{"id":"f09c503e-9088-4acc-a339-66aea30fa1cc","arxiv_id":"2505.24351","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A super-pixel particle swarm optimization method for 2D-3D coronary artery registration is claimed to outperform four baselines, but the supporting results are incomplete and circularly evaluated.","lead":"This paper proposes a particle swarm optimization method to align 2D X-ray angiography with 3D CT angiography of coronary arteries for guiding heart procedures. The authors claim their method beats four state-of-the-art baselines, but the evidence presented is internally inconsistent and lacks a direct comparison.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that SPSO outperformed selected SOTA baselines in all aspects is unverifiable: the manuscript reports no comparative results against any of the nine listed baselines; Table 2 compares only segmentation backbones, not registration methods.","rationale":"The reader's REJECT verdict is correct, but the most load-bearing concern is slightly more specific than the reader's stated weakest assumption. The reader focused on the validity of the evaluation protocol: using mPD with known correspondences and manually trimmed CTA data may not predict clinical performance. That is a serious concern, but it presumes that a comparison exists to validate. In fact, the manuscript contains no comparative results at all against the nine baselines it names. Section IV-C announces a comparison, but the results are never shown; Table 2 compares six deep-learning segmentation backbones, not registration methods. The abstract's central claim therefore fails at the most basic evidentiary level. I also note the manuscript's internal contradictions and nomenclature inconsistencies: the abstract says 28 pairs from 10 patients while Section IV-A says 28 patients; Section III-C describes six degrees of freedom as 'three rotation angles and one translation vector'; and 'super-pixel' and 'super-voxel' are used interchangeably. These reinforce the impression of an incomplete manuscript, but they are secondary to the missing comparison. To be fair, the paper does describe a concrete pipeline with several implemented modules: UNet-based DSA segmentation, skeletonization via K3M, Shi-Tomasi corner detection, CTA centerline extraction, and PSO-based matching. If the missing comparison table were provided and the protocol were made sound, the method might still be viable as a preliminary technical report. However, as submitted, the evidence for the central claim is non-existent. The reader's verdict of REJECT with high confidence remains appropriate, so I do not change the verdict.","tokens_in":12437,"tokens_out":2334,"duration_ms":30194,"concrete_test":"Request or reconstruct the missing comparative experiment: run SPSO and the nine listed baselines (ICP-BP, ICP-PnP, Tree, DT, CS, GMM, OGMM, PSO-EM, Grad) on the same 28 CTA/XRA pairs using the Section IV-B protocol, recording per-pair mPD, robustness (e.g., failure count or standard deviation), and runtime. If SPSO's mean mPD is not lower than every baseline with comparable variance and runtime, or if no such comparative table can be produced from the claimed experiments, the abstract's 'outperformed the selected SOTA baselines in all aspects' claim is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that SPSO 'outperformed the selected SOTA baselines in all aspects' of registration accuracy, robustness, and efficiency. Section IV-C says the method was compared with nine competing methods across four categories (ICP-BP, ICP-PnP, Tree, DT, CS, GMM, OGMM, PSO-EM, Grad). Yet no table or figure in the manuscript reports accuracy, robustness, or runtime for these methods. Table 1 only lists parameter settings; Table 2 reports mPD values for different DSA segmentation backbones (R2Net, ResUNet++, TransResUNet, U2Net, UNet, VGGNet), not for any of the nine registration baselines. The only quantitative registration result, 'The average registration error of the method on the dataset is 2.67mm' (Section V), coincides with the UNet row of Table 2 and is not accompanied by any baseline mPD for comparison. This is not merely a weakness in the evaluation protocol, though the protocol itself is also problematic: the mPD loss is optimized using known point correspondences (Section IV-B), and imperfect clinical data are handled by manually removing corresponding CTA parts to force 1:1 matching (Section V). The decisive issue is that the experimental evidence needed to support the central comparative claim is absent from the manuscript. If the authors possess such a comparison, it must be presented; otherwise the claim that SPSO outperforms SOTA in all aspects is unsupported and cannot be assessed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a 3D-2D coronary artery registration method, SPSO, that combines deep-learning-based DSA segmentation with a particle swarm optimization search over virtual camera poses, using discretized 'super-voxel' point sets extracted from DSA and CTA images. The authors report on a pilot dataset of 28 XRA/CTA pairs from 10 