{"id":"abde9075-492c-43b9-8db5-f720c7c5c4ae","arxiv_id":"2508.16852","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GPO, a sparse Gaussian-primitive optimization framework, reports 2.35 px TRE and 0.938 AUC@25 on FIRE, outperforming compared baselines for deformable retinal registration.","lead":"Gaussian Primitive Optimization aligns retinal images by placing sparse, adjustable Gaussian anchor points on vessels and optimizing their displacements. On the FIRE benchmark it reports cutting average landmark error from 6.20 to 2.35 pixels and raising AUC at 25 pixels from 0.770 to 0.938.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline coordinate frame may differ: GPO evaluated at 1024×1024 while GeoFormer/RetinaRegNet numbers appear to be from native 2912×2912; the headline 3.8 px improvement could reverse.","rationale":"The reader's weakest_assumption focuses on whether the KNN Gaussian blend can represent all relevant deformations (Eq. 3). That is a legitimate theoretical concern, but it is not the most load-bearing threat to the paper's central claim. The central claim is a quantitative comparison on FIRE, and that comparison is only meaningful if all numbers are computed in the same coordinate frame. The paper's wording creates a strong possibility that GPO results are at 1024×1024 while baseline numbers are at 2912×2912; if so, the headline improvement inverts. This is concrete, checkable, and more decisive than the representational-capacity question. I nevertheless recommend keeping the reader's CONDITIONAL verdict, because the reader already listed 'clarify which baselines were retrained' and 'unverified code' as conditions. My concern sharpens that condition by pointing to a specific potential flaw, but it does not move the category: the paper should be either accepted after the coordinate-frame issue is resolved or rejected if the numbers turn out to be incomparable. Since the required check has not been performed, 'UNCHANGED' (i.e., still CONDITIONAL) is the appropriate recommendation.","tokens_in":8265,"tokens_out":9360,"duration_ms":108452,"concrete_test":"Ask the authors to supply the exact evaluation script and recompute every row of Table 1 in one explicitly stated coordinate frame. Concretely: (1) scale the FIRE landmark coordinates to 1024×1024 and run GeoFormer and RetinaRegNet under exactly the same preprocessing as GPO, reporting their TRE and AUC@25; (2) report GPO-DCN's TRE computed on the original 2912×2912 landmark coordinates (e.g., by upsampling the predicted displacement field). If GPO-DCN's advantage over GeoFormer/RetinaRegNet disappears or reverses in either check, the paper's central claim is a resolution artifact rather than a registration improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing issue is the coordinate frame of Table 1. Sec. 3.1 says images are resized to 1024×1024 'for all experiments', but the paper never states whether the GeoFormer/RetinaRegNet/SuperPoint etc. baseline TREs were recomputed at this resolution (with landmarks scaled) or taken from published values on FIRE's native 2912×2912 images. The numbers in Table 1 (e.g., GeoFormer 6.201 px, RetinaRegNet 2.766 px) are in the range of published native-resolution FIRE results. If GPO-DCN's 2.352 px is measured at 1024×1024 while the baselines are at 2912×2912, the comparison is invalid: 2.352 × (2912/1024) ≈ 6.69 px, i.e., GPO would not beat GeoFormer's 6.201 px. Similarly, the AUC@25 px threshold at 1024 corresponds to ~71 px at native resolution, so the 0.938 vs 0.770 AUC comparison is incomparable. This is not a minor statistical issue; it determines whether the headline claim is true. The text's silence on this point, combined with no released code or evaluation script, leaves the central quantitative claim unverifiable.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents Gaussian Primitive Optimization (GPO), an iterative deformable registration framework for retinal images. The pipeline first applies a coarse alignment (using GeoFormer) and extracts matched keypoints that serve as descriptor-based control nodes (DCN). Each node is modeled as a Gaussian primitive with learnable position, displacement, and radius; a K-nearest-neighbor Gaussian blending step converts the sparse node displacements into a dense displacement field. The node parameters are refined iteratively under a multi-term loss combining control-node cross-correlation and image normalized cross-correlation. On the FIRE dataset, the authors report that GPO-DCN reduces target registration