{"id":"4e587c3f-a133-49c0-bdc6-2ed6bfddf37e","arxiv_id":"2412.10106","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An EfficientViT variant with a cascaded atrous attention module reports up to 98% accuracy on Mpox lesion datasets with 37.5% fewer parameters, but the SOTA claim is not statistically robust.","lead":"This paper proposes a new attention module for Mpox skin lesion classification that combines dilated convolutions with cascaded group attention, reporting 98% accuracy on one small dataset with fewer parameters than the base model. A generalist reader might care because efficient and accurate automated screening could aid Mpox triage in settings where lab tests are slow.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MSLD evaluation splits after augmentation, so train and test contain augmented variants of the same images; the 0.9969 accuracy and 'consistently outperforms' claim may reflect memorization rather than generalization.","rationale":"The paper's central claim is that EfficientViT-CAGA achieves state-of-the-art accuracy across all three datasets while reducing parameters. The efficiency pillar is plausible on the reported parameter and FLOP counts, and the MCSI and MSID protocols appear to apply augmentation only to training data. The MSLD protocol, however, explicitly augments the full dataset before applying the 70:20:10 split, which permits optimistic accuracy estimates from train/test near-duplicates. This is a concrete, internally identifiable flaw rather than a disagreement with external consensus, and it directly affects one of the three datasets used to claim consistent superiority. The reader's conditional verdict already identifies this same weakness, so the stress-test pass does not change the verdict. Other concerns, such as the absence of paired significance testing on MCSI and the small ablation study, are real but secondary: they would weaken the strength of the claims without invalidating an entire supporting dataset. A corrected MSLD evaluation is the single check most likely to determine whether the central claim survives.","tokens_in":11881,"tokens_out":7127,"duration_ms":74204,"concrete_test":"Re-run the MSLD comparison with a strict split-before-augmentation protocol: partition the original 228 images into 70/20/10 before any augmentation, apply augmentation only to the training partition, and retrain all Table IV models with the same hyperparameters and seeds, testing on the unaugmented held-out partition. If EfficientViT-CAGA no longer ranks first on MSLD, or its margin over DeiT3-Medium and EfficientViT-L1 is within one standard deviation, the three-dataset state-of-the-art claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.C.2 states that for MSLD the dataset was first expanded by augmentation (rotation, translation, reflection, shear, hue, saturation, contrast, brightness jitter, noise, and scaling) and that 'the augmented dataset was then divided into training, validation, and testing sets using a 70:20:10 split ratio.' Splitting after augmentation places augmented copies of the same original image on both sides of the train/test boundary. The near-perfect MSLD results (EfficientViT-CAGA 0.9969 accuracy, 1.0 on the Mpox class) can therefore be explained by the model recognizing near-duplicate variants encountered during training rather than by generalization to new lesion images. MSLD is one of the three datasets used to support the 'consistently outperforms existing approaches' claim, and the reported margin over DeiT3-Medium (0.9969 vs 0.9938) is far smaller than the inflation such leakage can produce. Comparing all models under the same leaky protocol does not neutralize the problem, because models with different capacities and inductive biases memorize augmented duplicates to different degrees. Until a split-before-augmentation protocol is applied, MSLD cannot serve as valid evidence for the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a Cascaded Atrous Group Attention (CAGA) module that combines a cascaded atrous attention (CAA) mechanism with cascaded group attention (CGA), and integrates it with the EfficientViT-L1 backbone for Mpox lesion classification. The authors report state-of-the-art accuracy on three datasets (MCSI, MSID, MSLD) with reduced parameters and FLOPs relative to several baseline ViT and CNN models, and provide an ablation study and Grad-CAM visualizations. The central claim is that EfficientViT-CAGA achieves higher accuracy than existing methods across all evaluated datasets while being computationally efficient.","tokens_in":12148,"tokens_out":2319,"duration_ms":24514,"significance":"If the claims are validated, the CAGA module would be a lightweight, accurate alternative for Mpox classification, which is clinically useful given the need for fast and reliable screening. The paper's strengths include evaluation on three public datasets, comparison with a reasonable set of modern baselines, an ablation study, and an explicit efficiency analysis. However, the supporting evidence for the central SOTA claim is weakened by methodological concerns in the MSLD evaluation, a lack of statistical significance testing on the MCSI result, and inconsistencies