patients, claim that the method outperforms nine state-of-the-art baselines in accuracy, robustness, and efficiency, and report an average registration error of 2.67 mm. Section IV-C describes the baseline comparison protocol, Section III-D defines the mPD metric used for evaluation, and Section IV-D compares six segmentation backbones.","tokens_in":12795,"tokens_out":7818,"duration_ms":84961,"significance":"The clinical problem is relevant and a robust CTA-to-XRA registration method would be valuable for PCI guidance. The paper presents a coherent pipeline: segmentation, centerline extraction, discretization, and PSO-based pose optimization. The inclusion of a deep segmentation network and the use of a swarm optimizer are reasonable engineering choices. However, the central claims are not supported by the evidence presented: no results for the baselines are shown, the evaluation metric is the same objective used for optimization, and the test data are manually pruned to remove mismatches. As submitted, the manuscript does not yet establish that the method performs better than existing approaches, so the significance of the contribution cannot be assessed.","major_comments":[{"comment":"The abstract claims that SPSO 'outperformed the selected SOTA baselines in all aspects' of registration accuracy, robustness, and efficiency, and Section IV-C states that the method was compared with nine competing methods (ICP-BP, ICP-PnP, Tree, DT, CS, GMM, OGMM, PSO-EM, Grad). However, no table or figure in the manuscript reports accuracy, robustness, or runtime results for any of these baselines. Table 1 lists only parameter settings, and Table 2 reports mPD values and segmentation metrics for different DSA segmentation backbones (R2Net, ResUNet++, TransResUNet, U2Net, UNet, VGGNet), not for any of the nine registration methods. The only quantitative registration result, the 2.67 mm average in Section V, is the UNet row of Table 2 and has no baseline comparison. The central comparative claim is therefore unsupported and cannot be verified from the submitted manuscript.","section":"Section IV-C, Tables 1-2"},{"comment":"The evaluation protocol is circular: Eq. (13) defines mPD as the average distance from each 3D point to its projection on the 2D plane, and Section IV-B states that 'since the data is generated based on the known point correspondence relationship, mPD is used in the loss function instead of soft dtw when performing experiments on the dataset.' Section III-D then presents mPD as the registration accuracy metric. The reported mPD values, including the headline 2.67 mm average, are therefore values of the very same objective that was minimized during optimization. This is not an independent assessment of registration accuracy, and any conclusion drawn from these numbers about the quality of the registration is invalid.","section":"Sections III-D and IV-B"},{"comment":"The paper explicitly states that imperfect clinical data are handled by 'artificially removing the corresponding parts in the CTA images to achieve a 1:1 matching effect.' This manual pruning removes the vessel segments that do not have a counterpart in the XRA image, which are precisely the cases where occlusions or mismatches occur in real clinical data. As a result, the reported 'average registration error of the method on the dataset is 2.67mm' is an average over a dataset from which the difficult cases have been eliminated, and it cannot support the claim of robustness on real clinical images with unknown correspondences.","section":"Section V"},{"comment":"The text in Section III-D describes mPD as a 'minimum projected distance': 'for each point in the 3D vascular structure, we project it onto the 2D vascular plane along all possible directions and choose the direction that minimizes the projection distance.' However, Eq. (13) is a simple average of d(a, P_B(a)) over points a in the 3D structure, with no minimization over directions. These two definitions are inconsistent, so the reported numeric values cannot be reproduced. In addition, Table 2 shows that the method's mPD for the right coronary artery (2.66 mm) is lower than the manually annotated ground-truth mPD (2.99 mm), which indicates that mPD as computed is not a true registration error with respect to the manual reference.","section":"Section III-D, Eq. (13), and Table 2"}],"minor_comments":[{"comment":"The abstract says the method was compared with 'four state-of-the-art (SOTA) methods,' while Section IV-C says it was compared with 'nine competing methods.' Please reconcile the number.","section":"Abstract and Section IV-C"},{"comment":"The section title is 'CTA Image Keypoint Extraction Module,' but the first paragraph describes a 'DSA Image Keypoint Extraction Module.' The heading and the content need to be aligned.","section":"Section III-B"},{"comment":"The paper has two consecutive discussion sections, 'Discussions and Conclusions' and 'Discussions and Future Work,' with overlapping content. They should be merged or clearly separated.","section":"Sections V and VI"},{"comment":"There are placeholders in the text: 'Figure