{"id":"24bcfeef-08fc-4a9a-93c5-36a2c0fbd15a","arxiv_id":"2607.27620","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A frequency-based attention module plus adaptive margins raises generalized category discovery accuracy on four medical imaging benchmarks by an average of 8.5 points over prior methods.","lead":"MedXplore is a method for discovering new disease categories in unlabeled medical images while still recognizing known ones. It combines frequency-domain attention with adaptive classification margins, and reports large accuracy gains across four medical imaging benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on an uncontrolled baseline comparison: Table 1 mixes imported numbers, omits MedGCD/ClearGCD, and reports no error bars.","rationale":"The reader's weakest assumption already identifies baseline evaluation fairness and the absence of error bars as the core fragility. My review confirms that this is the single most load-bearing concern: the empirical SOTA claim is the central contribution, and Table 1 as presented does not allow a reader to verify that the comparison is controlled. The missing explicit statement about SimGCD/LegoGCD re-runs, the imported starred baselines, the omission of MedGCD and ClearGCD, and the absence of multi-seed statistics all bear directly on whether the +8.5% average gain is real or an artifact of protocol mismatch. This is a correctness risk, not a disagreement with consensus. It is not internally inconsistent, and the internal ablations (Table 2) are plausible evidence that FAAC and ACAM help. But because the headline claim is empirical and the comparison is under-specified, the conditional verdict remains appropriate. No new concern beyond the reader's was found, so the verdict should stay UNCHANGED: conditional acceptance pending code, multi-seed results, and a clearly controlled baseline protocol.","tokens_in":16244,"tokens_out":3485,"duration_ms":36531,"concrete_test":"Request the exact evaluation scripts and run logs for every Table 1 entry. Re-run SimGCD and LegoGCD under the MedXplore protocol (ResNet-18, fine-tune only last stage, 200 epochs, batch size 128, initial lr 0.1, lambda 0.35, same augmentations and splits) with at least 3 seeds. Also run MedGCD and ClearGCD under the same protocol. If the average All-accuracy gain over the strongest fairly-run baseline falls below roughly 2 points or within one seed standard deviation, the SOTA claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — 'consistently outperforms all previous methods on all four datasets, establishing a new state of the art in medical GCD' — requires Table 1 to be a fair and controlled comparison. Section 5.1 states only: 'Following [11], we employ a ResNet-18 backbone ... fine-tune only its last stage', and says this 'avoids any architecture advantage over the ResNet-18-based baselines.' It does not explicitly state that SimGCD and LegoGCD were re-run under the same protocol (ResNet-18, last-stage fine-tuning, 200 epochs, batch 128, lr 0.1, lambda 0.35, same augmentations). Their original formulations use ViT-B/16 with DINO pretraining, so the unstarred rows in Table 1 are ambiguous: either the authors re-implemented them under the medical protocol, or they are reporting numbers from a different setting. The starred rows are imported from NGUF, and no command-level protocol is given for any baseline. In addition, MedGCD [5] and ClearGCD [28] are discussed in related work but absent from Table 1; a claim of 'all previous methods' cannot exclude those recent methods. Finally, no variance or seed information is reported — the Kvasir false-old reduction from 14.50% to 0.80% is presented as a point estimate. These issues do not invalidate the method, but they make the headline empirical claim unverified as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MedXplore, a framework for Generalized Category Discovery (GCD) in medical imaging. It consists of two modules: FAAC, which operates in the frequency domain (learnable full-spectrum filtering, global-local energy contrast activation, and patch consistency), and ACAM, which adaptively applies angular and cosine margins to classification logits based on a fused confidence score. The method is trained on top of a SimGCD/NGUF-style parametric GCD baseline with a ResNet-18 backbone and is evaluated on four medical datasets: Kvasir, OrganAMNIST, OrganCMNIST, and NCT-CRC-HE-100K. The authors report an average 8.5% improvement in All accuracy over the strongest competing methods and a reduction of false-old errors on Kvasir from 14.50% to 0.80%. They also provide ablations showing that both FAAC and ACAM contribute to the reported gains.","tokens_in":16666,"tokens_out":4826,"duration_ms":49958,"significance":"If the reported results hold under a controlled comparison, the paper makes a useful contribution to medical GCD. The perceptual/decision decomposition is sensible, and the paper gives credit for several good practices: FAAC and ACAM are training-time only, so inference matches