{"id":"82c44e3d-2986-4499-9c17-d77cc4b30dc4","arxiv_id":"2607.03959","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Bidirectional cross-modal attention fusion of OCT B-scans with single-channel enface OCTA (real or diffusion-translated) consistently beats a ConvNeXt V2 OCT-only baseline for binary DR classification across two cohorts and cross-dataset tests.","lead":"Fusing structural OCT scans with vascular OCTA maps via cross-attention improves automated diabetic retinopathy detection over OCT alone. Synthetic OCTA generated from ordinary OCT often matches real OCTA, which could cut hardware cost for screening.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"TR>GT gains on OCT500 may reflect translator leakage/denoising rather than true vascular signal, undermining the hardware-replacement claim.","rationale":"The Reader correctly isolates the load-bearing premise: TR OCTA must supply real vascular signal, not OCT500-specific artifacts. The paper’s own numbers make this the softest point—the largest TR>GT lift occurs precisely where the translator was trained and where class imbalance is worst—while the clinical narrative (Discussion) treats TR as a drop-in hardware substitute. Metric anomalies (specificity ≡ recall in nearly every cell of Tables I–III) and private UIC data further reduce confidence but are secondary; even with perfect metrics the TR>GT pattern would still require the hold-out test above. The directional multimodal-fusion benefit over ConvNeXt is plausible and repeatedly shown, so the verdict stays CONDITIONAL rather than REJECT: accept the experimental contribution once leakage is ruled out and metrics/code are clarified; do not treat hardware replacement as settled. No stronger internal inconsistency appears; the concern is empirical and directly testable.","tokens_in":14599,"tokens_out":651,"duration_ms":6457,"concrete_test":"Hold out a patient-level OCT500 subset never seen by the diffusion translator; re-generate TR OCTA only on that hold-out; re-run the exact CRD-Net fusion protocol of Table III (same splits, same bootstrap CIs). If TR accuracy/recall on the hold-out collapses toward or below GT (or toward the unimodal OCT baseline), the TR>GT gains are leakage/denoising artifacts and the replacement claim weakens. Optionally ablate by feeding pure noise or OCT-derived non-vascular maps as the second modality; if performance remains high, the second stream is not supplying vascular signal.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central clinical claim (Abstract, contributions, Discussion) is that diffusion-translated OCTA (TR) can replace hardware GT OCTA because TR fusion matches or exceeds GT fusion and improves sensitivity/cross-dataset robustness. That claim requires TR to inject genuine, generalizable vascular disease signal. The translator was trained only on OCT500 (authors’ prior work [12]; §III.A). On the same OCT500 cohort, TR fusion accuracy jumps from 0.8788 (GT) to 0.9697 and recall from 0.7143 to 0.9286 (Tables II–III), while the Normal:DR split is severely imbalanced (251:64). Discussion attributes this to TR being a “denoised, average-representative vascular prediction,” but that explanation is consistent with the translator simply re-emitting OCT500-specific texture or label-correlated artifacts that the fusion head can exploit, rather than true microvascular biomarkers. Cross-dataset results are mixed (TR helps OCT500→UIC recall but accuracy falls to 0.7005 on UIC→OCT500), so the strongest TR>GT numbers sit inside the translator’s training domain. Without an independent check that TR carries disease-relevant vascular information beyond what OCT already provides, the hardware-replacement conclusion does not follow from the tables.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes bidirectional cross-modal attention fusion (CRD-Net) of OCT B-scans with single-channel enface OCTA for binary DR vs normal classification. Using OCT500 and UIC (730 subjects total), it evaluates five scenarios: within-dataset, combined, and bidirectional cross-dataset tests. A ConvNeXt V2 OCT-only model is the unimodal baseline. In addition to ground-truth (GT) OCTA, the authors substitute translated (TR) OCTA generated from OCT by a conditional diffusion model trained on OCT500. Tables I–III report that fusion consistently outperforms the unimodal baseline; TR OCTA matches or exceeds GT OCTA in most settings and improves recall in the OCT500→UIC direction. The central clinical claim is that attention-based OCT–OCTA enface fusion improves DR detection and that computationally generated OCTA can replace hardware OCTA for screening in resource-limited settings.","tokens_in":14943,"tokens_out":1026,"duration_ms":8287,"significance":"If the directional fusion benefit and the TR-as-surrogate claim hold under scrutiny, the work is clinically relevant: OCTA hardware is expensive and scarce, while standard OCT is widely available. The multi-scenario design (patient-level splits, combined training, bidirectional cross-dataset tests) and bootstrap 95% CIs on fusion metrics are strengths relative to many single-cohort multimodal OCT papers. Using only single-channel