{"id":"7d0c4123-f77c-46e6-a37f-20591103aeee","arxiv_id":"2602.00995","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"VAMOS-OCTA adds axial and lateral MIP/AIP projection losses to vessel-weighted MSE training, improving motion-corrupted B-scan inpainting in handheld OCTA.","lead":"A deep-learning method for filling in motion-blurred blank stripes in retinal OCT angiography scans combines two existing loss ideas: vessel-weighted pixel reconstruction and orthogonal projection consistency. It reports sharper vessel detail and cleaner en face images than its direct baseline, though on a small private dataset and with metrics that overlap its training objectives.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-artifact transfer is unvalidated: synthetic corruption model (§2.1.2) only removes whole B-scans, while real handheld motion artifacts are more varied; quantitative evidence is limited to synthetic data.","rationale":"The reader's weakest_assumption identifies the synthetic corruption model as the least secure condition for the central claim. I agree: the central claim—that VAMOS-OCTA consistently outperforms prior methods on motion-corrupted OCTA volumes—depends on the synthetic corruptions in §2.1.2 being representative of real handheld motion artifacts. The paper provides only qualitative real-world results (Figure 4) and no quantitative validation on real data, so the general claim is not yet supported. Other concerns exist, such as the overlap between training losses (axial MIP/AIP L1) and evaluation metrics (MIP L1, MIE) and the unaddressed B-scan PSNR drop, but these affect the interpretation of the synthetic results rather than the fundamental question of real-world transfer. A concrete test with realistic simulated artifacts or real data with ground truth would directly determine whether the reported improvements hold. Given this concern, the conditional verdict remains appropriate; no verdict change is needed.","tokens_in":7471,"tokens_out":6691,"duration_ms":74535,"concrete_test":"Use motion-tracking data from a handheld OCTA probe to generate realistic corruptions (partial B-scan blur, shearing, non-contiguous dropouts) applied to clean volumes, then retrain/evaluate both models on this more realistic test set. If VAMOS-OCTA's quantitative advantage over SOAD decreases or disappears, the synthetic corruption model in §2.1.2 is the reason and real-world transfer is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1.2 simulates motion artifacts by dropping the center B-scan plus 1–6 contiguous neighbors sampled from a geometric distribution (p=0.4). This models only complete-slice dropouts. Real handheld OCTA motion artifacts, however, also include partial-slice blur, shearing, intraslice distortion, and non-contiguous missing regions (cf. refs [1,14,15]). The network is trained only on the dropout model, so it may learn to interpolate from neighboring slices but not to handle partial or deformed slices. The only real-world evaluation is qualitative (Figure 4) with no ground truth, so the abstract's claim that VAMOS-OCTA 'consistently outperforms prior methods' for motion-corrupted OCTA volumes in general is unsupported beyond the synthetic setting. Moreover, the abstract states training on 'real-world corrupted volumes' but Methods only describe synthetic corruption; no real-world training or quantitative validation is provided. If the synthetic-to-real gap is large, the reported gains (e.g., SSIM 0.895 vs 0.763 in Table 1b) may not persist in clinical data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VAMOS-OCTA, a 2.5D U-Net for inpainting motion-corrupted B-scans in handheld OCTA volumes. The method augments SOAD's vessel-weighted MSE loss with axial and lateral projection losses defined as per-B-scan 1D MIP/AIP L1 terms. The authors evaluate on a synthetic corruption model that drops contiguous blocks of B-scans and report improvements in B-scan perceptual metrics (LPIPS, Laplacian blur, Sobel edge preservation) and en face MIP metrics (L1, MIE, SSIM, NCC, PSNR) relative to standard MSE, SOAD, and an axial-only ablation. Qualitative results on real-world corrupted volumes are also shown.","tokens_in":7839,"tokens_out":4945,"duration_ms":53174,"significance":"If the central claims hold, the idea of using multi-axis projection supervision for slice-based volumetric inpainting is simple, practical, and potentially generalizable to other modalities. Strengths of the manuscript include the public code release, the dynamic synthetic corruption strategy during training, and the systematic ablation adding axial and lateral projection terms. The authors also correctly motivate the need to go beyond pixel-wise metrics by including perceptual and edge-aware B-scan evaluations. However, several load-bearing issues—especially a mismatch between the claimed en face supervision and the actual loss definition, circularity between the training loss and reported MIP metrics, and unsupported real-world