error from 6.201 px (GeoFormer) to 2.352 px and increases AUC@25px from 0.770 to 0.938, with additional ablations over the number of nodes N, the number of neighbors K, and the number of iterations tau.","tokens_in":8600,"tokens_out":6216,"duration_ms":73595,"significance":"The core idea is well motivated by the gradient-dilution problem in retinal registration, and the use of sparse, anatomically anchored Gaussian primitives refined with iterative optimization is both intuitive and potentially useful. If the reported quantitative gains are obtained under a common evaluation protocol, the improvement over strong baselines is substantial. The paper also includes an ablation study and provides a public code repository. However, the current experimental description leaves the comparability of Table 1 unverified, and the lack of statistical confidence measures on a small test set makes the significance conditional on the requested clarifications.","major_comments":[{"comment":"The paper states 'For all experiments, we resized images to 1024×1024' but does not state whether the GeoFormer, RetinaRegNet, SuperPoint, RoMa, and other baseline numbers in Table 1 were recomputed at this resolution or taken from published native-resolution (2912×2912) FIRE papers. The listed baseline values are in the range of published native-resolution results. If GPO is evaluated at 1024×1024 while the baselines are at 2912×2912, the comparison is invalid: a GPO TRE of 2.352 px at 1024×1024 would scale to roughly 6.69 px at native resolution, which does not beat GeoFormer's 6.201 px; similarly, AUC thresholds are not resolution-invariant. Please state the exact evaluation resolution for every row, how landmarks were handled across resolutions, and provide the evaluation script or raw per-pair errors. This is required to make the central claim verifiable.","section":"§3.1, Table 1"},{"comment":"With a 7:1:2 split on the 134 FIRE pairs, the test set is only about 27 pairs. Table 1 reports point estimates only, with no standard deviations, confidence intervals, or significance tests. The reported improvement of GPO-DCN (2.352 px) over RetinaRegNet (2.766 px) is about 0.4 px and could be within noise on this test size. Please report per-pair TRE distributions, error bars, paired significance tests, and per-category (S/A/P) results, especially for the four anatomical-change pairs, which stress the smoothness assumption of Eq. (3).","section":"§3.2, Table 1"},{"comment":"The final hyperparameters N=1000, K=10, tau=100, alpha_gcc=0.4, and alpha_ncc=1.0 appear to be selected from the ablations in Fig. 4, but the paper does not state whether those ablations were performed on the validation split or on the test split. Without a clearly held-out validation procedure, the reported test numbers may be a result of selection on the test set. Please specify which split was used for the ablations, and confirm that the final Table 1 numbers were obtained with hyperparameters fixed on validation only.","section":"§3.2, Fig. 4 and Implementation Details"},{"comment":"The displacement field is computed by normalizing Gaussian weights only over the K nearest nodes at each pixel. Because the set of K nearest neighbors changes with x, u(x) can be discontinuous at locations where a neighbor enters or leaves the K-nearest set, even if the Gaussian weights themselves are continuous. This appears to contradict the claim of a 'smoothly varying displacement field' and may create seams or artifacts in the deformation. Normalizing over all N nodes, or otherwise constraining the blending to be partition-of-unity over a fixed set, would avoid this issue; at minimum, the authors should analyze or quantify the effect of this non-smoothness on the final registration.","section":"§2.2, Eq. (3)"}],"minor_comments":[{"comment":"The paper says the descriptor network provides 'N matched keypoints', but does not explain how exactly N=1000 matches are selected when the network produces more or fewer matches. Please describe the matching and selection procedure.","section":"§2.1"},{"comment":"The displayed sum contains a typographical 'KX' notation; it should be a standard sum over i=1..K. Also, the radius parameterization includes rmin and rmax but their numerical values are not given in the Implementation Details.","section":"§2.2, Eq. (3)"},{"comment":"For baselines, it is unclear which methods were retrained on FIRE, which were used off-the-shelf, and which were evaluated from published numbers. This information is necessary for reproducibility and for judging the fairness of the comparison.","section":"§3.1"},{"comment":"The gradient heatmaps lack a color scale and a detailed description of the preprocessing and normalization used. This makes the claimed difference in gradient propagation difficult to assess.","section":"Figure 1"},{"comment":"The ablation text reports 'median TRE' in several places, while Table 1 and the main text report 'TRE' as an average. Please clarify which summary statistic is used in each figure and table.