{"id":"22e2d82d-deef-49fe-a20f-c6f9c2d98271","arxiv_id":"2509.02918","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"KG-DG fuses YOLO-derived lesion features with a frozen ViT via confidence-based fusion and claims gains in diabetic retinopathy domain generalization, but the central KL-divergence mechanism and the MDG headline are contradicted by the paper's own tables.","lead":"A neuro-symbolic framework for diabetic retinopathy that fuses lesion-detection features with vision transformers is claimed to improve cross-dataset generalization. The paper's own tables contradict key headline results, and the central KL-divergence alignment mechanism is not implemented in the methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central cross-domain results depend on an unstated provenance assumption: the lesion/vessel extractors in Sec. 3.2 were fine-tuned on approx. 500 images of unspecified origin; if any target-domain images (e.g., APTOS) are in that set, the unseen-target SDG/MDG claims are contaminated.","rationale":"The reader's verdict is REJECT, and my stress-test does not change that. I agree with the reader's broad conclusion that the paper as written cannot support its central claims. However, I am singling out feature-extractor provenance rather than the label-remapping issue from the reader's weakest_assumption: if the approx. 500 fine-tuning images include APTOS, the most striking SDG result (MESSIDOR-trained model reaching 74% on APTOS in Table 3) becomes an artifact of feature-level leakage, not domain-invariant symbolic reasoning. This is load-bearing and testable. The reader also identified the missing KL-divergence mechanism and the self-contradictory MDG numbers (63.67 vs 65.5; ViT+KL 58.4 in Table 6); those are direct textual errors and already justify rejection. My added concern is about the hidden data-dependence of the symbolic extractors, which would remain a problem even if the arithmetic were corrected. Hence the verdict stays REJECT; if the provenance check came back clean, a corrected paper focused on the fusion result might be salvageable.","tokens_in":15612,"tokens_out":9189,"duration_ms":104444,"concrete_test":"Ask the authors for the composition of the approx. 500-image fine-tuning set used for YOLOv11 and the vessel segmenter (Sec. 3.2), or inspect the released annotation metadata and data splits. If the set includes any APTOS, EyePACS, Messidor-1, or Messidor-2 images, retrain the two extractors without those target-domain images and rerun Tables 2-6. If the KL-only or DL+KL accuracy on the corresponding target (e.g., APTOS under MESSIDOR training) drops by more than the reported cross-domain gains, the central generalization claim is contaminated by feature-level leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that symbolic lesion features generalize to unseen domains, with the strongest effect on APTOS (e.g., KL 74.0% in Table 3 with MESSIDOR as source, vs 49.8% for ViT). This claim rests on the extractors being domain-neutral. Section 3.2 says YOLOv11 and the vessel segmenter were fine-tuned on approximately 500 expert-annotated images but never states which dataset those images came from. If that fine-tuning set includes APTOS (a natural choice for lesion-rich annotations), then APTOS is not truly unseen for the symbolic branch: the feature extractor has been optimized on APTOS images even though the Gradient-Boosting classifier was trained on MESSIDOR features. The 74.0/74.1% APTOS entries in Tables 3-4 and the KL-only MDG average (63.67% in Table 6) would be partially in-domain at the feature level, directly undermining the +5.2% Messidor2->APTOS gain highlighted in Sec. 5.4 and the abstract's domain-invariant claim. This is a load-bearing unstated premise because the paper provides no way to verify that the extractors never saw target-domain data. Separately, the MDG claim is already internally inconsistent: Sec. 4.2 calls 63.67% the best while listing 65.5% methods, and Table 6's ViT+KL is 58.4%, below both its own branches and the 65.5% baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes KG-DG, a neuro-symbolic framework for diabetic retinopathy (DR) classification that combines a Vision Transformer (ViT) branch with a symbolic branch based on lesion detection (YOLOv11) and retinal vessel segmentation features. The symbolic features are classified by a Gradient Boosting model, and the two branches are fused via confidence-based strategies. The authors claim that the framework addresses both single-domain generalization (SDG) and multi-domain generalization (MDG) by minimizing KL divergence between domain embeddings, and report accuracy gains up to 5.2% over baselines on APTOS, EyePACS, Messidor-1, and Messidor-2, with a