{"id":"88f5fee3-b30d-40c6-a3d4-8754a5f735a5","arxiv_id":"2507.21922","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SwinECAT combines Swin Transformer with ECA channel attention and reports 88.29% accuracy on 9-class fundus disease classification on EDID, though the claim rests on a single data split.","lead":"This paper inserts the lightweight Efficient Channel Attention module into each stage of a Swin Transformer backbone to classify fundus images into nine categories, reporting 88.29% accuracy on the public EDID dataset. A generalist might read it as a routine architecture tweak in medical AI, but its state-of-the-art claim rests on a single split with no error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains depend on untuned baselines: fixed lr/patience for all models and no stated pretrained initialization, so the 1.7-point lead over Swin may not survive per-model hyperparameter tuning.","rationale":"The paper's headline is a comparative claim: SwinECAT 'significantly outperforms' Swin and other baselines. For that claim to be true, the comparison must be at representative operating points. Section 4.1 declares all models share lr=1e-5, batch=32, and patience=3; it never states whether pretrained ImageNet weights were loaded for each architecture. The CNN-Transformer baseline at 79.93% is far below typical performance of its components on similar tasks, suggesting it may not have converged. A single run with no error bars cannot separate a true advantage from training noise (test n=1614 gives SE≈0.8%). The proposed test—per-model lr sweep plus repeated seeds—directly checks whether the advantage persists under minimally fair conditions. If it does, the central claim survives; if not, the conclusions reduce to an artifact of experimental setup. This is exactly the condition the reader identified, so I agree with their weakest assumption and keep the CONDITIONAL verdict until the test is run.","tokens_in":13035,"tokens_out":9562,"duration_ms":102618,"concrete_test":"Using the same data split, run each baseline and SwinECAT with a per-model learning-rate search (e.g., 3e-6, 1e-5, 3e-5, 1e-4) and early-stopping patience 10, selecting the checkpoint with the lowest validation loss; load the official pretrained weights for each architecture and run five random seeds per configuration. Report mean and standard deviation. If SwinECAT does not have the highest mean accuracy and F1 with a gap beyond the pooled standard error, the 'significantly outperform' and state-of-the-art claims fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 states that all models use learning rate 1e-5, batch size 32, and early stopping with patience 3, and that hyperparameters are kept consistent. It does not say which pretrained weights, if any, are loaded for each architecture. The 79.93% accuracy of the CNN-Transformer baseline (76.07M parameters) sits far below what its components (Inception-ResNet-v2 + ViT) typically achieve on fundus tasks, and BEiT at 85.32% is also suspicious given its stronger pretraining. With a fixed low learning rate and patience 3, large models can terminate prematurely. The central comparative claim—SwinECAT 'significantly outperforms' these baselines—therefore rests on the unstated assumption that every baseline was at a representative operating point. If that assumption is false, the 1.73-point gain over Swin and the state-of-the-art claim are artifacts of undertrained competitors. The single test set of 1,614 images provides no error bars, so neither the magnitude nor the significance of the gap is established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SwinECAT, a modification of the Swin Transformer that inserts Efficient Channel Attention (ECA) modules into each of the four hierarchical stages, and evaluates it on the EDID fundus dataset for 9-category disease classification. The authors report 88.29% accuracy, macro F1 0.90, and weighted F1 0.88 on a single 80/10/10 split, claiming that these results significantly outperform the Swin Transformer baseline and other general-purpose and fundus-specific models, and that this is the highest reported accuracy for this task on this dataset.","tokens_in":13204,"tokens_out":2751,"duration_ms":29203,"significance":"If the empirical claims are correct, the paper would provide a simple, parameter-efficient enhancement of a widely used vision transformer backbone for a clinically relevant 9-class medical image classification task, with a modest gain over the Swin Transformer baseline. The use of a public dataset and the inclusion of parameter counts for all compared models are strengths. However, the significance is substantially tempered by the lack of statistical rigor: all results come from a single data split with no repeated runs, no standard deviations, and no significance testing, and the baseline comparisons appear to rest on fixed hyperparameters without evidence that each model was at a representative operating point. As presented, the central claim of 'significant' improvement and state-of-the-art performance is not adequately supported, although the underlying architecture is reasonable and the claimed direction of improvement is plausible.","major_comments":[{"comment":"The claim that SwinECAT 'significantly outperform[s]' the baselines is not supported by the experimental protocol. All results are from a single 80/10/10 