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The coordinate-frame issue is the decisive point: if the baseline rows in Table 1 are not recomputed at 1024×1024 with appropriately scaled landmarks, the headline improvement is not established. During revision I would ask the authors to provide the evaluation script and the raw per-pair TRE/AUC values for every method, along with an explicit statement of the resolution used for each row."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The method is a clean, sensible integration of sparse descriptor-based control nodes with KNN Gaussian blending and iterative intensity-based refinement. Eq. 3 is just scattered-data interpolation with trainable centers, displacements, and radii, but combining it with anatomically anchored nodes for retinal registration is a reasonable and well-targeted contribution. The motivation is real: gradient dilution in textureless regions is a genuine problem, and the design directly addresses it. The ablations on N, K, and tau are clear and show expected trade-offs. On its own terms, the paper is coherent and the central mechanism is plausible.\n\nThe load-bearing issue is the coordinate frame of Table 1. Section 3.1 says images are resized to 1024x1024 for all experiments, but nowhere does it state whether the baseline results were recomputed at that resolution with scaled landmarks, or taken from published native-resolution (2912x2912) numbers. The reported GeoFormer TRE of 6.201 px and RetinaRegNet of 2.766 px look exactly like published native-resolution values. If GPO is measured at 1024x1024 while the baselines are at native resolution, the comparison is invalid: 2.352 px at 1024 corresponds to roughly 6.7 px at 2912, which would not beat GeoFormer. Likewise, an AUC@25px threshold at 1024 maps to about 71 px at native resolution. This is not a minor statistical issue; it determines whether the headline claim is true. The paper must explicitly clarify which baselines were re-evaluated and provide scaled landmarks.\n\nOther concerns are more standard: the test set is small (~27 pairs) with no error bars or significance tests, hyperparameters are chosen from ablations without a clearly described held-out protocol, results are not broken out by FIRE category (particularly the four anatomical-change pairs), and no code or evaluation script is provided. These are not fatal by themselves, but combined with the resolution ambiguity, the quantitative claims should be treated as provisional.\n\nThe paper deserves a serious referee, not a desk reject. The method is plausible, the ablations are informative, and the problem is relevant. But the authors must resolve the evaluation protocol before the numbers can be trusted. If the baselines were fairly re-evaluated at 1024x1024, this is a solid contribution; if not, the core comparison collapses.\n\nRecommendation: send to peer review with explicit instructions to clarify the coordinate frame and release reproducible code.","headline":"A sensible sparse-node registration method whose headline numbers may be comparing apples to oranges unless the baselines were re-evaluated at the same resolution.","tokens_in":9083,"tokens_out":2414,"would_cite":false,"duration_ms":28710,"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":"Gaussian-blended control nodes anchored at retinal vessels cut FIRE target registration error from 6.2 px to 2.35 px, the paper reports.","keywords":["retinal image registration","deformable registration","Gaussian primitives","sparse feature propagation","keypoint control nodes","FIRE dataset","iterative optimization"],"falsifier":"Register the four FIRE Category A pairs (anatomical changes) with GPO-DCN and compare per-pair TRE to the reported 2.35 px average; if the sparse weighted-average representation cannot capture true structural change, those pairs should show TRE far above the average, directly testing the representation's coverage.","tokens_in":8162,"feed_emoji":"👁️","tokens_out":7409,"duration_ms":73505,"temperature":0.7,"pith_summary":"The paper sets