enface OCTA also matches a realistic clinical data constraint. The TR pathway, if it truly injects generalizable vascular signal rather than domain-specific texture, would be a practical bridge for high-sensitivity DR screening without OCTA devices. These contributions are incremental but useful for the eess.IV / ophthalmic imaging community.","major_comments":[{"comment":"Tables II–III and Discussion (TR OCTA paragraphs): The hardware-replacement claim rests on TR OCTA supplying genuine, generalizable vascular disease signal. The translator was trained only on OCT500 (§III.A, ref. [12]). On that same cohort, TR fusion accuracy rises from 0.8788 (GT) to 0.9697 and recall from 0.7143 to 0.9286, under a severe 251:64 Normal:DR imbalance. Discussion attributes this to TR being a “denoised, average-representative” map, but that is equally consistent with re-emitted OCT500-specific texture or label-correlated artifacts that the fusion head can exploit. Cross-dataset results are mixed (TR helps OCT500→UIC recall; UIC→OCT500 accuracy falls to 0.7005). Without an independent check—e.g., vessel-density / FAZ correlation of TR vs GT, ablation of TR quality, or a translator trained off OCT500—the claim that TR can replace hardware OCTA does not follow from the tables","section":null},{"comment":"Tables I–III: Specificity equals Recall in essentially every row (e.g., OCT500 unimodal 0.6872/0.6872; GT OCT500 0.7143/0.7143; TR OCT500 0.9286/0.9286; same pattern for UIC and Combined). For binary classification this is only expected under balanced classes or a particular threshold/definition; with the reported 251:64 imbalance it is anomalous and suggests either a reporting error (specificity copied from recall), macro-averaging over classes without stating it, or an unintended operating point. This undermines trust in the metric suite that supports the “consistently outperforms” claim. Please recompute and report standard binary specificity (TN/(TN+FP) for the DR-positive definition) and clarify the positive class and averaging convention.","section":null},{"comment":"§III.B / Results: The unimodal baseline is described as “ConvNeXt Large” with ImageNet-1k weights, while the abstract, contributions, and Table I caption say “ConvNeXt V2.” These are different architectures. The baseline must be named and configured consistently; otherwise the fusion gains relative to a “strong unimodal ConvNeXt V2 baseline” cannot be verified. Please also state whether the same backbone family and capacity are used inside CRD-Net’s two streams.","section":null},{"comment":"§III.A and Tables I–III: How multiple (~80) fovea-centered OCT B-scans per patient are aggregated into a single patient-level prediction is not specified (slice-wise majority vote, feature pooling, multi-instance head, etc.). Patient-level stratified splits are claimed, but without the aggregation rule the reported accuracies and cross-dataset numbers are not reproducible and the comparison to a single enface OCTA is ambiguous. This is load-bearing for all five scenarios.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core here is empirical, not architectural. They take CRD-Net-style bidirectional cross-attention (Liu et al.), pair ~80 fovea-centered OCT B-scans with one enface OCTA channel, and run a clean five-scenario design: within OCT500, within UIC, combined, and both cross-dataset directions, with a ConvNeXt V2 OCT-only baseline and bootstrap CIs on the fusion numbers. Across 730 patients the directional result is consistent: fusion lifts accuracy, F1, and especially AUC over unimodal OCT. That is the part worth knowing.\n\nWhat is actually new is the evaluation frame plus the head-to-head of ground-truth versus diffusion-translated OCTA (their own prior translator, trained only on OCT500). Using only single-channel enface rather than multi-layer OCTA is a realistic clinical constraint and is stated clearly. Patient-level stratified splits and the dual-dataset setup are done properly. The citation trail is standard and the math is ordinary supervised classification; nothing is circular.\n\nSoft spots, in proportion. Specificity is reported identical to recall in essentially every cell of Tables I–III; that is almost certainly a reporting or macro-averaging error and needs fixing before anyone trusts the numbers. The stronger clinical claim—that TR OCTA can replace hardware OCTA—rests on TR often beating GT, most dramatically on OCT500 itself (0.97 vs 0.88 accuracy, large recall jump) where the translator was trained and the class balance is 251:64. Discussion’s “denoised average vascular map” story is plausible but also consistent with domain-specific texture or label leakage; cross-dataset results are mixed (TR helps OCT500→UIC recall, UIC→OCT500 stays hard). Private UIC data and no released code/splits keep the claim provisional. Novelty is extension-level, not a new fusion principle.