split with no repeated runs, no standard deviations, and no statistical significance test. The reported gain over the Swin Transformer baseline is 1.73 percentage points (88.29 vs 86.56), which could easily be within run-to-run variance for deep models on a test set of 1,614 images. The paper should report results across multiple seeds or cross-validation folds, and should apply an appropriate significance test (e.g., a paired test across seeds) before using the word 'significantly'.","section":"§4.1, §4.2, Table 3"},{"comment":"The baseline comparison is confounded by the fixed hyperparameter protocol. The paper states that all models use learning rate 1e-5, batch size 32, and early stopping with patience 3, but it does not state whether each model was tuned for this task or which pretrained weights (if any) were used. The CNN-Transformer baseline achieving only 79.93% accuracy, far below what its components (Inception-ResNet-v2 plus ViT) typically achieve on fundus tasks, strongly suggests undertraining or suboptimal initialization. Without evidence that every baseline was at a representative operating point—for example, by tuning learning rate and early stopping per model or by showing consistent results across a small hyperparameter sweep—the comparative claim that SwinECAT outperforms these baselines is not established.","section":"§4.1, Table 3"},{"comment":"The state-of-the-art claim ('the highest reported performance for 9-category classification on this public dataset') is not substantiated within the manuscript. The comparison is limited to the authors' own re-implementations of a small set of baselines; there is no systematic comparison with published results on the same EDID dataset, and no discussion of whether those published results used the same data split, preprocessing, or class definitions (e.g., whether pterygium was excluded). To support the state-of-the-art claim, the authors should compare against existing published numbers on this dataset under a consistent evaluation protocol, or at least clearly state that no such comparison was performed.","section":"Abstract, §4.2.1, Table 3"}],"minor_comments":[{"comment":"The sentence 'On this basis, ECA module is added after the four stages' is ambiguous: it could mean after each stage or only after the final stage. Given Figure 1 and the surrounding text, it likely means 'after each of the four stages,' but the wording should be corrected for reproducibility.","section":"§3.2.1"},{"comment":"In the paragraph discussing Figure 3(g-i), the sentence comparing SwinECAT's validation loss with 'the combined model method of Figure 3(g) and Figure 3(i)' appears to contain a typo: the comparison should be with Figure 3(g) and Figure 3(h), not with itself. Please verify and correct.","section":"§4.2.1"},{"comment":"The experimental setup should explicitly state whether pretrained weights were used for each model and, if so, which checkpoint (e.g., ImageNet-1k or ImageNet-21k). This detail is essential for reproducibility and for interpreting the baseline results.","section":"§4.1"},{"comment":"The table formatting is difficult to read because the column headers are run together (e.g., 'Parameters(M)ViT' instead of a separate column). Please format the table so that each column is clearly separated.","section":"Table 3"},{"comment":"There is a duplicated reference in the introduction: '[15, 15, 16]' should likely be '[14, 15, 16]' or similar. Also, there are several minor grammatical and punctuation issues throughout (e.g., inconsistent capitalization of 'SwinECAT'/'swinECAT' and 'Transformer' in places like 'Swin T ransformer').","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal but the empirical evaluation is currently too weak to support the stated conclusions. The authors should be asked to address the statistical rigor and baseline fairness concerns; if they can provide repeated-run results with error bars and a more careful baseline tuning protocol, the contribution may be publishable. I also note that no code or data availability statement is provided, which is increasingly expected for empirical papers in this area."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest, well-meaning benchmark paper. The new thing is the specific insertion of ECA after each Swin stage for 9-class fundus classification on EDID, and the authors show it beats a Swin baseline by about 1.7 accuracy points. That gain is plausible, and the paper is honest enough to report a broad set of baselines and an ablation.\n\nThe strengths: the architecture is clearly described, the comparison includes the two most relevant prior fundus classifiers (MaxViT+ResNet and CNN-Transformer), and the parameter count (28.3M) is reasonable. The ablation against Swin is the cleanest experiment in the paper, since the only change is the ECA modules.\n\nThe soft spots are methodological and they matter. Everything rests on a single 80/10/10 split with no repeated runs, no standard deviations, and no significance test. The 1.73-point gap over Swin could easily be run-to-run noise. Second, the paper fixes learning rate (1e-5), batch size, and early stopping patience (3) for every model, which is a poor fit when the baselines range from 21M to 85M parameters. The CNN-Transformer at 79.93% (with 76M params) is far below what that architecture should deliver, and BEiT at 85.32% also looks undertrained; the stress-test note that the comparison may rest on undertrained baselines is fair. Third, the paper does not say whether any model is pretrained or which weights are loaded, which is essential for reproducing the numbers. Finally, the 'expands to 9 categories' claim is overstated, since reference [12] already does nine-class infantile fundus classification; the authors cite it, so they know.