{"id":"075de2f8-5509-4b58-a623-402a5161e7a1","arxiv_id":"2507.00051","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Siamese network with edge-enhanced attention tracks guidewire tips in DSA video at 57 FPS, with 0.421 mm mean error on three test sequences and 0.148-0.708 mm error on two robotic tasks.","lead":"This paper trains a Siamese network with attention modules to track the tip of a guidewire in real-time X-ray images during endovascular procedures. It reports sub-millimeter localization errors on three clinical sequences and in two robotic phantom navigation tests.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The millimetre accuracy figures are not reproducible because no pixel-to-mm calibration or per-sequence pixel spacing is reported; this single omission underpins every reported error value.","rationale":"The reader's weakest assumption identifies the core issue correctly: the millimetre errors are the paper's headline quantitative claim, and they cannot be verified without an explicit pixel-to-mm calibration. The paper gives bounding-box annotations only in pixels (Section 3.1), then evaluates and reports center error in millimetres (Section 3.3, Table 1, Fig. 3b). The conversion factor is never stated, and the three test sequences are drawn from different clinical procedures, so assuming a single global factor is not justified. Every reported accuracy figure in the abstract and Section 4.1 propagates this assumption. The IoU and FPS numbers are less affected, but the central claim of sub-pixel accuracy in Section 6 depends directly on the unknown pixel size. I also note the robotic ground-truth inconsistency between Section 3.1 and Fig. 5; this is secondary but relevant to the transfer claim. These issues do not, by themselves, show the method is wrong; the architecture is plausible and the experiments are described in enough detail to be re-run. They do make the numerical claims conditional on reporting that calibration and clarifying the annotation protocol. Since the reader's verdict is already CONDITIONAL, my analysis does not change the recommended verdict; it sharpens the condition that must be satisfied for acceptance.","tokens_in":13270,"tokens_out":4023,"duration_ms":48273,"concrete_test":"Request, for each of the three test DSA sequences, the per-sequence pixel spacing (e.g., DICOM PixelSpacing or equivalent detector calibration), and recompute the center error in Table 1 from raw pixel coordinates using each sequence's own spacing. If no spacing can be supplied, or if the sequence-wise mm conversion changes the pooled mean by more than one reported standard deviation (0.138 mm), the headline 0.421 mm result is not currently established. Separately, ask the authors to state whether the robot-task ground truth in Fig. 5 is manual annotation or robot registration; if it is manual, the claimed ±0.1 mm reference does not apply.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the reported accuracy is that pixel coordinates can be converted to millimetres in a well-defined, reproducible way. The paper never states that conversion. Section 3.1 describes ground-truth boxes only as 12×12 px (±3 px); Section 3.3 defines center error as Euclidean distance in mm; Table 1 and Fig. 3b report mm values. Yet no DICOM pixel spacing, detector calibration, source-to-detector distance, or mm-per-pixel factor is given for any of the three test sequences. If the three DSA runs come from different patients or procedures, their magnification and spacing may differ, so there may not even be a single scale factor. Without this, the 0.421±0.138 mm mean, the 1.736 mm maximum, and the \"sub-pixel accuracy\" claim in Section 6 cannot be reproduced or compared with prior work. A secondary but related inconsistency affects the robotic transfer claim: Section 3.1 says robot registration provides ground truth to ±0.1 mm, while Fig. 5's caption says the orange boxes are manually labeled ground truth; the 0.708/0.148 mm robot errors therefore have an unclear reference standard.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Siamese-network tracker with self- and cross-attention mechanisms and a Directional Edge-Enhanced Attention Network (DEAN) for real-time guidewire tip tracking in DSA images. The method is validated on three clinical DSA sequences, reporting a mean localization error of 0.421 +/- 0.138 mm, a maximum error of 1.736 mm, a mean IoU of 0.782, and 57.2 FPS. It further reports zero-shot transfer to a robotic vascular phantom with errors of 0.708 +/- 0.695 mm and 0.148 +/- 0.057 mm in two navigation tasks. The central claim is that the framework is a real-time, segmentation-free perception module for image-guided endovascular procedures and robot-assisted navigation.","tokens_in":13482,"tokens_out":5521,"duration_ms":55028,"significance":"If