the baseline; the ablation study isolates the contributions of each module; the Kvasir error decomposition explicitly targets old-class bias; and the hyperparameter sensitivity analysis is reported. The methodological novelty is moderate but, together with the strong empirical gains, would be of interest to the medical image analysis and GCD communities. However, the empirical claims currently rest on an uncontrolled baseline table and on point estimates without variance, so the significance is conditional on the authors being able to substantiate the comparison.","major_comments":[{"comment":"The headline claim of an 8.5% average gain and 'state of the art in medical GCD' rests on an uncontrolled baseline table. Section 5.1 states only 'Following [11], we employ a ResNet-18 backbone ... fine-tune only its last stage,' and explicitly says the ViT-B/16 description in Sec. 3.2 is not the experimental backbone. However, SimGCD [43] and LegoGCD [2] appear as unstarred rows without any statement that they were re-run with ResNet-18, last-stage fine-tuning, 200 epochs, batch size 128, lr 0.1, and lambda=0.35. Both original methods use ViT-B/16 with DINO pretraining. Several other rows (RS+, UNO+, GCD, DCCL, CMS, NGUF) are imported from NGUF [11] without evidence that NGUF used the same backbone and training protocol. Without a controlled re-run or a clear demonstration of comparability, the average gain is not a fair comparison.","section":"§5.1, Table 1"},{"comment":"The 'all previous methods' claim is contradicted by the paper's own related-work discussion. MedGCD [5] and ClearGCD [28] are described as recent GCD methods but are absent from Table 1. If their numbers are not available, the claim should be narrowed or the methods should be included. Additionally, no results are reported with multiple seeds or error bars; the Kvasir false-old reduction (14.50% to 0.80%) and all ACC values are point estimates. Clustering accuracy is sensitive to initialization and data split, so the reported robustness and 'consistently outperforms' claims are not yet supported.","section":"§5.1, Table 1"},{"comment":"Hyperparameters are set on OrganAMNIST and then reused on the other datasets, but OrganAMNIST is itself one of the four evaluation datasets. Thus the OrganAMNIST row in Table 1 and the hyperparameter studies in Fig. 6 are not independent of the hyperparameter selection. The authors should either select hyperparameters on a held-out validation split or explicitly acknowledge and justify this optimistic-bias risk. This does not invalidate the method, but it weakens the claim that the reported numbers transfer without dataset-specific tuning.","section":"§5.1, 'Implementation details'; §5.4"}],"minor_comments":[{"comment":"The caption for 'Table 3: Incremental internal ablation of FAAC' appears in the manuscript but the table body is missing from the provided text. Please include the actual table or provide a precise reference to the supplementary material with the numbers.","section":"§5.4, Table 3"},{"comment":"The figure showing alpha/beta sensitivity and K/delta sensitivity is not legible in the current version; the labels appear garbled. Please replace it with a clean figure and ensure the axis labels and curves are readable.","section":"Fig. 6"},{"comment":"The disclaimer that the ViT-B/16 encoder is only a summary of SimGCD and not the experimental backbone is unusual and could be confusing. Consider moving this clarification to Sec. 5.1 or to a footnote.","section":"§3.2"},{"comment":"In the ACAM loss, the symbol C (total number of classes) and the notation B_l are used without a definition in Sec. 4.2. Clarify whether B_l is the same as B_ell introduced in Sec. 3.2 and define C explicitly.","section":"Eq. (24)"},{"comment":"Please specify the old/new class split sizes and the labeled/unlabeled sample counts for each dataset in the main text, not only in the appendix. This would help readers assess the difficulty of each benchmark.","section":"§5.1, Datasets"}],"recommendation":"major_revision","confidential_remarks":"The omission of ClearGCD [28] from Table 1 is conspicuous because it shares authors with the current submission. This may be a legitimate scope or availability decision, but the editor should ask the authors to justify the omission and to disclose any relationship. The baseline-comparison issue and the absence of seed/error-bar information are the primary reasons for the major-revision recommendation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is the combination of a training-time frequency-domain perception module (FAAC) with an adaptive angular/cosine margin loss (ACAM). FAAC does something non-obvious: it filters in the frequency domain with learnable depthwise convolutions on real and imaginary parts, then uses global-local energy contrast to pick semantic anchors for patch consistency. ACAM adapts margins using a fused confidence signal. Both are discarded at inference, so the deployed model is exactly the baseline architecture. That is a clean, honest design, and the ablation study (Table 2) shows each component helps, with the two together beating either alone. The Kvasir false-old error drop from 14.50% to 0.80% is striking, and the paper's framing of that as a bias problem is well argued. I also credit the explicit note in Sec. 3.2 clarifying that the ViT-B/16 description refers to the original SimGCD and not their setting; that preempts one obvious misunderstanding.