generalization—currently prevent the paper from being accepted.","major_comments":[{"comment":"The loss in Eq. (3) computes 1D projection profiles per B-scan: max over z gives a 1D profile in x, and max over x gives a 1D profile in z. The en face MIP evaluated in Table 1b is a 2D image formed by collapsing depth across the entire B-scan stack, i.e., a function of lateral position and B-scan index. The proposed loss never compares projections across B-scans, so the reported MIP improvements cannot be attributed to direct supervision of the en face MIP. If the authors intend to supervise en face projections, the loss should be computed on the 2D projection of the full volume or of the input stack; otherwise the mechanism described in the abstract and Section 2.3.2 should be revised.","section":"§2.3.2, Eq. (3) and Table 1b"},{"comment":"There is a circularity concern. The VAMOS loss directly minimizes the L1 distance between predicted and ground-truth axial/lateral MIP and AIP profiles. Table 1b then reports MIP L1 and MIE as primary evidence of MIP improvement. Since MIP L1 is exactly the training objective and MIE is essentially the AIP L1 objective, these metrics are not independent measurements. The paper should either report projection metrics that are not aligned with the loss (e.g., vessel segmentation agreement, vessel skeleton connectivity, or a manual reading study) or clearly state that Table 1b is a sanity check rather than evidence of generalization.","section":"§2.3, Eq. (1)–(3) and Table 1b"},{"comment":"The claim that VAMOS-OCTA produces \"consistent improvements ... without trade-offs\" is contradicted by the paper's own numbers. In Table 1a, VAMOS-OCTA has lower B-scan PSNR (26.122 ± 0.709) than SOAD (27.084 ± 0.712) and standard MSE (27.036 ± 0.764). In Table 1b, VAMOS-OCTA has lower MIP PSNR (27.772 ± 0.865) than the axial-only ablation (28.168 ± 1.044). These PSNR losses may be acceptable if the authors value LPIPS and edge preservation more, but the text must acknowledge this trade-off and explain why PSNR is less relevant for this clinical task.","section":"§3.1, Table 1 and Section 3.2"},{"comment":"The abstract states that the model was \"trained on both synthetic and real-world corrupted volumes,\" but the Methods section only describes synthetic corruption (contiguous whole-B-scan dropout). No real-world training data or procedure is described. Figure 4 shows qualitative real-world examples without ground truth, so the claim that VAMOS-OCTA \"consistently outperforms prior methods\" on real handheld motion artifacts is unsupported. Furthermore, the synthetic model only removes entire B-scans; real artifacts often include partial-slice blur, shearing, and non-contiguous missing regions. The authors should either add quantitative real-world validation (e.g., expert grading or a proxy metric on stable regions) or substantially temper the real-world claims in the abstract and conclusion.","section":"§2.1.2 and Abstract"}],"minor_comments":[{"comment":"The table is titled \"perceptual quality metrics\" but includes a PSNR column, which is a pixel-wise fidelity metric. Please clarify the categorization or move PSNR elsewhere.","section":"Table 1a"},{"comment":"The statistical testing is not described in detail. Paired t-tests over 7 cross-validation folds are weak with n=7, and no multiple-comparison correction is mentioned. Please report effect sizes, confidence intervals, or per-fold results.","section":"§3.1"},{"comment":"The architecture description is minimal: no number of layers, channels, normalization, or training hyperparameters (learning rate, epochs, optimizer) is given. The code link helps, but the paper should still include sufficient architectural detail for reproducibility.","section":"§2.2"},{"comment":"MIE is never defined in the text. Please define Mean Intensity Error and state how it is computed.","section":"§3.1, Table 1b"},{"comment":"The terms \"axial\" and \"lateral\" projections may confuse readers. Since each B-scan is 2D, \"axial projection\" collapses the depth dimension to produce a 1D profile along x, while \"lateral projection\" collapses x to produce a 1D profile along z. Please use more explicit names such as \"depth-collapsed\" and \"lateral-collapsed\" profiles.","section":"§2.3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a medical imaging venue, and the multi-axis supervision idea is interesting. However, the gap between the abstract's strong claims and the evidence is significant, especially with respect to real-world training/validation and the circular evaluation of MIP metrics. I would encourage the editor to request a revision that addresses the loss-mechanism mismatch and adds independent metrics before reconsidering."