out to solve a specific failure mode in deformable retinal image registration: most pixels are flat, textureless retina, while the informative vascular features occupy a small fraction of the image, so standard learning-based losses drown the alignment signal from vessels in noise. The proposed GPO framework places a sparse set of control nodes at keypoints on major vessels, models each node as a Gaussian primitive with learnable position, displacement, and radius, and propagates the nodes' displacements through a K-nearest-neighbour Gaussian blend to build a dense, globally coherent displacement field. It then iterates the node parameters with a loss that combines keypoint cross-correlation and global intensity alignment. On the FIRE dataset the paper reports a target registration error of 2.35 px and an AUC@25 px of 0.94, down from 6.20 px and 0.77 for the best descriptor-only baseline, and better than the strongest learning-based baseline. The significance is that the method shows how to keep gradient information flowing from sparse anatomical features into homogeneous regions, a problem that limits both classical and deep registration systems.","feed_headline":"Gaussian control nodes cut retinal registration error to 2.35 px","feed_subtitle":"Sparse nodes anchored at vessels carry alignment into texture-poor retina, beating prior methods on FIRE.","key_machinery":"The load-bearing mechanism is KNN-based Gaussian blending: a displacement field u(x) = sum over the K nearest control nodes of a normalized Gaussian weight times that node's displacement vector, where each weight depends on distance from x to the node's position and on the node's learned radius. Each node is a Gaussian primitive with trainable position, displacement, and radius, so the spatial influence of each anchor can adapt to local deformation scale. This construction performs structured message passing: high-gradient pixels near vascular keypoints feed gradients back to their nearby nodes through the same weights, and the K-nearest restriction keeps the field locally detailed while lim","core_discovery":"The paper's central claim is that deformable retinal registration can be reduced to optimizing a sparse set of Gaussian primitives rather than predicting a dense flow. Each primitive is a control point located on a salient vessel or grid site, with a learnable position, displacement vector, and radius. A K-nearest-neighbour Gaussian interpolation turns the primitives' displacements into a globally coherent displacement field, and gradient-based iteration over the node parameters, guided by a loss that combines keypoint cross-correlation with intensity alignment, refines the warp. On the FIRE benchmark this reaches 2.352 px mean target registration error and 0.938 AUC at 25 px, surpassing bot","pith_inferences":["Not explored in the paper: the same Gaussian-primitive parametrization could be applied to other sparse-feature medical images, such as OCT, X-ray, or microscopy, where textureless backgrounds dominate and dense descriptors are unreliable.","Because FIRE Category A (anatomical change) contains only 4 pairs, the paper's pooled numbers may not reflect those cases; a per-category breakdown would test whether the sparse Gaussian blend can represent genuine structural change.","A direct ablation of the two loss terms (keypoint consistency only vs intensity only) would clarify which term actually carries the vessel-alignment signal, a claim the current experiments leave implicit.","The learnable node positions suggest an adaptive variant that grows or prunes primitives based on local gradient magnitude, potentially improving accuracy without fixing the node count N."],"forward_implications":["On the FIRE dataset, GPO-DCN achieves a target registration error of 2.352 px versus 6.201 px for GeoFormer and 2.766 px for RetinaRegNet, and higher AUC at 15, 25, and 50 px thresholds, indicating both better mean accuracy and fewer large outlier errors.","Anatomically placed descriptor-based nodes outperform uniform grid nodes (GPO-DCN 2.35 px vs GPO-GCN 2.65 px TRE), confirming that anchoring primitives at salient vascular structures matters.","The KNN Gaussian interpolation with K=10, N=1000 nodes, and 100 iterations provides a practical accuracy/runtime trade-off, roughly 30 seconds per FIRE pair at 1024x1024 resolution.","Because the displacement field is parameterized by sparse nodes rather than a dense per-pixel flow, the method does not need dense correspondence prediction and avoids the vanishing-gradient problem