\n\nThis is for people who care about practical multimodal retinal screening and domain-shift testing, not for readers hunting a new attention architecture. The fusion-beats-unimodal result is solid enough to engage; the hardware-replacement conclusion is not yet settled. I would send it to peer review with a clear request to correct the metric tables, release splits or code, and add an independent check that TR carries disease-relevant vascular signal beyond OCT. Worth a careful read and a citation if the metrics hold after cleanup.","headline":"Solid multi-scenario evidence that OCT + single-channel enface OCTA attention fusion beats a strong OCT-only baseline; the hardware-replacement claim for diffusion TR OCTA is the soft spot and needs stress-testing, not rejection of the whole paper.","tokens_in":15646,"tokens_out":595,"would_cite":true,"duration_ms":5629,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Attention fusion of OCT with single-channel enface OCTA beats OCT-only DR detection, and OCT-translated OCTA can replace hardware OCTA.","keywords":["diabetic retinopathy","optical coherence tomography","OCT angiography","cross-modal attention","multimodal fusion","translated OCTA","diffusion model","medical image classification"],"falsifier":"Train the identical fusion pipeline on a third independent multi-scanner cohort that includes mild DR and measure whether TR-OCTA fusion still matches or exceeds GT-OCTA fusion in recall and cross-dataset AUC; a clear drop for TR relative to GT would falsify the substitutability claim.","tokens_in":15412,"feed_emoji":"👁️","tokens_out":566,"duration_ms":5402,"temperature":0.7,"pith_summary":"This paper aims to show that fusing structural OCT B-scans with a single-channel enface OCTA map through bidirectional cross-modal attention reliably improves automated detection of diabetic retinopathy versus normal retina. The authors compare a strong OCT-only ConvNeXt V2 baseline against the fused model on two independent cohorts (OCT500 and UIC, 730 subjects total) under within-dataset, pooled, and bidirectional cross-dataset tests. They further claim that OCTA maps synthesized from OCT alone by a conditional diffusion model match or exceed real OCTA fusion in most settings, raise sensitivity, and transfer better across scanners. A sympathetic reader cares because OCTA hardware is expensive and scarce outside specialist clinics, while standard OCT is far more common; if the claim holds, high-performance multimodal screening becomes possible without extra hardware, especially in primary-care and resource-limited settings where most DR surveillance must occur.","feed_headline":"OCT fused with software OCTA beats OCT alone for DR screening","feed_subtitle":"Attention fusion works; maps translated from OCT can replace costly hardware OCTA","key_machinery":"Bidirectional cross-modal attention network (CRD-Net backbone): independent CNN streams extract features from OCT and enface OCTA, self-attention focuses each stream, then each modality queries the other via scaled dot-product attention before multi-branch classification; the same pipeline accepts either ground-truth or diffusion-translated OCTA.","core_discovery":"Cross-modal attention fusion of OCT B-scans with single-channel enface OCTA consistently and substantially outperforms a strong unimodal OCT baseline for binary DR-versus-normal classification across all five evaluation scenarios, and computationally translated OCTA can substitute for hardware-acquired OCTA while preserving or improving accuracy, sensitivity, and cross-dataset robustness.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["OCT-OCTA attention fusion tops OCT alone for DR classification","Software OCTA fused with OCT beats unimodal OCT in DR screening","Cross-modal OCT and enface OCTA fusion lifts DR detection accuracy","Translated OCTA substitutes for hardware in superior DR diagnosis","Attention OCT plus OCTA outperforms OCT baseline across datasets"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The assumption that OCT-to-OCTA translation produces genuine, generalizable vascular disease signal rather than dataset-specific texture or denoising patterns that merely inflate metrics.","fun_headline_variants_meta":{"raw":{"variants":["OCT-OCTA attention fusion tops OCT alone for DR classification","Software OCTA fused with OCT beats unimodal OCT in DR screening","Cross-modal OCT and enface OCTA fusion lifts DR detection accuracy","Translated OCTA substitutes for hardware in superior DR diagnosis","Attention OCT plus OCTA outperforms OCT baseline across datasets"]},"model":"grok-4.5","effort":"low","cost_usd":0.005222,"raw_usage":{"total_tokens":1459,"prompt_tokens":835,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":52220000,"prompt_tokens_details":{"text_tokens":835,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":556,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":835,"tokens_out":68,"duration_ms":38681,"temperature":1.0,"reasoning_tokens":556,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T22:44:20.256465+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the identical fusion pipeline on a third independent multi-scanner cohort that includes mild DR and measure whether TR-OCTA fusion still matches or exceeds GT-OCTA fusion in recall and cross-dataset AUC; a clear drop for TR relative to GT would falsify the substitutability claim.","supporting_citations":[],"review_version":1}