{"id":"e7878b34-084b-4b28-a62a-b4c3a02718d7","arxiv_id":"2412.00374","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A modified ViT-Adapter with learnable queries improves gallbladder cancer bounding-box detection on ultrasound images, beating DINO and FocalNet-DINO.","lead":"This paper introduces LQ-Adapter, a lightweight modification for vision transformer adapters that improves gallbladder cancer detection in ultrasound images by adding learnable queries. It reports higher localization accuracy than several existing detectors while using fewer trainable parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's SOTA claim rests on uncontrolled comparisons: LQ-Adapter uses a frozen Uni-perceiver backbone while the ViT-Adapter baseline's backbone is unspecified, and no training recipe is given for DINO/FocalNet-DINO.","rationale":"The reader's weakest assumption already identifies the core issue: Table 1's baselines are not demonstrably fair or controlled. My stress-test confirms this is the most load-bearing concern and sharpens it: Section 5.1 pins LQ-Adapter to a frozen Uni-perceiver backbone but never states the backbone used for the ViT-Adapter baseline, and Section 5.3 omits the entire training recipe for DINO and FocalNet-DINO, including whether they were initialized from COCO weights. Because DETR variants are known to be sensitive to training schedule and query design, and because the paper also fails to state the detection head and loss for LQ-Adapter, the reported mIoU gaps cannot currently be attributed to the proposed learnable queries. This does not mean the claim is false; the released code and a re-run with a matched protocol could confirm it. The reader's CONDITIONAL verdict is the appropriate disposition: accept only after the missing experimental controls are documented and the comparison is shown to be apples-to-apples. Hence no verdict change, and agreement with the reader's weakest assumption.","tokens_in":12729,"tokens_out":8254,"duration_ms":81846,"concrete_test":"Re-run Table 1 under a matched protocol: instantiate ViT-Adapter with the exact same frozen Uni-perceiver backbone and the same adapter-training budget as LQ-Adapter; fine-tune DINO and FocalNet-DINO from their official COCO checkpoints with the same patient-level 5-fold splits, input resolution, epochs, and optimizer settings; and verify from the released code that LQ-Adapter and ViT-Adapter use the identical detection head and loss. If the 5.4/5.8/2.7 point mIoU gaps persist under this protocol, the SOTA claim is supported; if they shrink, flip, or become non-significant, the published comparison is not.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the 5.4/5.8/2.7 mIoU gains in Table 1. For that claim to hold, LQ-Adapter and each baseline must be matched except for the proposed learnable-queries mechanism. This is not established. Section 5.1 says only 'We used a Uni-perceiver backbone pre-trained on ImageNet-1k' and the backbone was frozen; it does not say whether the ViT-Adapter row in Table 1 uses the same Uni-perceiver backbone, a standard ViT-B/L, or the official ViT-Adapter settings. Since ViT-Adapter's design is entirely about adapting a frozen ViT, a backbone mismatch would mean the 5.4% improvement over ViT-Adapter could come from the backbone rather than from Eqs. (5)-(6). Similarly, Section 5.3 gives no training protocol for DINO or FocalNet-DINO: no epochs, no batch size, no resolution, no learning-rate schedule, no augmentation, and no statement of whether COCO-pretrained weights were fine-tuned. DETR variants are highly schedule-sensitive; if these baselines were undertrained or under-tuned, the reported 0.661/0.692 mIoU would be underestimates, inflating the claimed SOTA. The method also omits the loss function and the detection/decoding head that converts F_sp to a bounding box, so the comparison cannot be reproduced from the text. The unsupported DDSM claim in the Introduction and Conclusion (no DDSM experiments appear in Section 5) further indicates that the experimental reporting is not yet reliable enough to support the headline numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LQ-Adapter, a lightweight adapter for frozen ViT backbones that augments the ViT-Adapter design with learnable content queries in the extractor blocks. The method is evaluated for gallbladder cancer (GBC) detection on the GBCU ultrasound dataset, reporting mIoU gains of 5.4% over ViT-Adapter, 5.8% over DINO, and 2.7% over FocalNet-DINO (Table 1). The authors further integrate LQ-Adapter as the ROI generator in GBCNet and report improved GBC classification accuracy over the prior SOTA (Table 2), and they