A shows the basic registration results' and 'The registration results of the six competitive methods are shown in Figure XXX.' These need to be replaced with actual figure references.","section":"Section IV-C and IV-D"},{"comment":"The table title 'PARAMETERS OF 3D-2D REGISTRATION METHODS' is incorrect; the table reports mPD values and segmentation performance metrics, not method parameters.","section":"Table 2"},{"comment":"The notation in Eq. (13) is underspecified: d is not defined as a distance metric, and P_B(a) is called a projection onto plane B even though B is defined as a set of points in the 2D vascular plane. Please clarify the mathematical objects.","section":"Eq. (13) and Section III-D"},{"comment":"The list of contributions contains duplicated bullet markers ('• •') on two items.","section":"Section I, Contributions"},{"comment":"The terms 'super-pixel,' 'super-voxel,' 'hypergraph,' and 'hyperedge' are used interchangeably. Choose one consistent terminology and define it at first use.","section":"Throughout"},{"comment":"The dataset section says 1571 XRA images were annotated and split 8:1:1, while Section IV-D mentions training on '968 pairs of DSA images.' Please clarify the relationship between these numbers.","section":"Section IV-A"}],"recommendation":"reject","confidential_remarks":"Dear Editor, this manuscript is not ready for publication. The central claims are unsupported by the experiments as presented, and the evaluation protocol is circular. The problems are not minor stylistic issues; they would require a substantial new experimental study (baseline comparisons, an independent error metric, and evaluation without manual data removal) to address. I therefore recommend rejection. There is no indication of misconduct; the issues appear to stem from an incomplete draft, as evidenced by multiple placeholders ('Figure A', 'Figure XXX', 'Year' in reference [16])."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an incremental engineering contribution with an unsupported headline. The core idea—PSO-based 2D/3D coronary registration with weighted super-pixels and UNet segmentation—is a modest combination of known components. PSO-based registration already exists in the cited PSO-EM baseline, and the super-pixel weighting is a small variant. No new theoretical mechanism is presented.\n\nWhat the paper does well: it uses real CTA/XRA pairs from a clinical pilot, spells out the preprocessing and optimization steps in enough detail to be reproduced in principle, and is candid about limitations: manual removal of CTA parts to force 1:1 matching, occasional local optima, and segmentation robustness. Those admissions are useful and not typical.\n\nThe soft spots are not minor. The abstract claims the method outperformed selected SOTA baselines in all aspects, but no table or figure reports accuracy, robustness, or runtime for any of the nine listed baselines. Table 2 compares segmentation backbones, not registration methods; the quantitative headline (2.67 mm) is just the UNet row of that table. That alone undermines the central claim.\n\nThe evaluation is also circular: Section IV-B says mPD is used as the loss because the data have known correspondences, and Section III-D reports mPD as the accuracy metric. So the reported error is the optimized objective, not an independent measure. Eq. 13 defines a mean per-point projection distance, not the \"minimum projected distance\" described in the text. The manual trimming of CTA data to force 1:1 vessel matching further distances the evaluation from real occluded clinical cases.\n\nThere are smaller inconsistencies: the abstract says four SOTA methods, Section IV-C says nine; the discussion mentions a hypergraph PSO though no hypergraph is defined; a figure placeholder remains; and dataset numbers shift (28 pairs vs 968 pairs vs 60 pairs). The citations themselves look appropriate for the area.\n\nBottom line: the domain is relevant and the authors are honest about some limitations, but the main comparative claim cannot be assessed because the comparison is absent. For the PCI registration community, this is currently more useful as a cautionary example than a technical reference. I would not send this to peer review in current form. I'd ask for the actual baseline results, an independent evaluation metric, and a clearer data-handling description. If those appear, it could become a modest engineering contribution.","headline":"A clinically motivated but methodologically broken registration paper: the claimed SOTA comparison never actually appears in the manuscript.","tokens_in":13329,"tokens_out":4067,"would_cite":false,"duration_ms":54741,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims a super-pixel particle swarm optimization pipeline registers preoperative 3D coronary CTA to intraoperative 2D X-ray angiography with an average error of 2.67 mm, outperforming four state-of-the-art baselines.","keywords":["multimodal image registration","coronary artery","CT angiography","X-ray angiography","particle swarm optimization","super-pixel","vessel segmentation","percutaneous