symbolic-only model achieving 63.67% average accuracy in MDG. The paper includes SDG tables per source domain, an MDG comparison table, and an ablation study on lesion versus vessel features.","tokens_in":15924,"tokens_out":2574,"duration_ms":29478,"significance":"If the central claims were supported, the paper would make a useful contribution: it would demonstrate that structured clinical lesion features, when combined with deep representations, improve cross-domain generalization in DR classification—a practically important problem. The idea of using knowledge-guided symbolic features as regularizers is plausible and aligned with prior work on concept bottleneck models and knowledge-guided medical imaging. However, the significance is undermined by the fact that the main proposed mechanism (KL divergence alignment of domain embeddings) is never actually defined or implemented in the methodology, and the headline MDG result is contradicted by the paper's own table. The experimental comparisons also suffer from unresolved questions about feature-extractor provenance and label remapping, making the empirical claims unverifiable as stated.","major_comments":[{"comment":"The abstract and related work claim that KG-DG 'addresses both SDG and MDG by minimizing the KL divergence between domain embeddings, thereby enforcing alignment of high-level clinical semantics.' However, no KL-divergence loss, domain-embedding definition, or alignment training step appears anywhere in the Methodology (Sec. 3). The only occurrence of 'KL' in the experimental sections is as an abbreviation for the knowledge-based classifier (e.g., 'Knowledge (KL)' in Tables 2-5 and 'ViT+KL' in Table 6). Thus the central alignment mechanism is absent; the paper conflates a classifier label with a domain-alignment objective.","section":"Abstract, Sec. 3, Sec. 5.2"},{"comment":"The MDG claim is internally contradicted. Sec. 4.2 states that 'the knowledge-guided symbolic model (KL only) achieved the best average accuracy (63.67%),' but Table 6 lists ERM-ViT (CvT-13) and SPSD-ViT with averages of 65.5%, i.e., above 63.67%. Additionally, the fused ViT+KL method (Ours) achieves only 58.4% average accuracy, which is below both its own neural-only ViT (61.18%) and the KL-only model (63.67%), and below nearly all baselines. The abstract's assertion that 'the complete neuro-symbolic integration achieves the highest accuracy compared to existing published baselines' is therefore not supported by the reported data.","section":"Sec. 4.2, Table 6"},{"comment":"The provenance of the fine-tuning images for the feature extractors is unspecified. The paper says that YOLOv11 and the vein segmentation module were fine-tuned on 'approximately 500' expert-annotated images but does not state which dataset those images came from. If any target-domain images (e.g., APTOS) were included, then the symbolic features for that target are partially in-domain at the feature-extractor level, and the cross-domain generalization claim—particularly the large APTOS gains in Tables 3 and 4—is contaminated. The authors must disclose the exact source of these fine-tuning images and confirm that no target-domain data were used in training the knowledge extractors.","section":"Sec. 3.2"},{"comment":"The paper treats APTOS, EyePACS, Messidor-1, and Messidor-2 as interchangeable classification tasks with the same label space, but it never states how the Messidor grading scale was remapped to the 5-class APTOS/EyePACS labels. Messidor uses a 0-3 severity scale (or 0-4 in some versions), while APTOS/EyePACS use 0-4. Without an explicit remapping procedure, the symbolic rules and the cross-domain accuracy numbers are not well defined. The authors should specify the label alignment and report per-class or per-dataset label distributions to confirm that the comparison is meaningful.","section":"Sec. 5.1"},{"comment":"Fusion hyperparameters are 'empirically tuned' with no stated validation procedure. The weighted fusion weights (α_DL, α_KL) and the selection among the four fusion strategies appear to be chosen based on the test domains, as no held-out validation set is described. This constitutes a form of test-set fitting for the reported 'best' results. The authors should specify how these choices were made, ideally with a validation split or a sensitivity analysis across the full range of weights.","section":"Sec. 3.1"}],"minor_comments":[{"comment":"The notation 'KL' for the knowledge classifier is confusing because the same letters are used for Kullback-Leibler divergence in the text. Please rename the symbolic classifier (e.g., 'Sym' or 'KB') and reserve 'KL' for the divergence measure.","section":"Throughout"},{"comment":"The paper alternates between 'vein segmentation' and 