report polyp detection results on Kvasir-Seg (Table 3). The central claim is that learnable queries provide a more parameter-efficient detector that outperforms DETR-based SOTA on GBCU.","tokens_in":13084,"tokens_out":3440,"duration_ms":31771,"significance":"If the headline GBCU results hold under fair comparison, LQ-Adapter is a meaningful contribution: it offers an adapter-based alternative to DETR variants for medical detection, with roughly half the trainable parameters and strong localization on small, low-quality ultrasound pathologies. The paper's use of patient-level 5-fold cross-validation and reported error bars on the main GBCU results are good empirical practices, and the public code release supports reproducibility. The main significance currently hinges on whether the Table 1 comparisons are controlled, because the proposed gains are modest in absolute terms (e.g., mIoU 0.719 vs. 0.692) and a backbone or training-recipe discrepancy could easily explain them.","major_comments":[{"comment":"The paper explicitly claims DDSM validation twice: Contribution (4) says 'We also experimentally demonstrate the applicability of LQ-Adapter on DDSM dataset for detecting breast lesions in mammography,' and the Conclusion says LQ-Adapter 'extends applicability to breast lesion detection.' However, Section 5 contains no DDSM experiments, no DDSM results table, and no description of a mammography setup. This is a stated contribution that is entirely unsupported in the manuscript. The authors should either add the DDSM experiments with the same level of reporting as the other datasets or remove these claims from the contributions and conclusion.","section":"§1 Contribution (4) and §6 Conclusion"},{"comment":"The headline SOTA claim rests on uncontrolled baseline comparisons. Section 5.1 states only that 'We used a Uni-perceiver backbone pre-trained on ImageNet-1k' and that the backbone was frozen; it does not state what backbone the ViT-Adapter baseline uses, nor does Section 5.3 give any training protocol for DINO or FocalNet-DINO (epochs, batch size, resolution, learning-rate schedule, augmentations, or whether COCO-pretrained weights were fine-tuned). Since ViT-Adapter's design is centered on adapting a frozen ViT, a backbone mismatch between LQ-Adapter and the ViT-Adapter row in Table 1 could produce the reported 5.4% mIoU difference without any benefit from the learnable queries in Eqs. (5)-(6). Similarly, DETR variants are schedule-sensitive, so the reported 0.661/0.692 mIoU for DINO/FocalNet-DINO may be underestimates if they were undertrained. The authors must specify the exact backbone and training recipe for every baseline, or compare all methods under matched settings, to make the claimed gains credible.","section":"§5.1, §5.3, Table 1"},{"comment":"The text overstates the Kvasir-Seg results relative to what is shown. Section 5.5 says LQ-Adapter 'holds comparable performance against FocalNet-DINO and DINO DETR,' yet Section 5.3 claims 'we achieve superior localization' and the abstract says 'Superior performance of our design on this problem.' Table 3 shows LQ-Adapter at 0.850 mIoU, exactly tied with FocalNet-DINO at 0.850 and 0.002 above DINO at 0.848, and the table has no error bars, so 'superior' is not supported. Because the generality claim is a contribution, Table 3 should report means with standard deviations across folds (or another variance estimate) and the wording should be adjusted to describe the result as comparable with a parameter-count advantage, unless statistical significance is established.","section":"§5.3, §5.5, Table 3"},{"comment":"The method description is not reproducible as written. The paper never specifies the detection/decoding head that converts the adapter's multi-scale features F_sp^i into a bounding box, nor does it state the loss function used for localization. Equations (5) and (6) define updates to learnable queries, but the connection between the final LQ output and the predicted box coordinates is absent. Section 5.1 describes optimizer and epoch settings but no training objective or inference procedure. The authors should add a description of the output head, the loss (e.g., L1 + GIoU), and how the predicted box is derived from the final features, because without this the central detection result cannot be reproduced from the manuscript.","section":"§3.5, §5.1"}],"minor_comments":[{"comment":"The figure caption contains a typo: 'backone' should be 'backbone.' In Section 3.3, the sentence 'This approach could prove to be beneficial for In medical imaging tasks' has a stray 'In' and should be reworded.","section":"Figure 2 caption and Section 3.3"},{"comment":"The