the reported figures are reproducible, the contribution is practically useful: a real-time guidewire-tip tracker that avoids a segmentation dependency, with evidence from clinical DSA sequences and a robotic phantom. The architecture is described in enough detail to be reimplemented, and the inclusion of a DEAN ablation baseline is a useful check. The paper also acknowledges important limitations, including the small proprietary dataset and the clinical-to-robotic domain gap. However, the central quantitative claims are currently not fully verifiable because the pixel-to-millimeter conversion is not reported and because the robotic ground-truth reference standard is described inconsistently. These issues are local and fixable, but they must be resolved before the accuracy figures can be accepted.","major_comments":[{"comment":"The reported millimeter accuracies are not reproducible because no pixel-to-millimeter conversion is provided. Section 3.1 states that ground-truth boxes average 12x12 pixels with a standard deviation of +/-3 pixels, while Section 3.3 and Table 1 report center error in millimeters; no DICOM pixel spacing, detector calibration factor, or mm-per-pixel value is given for any of the three DSA sequences. If the sequences come from different patients or procedures, their magnification and pixel spacing may differ, so even a single scale factor may not be well defined. This omission underpins every error value in the paper, including the 0.421 mm mean, the 1.736 mm maximum, the 0.708/0.148 mm robotic errors, and the 'sub-pixel accuracy' claim in Section 6. Please report the per-sequence pixel spacing or another explicit, verifiable conversion and discuss whether a single conversion factor is valid across all three sequences.","section":"Section 3.1, Section 3.3, Table 1"},{"comment":"The reference standard for the robotic experiments is internally inconsistent. Section 3.1 says that the robot's registration system provides ground truth with +/-0.1 mm accuracy, but Section 4.4 and the Figure 5 captions describe the orange boxes as manually labeled or manually annotated ground truth. The claimed 0.708 and 0.148 mm errors therefore have an unclear reference: if the ground truth is manual labels in RGB images, the errors are not measured against the robot-registration ground truth, and the +/-0.1 mm accuracy statement is irrelevant. Please state which reference standard was used for each robotic task and explain how the manual annotations relate to the robot registration system.","section":"Section 3.1, Section 4.4, Figure 5"},{"comment":"The comparison is not sufficient to support the claim that the method outperforms existing approaches. The baselines in Table 1 are Kalman, extended Kalman, particle, and unscented Kalman filters, plus a single 'Baseline' that appears to be the proposed network without DEAN and augmentation; no recent Siamese-based or other deep-learning trackers are included. Because the contribution is a learned Siamese tracker, the 14-35% reduction in mean error relative to classical filters may reflect the use of deep features rather than the proposed attention and DEAN components. Please add at least one modern deep tracker, or alternatively restrict the claim to comparisons against the included baselines and make the ablation of DEAN and attention components more explicit.","section":"Section 4.1, Table 1"},{"comment":"The statistical support for the headline accuracy figures is thin. The evaluation uses three DSA sequences totaling 269 frames, and Table 1 reports only averaged metrics over these sequences with no per-sequence breakdown for IoU, no confidence intervals, and no significance testing. The paper itself acknowledges in Section 5 that validation 'currently spans only three clinical sequences' and that testing on a larger dataset would strengthen generalizability. Please report per-sequence errors and IoU values, and temper the conclusions accordingly or add a statistical test if any claim of superiority is retained.","section":"Section 4.1, Section 5"}],"minor_comments":[{"comment":"The Figure 3(b) caption says 'IoU error curves,' but Section 4.2 describes these as frame-wise localization errors in millimeters; the caption and the text should be made consistent.","section":"Figure 3 caption, Section 4.2"},{"comment":"There are unresolved cross-references in Section 4.4, including 'Fig. ref fig:Deployment(b)' and 'Fig. ref fig:Deployment(c)'; these should be corrected to the actual figure labels.","section":"Section 4.4"},{"comment":"Equation (8) uses Fourier-transformed feature maps \\hat{F}_z and \\hat{F}_x, but it is not defined how the multi-scale features from Equation (3) are aggregated before the frequency-domain cross-correlation; please clarify the feature combination step.","section":"Equation (8)"},{"comment":"The 'Composite Anchors Regression' mentioned in Section 3.2 is not described further; please define this head or give a citation, since it appears to be a nontrivial design choice.","section":"Section 3.2"},{"comment":"The phrase 'sub-pixel accuracy' in Section 6 is not supported by the reported metrics unless a pixel-domain error is also given; the pixel-to-millimeter issue in the major comments applies here as well.