'vessel segmentation'; retinal imaging typically refers to arteries and veins jointly as vessels. Please use consistent terminology.","section":"Sec. 3.2"},{"comment":"The text says 'MESSIDOR' while elsewhere 'Messidor' is used. Please unify capitalization and dataset name formatting across tables and text.","section":"Sec. 5.1"},{"comment":"Some references appear mismatched: e.g., [59] is cited as TransDR but the reference list entry [59] is a diffusion-based domain augmentation paper. Please verify all citation-to-reference mappings.","section":"Sec. 2"},{"comment":"The backbone columns for the proposed methods list parameter counts but not the specific ViT variant consistently (e.g., 'Vit (22M)' vs 'Vit(21.5M)'). Clarify which architecture was used for each row.","section":"Table 6"}],"recommendation":"reject","confidential_remarks":"The paper's central proposed mechanism is absent from the methodology, and its headline MDG result is contradicted by its own table. These are not merely presentational issues; they concern the scientific validity of the claims. The provenance of the feature-extractor training data and the label remapping for Messidor are also unstated, so the experimental results cannot be independently verified. Even with revision, the missing KL-alignment mechanism and the internal contradictions in the experimental reporting would require substantial re-analysis and likely new experiments, which is beyond a standard revision cycle. I would recommend rejection. I also note that the paper's introduction mentions generalization to seizure-onset detection, but no such experiments are reported elsewhere in the manuscript; this claim should be removed or substantiated if the paper is resubmitted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll skip the niceties. The paper as written does not support its abstract. The claimed KL-divergence-minimization mechanism never appears in the methods: no loss, no equation, no training step. The 'KL' in the tables refers to the knowledge classifier, not Kullback-Leibler. In Table 6, the fused ViT+KL gets 58.4% average, lower than either branch and lower than several baselines, yet section 4.2 calls the symbolic-only 63.67% 'the best' while listing 65.5% baselines. There are also arithmetic issues: the ViT+KL average should be 58.2, and the KL average 63.87, not 63.67. These are not typos—they are load-bearing contradictions with the central claims.\n\nThat said, the SDG experiments (Tables 2–5) do show something: in three of four source domains, either the symbolic-only or fused model beats the published baselines. The ablation also gives credit: lesion-based features (84.65% on APTOS) beat vessel-inclusive features. That is a concrete, falsifiable empirical pattern, and if it holds up it would be worth publishing.\n\nThe soft spots, in proportion:\n\n1. The provenance of the 500 fine-tuning images is never stated. If any target-domain images (APTOS, for instance) were used to fine-tune YOLOv11 or the vessel segmenter, the 'unseen domain' claim is contaminated at the feature level. This is the first question a referee should ask.\n\n2. The label spaces are not reconciled. Messidor has four grades, APTOS/EyePACS five. The paper never says how the remap works. Without that, the cross-domain accuracy comparisons are not interpretable.\n\n3. The fusion weights are 'empirically tuned' with no validation protocol described. That's a minor-to-moderate concern, but it means the 'best' results may be partly fitted.\n\nWho is this for? A skeptic working on neuro-symbolic DG in medical imaging might find the SDG pattern interesting enough to chase down. But as it stands, the paper would not survive a competent referee without major revision.\n\nMy recommendation: send it back for major revision, not desk-reject. The empirical core may be salvageable, but the abstract and the MDG section need to be rewritten to match the actual methods and numbers, and the provenance and label-remapping questions must be answered.","headline":"Abstract and MDG claims don't survive contact with the paper's own tables, but the SDG results may hide a salvageable empirical core if provenance questions are answered.","tokens_in":16542,"tokens_out":4028,"would_cite":false,"duration_ms":41112,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"KG-DG claims that combining clinical lesion rules with a vision transformer via confidence fusion improves diabetic retinopathy generalization to unseen domains.","keywords":["domain generalization","neuro-symbolic learning","diabetic retinopathy","lesion features","vision transformers","confidence fusion","fundus imaging","out-of-distribution robustness"],"falsifier":"Retrain the YOLOv11 lesion detector from scratch