not 'clear improvement.'\n\nWho is it for: people working on breast MRI CAD or foundation-model transfer for medical imaging might want to read it. It deserves a serious referee because the question is clinically relevant and the cohort is substantial. I would not cite it yet; I would wait for a version that reports the baseline, describes the patient-level aggregation, and ideally external validation.","headline":"Plausible incremental transfer-learning result, but the central comparison is under-documented and the slice-level labeling may inflate the AUC.","tokens_in":5720,"tokens_out":2324,"would_cite":false,"duration_ms":25995,"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 frozen medical-imaging transformer plus a Kolmogorov-Arnold classifier reaches AUC 0.80 for classifying benign versus malignant breast lesions on DCE-MRI in a high-risk screening cohort.","keywords":["breast DCE-MRI","lesion classification","Medical Slice Transformer","Kolmogorov-Arnold Networks","DINOv2","attention heatmaps","high-risk screening","transfer learning"],"falsifier":"Take a held-out set of DCE-MRI studies with manually annotated lesion masks. Compare model AUC on slices that contain lesion versus slices that do not; if the classifier keeps high AUC on lesion-free slices, its decision is not lesion-detection. Additionally compute overlap of attention heatmaps with annotated masks; low Dice on true positives would contradict the localization claim.","tokens_in":4815,"feed_emoji":"🩻","tokens_out":3928,"duration_ms":43738,"temperature":0.7,"pith_summary":"This paper asks whether a frozen medical-imaging foundation model can be repurposed for breast lesion classification in a high-risk screening population by swapping its final classifier for a Kolmogorov-Arnold network. The authors report that the adapted pipeline reaches an AUC of 0.80 ± 0.02, a clear improvement over the original slice-wise transformer classifier, while attention heatmaps still point to lesion sites. If the result holds, it offers a lightweight route to adapt self-supervised MRI features to a new clinical task and population. The paper's central claim is the empirical performance gain itself, not a new theory.","feed_headline":"Frozen MRI embeddings plus KAN lift lesion AUC to 0.80","feed_subtitle":"Frozen slice-wise transformer features plus a spline classifier improve MRI lesion discrimination while keeping attention heatmaps.","key_machinery":"The load-bearing components are: the Medical Slice Transformer (MST), a slice-wise vision transformer pretrained with DINOv2 that converts each 256x256 axial subtraction slice into a feature embedding; and the Kolmogorov-Arnold Network (KAN), whose learnable B-spline activations replace fixed nonlinearities and are trained on the frozen embeddings. The attention mechanism provides the interpretability claim.","core_discovery":"On a biopsy-confirmed cohort of 4,900 benign and 1,214 malignant breast DCE-MRI studies from high-risk screening, the paper's MST+KAN pipeline classifies each axial slice by extracting frozen DINOv2-based embeddings from the Medical Slice Transformer and training a two-hidden-layer KAN with adaptive B-spline activations. In five-fold cross-validation, it reports AUC=0.80±0.02, outperforming the standard MST classifier; attention heatmaps concentrate on lesions, including small lesions and dense tissue, even though the model is never given segmentation masks.","pith_inferences":["Because slices are labeled by whole-breast biopsy outcome, the reported AUC may partly reflect breast-level confounds rather than lesion detection; testing on lesion-free slices would separate these.","The improvement is attributed to KAN, but the paper does not compare against an MLP trained on identical embeddings; that comparison would isolate the KAN contribution.","If attention localization is accurate, the same embeddings could support lesion segmentation without additional supervision.","The method's generality across scanner types is asserted from attention maps; a multi-site external validation would turn that assertion into measurable generalization."],"forward_implications":["An AUC of 0.80 would support using frozen MST embeddings as general-purpose breast MRI features for new classification heads.","The KAN head could be retrained quickly on other cohorts or protocols without recomputing embeddings.","Attention heatmaps could serve as weakly supervised lesion localization for triage in BI-RADS 4 cases.","The pipeline addresses class imbalance via SMOTE and focal loss, making it more applicable to real screening data."],"supporting_citations":[{"why":"Provides the MST slice-wise transformer model and the DCE-MRI preprocessing pipeline that this work adapts.","marker":"[13]"},{"why":"Supplies the DINOv2 pretrained visual features that initialize and power the MST encoder.","marker":"[14]"},{"why":"Provides the breast MRI pretraining dataset used to train the MST model.","marker":"[15]"},{"why":"Defines the Kolmogorov-Arnold network architecture with adaptive B-spline activations used as the new classifier.","marker":"[16]"},{"why":"Supplies Borderline-SMOTE, the oversampling method used to mitigate class imbalance.","marker":"[17]"},{"why":"Supplies focal loss, the loss function that emphasizes misclassified samples during training.","marker":"[18]"}],"fun_headline_variants":["High-risk breast MRI: frozen embeddings plus KAN hit AUC 0.80","Spline classifier on frozen MRI embeddings clears 0.80 AUC","MST+KAN: interpretable lesion detection reaches 0.80 AUC","Frozen transformer features plus KAN outscore baseline in breast MRI","Breast DCE-MRI: KAN on frozen features boosts AUC to 0.80"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Each 2D axial slice is labelled with the biopsy result of the whole breast, even though many slices contain no lesion, so high accuracy may come from whole-breast signals instead of the lesion.","fun_headline_variants_meta":{"raw":{"variants":["High-risk breast MRI: frozen embeddings plus KAN hit AUC 0.80","Spline classifier on frozen MRI embeddings clears 0.80 AUC","MST+KAN: interpretable lesion detection reaches 0.80 AUC","Frozen transformer features plus KAN outscore baseline in breast MRI","Breast DCE-MRI: KAN on frozen features boosts AUC to 0.80"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000507,"raw_usage":{"total_tokens":2286,"prompt_tokens":698,"completion_tokens":1588,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":1496}},"tokens_in":442,"tokens_out":1588,"duration_ms":10652,"temperature":1.0,"reasoning_tokens":1496,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:28:09.226980+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of DCE-MRI studies with manually annotated lesion masks. Compare model AUC on slices that contain lesion versus slices that do not; if the classifier keeps high AUC on lesion-free slices, its decision is not lesion-detection. Additionally compute overlap of attention heatmaps with annotated masks; low Dice on true positives would contradict the localization claim.","supporting_citations":[{"cited_title":"Medical slice transformer for improved diagnosis and explainability on 3d medical images with dinov2,","cited_arxiv_id":null,"evidence_quote":"Provides the MST slice-wise transformer model and the DCE-MRI preprocessing pipeline that this work adapts."},{"cited_title":"Dinov2: Learning robust visual features without supervision,","cited_arxiv_id":null,"evidence_quote":"Supplies the DINOv2 pretrained visual features that initialize and power the MST encoder."},{"cited_title":"Borderline-smote: a new over-sampling method in imbalanced data sets learning,","cited_arxiv_id":null,"evidence_quote":"Supplies Borderline-SMOTE, the oversampling method used to mitigate class imbalance."},{"cited_title":"Focal loss for dense object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies focal loss, the loss function that emphasizes misclassified samples during training."}],"review_version":1}