tensor-shape notation is garbled; for example, Section 3.5 writes 'F i vit ∈ R H×W 162 ×D' where the intended shapes appear to be collapsed subscripts. Please define all shapes cleanly with standard notation (e.g., H/16 × W/16 × D) and ensure the dimensions of LQ^i and the cross-attention operands in Eqs. (5)-(6) are compatible.","section":"Equations (2)-(6)"},{"comment":"References [17] and [18] are the same paper, and references [46] and [47] are also the same paper (DINO). The duplicate entries should be merged or distinguished correctly.","section":"References"},{"comment":"The paper states 'Focal Modulation Backbone-based DETR [35, 45]' but [45] is the Focal Modulation Networks paper and [35] is 'A strong and reproducible object detector with only public datasets'; neither is a direct DINO+FocalNet citation. Please cite the exact detector variant used for the FocalNet-DINO baseline.","section":"§5.3"},{"comment":"The parameter-count comparison is misstated: 135.5 million is about 59% of 228.1 million, not 56%, and the phrase '135,530,325(130 Million)' mixes exact and rounded numbers. Please give consistent counts and percentages, and clarify whether the ViT-Adapter parameter count uses the same Uni-perceiver backbone as LQ-Adapter.","section":"§5.3, Figure 1"},{"comment":"The caption contains a stray fragment 'FocalNet-DINO' before the metrics line; this appears to be an editing artifact and should be removed.","section":"Table 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's most serious issue is the gap between the stated contributions and the reported experiments: the DDSM claim is listed as a contribution but has no corresponding experiments, and the GBCU SOTA comparison lacks the baseline control needed to support the headline numbers. The authors should be required to either add DDSM results or remove the claim, and to document the full training recipe for all baselines. The core idea is plausible and the GBCU results with error bars are promising, so I see this as fixable within revision rather than as a fatal flaw. I would also encourage the editor to ask the authors to report error bars on the Kvasir-Seg table, since the generality claim currently rests on a tie."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea is a small, plausible extension of ViT-Adapter—adding learnable content queries to its extractor blocks—but the headline SOTA claim rests on comparisons that aren't controlled enough to believe yet.\n\nWhat's actually new: applying the learnable-query trick from DAB-DETR/DINO inside the adapter branch rather than in a detection decoder. The ablations on block count and zero-vs-random initialization are sensible and show the design choice matters. The downstream result—LQ-Adapter as a ROI generator for GBC classification (93.4% vs 92.1% for RadFormer)—is a useful practical contribution if it holds.\n\nWhere it gets shaky: the experimental reporting underdetermines the main claim. Table 1 gives no training recipe for DINO or FocalNet-DINO—no epochs, batch size, resolution, schedule, or whether COCO-pretrained weights were fine-tuned. DETR variants are notoriously schedule-sensitive, so under-tuning could easily explain the reported gaps. More importantly, LQ-Adapter uses a frozen Uni-perceiver backbone, while the ViT-Adapter baseline's backbone is never stated; if it's the standard ViT-B, the 5.4% gain over ViT-Adapter may be backbone, not the proposed mechanism. The detection head and loss function are also absent, so the comparison can't be reproduced from the text.\n\nThe unsupported DDSM claim in the contributions and conclusion (no DDSM experiment appears anywhere) is a clear overreach, and the Kvasir result is a tie with FocalNet-DINO, not 'superior performance' as the abstract says. Table 3 also lacks error bars. Smaller points: the design choices (block count, initialization) are tuned on the same dataset, which is a minor selection concern given the patient-level splits, and the 'first attempt at using foundational models for GBC' is questionable given the cited DETR-based weakly supervised work. Minor writing issues: a broken sentence in Section 2.2 and a duplicated DINO reference.