in flat retinal regions.","The framework also works with grid-based control nodes when descriptors are unavailable, at a modest accuracy cost, so it does not depend on a specific keypoint detector."],"supporting_citations":[{"why":"Provides the coarse homography alignment and the descriptor-based control nodes (DCN) used to initialize GPO in the main experiments.","marker":"[14]"},{"why":"Supplies the FIRE benchmark, the 134 image pairs, category splits, and the 10 expert landmarks used for TRE and AUC evaluation.","marker":"[9]"},{"why":"RetinaRegNet, the strongest learning-based baseline, whose TRE of 2.766 px GPO-DCN surpasses at 2.352 px.","marker":"[22]"},{"why":"GraDIRN, a gradient-embedding registration network baseline that frames the learning-and-iterative-optimization class compared in Table 1.","marker":"[20]"},{"why":"PDD-Net, a probabilistic dense displacement baseline used for comparison against discrete and probabilistic deformation prediction.","marker":"[6]"},{"why":"VoxelMorph++, a keypoint-supervised and instance-optimized deformable registration baseline.","marker":"[7]"},{"why":"RoMa, a robust dense feature matching method representative of the descriptor-based class.","marker":"[5]"},{"why":"SuperPoint, a self-supervised keypoint detector and descriptor baseline used in the descriptor-based comparison.","marker":"[4]"}],"fun_headline_variants":["Vessel-anchored Gaussian nodes slash retinal registration error","Retinal alignment error cut from 6.2 to 2.4 px by sparse Gaussian nodes","Gaussian primitives boost retinal registration AUC to 0.938","From 6.2 to 2.35 px: Gaussian nodes sharpen retinal alignment","Gaussian control points beat dense flows in retinal registration"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method assumes every retinal deformation can be captured as a locally smooth weighted average of the translations carried by a few hundred control points, so deformations involving true anatomical change or very large undetected warps may not fit that representation.","fun_headline_variants_meta":{"raw":{"variants":["Vessel-anchored Gaussian nodes slash retinal registration error","Retinal alignment error cut from 6.2 to 2.4 px by sparse Gaussian nodes","Gaussian primitives boost retinal registration AUC to 0.938","From 6.2 to 2.35 px: Gaussian nodes sharpen retinal alignment","Gaussian control points beat dense flows in retinal registration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0008,"raw_usage":{"total_tokens":3380,"prompt_tokens":792,"completion_tokens":2588,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":2491}},"tokens_in":536,"tokens_out":2588,"duration_ms":24974,"temperature":1.0,"reasoning_tokens":2491,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:08:01.545671+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Register the four FIRE Category A pairs (anatomical changes) with GPO-DCN and compare per-pair TRE to the reported 2.35 px average; if the sparse weighted-average representation cannot capture true structural change, those pairs should show TRE far above the average, directly testing the representation's coverage.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Provides the coarse homography alignment and the descriptor-based control nodes (DCN) used to initialize GPO in the main experiments."},{"cited_title":"Modeling and Artificial Intel- ligence in Ophthalmology1(4), 16–28 (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the FIRE benchmark, the 134 image pairs, category splits, and the 10 expert landmarks used for TRE and AUC evaluation."},{"cited_title":"Computers in Biology and Medicine186, 109645 (2025)","cited_arxiv_id":null,"evidence_quote":"RetinaRegNet, the strongest learning-based baseline, whose TRE of 2.766 px GPO-DCN surpasses at 2.352 px."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"GraDIRN, a gradient-embedding registration network baseline that frames the learning-and-iterative-optimization class compared in Table 1."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"PDD-Net, a probabilistic dense displacement baseline used for comparison against discrete and probabilistic deformation prediction."},{"cited_title":"In: International workshop on biomedical image registration","cited_arxiv_id":null,"evidence_quote":"VoxelMorph++, a keypoint-supervised and instance-optimized deformable registration baseline."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"RoMa, a robust dense feature matching method representative of the descriptor-based class."}],"review_version":1}