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The missing pixel-to-millimeter calibration and the inconsistent robotic ground-truth description are fixable in revision but currently prevent verification of the paper's central accuracy claims. I would support publication after these are resolved and after the claims are either compared with a modern deep tracker or explicitly scoped to the included baselines. The authors' own limitation paragraph is welcome, but it does not substitute for the missing measurement details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate engineering contribution, but you can't check the main numbers as written. The paper reports localization error in millimeters but never states the pixel spacing or any calibration for the DSA sequences. Without that, the 0.421±0.138 mm mean, the 1.736 mm maximum, and the 'sub-pixel accuracy' claim have no reproducible basis. That is the load-bearing problem.\n\nWhat's actually new: the specific combination is new — Siamese ResNet-50 with multi-scale features, a transformer encoder, a directional edge-enhanced attention module (DEAN), and frequency-domain cross-correlation for tip tracking. The robotic platform validation is a genuine plus and gives the work a practical angle that pure tracking papers often miss. The architecture and loss design are clearly described.\n\nThe soft spots, in order of seriousness. First, the calibration issue. The stress-test note is correct: ground-truth boxes are 12×12 px, errors are reported in mm, and no DICOM spacing or mm-per-pixel factor appears anywhere. If the three test sequences come from different patients, there may not even be a single scale factor. This has to be fixed with per-sequence pixel spacing or an explicit calibration method. Second, the baselines. Only classical filters (Kalman, EKF, particle, UKF) are compared. Modern deep trackers are absent, so the superiority claim is not established against current practice. Third, the test set is small — 269 frames across three sequences. The paper acknowledges this, but it still limits the strength of the conclusions. Fourth, there are minor inconsistencies: the abstract says self/cross attention while the method describes channel/spatial attention, and the robotic ground truth is described as robot registration in one place and manual labels in another. These are fixable but should be corrected. Fifth, the 'first' and 'sub-pixel' claims are unsupported as stated.\n\nWho this is for: researchers in interventional imaging and robotic guidance who want a concrete architecture for real-time tool tracking. They'll get useful ideas and a reminder of what must be reported for clinical tracking metrics.\n\nPeer review: yes, it deserves serious referee time, but with the clear expectation of major revision. The core approach is sound and the clinical motivation is real; the quantitative claims need calibration details, stronger baselines, and more careful wording before they can be trusted.","headline":"A sensible Siamese tracker for guidewire tips with a real robotic validation, but every mm accuracy number is unverifiable because the pixel-to-mm conversion is never reported.","tokens_in":14024,"tokens_out":3341,"would_cite":false,"duration_ms":34892,"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":"A Siamese network with dual attention tracks guidewire tips in real-time DSA images with 0.421 mm mean error and transfers zero-shot to a robotic phantom.","keywords":["Guidewire tip tracking","Siamese network","Digital subtraction angiography","Image-guided therapy","Robot-assisted endovascular procedures","Dual attention mechanism","Real-time tracking","Edge-enhanced attention network"],"falsifier":"Take the three test DSA sequences, read the pixel spacing from their DICOM headers, and recompute center errors from the predicted and ground-truth box centers; if the spacing varies by sequence or is absent, the claimed $0.421$ mm mean cannot be reproduced. A second check is annotating a public DSA dataset with millimeter ground truth and running the released model; the method would be falsified if the observed mean error exceeds the reported value by more than the standard deviation.","tokens_in":13047,"feed_emoji":"🎯","tokens_out":5538,"duration_ms":48240,"temperature":0.7,"pith_summary":"This paper aims to establish that guidewire tip tracking