\n\nNet: the idea is worth a look and the application is meaningful, but the current evidence doesn't support 'new SOTA.' I'd send it to a serious referee, with expectation of major revisions: controlled backbones, full baseline recipes, a described head/loss, and removal or substantiation of the DDSM claim.","headline":"Plausible incremental adapter tweak whose headline SOTA claim isn't yet supported by controlled baselines.","tokens_in":13608,"tokens_out":4818,"would_cite":false,"duration_ms":40980,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A lightweight adapter with learnable queries raises gallbladder cancer detection mIoU by 5.4% over ViT-Adapter on ultrasound, and beats DETR-based DINO and FocalNet-DINO by 5.8% and 2.7%.","keywords":["gallbladder cancer detection","ultrasound imaging","ViT adapter","learnable queries","object detection","medical imaging","parameter-efficient fine-tuning","foundation models"],"falsifier":"Retrain DINO and FocalNet-DINO on GBCU using the same data splits, augmentation, learning-rate schedule, and evaluation protocol as LQ-Adapter, and release the exact hyperparameters; if their mIoU rises to match or exceed LQ-Adapter's 0.719, the claimed state-of-the-art advantage would be refuted.","tokens_in":12549,"feed_emoji":"🩺","tokens_out":2347,"duration_ms":20954,"temperature":0.7,"pith_summary":"This paper claims that the limiting factor in adapting a frozen Vision Transformer (ViT) to detect gallbladder cancer in ultrasound images is not the backbone but the adapter's spatial prior module, which is too primitive to capture small, variable pathological regions. The authors propose LQ-Adapter, a modified ViT-Adapter design that adds learnable content queries to the extractor blocks, letting the model attend to task-relevant features through cross-attention. On the GBCU dataset, LQ-Adapter improves mean Intersection-over-Union by 5.4% over ViT-Adapter, 5.8% over DINO, and 2.7% over FocalNet-DINO, while using roughly half the trainable parameters of the DETR baselines. The same design also yields comparable polyp-detection performance on Kvasir-Seg and, when used as a region-of-interest generator in GBCNet, pushes classification accuracy to 93.4%, surpassing the prior state of the art. If these results hold, they suggest that parameter-efficient adapter tuning of a frozen foundation model can outperform heavy task-specific detection architectures on small medical datasets.","feed_headline":"Learnable-query adapter lifts gallbladder cancer detection by 5.4%","feed_subtitle":"A lightweight ViT adapter beats DINO and FocalNet-DINO on ultrasound mIoU with far fewer trainable parameters.","key_machinery":"The key mechanism is the learnable content query (LQ) block inserted into the adapter's extractor modules. These queries, implemented as tensors matching the spatial-prior dimensionality, are first cross-attended with frozen ViT backbone features and then with multi-scale spatial features, producing updated queries that refine localization cues. The paper's equations (5) and (6) define this two-step cross-attention update, which the authors argue couples information-rich backbone features with spatial priors more effectively than the primitive CNN prior alone.","core_discovery":"The central claim is that ViT-Adapter's CNN-based spatial prior module is the bottleneck for localizing small, variably appearing pathologies in ultrasound, and that replacing its role with learnable content queries—zero-initialized tensors co-optimized through cross-attention with backbone features and spatial priors—substantially improves localization. LQ-Adapter achieves a mean IoU of 0.719 on GBCU, surpassing ViT-Adapter (0.665), DINO (0.661), and FocalNet-DINO (0.692), while keeping the ViT backbone frozen and training only the adapter. The paper further demonstrates that the adapter's regions-of-interest improve a downstream GBC classifier (93.4% accuracy) and that the design transfers to polyp detection in colonoscopy, where it matches FocalNet-DINO's mIoU (0.85) with about 56.5% fewer trainable parameters.","pith_inferences":["The learnable-query mechanism may generalize to other small-lesion detection tasks in ultrasound or mammography, where flexible spatial attention matters more than a fixed CNN prior.","The parameter-efficiency result suggests that foundation-model adapters could be a default starting point for medical detection when annotated data are scarce, potentially reducing the carbon and compute cost of medical AI development.","If the LQ design proves robust across more modalities, it might be combined with weakly supervised or video-based methods (like masked autoencoders for ultrasound videos) to propagate localization knowledge across frames.","A direct extension the paper does not test is whether LQ-Adapter's advantage persists when the ViT backbone is unfrozen or when the spatial prior module is replaced by a stronger but still lightweight CNN; those ablations would clarify what exactly the queries contribute."],"forward_implications":["LQ-Adapter establishes a new state of the art for GBC detection on the GBCU ultrasound dataset, with mIoU 0.719 