\n\nThe soft spots are real, though. The biggest is the baseline protocol in Table 1. Rows marked with an asterisk are taken from NGUF, but the unmarked rows (SimGCD, LegoGCD, ORCA) carry no statement that they were re-run under the same ResNet-18, last-stage fine-tuning, 200-epoch, batch-128, lr-0.1 protocol. Since SimGCD and LegoGCD were originally published with ViT-B/16, the unstarred numbers are ambiguous. That makes the average +8.5% All gain and the per-dataset wins an uncontrolled comparison as reported. Relatedly, MedGCD and ClearGCD are cited in related work but absent from Table 1, so \"consistently outperforms all previous methods\" overreaches. No error bars or seeds anywhere, which is a minor-to-moderate worry for the Kvasir point estimate. And hyperparameters were tuned on OrganAMNIST, so that dataset's result is mildly optimistic even if the reuse across datasets is good practice. The paper itself admits a scope limitation in the conclusion (localized lesions, not diffuse low-frequency pathologies), which is honest but narrows the claim.\n\nThe circularity concern from the stress-test does not land hard: ACAM's confidence signal uses the model's own logits plus the FAAC attention, but the margin is applied only to labeled samples with ground truth, so there is no pseudo-label confirmation loop. That is fine.\n\nBottom line: the method deserves a serious referee. I would send this to review, but the acceptance should be conditional on code release, multi-seed results, an explicit statement of which baselines were re-run and under what protocol, and at least one comparison against MedGCD and ClearGCD. Until then, treat the numeric SOTA as provisional.","headline":"MedXplore is a thoughtful method with believable internal ablations, but the headline SOTA claim is built on an uncontrolled baseline table and no code; worth refereeing, with revisions.","tokens_in":809,"tokens_out":877,"would_cite":true,"duration_ms":30876,"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":"This paper claims that MedXplore, a frequency-domain attention module coupled with confidence-scaled angular margins, makes generalized category discovery in medical images more accurate and less biased toward known classes.","keywords":["generalized category discovery","medical imaging","frequency-domain filtering","adaptive margin loss","old-class bias","open-world recognition","lesion-sensitive representation","clustering accuracy"],"falsifier":"Run every baseline method and MedXplore under identical conditions—same backbone, same 200-epoch schedule, same data splits, and multiple random seeds—then compute the mean and variance of All accuracy. If the average gain over the strongest baseline falls below about two points or overlaps within one standard deviation, the claim of consistent superiority is falsified. Separately, recompute the Kvasir false-old error across at least five seeds; if it does not reproduce near 0.80%, the bias-reduction claim is not robust.","tokens_in":16177,"feed_emoji":"🩺","tokens_out":5363,"duration_ms":47261,"temperature":0.7,"pith_summary":"The paper argues that medical images, unlike natural images, contain subtle lesion cues and weak semantic separation, so standard generalized category discovery methods drift attention and over-assign ambiguous novel samples to old, known classes. To fix this, MedXplore couples a perception module (FAAC) that learns to filter the full frequency spectrum and activate local energy contrasts—highlighting the small abnormal regions that distinguish disease subtypes—with a decision module (ACAM) that adaptively sharpens class boundaries based on sample confidence. On four medical benchmarks, the combination raises average All-class clustering accuracy by 8.5 percentage points over the strongest competitor, and on the Kvasir endoscopy dataset it cuts false-old errors (novel samples mislabeled as known) from 14.5% to 0.8%. If correct, this means open-world medical image recognition can discover new disease categories reliably without retraining or manual annotation of those categories.","feed_headline":"8.5% gain in medical category discovery via two modules","feed_subtitle":"Frequency-aware attention and confidence-scaled margins cut false-old errors on endoscopy images from 14.5% to 0.8%.","key_machinery":"The key machinery is a two-part training-time addition: the FAAC perception branch, which performs learnable full-spectrum filtering via depth-wise convolutions on FFT components, followed by global-local energy contrast activation to select reliable semantic anchors for patch consistency; and