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a competent, incremental paper with one genuinely new piece — the VAMOS loss, which adds axial and lateral MIP/AIP projection terms to SOAD's vessel-weighted MSE for 2.5D OCTA B-scan inpainting. That specific recipe is not in the cited prior work, and the ablation (wMSE vs. wMSE+Axial vs. VAMOS) is the right way to show it. On LPIPS, Sobel edge preservation, Laplacian blur, and most MIP metrics, VAMOS clearly beats the baselines, and the qualitative figures look plausible. Credit where due: the experiments are clean, the ablations are properly staged, the corruption simulation is dynamic during training, and the code release is promised. The paper is honestly written about what it does.\n\nSoft spots, in order of size. First, the synthetic corruption model only drops whole contiguous B-scans. Real handheld OCTA artifacts also include partial-slice blur, shearing, and intraslice distortion. The paper says training used 'synthetic and real-world corrupted volumes,' but Methods describe only synthetic corruption, and the real-world evidence is a single qualitative figure with no ground truth. So the abstract's 'consistently outperforms prior methods' is only supported for the dropout model. That is a real limitation, not a manufactured one.\n\nSecond, the paper's own Table 1 contradicts the 'without trade-offs' claim in Section 3.1. VAMOS is worse than SOAD on B-scan PSNR (26.122 vs. 27.084) and worse than axial-only on MIP PSNR (27.772 vs. 28.168). The authors emphasize perceptual metrics precisely because pixel-wise metrics favor blurring — that is a defensible choice — but they should not then claim no trade-offs. The PSNR dips are small but real, and ignoring them looks like cherry-picking.\n\nThird, there is a partial circularity concern: the VAMOS loss directly minimizes MIP/AIP L1, and Table 1b reports MIP L1 and MIE as key gains. That does not invalidate the SSIM/NCC/PSNR results, which are not part of the loss, but the MIP L1 improvement is partly the training objective measuring itself.\n\nVerdict: this paper deserves a serious referee. It is a solid incremental contribution to a narrow clinical problem, with sound methodology and honest ablations — the synthetic-to-real gap and the PSNR trade-offs are fixable with more careful claims and external validation. I would cite it if I worked on OCTA restoration, and I would send it to review, expecting revision rather than acceptance as-is.\n\nWould I bring it to reading group? Maybe, if the group cares about medical-image inpainting loss design. Otherwise it is a 'skim the tables' paper.","headline":"A fairly honest incremental paper: a new loss recipe for OCTA B-scan inpainting that shows real gains on perceptual and MIP metrics, but the central claim of 'consistent outperformance' is undercut by the paper's own PSNR tables and by unvalidated synthetic-to-real transfer.","tokens_in":8313,"tokens_out":739,"would_cite":true,"duration_ms":9005,"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":"Adding axial and lateral projection-consistency losses to a vessel-weighted reconstruction loss lets a 2.5D U-Net inpaint motion-corrupted OCTA B-scans, delivering sharp capillaries, restored vessel continuity, and clean en face projections","keywords":["OCTA inpainting","handheld retinal imaging","motion artifact correction","vessel-aware loss","multi-axis projection supervision","2.5D U-Net","en face MIP restoration","medical image restoration"],"falsifier":"Present the same trained model with volumes containing partial-slice blur, sheared, or non-contiguous corruption instead of whole contiguous B-scan dropouts. If VAMOS-OCTA's advantage over the vessel-weighted baseline disappears or its en face maps show the same banding and fragmentation, then the multi-axis supervision is tuned to the synthetic dropout model rather than to real handheld motion.","tokens_in":7417,"feed_emoji":"👁️","tokens_out":4962,"duration_ms":54004,"temperature":0.7,"pith_summary":"The paper aims to show that motion-related blind spots in handheld OCTA volumes—entire B-scans lost to sudden motion—can be filled in by a 2.5D U-Net trained with a composite loss that watches not only the slices themselves but also their one-dimensional projections along depth and across slices. The proposed VAMOS loss combines a vessel-weighted MSE with axial and lateral maximum/average intensity projection losses. The authors argue that these orthogonal projection constraints force the network to keep vessels continuous in 3D without building an explicit 3D model. If the claim holds, clinicians using handheld OCTA on children or uncooperative patients get sharper cross-sectional images and artifact-free en face maps from the same acquisition, and slice-based 3D imaging generally gains a simple loss-level tool for volumetric consistency.","feed_headline":"Projection losses restore motion-blanked OCTA volumes","feed_subtitle":"Handheld retinal scans lose whole slices to motion; these losses refill them while keeping vessel structure sharp.","key_machinery":"The load-bearing object is the VAMOS loss: L = wMSE + λ_proj (L_ax_MIP + L_lat_MIP + L_ax_AIP + L_lat_AIP). The wMSE weights each pixel by a combination of the target intensity and