coronary intervention"],"falsifier":"Run the SPSO pipeline on a held-out set of clinical CTA/XRA pairs without manually removing CTA segments that have no matching XRA vessel, and measure mPD against expert-annotated ground truth. If the average error rises to the level of the baselines or above, the reported superiority depends on the manual preprocessing rather than on the registration method itself.","tokens_in":12262,"feed_emoji":"🫀","tokens_out":6406,"duration_ms":73933,"temperature":0.7,"pith_summary":"Percutaneous coronary intervention (PCI) is guided in real time by 2D X-ray angiography, which lacks depth, while preoperative CT angiography provides 3D coronary anatomy. This paper proposes a registration method that aligns the two by turning both into weighted discrete point sets and searching for the 3D-to-2D pose with particle swarm optimization. On a pilot dataset of 28 CTA/XRA image pairs from 10 PCI patients, the method reports an average registration error of 2.67 mm measured as minimum projection distance, and the authors claim it beats four competing baseline families in accuracy, robustness, and efficiency. The value, if true, is a practical way to overlay 3D CTA context onto live X-ray images during PCI, potentially improving navigation through occluded or twisted vessels.","feed_headline":"Coronary CTA-to-X-ray registration hits 2.67 mm error","feed_subtitle":"A super-pixel particle swarm pipeline beats four baselines on a 28-pair pilot dataset, aiding PCI guidance.","key_machinery":"The load-bearing object is the super-pixel (called super-voxel when projected) representation: the vascular tree is reduced to weighted discrete points at edges, centerline locations, and bifurcations, so that registration becomes a particle-swarm search over a six-DOF virtual camera pose. DSA features come from UNet segmentation, K3M skeletonization, Shi-Tomasi corner detection, and moment-based centroid extraction; CTA features come from dynamic threshold segmentation of the 3D vessel tree. The particle swarm updates particle velocities and positions, projects the 3D points to 2D, and uses normalized cross-correlation as the stopping criterion, with weights emphasizing bifurcation points.","core_discovery":"The paper's central claim is that coronary CTA-to-XRA registration is best handled as a weighted discrete-point matching problem: vessel centerlines, edges, and bifurcations are extracted from both modalities, discretized into 'super-voxels,' and aligned by a particle swarm optimizer that searches the six-degree-of-freedom camera pose and scores each pose by normalized cross-correlation. The authors report that this SPSO pipeline achieves an average registration error of 2.67 mm on the pilot dataset, outperforming the ICP-based, topology-based, Gaussian mixture model-based, and feature point-based baselines compared in Table 1. They also claim that UNet is the best of the six deep segmentation models they tested for the DSA preprocessing module, with a Dice coefficient of 0.7166 and an overall matching accuracy of 0.9779, and that using bifurcation features in addition to centerlines is necessary to avoid matching errors at branch starting points.","pith_inferences":["The authors do not test registration on untrimmed CTA data; a natural next experiment is to hold out cases with occluded or stenotic vessels and compare mPD with and without manual CTA trimming. If the gap is large, the clinical advantage depends on automatic handling of missing segments.","A direct extension is to replace the PSO search with a differentiable pose estimator trained on the same weighted super-voxel representation, which could remove the need for hand-set PSO parameters and speed up convergence.","The same discretize-and-swarm strategy could apply to other 3D/2D vessel registration problems, such as cerebral or peripheral angiography, wherever centerlines and bifurcations can be extracted.","Reported mPD values near 2.7 mm should be compared against coronary vessel diameter and clinical tolerance; whether 2.67 mm is adequate for guiding PCI is a clinical question the paper does not answer."],"forward_implications":["If the 2.67 mm average mPD transfers to clinical use, the pipeline could furnish real-time 3D CTA overlay during PCI with error below typical coronary vessel diameters.","The weighted discrete-point formulation lets the optimizer prioritize bifurcations, so registration accuracy concentrates on the landmarks surgeons use for navigation.","Because the method uses simple, fast PSO rather than a learned deformation model, it may be easier to deploy or adapt to new imaging setups without retraining.","The UNet segmentation experiment implies that better DSA vessel segmentation directly translates into better registration, making segmentation quality a primary lever for further gains.","The reported consistency across left and right coronary arteries suggests the method is not tied to one vascular geometry."],"supporting_citations":[{"why":"Supplies the UNet architecture used to segment DSA vessels in the preprocessing module.","marker":"[14]"},{"why":"Supplies