in reported parameter counts. The architecture itself is a plausible incremental improvement over existing attention mechanisms, but the empirical validation currently does not meet the bar for a definitive SOTA claim.","major_comments":[{"comment":"Section IV.C.2 states that for MSLD the dataset was first expanded by augmentation and then divided into training, validation, and testing sets using a 70:20:10 split. Augmentations of the same original image (rotation, translation, reflection, shear, etc.) can therefore appear in both training and test partitions, creating near-duplicate leakage. The near-perfect MSLD accuracy (0.9969, 1.0 for the Mpox class) and the claimed consistent outperformance on MSLD are not reliable evidence of generalization under this protocol. A split-before-augmentation protocol is required before MSLD can support the central claim.","section":"IV.C.2"},{"comment":"On the MCSI dataset, EfficientViT-CAGA achieves 0.98 ± 0.0229 accuracy, while the EfficientViT-L1 backbone achieves 0.9725 ± 0.018. The difference of 0.0075 is well within one standard deviation of either result, so the reported improvement is not statistically significant. The manuscript should provide a paired significance test (e.g., a paired t-test or Wilcoxon test across the 10 folds) or confidence intervals before claiming improved accuracy over the backbone.","section":"Table II"},{"comment":"The parameter counts for the same architectures are inconsistent across tables: EfficientViT-L1 is reported as 58.9M in Tables I and II but 42M in Table IV, and EfficientViT-CAGA is reported as 37.8M in Table I but 36.8M in Tables II and IV. These inconsistencies (also affecting the stated 37.5% and 35.8% parameter reductions) make the efficiency claims difficult to verify. The authors should state the exact parameter counting method and ensure all tables use consistent model configurations.","section":"Tables I, II, and IV"},{"comment":"Results on MSID (Table I) and MSLD (Table IV) are reported as point estimates with no standard deviations or confidence intervals, even though MCSI uses 10-fold cross-validation. Without error bars, the claimed 'consistent outperformance' over DeiT3-Medium (0.9545 vs 0.9481 on MSID; 0.9969 vs 0.9938 on MSLD) cannot be assessed for statistical significance. The authors should report variance across multiple splits or folds for these datasets as well.","section":"Tables I and IV"},{"comment":"The manuscript states that dilation rates are 'adjusted for each dataset' and that focal loss uses 'class-specific weights (alpha),' implying per-dataset tuning of these hyperparameters on the same benchmark test sets. Reporting tuned results as SOTA without external validation or a nested cross-validation procedure risks selection bias. The authors should either fix hyperparameters a priori, provide sensitivity analyses, or validate on a held-out external dataset.","section":"IV.B"}],"minor_comments":[{"comment":"The conclusion says 'the approach highlighted state-of-the-art results,' which is awkward phrasing; consider rewording to 'achieved state-of-the-art results.'","section":"V"},{"comment":"The implementation details state that the learning rate is 'on the order of 10^-5,' which is imprecise; the exact value should be reported for reproducibility.","section":"IV.B"},{"comment":"Equation (2) uses the index i in the left-hand side but the cascading in d is not fully defined; clarify the indexing of heads and dilations to avoid ambiguity.","section":"III.A"},{"comment":"The phrase 'reducing model parameters by 37.5% compared to the original EfficientViT-L1' is only accurate for the MCSI configuration (58.9M vs 36.8M); the percentages differ for MSID and MSLD, so state the specific comparison.","section":"Abstract"},{"comment":"The statement that 'runtime data augmentation techniques were applied to the training set' and caused 'a marginal decrease in performance' is vague; specify which augmentation techniques were used and how this was applied consistently across models.","section":"IV.C.1"},{"comment":"The code availability line says 'The code will be available here' but no link is provided; please include a URL or state a clear availability plan.","section":"I"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely application and the proposed module is a reasonable incremental design. The main concerns are methodological: the MSLD leakage, lack of significance testing, and inconsistent parameter counts. These issues are fixable within the manuscript's scope by re-running the experiments with proper protocols and adding statistical validation. I would not recommend rejection, but the revised version must address the load-bearing concerns before the SOTA claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a reasonable engineering paper, not a breakthrough. The CAGA module is a specific combination of dilated convolutions (from ASPP) and cascaded group attention (from EfficientViT), and the paper cites both sources. That is legitimate—no overclaiming of component novelty. The module is clearly described, and the experiments include class-wise accuracy, FLOPs, parameters, ablation, and Grad-CAM. The parameter reduction relative to the EfficientViT-L1 backbone is real, even if the exact numbers differ between tables (58.9M vs 42M for the backbone, 37.8M vs 36.8M for CAGA).