the predicted intensity raised to 1/3 (with constants 100 and 0.5), so bright vessel pixels dominate and hallucinated bright spots are penalized. The projection terms collapse each B-scan into 1D profiles—maximum and average along depth (axial) and along the lateral axis—and compare the reconstruction's profiles to ground truth with L1. This makes the network judge each slice by how it aggregates into en face views and by how vessel structure lines up across neighboring slices, which is how the method avoids bandi","core_discovery":"The paper's central claim is that vessel-weighted intensity reconstruction alone over-smooths OCTA slices and fragments vessels in en face views, and that adding projection-consistency losses along two orthogonal axes fixes both defects. With axial projection supervision alone, en face maximum-intensity-projection accuracy jumps (SSIM from 0.763 to 0.888, NCC from 0.812 to 0.914, L1 roughly halved), but B-scan sharpness barely changes; adding lateral projection supervision is what removes horizontal banding and lifts perceptual B-scan metrics (LPIPS from 0.608 to 0.510, Sobel edge preservation from 0.313 to 0.427). The same training recipe, learned on synthetic contiguous slice dropouts, is","pith_inferences":["A natural testable extension is to replace the L1 projection terms with perceptual or structural losses on the 1D profiles; if projection consistency is the operative constraint, stronger profile metrics could sharpen recovered vasculature further.","Because the corruption model only drops whole contiguous B-scans, the method's real-artifact performance would be clarified by stress-testing on partial-slice blur or shearing; that scenario is unaddressed in the paper.","The same loss recipe could be dropped into other slice-acquired 3D modalities, such as MRI or ultrasound volumes with out-of-plane motion, since it only requires collapsing volumes along two orthogonal axes; the paper hints at generalizability but does not demonstrate it.","A reader wanting clinical assurance might verify restored vessels with graph-connectivity or segmentation-continuity metrics, because the reported sharpness gains are perceptual and edge-based rather than anatomical."],"forward_implications":["En face MIP quality can be improved substantially by projection losses alone: with axial supervision, SSIM rises to 0.888 and NCC to 0.914 relative to the vessel-weighted baseline.","Lateral projection supervision is the component that removes horizontal banding and brings B-scan sharpness gains, so a complete volumetric fix needs both axes.","A 2.5D network with per-slice losses can produce 3D-consistent volumes, meaning explicit 3D architectures are not required for this inpainting task.","Performance degrades gracefully with corruption severity: mean intensity error stays low even when large contiguous blocks of slices are missing, where baseline methods' errors climb sharply.","The approach transfers to volumes with real-world motion corruptions, at least qualitatively, after training on synthetic contiguous slice dropouts."],"fun_headline_variants":["Orthogonal losses fix motion-blanked OCTA in both planes","Multi-axis supervision restores motion-corrupted OCTA volumes","Vessel-aware losses refill blank B-scans and sharpen en face","Two-axis projection loss repairs motion-corrupted retinal scans","Axial plus lateral losses revive blank OCTA volumes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that dropping a center B-scan plus a random contiguous block of 1–6 neighboring B-scans, with block size drawn from a geometric distribution with p=0.4, faithfully reproduces the motion artifacts seen in real handheld OCTA; if real artifacts involve partial-slice blur or shearing rather than whole-slice dropouts, the claimed transfer to clinical volumes is unproven.","fun_headline_variants_meta":{"raw":{"variants":["Orthogonal losses fix motion-blanked OCTA in both planes","Multi-axis supervision restores motion-corrupted OCTA volumes","Vessel-aware losses refill blank B-scans and sharpen en face","Two-axis projection loss repairs motion-corrupted retinal scans","Axial plus lateral losses revive blank OCTA volumes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3089,"prompt_tokens":854,"completion_tokens":2235,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":2152}},"tokens_in":598,"tokens_out":2235,"duration_ms":16814,"temperature":1.0,"reasoning_tokens":2152,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:48:17.459437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Present the same trained model with volumes containing partial-slice blur, sheared, or non-contiguous corruption instead of whole contiguous B-scan dropouts. If VAMOS-OCTA's advantage over the vessel-weighted baseline disappears or its en face maps show the same banding and fragmentation, then the multi-axis supervision is tuned to the synthetic dropout model rather than to real handheld motion.","supporting_citations":[],"review_version":1}