the K3M skeleton erosion algorithm used to thin segmented vessels to one-pixel-wide centerlines.","marker":"[15]"},{"why":"Supplies the particle swarm optimization algorithm that searches the six-DOF pose during matching.","marker":"[18]"},{"why":"Defines the minimum projected distance metric used as both the loss function and the registration error measure.","marker":"[19]"},{"why":"Tree-topology baseline that the proposed method must beat in the registration comparison.","marker":"[20]"},{"why":"Distance-transform-based baseline included in the topology-based comparison group.","marker":"[21]"},{"why":"Curvilinear structure baseline used in the comparison for coronary artery registration.","marker":"[22]"},{"why":"Gaussian mixture model baseline compared against the proposed method.","marker":"[23]"},{"why":"Oriented Gaussian mixture model baseline compared against the proposed method.","marker":"[24]"},{"why":"Feature-point PSO-EM baseline that the proposed method extends and outperforms.","marker":"[25]"}],"fun_headline_variants":["Super-pixel PSO aligns coronary CTA with X-ray at 2.67 mm error","Particle swarm registration of coronary arteries hits 2.67 mm error","Coronary CTA-XRA registration error down to 2.67 mm with SPSO","SPSO beats four baselines in coronary CTA-XRA registration","Super-pixel swarm optimization cuts coronary registration error to 2.67 mm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that an error metric computed on manually trimmed data with known point correspondences predicts how well the method will register real clinical images, where correspondences are unknown and some vessel segments may be occluded or missing.","fun_headline_variants_meta":{"raw":{"variants":["Super-pixel PSO aligns coronary CTA with X-ray at 2.67 mm error","Particle swarm registration of coronary arteries hits 2.67 mm error","Coronary CTA-XRA registration error down to 2.67 mm with SPSO","SPSO beats four baselines in coronary CTA-XRA registration","Super-pixel swarm optimization cuts coronary registration error to 2.67 mm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000927,"raw_usage":{"total_tokens":4023,"prompt_tokens":1048,"completion_tokens":2975,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":2881}},"tokens_in":664,"tokens_out":2975,"duration_ms":25546,"temperature":1.0,"reasoning_tokens":2881,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:24:25.492279+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the SPSO pipeline on a held-out set of clinical CTA/XRA pairs without manually removing CTA segments that have no matching XRA vessel, and measure mPD against expert-annotated ground truth. If the average error rises to the level of the baselines or above, the reported superiority depends on the manual preprocessing rather than on the registration method itself.","supporting_citations":[{"cited_title":"Oriented gaussian mixture models for nonrigid2d/3d coronary artery registration,","cited_arxiv_id":null,"evidence_quote":"Oriented Gaussian mixture model baseline compared against the proposed method."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the UNet architecture used to segment DSA vessels in the preprocessing module."},{"cited_title":"K3m: A universal algorithm for image skeletonization and a review of thinning tech- niques,","cited_arxiv_id":null,"evidence_quote":"Supplies the K3M skeleton erosion algorithm used to thin segmented vessels to one-pixel-wide centerlines."},{"cited_title":"Particle swarm optimization: developments, applications and resources,","cited_arxiv_id":null,"evidence_quote":"Supplies the particle swarm optimization algorithm that searches the six-DOF pose during matching."},{"cited_title":"A minimum projected-distance test for parametric single-index berkson models,","cited_arxiv_id":null,"evidence_quote":"Defines the minimum projected distance metric used as both the loss function and the registration error measure."},{"cited_title":"A tree-topology preserv- ing pairing for 3d/2d registration,","cited_arxiv_id":null,"evidence_quote":"Tree-topology baseline that the proposed method must beat in the registration comparison."},{"cited_title":"3d-2d registration of cerebral angiograms: A method and evaluation on clinical images,","cited_arxiv_id":null,"evidence_quote":"Distance-transform-based baseline included in the topology-based comparison group."},{"cited_title":"Iterative closest curve: A framework for curvilinear structure registration application to 2d/3d coronary arteries registration,","cited_arxiv_id":null,"evidence_quote":"Curvilinear structure baseline used in the comparison for coronary artery registration."},{"cited_title":"Robust point set registration using gaussian mixture models,","cited_arxiv_id":null,"evidence_quote":"Gaussian mixture model baseline compared against the proposed method."},{"cited_title":"Nonrigid 2d/3d regis- tration of coronary artery models with live fluoroscopy for guidance of cardiac interventions,","cited_arxiv_id":null,"evidence_quote":"Feature-point PSO-EM baseline that the proposed method extends and outperforms."}],"review_version":1}