on a training domain with no test-domain images, then run the full KG-DG fusion on a held-out test domain: if the symbolic branch's cross-domain accuracy matches the reported 74.0% on APTOS when trained on Messidor, the claim survives; if it collapses, the symbolic features are not domain-invariant. Separately, add the KL-divergence loss to training and measure embedding distances; if removing it changes no accuracy, the alignment mechanism is not load-bearing.","tokens_in":15420,"feed_emoji":"👁️","tokens_out":9886,"duration_ms":108375,"temperature":0.7,"pith_summary":"The paper sets out to show that diabetic retinopathy grading can generalize to unseen hospital domains when a vision transformer is paired with a symbolic branch that reasons over clinical lesion features. The proposed KG-DG framework detects hemorrhages, exudates, and cotton-wool spots, encodes them into a structured feature vector, classifies them with a gradient-boosting model, and fuses the symbolic prediction with the transformer's softmax output using confidence-based rules. The authors report that this integration beats published single-domain and multi-domain generalization baselines, with up to 5.2% cross-domain accuracy gains, a 6% gain over a ViT baseline on APTOS, and a symbolic-only MDG average of 63.67%. If true, the result would mean that explicit clinical knowledge, not just more data or bigger models, can make medical image classifiers more portable across devices and populations, and that interpretable lesion features can act as a regularizer rather than a bottleneck.","feed_headline":"Fusing retinal lesion rules with ViT lifts cross-domain DR accuracy","feed_subtitle":"Clinical lesion rules plus confidence fusion aim to keep DR grading stable across unseen datasets.","key_machinery":"The engine is a dual-branch classifier. One branch is an ImageNet-pretrained ViT (DeiT-Small, CvT-13, or T2T-ViT) whose classification head is trained with class-weighted cross-entropy. The other branch turns clinical rules into numbers: YOLOv11 detects lesion bounding boxes, a U-Net vessel segmenter optionally measures tortuosity, caliber, and branching angle, and these measurements form a structured feature vector F* that a Gradient Boosting classifier maps to a severity label. A fusion step, selecting the highest-confidence prediction, the class-wise max, or a weighted sum of softmax scores, produces the final output. The framework also states that KL divergence between symbolic domain em","core_discovery":"KG-DG's central claim is that the best way to generalize a DR classifier to an unseen fundus-image domain is not to make the neural network alone more robust, but to give it a parallel, expert-driven reasoning path. The symbolic path encodes clinical rules from Table 1 (e.g., neovascularization implies proliferative DR) into quantitative lesion features, learns a Gradient Boosting classifier on those features, and then combines the two branches by a confidence-weighted fusion rule. On four public datasets treated as four domains, the authors report that the fused model outperforms transformer and CNN baselines in most single-source settings and that the symbolic-only model reaches the best a","pith_inferences":["The fusion logic is modular, so swapping YOLOv11 for a lighter detector or using lesion counts alone would isolate whether the gain comes from the clinical prior or from detector quality.","The vein-feature ablation suggests not all clinical biomarkers transfer; ranking symbolic features by their cross-domain embedding distance could identify which ones should be dropped at deployment time.","A natural test of the alignment claim is to insert an explicit KL-divergence loss on the symbolic domain embeddings and compare cross-domain accuracy with and without it.","The same two-branch design could be tried on OCT or chest X-ray tasks where countable lesions and expert rules exist, with the symbolic branch providing an auditable second opinion."],"forward_implications":["If the paper's claims hold, DR grading systems could be deployed to a new clinic with a single source dataset, because lesion-based reasoning carries over even when pixel statistics do not.","Clinical lesion counts such as exudate count and hemorrhage density become a cheap, interpretable regularizer; models that use them need not trade away accuracy for explainability.","The symbolic-only result in MDG suggests that, for DR, a small curated set of biomarkers can outperform large-image classifiers under domain shift, reducing the annotation load for new deployment sites.","The same dual-branch pattern could transfer to other medical imaging tasks where expert rules identify countable lesions or structures, such as OCT biomarkers or chest X-ray