and near-perfect precision/recall.","Using LQ-Adapter's ROI as input to GBCNet's MS-SoP classifier yields 93.4% classification accuracy, surpassing both the original Faster-RCNN-based GBCNet and RadFormer.","The method transfers to polyp detection on Kvasir-Seg, matching FocalNet-DINO's mIoU while using roughly half the trainable parameters, indicating generalizability across medical imaging modalities.","The frozen-backbone, adapter-only training regime requires substantially fewer trainable parameters than DETR-based detectors, making it more practical for small medical datasets and limited compute.","Zero-initialization of the learnable queries is empirically better than random initialization, and adding LQ at all adapter blocks yields the largest gain."],"supporting_citations":[{"why":"The ViT-Adapter design that LQ-Adapter modifies; supplies the baseline spatial-prior module, injector, and extractor architecture.","marker":"[10]"},{"why":"DINO, a DETR-based state-of-the-art detector used as a primary baseline and source of the learnable-query inspiration.","marker":"[46]"},{"why":"FocalNet-DINO, the current DETR-based SOTA baseline that LQ-Adapter outperforms on mIoU.","marker":"[45]"},{"why":"GBCNet, the previous GBC detection pipeline whose ROI generator LQ-Adapter replaces and whose MS-SoP classifier is reused for the classification experiment.","marker":"[3]"},{"why":"RadFormer, the previous SOTA for GBC classification that LQ-Adapter's ROI-based pipeline surpasses.","marker":"[4]"},{"why":"Uni-perceiver, the frozen pre-trained backbone used in all LQ-Adapter experiments.","marker":"[51]"},{"why":"Kvasir-Seg, the colonoscopy polyp dataset used to demonstrate cross-domain generalizability.","marker":"[23]"}],"fun_headline_variants":["Learnable queries sharpen ultrasound cancer detection by 5.4%","Queries beat CNN priors: LQ-Adapter tops GBC ultrasound mIoU","Tiny adapter, big gain: learnable queries boost GBC detection","ViT-Adapter upgraded: 5.4% mIoU jump via learned queries","Smarter queries for ultrasound: LQ-Adapter leads GBC detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported improvements over DINO and FocalNet-DINO rest on the assumption that those baselines were trained or fine-tuned fairly and comparably on the GBCU dataset, but the paper does not describe their training recipe.","fun_headline_variants_meta":{"raw":{"variants":["Learnable queries sharpen ultrasound cancer detection by 5.4%","Queries beat CNN priors: LQ-Adapter tops GBC ultrasound mIoU","Tiny adapter, big gain: learnable queries boost GBC detection","ViT-Adapter upgraded: 5.4% mIoU jump via learned queries","Smarter queries for ultrasound: LQ-Adapter leads GBC detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3275,"prompt_tokens":1104,"completion_tokens":2171,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":720,"completion_tokens_details":{"reasoning_tokens":2078}},"tokens_in":720,"tokens_out":2171,"duration_ms":11356,"temperature":1.0,"reasoning_tokens":2078,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:26:14.522670+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain DINO and FocalNet-DINO on GBCU using the same data splits, augmentation, learning-rate schedule, and evaluation protocol as LQ-Adapter, and release the exact hyperparameters; if their mIoU rises to match or exceed LQ-Adapter's 0.719, the claimed state-of-the-art advantage would be refuted.","supporting_citations":[{"cited_title":"Focal modulation networks, 2022","cited_arxiv_id":null,"evidence_quote":"FocalNet-DINO, the current DETR-based SOTA baseline that LQ-Adapter outperforms on mIoU."},{"cited_title":"Surpassing the human accuracy: Detecting gallbladder cancer from usg images with curricu- lum learning","cited_arxiv_id":null,"evidence_quote":"GBCNet, the previous GBC detection pipeline whose ROI generator LQ-Adapter replaces and whose MS-SoP classifier is reused for the classification experiment."},{"cited_title":"Radformer: Transformers with global–local attention for interpretable and accurate gallbladder cancer detection","cited_arxiv_id":null,"evidence_quote":"RadFormer, the previous SOTA for GBC classification that LQ-Adapter's ROI-based pipeline surpasses."},{"cited_title":"Uni-perceiver: Pre- training unified architecture for generic perception for zero- shot and few-shot tasks","cited_arxiv_id":null,"evidence_quote":"Uni-perceiver, the frozen pre-trained backbone used in all LQ-Adapter experiments."},{"cited_title":"Kvasir-seg: A segmented polyp dataset","cited_arxiv_id":null,"evidence_quote":"Kvasir-Seg, the colonoscopy polyp dataset used to demonstrate cross-domain generalizability."}],"review_version":1}