\n\nNone of this kills the core idea. Swin+ECA is a sensible combination and likely yields a small improvement on this task. But as presented, the headline 'significantly outperforms' is not supported. The paper needs repeated runs, per-model hyperparameter settings or a sensitivity analysis, pretraining disclosure, and a softer state-of-the-art claim.\n\nWho it's for: people working directly on EDID or on lightweight attention additions to Swin for medical imaging. It deserves a serious referee turn, but with the expectation of major methodological revision before acceptance. I'd lean toward sending it to review rather than desk-rejecting.","headline":"A straightforward Swin+ECA combination on a 9-class fundus benchmark; the 1.7-point gain over Swin is plausible but the paper's single-split, fixed-hyperparameter evaluation leaves the 'significantly outperforms' claim unsupported.","tokens_in":13800,"tokens_out":2355,"would_cite":false,"duration_ms":27043,"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 adding an efficient channel attention module after each stage of a Swin Transformer reaches 88.29% accuracy on nine-class fundus disease classification, the highest reported on the EDID dataset.","keywords":["fundus disease classification","Swin Transformer","efficient channel attention","shifted window attention","medical image classification","EDID dataset","nine-class classification","channel attention"],"falsifier":"Retrain Swin Transformer and SwinECAT from scratch on the same EDID split across multiple random seeds and compute the distribution of test accuracy; the central claim of an ECA-driven improvement collapses if the accuracy distributions overlap substantially (e.g., the 1.73-point gap is within one standard deviation).","tokens_in":12799,"feed_emoji":"🩺","tokens_out":5536,"duration_ms":58812,"temperature":0.7,"pith_summary":"Paper claim: grafting an Efficient Channel Attention module onto each stage of a Swin Transformer produces the best published nine-class fundus-disease classifier on the EDID benchmark, at 88.29% accuracy, macro F1 of 0.90, and weighted F1 of 0.88. The proposed SwinECAT model is designed to tackle two well-known difficulties in fundus images: lesions that occupy tiny regions and diseases whose visual differences are subtle. Because the ECA block adds only about 0.77M parameters over the 27.53M Swin baseline, the paper argues that channel attention is a cheap and effective complement to shifted-window spatial attention. If the result holds, it means a modest architectural change to a standard transformer can move the needle on a fine-grained medical classification task.","feed_headline":"SwinECAT hits 88.29% accuracy on nine-class fundus disease task","feed_subtitle":"A lightweight channel attention module lifts Swin Transformer past prior fundus classifiers on the EDID benchmark.","key_machinery":"The model's central object is the Swin Transformer backbone, a hierarchical four-stage vision transformer whose attention is computed inside local windows and shifted between consecutive layers to allow cross-window information exchange, augmented with one Efficient Channel Attention (ECA) block after each stage. ECA compresses each feature map by global average pooling, applies a one-dimensional convolution with an adaptively chosen kernel size along the channel dimension, and uses a sigmoid to produce per-channel weights that rescale the features. The argument is that the shifted-window attention supplies multi-scale spatial modeling from local details to global context, while ECA, with only a tiny parameter cost, steers the network toward the most informative channels, so the two mechanisms jointly improve discrimination in fundus images.","core_discovery":"The central discovery is that fusing the Swin Transformer backbone with ECA modules, one after each of its four stages, improves classification performance on a nine-class fundus disease task: 88.29% accuracy, macro F1 of 0.9000, and weighted F1 of 0.8830, versus 86.56% accuracy and macro F1 of 0.8849 for the Swin Transformer baseline, with about 0.77M added parameters. The paper reports that this accuracy is higher than six general vision baselines (ViT, ConViT, MaxViT, BEiT, ResNet50, Swin Transformer) and two recent fundus-specific models (MaxViT+ResNet18 and CNN-Transformer), while using fewer parameters than both recent fundus-specific models. The authors attribute the gain to the complementary fusion of spatial attention from shifted windows and channel attention from ECA, which makes the model more sensitive to small lesion regions and fine inter-class differences.","pith_inferences":["Beyond the paper: a single 80/10/10 split underlies every number in Table 3, so the 1.73-point gain over Swin Transformer should be checked for seed-to-seed variance before treating it as stable.","Beyond the paper: placing ECA inside each Swin block rather than after each stage, or combining it with class-balanced sampling for the