\n\nThe soft spots are real and should be fixed before publication. The biggest one is the MSLD protocol: Section IV.C.2 says the dataset was expanded by augmentation and then split 70:20:10. That puts augmented copies of the same source images on both sides of the train/test boundary. The near-perfect accuracy (0.9969) is exactly what you'd expect from memorization, and comparing all models under the same leaky protocol does not neutralize the problem because capacities and inductive biases differ. MSLD is one of three datasets used to support the \"consistently outperforms\" claim, so this is a load-bearing flaw for that claim.\n\nOn MCSI, the 0.98 ± 0.0229 result overlaps with the backbone's 0.9725 ± 0.018, so the \"significant\" improvement in the abstract is not statistically supported. The MSID result is more credible—the split is before augmentation, and the class-wise balance is a genuine point in the model's favor. The per-dataset tuning of dilation rates and focal loss alpha on these same benchmarks is a mild circularity, not a fatal one. The ablation study uses randomly initialized weights and shows much lower accuracy, which makes it hard to interpret for the pretrained model.\n\nWho is this for? People working on efficient attention modules for small medical image datasets. It deserves a serious referee, not a desk reject, because the architecture is reproducible, the central issue (MSLD) is fixable, and the MCSI/MSID results may hold up under a re-run. I would not cite it in my own work yet. My recommendation: send it out, but expect a major revision that fixes the MSLD split, reports significance, reconciles parameter counts, and releases code.","headline":"A modest, honestly assembled architecture with one clean dataset result; the MSLD claim is undermined by augmentation-before-split leakage, and the MCSI 'SOTA' gap is not statistically significant.","tokens_in":12656,"tokens_out":2308,"would_cite":false,"duration_ms":24477,"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 cascaded dilated-attention module claims top accuracy on three Mpox lesion datasets while cutting parameters by over a third.","keywords":["Mpox classification","skin lesion classification","vision transformer","dilated convolution","atrous attention","self-attention","EfficientViT","medical image classification"],"falsifier":"Split the original 228 MSLD images into train, validation, and test sets first, then apply data augmentation only to the training partition, retrain EfficientViT-CAGA, and measure test accuracy; if it drops substantially from 0.9969, the cross-dataset SOTA claim loses the MSLD pillar.","tokens_in":11677,"feed_emoji":"🦠","tokens_out":7008,"duration_ms":65298,"temperature":0.7,"pith_summary":"This paper tries to show that a carefully designed attention module, called Cascaded Atrous Group Attention (CAGA), can make Mpox skin-lesion classification more accurate and more efficient at the same time. The module combines dilated convolutions at multiple rates with a cascaded multi-head attention design, and it is attached to the EfficientViT-L1 backbone. On three benchmark datasets the paper reports state-of-the-art accuracy, including 98% on the MCSI dataset, while cutting parameters by roughly 37.5% compared to the backbone alone. If true, the work would give clinicians an accurate model that is light enough to run on modest hardware, and it would provide a reusable attention block for other skin-lesion tasks.","feed_headline":"Mpox classifier hits 98% accuracy with 37.5% fewer parameters","feed_subtitle":"A cascaded dilated-attention module tops three skin-lesion benchmarks while trimming model size.","key_machinery":"The load-bearing object is the Cascaded Atrous Group Attention (CAGA) module, which stacks two mechanisms. The Cascaded Atrous Attention (CAA) part applies dilated convolutions (rates 1, 2, 3) to each attention head, runs self-attention per dilated map, and adds each map's output to the next to create multi-scale context. The Cascaded Group Attention (CGA) part splits the input into heads and gives each head the previous head's output added to it, so information flows across heads and head redundancy is reduced. This module replaces the standard self-attention in the EfficientViT-L1 backbone, turning it into EfficientViT-CAGA; the cascade and the atrous rates are what carry the multi-scale argument.","core_discovery":"The central claim is that CAGA, a module that nests Cascaded Atrous Attention (CAA) inside Cascaded Group Attention (CGA), lets an EfficientViT-L1 model beat all compared CNN and vision-transformer baselines on three Mpox lesion datasets. CAA runs dilated convolutions with rates 1, 2, 3 on each attention head, computes self-attention on each dilated feature map, then cascades the attention outputs across scales, while CGA adds each processed head's output to the next head to cut redundancy. Integrated with the EfficientViT-L1 backbone, the model reaches 98% accuracy on the MCSI dataset, outperforms baselines on MSID and MSLD, and uses 36.8M parameters and 4.86G FLOPs, about 