findings."],"supporting_citations":[{"why":"Supplies the DomainBed protocol that defines the single- and multi-domain generalization evaluation.","marker":"[18]"},{"why":"The DR-specific domain-generalization baseline that KG-DG is compared against in the single-source tables.","marker":"[3]"},{"why":"The SPSD-ViT baseline that reaches 65.5% in the multi-domain table and is the main transformer comparison.","marker":"[21]"},{"why":"The SD-ViT transformer baseline used in both SDG and MDG comparisons.","marker":"[51]"},{"why":"Provides the ViT backbone architecture that forms the neural branch.","marker":"[10]"},{"why":"The lesion-ontology GCN model that motivates encoding clinical lesion knowledge for DR.","marker":"[26]"},{"why":"Cited for the YOLO-family object detector used to localize hemorrhages, exudates, and cotton-wool spots.","marker":"[58]"},{"why":"The APTOS dataset used as a source domain and for the lesion-feature ablation.","marker":"[25]"},{"why":"The Messidor and Messidor-2 datasets used as domains and the source for one SDG setting.","marker":"[9]"},{"why":"The EyePACS dataset used as a source domain in the single-source setting.","marker":"[23]"}],"fun_headline_variants":["Neuro-symbolic fusion beats ViT alone on unseen DR domains","Expert rules plus ViT add up to 5.2% accuracy on unseen DR domains","Symbolic lesion features outperform neural-only DR classifiers","KG-DG neuro-symbolic model wins across four DR datasets","Clinical rules fused with ViT improve DR generalization to new domains"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that all four datasets can be treated as one five-class DR task, including Messidor's grading scale, and that the lesion detectors were fine-tuned on a small expert set that does not include the evaluation domain, so the symbolic features are genuinely domain-invariant.","fun_headline_variants_meta":{"raw":{"variants":["Neuro-symbolic fusion beats ViT alone on unseen DR domains","Expert rules plus ViT add up to 5.2% accuracy on unseen DR domains","Symbolic lesion features outperform neural-only DR classifiers","KG-DG neuro-symbolic model wins across four DR datasets","Clinical rules fused with ViT improve DR generalization to new domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000876,"raw_usage":{"total_tokens":3646,"prompt_tokens":782,"completion_tokens":2864,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":2783}},"tokens_in":526,"tokens_out":2864,"duration_ms":23085,"temperature":1.0,"reasoning_tokens":2783,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:17:00.613357+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the YOLOv11 lesion detector from scratch on a training domain with no test-domain images, then run the full KG-DG fusion on a held-out test domain: if the symbolic branch's cross-domain accuracy matches the reported 74.0% on APTOS when trained on Messidor, the claim survives; if it collapses, the symbolic features are not domain-invariant. Separately, add the KL-divergence loss to training and measure embedding distances; if removing it changes no accuracy, the alignment mechanism is not load-bearing.","supporting_citations":[{"cited_title":"Gulrajani and D","cited_arxiv_id":null,"evidence_quote":"Supplies the DomainBed protocol that defines the single- and multi-domain generalization evaluation."},{"cited_title":"Atwany and M","cited_arxiv_id":null,"evidence_quote":"The DR-specific domain-generalization baseline that KG-DG is compared against in the single-source tables."},{"cited_title":"Content Significance Distribution of Sub-Text Blocks in Articles and Its Application to Article-Organization Assessment","cited_arxiv_id":"2311.01673","evidence_quote":"The SPSD-ViT baseline that reaches 65.5% in the multi-domain table and is the main transformer comparison."},{"cited_title":"Sultana, M","cited_arxiv_id":null,"evidence_quote":"The SD-ViT transformer baseline used in both SDG and MDG comparisons."},{"cited_title":"Dosovitskiy et al","cited_arxiv_id":null,"evidence_quote":"Provides the ViT backbone architecture that forms the neural branch."},{"cited_title":"Khandelwal, R","cited_arxiv_id":null,"evidence_quote":"The lesion-ontology GCN model that motivates encoding clinical lesion knowledge for DR."},{"cited_title":"Kauppi et al","cited_arxiv_id":null,"evidence_quote":"The APTOS dataset used as a source domain and for the lesion-feature ablation."},{"cited_title":"Decencière et al","cited_arxiv_id":null,"evidence_quote":"The Messidor and Messidor-2 datasets used as domains and the source for one SDG setting."},{"cited_title":"Diabetic retinopathy detection","cited_arxiv_id":null,"evidence_quote":"The EyePACS dataset used as a source domain in the single-source setting."}],"review_version":1}