smallest classes, would isolate whether the gain comes from channel attention or from the extra processing at stage boundaries.","Beyond the paper: because EDID was captured with two Topcon cameras in one hospital network, external validation on other fundus datasets would test whether the channel-attention advantage carries across devices and populations."],"forward_implications":["A nine-class fundus screening system can reach roughly 88% accuracy with about 28 million parameters, which is lighter than the comparable MaxViT+ResNet fusion model.","Adding ECA to Swin Transformer yields a 1.7 percentage point accuracy gain with a minimal parameter increase, suggesting channel re-weighting is a high-leverage modification for medical image classification.","Pure global-attention transformers such as ViT, BEiT, and ConViT, and heavyweight fusion models such as CNN-Transformer, are less competitive than the hybrid spatial-channel design on this task.","The same four-stage SwinECAT recipe is a plausible starting point for other small-lesion medical imaging tasks like diabetic retinopathy grading or optic-disc analysis."],"supporting_citations":[{"why":"Supplies the Swin Transformer backbone with shifted-window attention that SwinECAT extends.","marker":"[23]"},{"why":"Supplies the efficient channel attention module, the paper's key addition and the source of the claimed gain.","marker":"[24]"},{"why":"Provides the EDID dataset and the nine-class benchmark on which SwinECAT is evaluated.","marker":"[31]"},{"why":"A recent fundus-specific feature-fusion baseline (MaxViT+ResNet18) that SwinECAT claims to outperform with fewer parameters.","marker":"[12]"},{"why":"A recent fundus-specific CNN-Transformer baseline that SwinECAT claims to outperform by a wide margin.","marker":"[13]"},{"why":"MaxViT, a hybrid local-global attention baseline used for comparison.","marker":"[29]"},{"why":"ResNet50, the CNN baseline used for comparison.","marker":"[30]"},{"why":"ViT, the global-attention transformer baseline used for comparison.","marker":"[26]"}],"fun_headline_variants":["SwinECAT fuses Swin and ECA attention for 9-class fundus diagnosis","SwinECAT: 88.29% accuracy on nine-class fundus disease task","Swin Transformer plus efficient channel attention boosts fundus classification","SwinECAT sets 9-class fundus disease record on EDID dataset","Lightweight attention lift: SwinECAT outperforms on fundus images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every compared model is trained to a representative performance level under the shared hyperparameters (learning rate 1e-5, batch size 32, early stopping with patience 3), so the reported gaps reflect architecture differences rather than undertrained baselines.","fun_headline_variants_meta":{"raw":{"variants":["SwinECAT fuses Swin and ECA attention for 9-class fundus diagnosis","SwinECAT: 88.29% accuracy on nine-class fundus disease task","Swin Transformer plus efficient channel attention boosts fundus classification","SwinECAT sets 9-class fundus disease record on EDID dataset","Lightweight attention lift: SwinECAT outperforms on fundus images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":3082,"prompt_tokens":1038,"completion_tokens":2044,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":1939}},"tokens_in":654,"tokens_out":2044,"duration_ms":13862,"temperature":1.0,"reasoning_tokens":1939,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:13:09.911292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain Swin Transformer and SwinECAT from scratch on the same EDID split across multiple random seeds and compute the distribution of test accuracy; the central claim of an ECA-driven improvement collapses if the accuracy distributions overlap substantially (e.g., the 1.73-point gap is within one standard deviation).","supporting_citations":[{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"ResNet50, the CNN baseline used for comparison."},{"cited_title":"Swin transformer: Hierarchical vision transformer using shifted windows","cited_arxiv_id":null,"evidence_quote":"Supplies the Swin Transformer backbone with shifted-window attention that SwinECAT extends."},{"cited_title":"ECA-Net: Efficient channel attention for deep convolutional neural net- works","cited_arxiv_id":null,"evidence_quote":"Supplies the efficient channel attention module, the paper's key addition and the source of the claimed gain."},{"cited_title":"A dataset of color fundus images for the detection and classification of eye diseases","cited_arxiv_id":null,"evidence_quote":"Provides the EDID dataset and the nine-class benchmark on which SwinECAT is evaluated."},{"cited_title":"Automated detection of nine infantile fundus diseases and conditions in retinal images using a deep learning system","cited_arxiv_id":null,"evidence_quote":"A recent fundus-specific feature-fusion baseline (MaxViT+ResNet18) that SwinECAT claims to outperform with fewer parameters."},{"cited_title":"Lesion identification in fundus images via convolutional neural network-vision transformer","cited_arxiv_id":null,"evidence_quote":"A recent fundus-specific CNN-Transformer baseline that SwinECAT claims to outperform by a wide margin."},{"cited_title":"Maxvit: Multi-axis vision transformer","cited_arxiv_id":null,"evidence_quote":"MaxViT, a hybrid local-global attention baseline used for comparison."}],"review_version":1}