the ACAM decision module, which maps a three-way confidence score into adaptive angular and cosine margins for the classification loss. The mechanism works by making the backbone learn texture-agnostic, lesion-sensitive features while constraining ambiguous samples from being collapsed into old classes.","core_discovery":"MedXplore's central claim is that old-class bias in medical generalized category discovery is driven by two fixable bottlenecks: attention that overfits to familiar anatomical textures, and decision boundaries that ignore sample reliability. FAAC addresses the first by transforming features into the frequency domain, applying learnable per-channel filtering to the real and imaginary parts of the FFT, then normalizing local variance against global statistics so that only genuinely anomalous regions are amplified; the resulting top patches become semantic anchors for cross-view consistency. ACAM addresses the second by computing a per-sample confidence from attention strength, cross-view patch","pith_inferences":["A critical test for this framework is whether the gains persist under distribution shift—e.g., CT scans from different scanners or histology from different stains—since the paper tunes hyperparameters on one dataset and reuses them elsewhere without cross-domain evaluation.","The adaptive margins depend on the confidence signals FAAC supplies; if the perception branch is noisy on diffuse or low-frequency pathologies (a limitation the paper acknowledges), the benefit of ACAM may shrink, suggesting a ceiling for this design on such cases.","One could directly test the FAAC hypothesis by measuring whether the learned frequency filters concentrate energy on radiologically-defined lesion regions and whether those regions align with clinician annotations.","The 0.8% false-old rate on Kvasir, if replicated across seeds, would make MedXplore a candidate for semi-automated case review: flagging samples that fall near decision boundaries for human inspection."],"forward_implications":["Existing medical imaging pipelines could discover novel disease categories from unlabeled data without manual annotation, which matters for rare or emerging conditions.","The sharp reduction in false-old errors on endoscopy images implies that CAD systems could flag genuinely new mucosal patterns instead of forcing them into familiar classes.","Because FAAC and ACAM are discarded at inference, the approach adds no computational cost at deployment, making it feasible for clinical settings with limited compute.","The frequency-domain treatment suggests that subtle, localized lesion cues, not just global appearance, carry the discriminative signal for disease recognition, which could inform other medical representation learning tasks.","The combination of perceptual and decision-level corrections may be directly portable to other open-set recognition problems where old-class bias dominates."],"fun_headline_variants":["Frequency-domain attention and adaptive margins slash false-old errors 14.5%→0.8%","MedXplore boosts medical GCD by 8.5%, cuts bias with FAAC and ACAM","Frequency-SNR adaptive attention and angular margins improve GCD accuracy by 8.5%","Two modules: frequency filtering and confidence-scaled margins for unbiased medical GCD"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim of consistent state-of-the-art performance rests on the assumption that the baseline numbers being compared against were measured under the same backbone, training epochs, and hyperparameter settings as MedXplore; if those baselines were tuned differently or run with different architectures, the reported gains may not reflect the method itself.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-domain attention and adaptive margins slash false-old errors 14.5%→0.8%","MedXplore boosts medical GCD by 8.5%, cuts bias with FAAC and ACAM","Frequency-SNR adaptive attention and angular margins improve GCD accuracy by 8.5%","Two modules: frequency filtering and confidence-scaled margins for unbiased medical GCD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000547,"raw_usage":{"total_tokens":2457,"prompt_tokens":753,"completion_tokens":1704,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":1608}},"tokens_in":497,"tokens_out":1704,"duration_ms":14287,"temperature":1.0,"reasoning_tokens":1608,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:28:31.760580+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run every baseline method and MedXplore under identical conditions—same backbone, same 200-epoch schedule, same data splits, and multiple random seeds—then compute the mean and variance of All accuracy. If the average gain over the strongest baseline falls below about two points or overlaps within one standard deviation, the claim of consistent superiority is falsified. Separately, recompute the Kvasir false-old error across at least five seeds; if it does not reproduce near 0.80%, the bias-reduction claim is not robust.","supporting_citations":[],"review_version":1}