37.5% fewer parameters than the backbone. The paper also reports balanced per-class accuracy on MSID and near-perfect Mpox recall on MSLD, and it offers an ablation study showing that the cascade and the CGA wrapper each contribute to the gain.","pith_inferences":["The MSLD evaluation splits the data after augmentation, so the test set can contain near-duplicates of training images; re-running the comparison with augmentation applied only after splitting would show whether the reported 99.69% accuracy reflects real generalization or memorization of augmented variants.","The dilation rates are fixed per dataset, so the module's success may depend on those specific rates; testing other rate schedules or making the rates learnable could separate the effect of the cascade structure from the effect of the chosen receptive fields.","Because the parameter counts given for the same baseline models differ slightly across tables (e.g., ResNet-101 has 43.5M parameters in the MSID table but 47.4M in the MCSI table), the exact 37.5% reduction depends on reproducible parameter counting, which an independent re-count could verify."],"forward_implications":["If the accuracy numbers hold, the model could run on edge devices for Mpox screening, since it uses only 4.86G FLOPs and 36.8M parameters, less than most compared architectures.","The balanced class-wise accuracy on MSID suggests the module counteracts the class-specific overfitting that the compared CNNs and ViTs show, which would be valuable in real-world triage where all lesion types must be recognized.","The module is defined independently of the task head, so it could be dropped into other vision-transformer backbones for skin-lesion classification or other fine-grained medical image tasks.","The paper's ablation attributes part of the gain to the cascading and the CGA wrapper, implying that the hierarchy itself, not just enlarged receptive fields, is responsible for the improvement."],"supporting_citations":[{"why":"Supplies the EfficientViT-L1 backbone whose convolutional body is reused and whose parameter count serves as the efficiency baseline.","marker":"[20]"},{"why":"Defines Cascaded Group Attention, the structure CAGA builds on to reduce multi-head redundancy.","marker":"[44]"},{"why":"DeepLab's Atrous Spatial Pyramid Pooling inspires the multi-rate dilated convolution design in Cascaded Atrous Attention.","marker":"[45]"},{"why":"Provides the MCSI dataset and the 10-fold cross-validation protocol that yields the headline 98% accuracy.","marker":"[17]"},{"why":"Provides the MSID dataset used as the second benchmark for the SOTA comparison.","marker":"[21]"},{"why":"Provide the MSLD dataset used as the third benchmark, including the augmentation-based split that the evaluation relies on.","marker":"[22] [23]"}],"fun_headline_variants":["Cascaded dilated attention hits 98% Mpox accuracy, cuts parameters 37.5%","Mpox classifier: 98% accuracy with 37.5% fewer parameters","New attention module trims model size while hitting 98% Mpox accuracy","CAGA framework: efficient Mpox detection at 98% accuracy","Dilated cascades boost Mpox diagnosis to 98% with leaner model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported state-of-the-art accuracy assumes each benchmark's split is clean, and in particular that the augmented MSLD images used for training do not reappear in the test set; if they do, the 99.69% accuracy on that dataset would measure memorization, not generalization to new patients.","fun_headline_variants_meta":{"raw":{"variants":["Cascaded dilated attention hits 98% Mpox accuracy, cuts parameters 37.5%","Mpox classifier: 98% accuracy with 37.5% fewer parameters","New attention module trims model size while hitting 98% Mpox accuracy","CAGA framework: efficient Mpox detection at 98% accuracy","Dilated cascades boost Mpox diagnosis to 98% with leaner model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1554,"prompt_tokens":1005,"completion_tokens":549,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":443}},"tokens_in":621,"tokens_out":549,"duration_ms":5749,"temperature":1.0,"reasoning_tokens":443,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:20:04.806437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Split the original 228 MSLD images into train, validation, and test sets first, then apply data augmentation only to the training partition, retrain EfficientViT-CAGA, and measure test accuracy; if it drops substantially from 0.9969, the cross-dataset SOTA claim loses the MSLD pillar.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DeepLab's Atrous Spatial Pyramid Pooling inspires the multi-rate dilated convolution design in Cascaded Atrous Attention."},{"cited_title":"A transfer learning and explainable solution to detect mpox from smartphones images","cited_arxiv_id":null,"evidence_quote":"Provides the MCSI dataset and the 10-fold cross-validation protocol that yields the headline 98% accuracy."},{"cited_title":"Monkeynet: A robust deep convolutional neural network for monkeypox disease detection and classification","cited_arxiv_id":null,"evidence_quote":"Provides the MSID dataset used as the second benchmark for the SOTA comparison."}],"review_version":1}