during endovascular procedures can be done accurately in real time without relying on image segmentation. The proposed tracker is a Siamese network whose shared ResNet-50 extracts template and search features, a transformer encoder models long-range context, and a Directional Edge-Enhanced Attention Network (DEAN) amplifies guidewire-like linear structures while suppressing artifacts. On three clinical DSA test sequences the tracker reports a mean localization error of $0.421\\pm 0.138$ mm, mean IoU of $0.782$, and 57.2 FPS. The same model transfers zero-shot to a robotic vascular phantom with errors of $0.708\\pm 0.695$ mm and $0.148\\pm 0.057$ mm on two navigation tasks. If true, this provides a segmentation-free, real-time perception module suitable for robot-assisted endovascular navigation.","feed_headline":"Guidewire tips tracked to 0.42 mm at 57 fps","feed_subtitle":"A segmentation-free Siamese dual-attention tracker runs in real time on DSA images and transfers to robotic navigation.","key_machinery":"The central mechanism is the Directional Edge-Enhanced Attention Network (DEAN) operating inside a Siamese tracker. DEAN computes directional edge maps with learnable Sobel-style kernels, applies channel attention that pools global context and spatial attention that fuses the original and edge-enhanced features, then multiplies both attention maps into the feature tensor. This selectively emphasizes the guidewire's thin, elongated structure and suppresses vascular background, artifacts, and noise. A transformer encoder provides self-attention across the template and search features, and a frequency-domain normalized cross-correlation yields a translation-equivariant response map that feeds classification, centerness, and regression heads; the multi-task loss combines GIoU, binary cross-entropy, and Smooth L1. Together these components convert the template-matching paradigm into a segmentation-free, real-time localization pipeline.","core_discovery":"On its own terms, the paper's central claim is that a Siamese template-matching tracker, augmented by edge-enhanced dual attention, can localize the guidewire tip in intraoperative DSA images with sub-millimeter accuracy at clinical frame rates. The network combines self-attention through a transformer encoder with channel and spatial attention in DEAN, and uses frequency-domain cross-correlation between the template and search patches to produce a dense similarity map. Guided by separate classification, centerness, and regression heads trained with GIoU, BCE, and Smooth L1 losses, the model achieves a mean error of $0.421\\pm 0.138$ mm and an IoU of $0.782$ on three held-out clinical sequences, outperforming Kalman, extended Kalman, particle, and unscented Kalman filters in mean error. The same weights, with no fine-tuning, track the tip in a robotic RGB phantom at 38 and 36 FPS, supporting the claim that the tracker can act as the perception input for autonomous endovascular navigation.","pith_inferences":["My inference: the large difference between the two robotic tasks (0.708 vs 0.148 mm) likely reflects imaging quality more than tracking difficulty, so the zero-shot numbers should be read as a range of domain-gap conditions rather than a single transfer accuracy.","My inference: because the paper reports errors in millimeters while annotations are described as about 12×12 pixel boxes, publishing the DICOM pixel spacing for each sequence would let other groups reproduce the comparison; without it, the headline accuracy cannot be independently checked.","My inference: the frame-to-frame template update may be the first failure point under prolonged occlusion; a natural extension is to add a re-detection or template-reinitialization mechanism and test recovery after the tip disappears behind a bifurcation.","My inference: the same Siamese-plus-edge-attention design should transfer to other thin instruments such as catheters, needles, or guidewire shafts, where the same linear-edge prior applies; a cheap test is to train on DSA frames with the shaft tip annotated instead of the distal tip."],"forward_implications":["At 57.2 FPS, the tracker exceeds the 15–30 FPS range the paper cites as needed for smooth intraoperative visualization, so it can feed continuous tip coordinates during an intervention.","Because tracking does not depend on a segmentation step, it avoids the failure mode the paper attributes to UNet-based endpoint localization when vessel boundaries are obscured by motion blur or contrast washout.","The mean IoU of 0.782 versus 0.641 for the no-DEAN baseline implies the edge-enhanced attention module is what closes most of the accuracy gap, not just the Siamese backbone.","Zero-shot transfer to the robotic RGB phantom with mean errors of 0.708 mm and 0.148 mm indicates the same model can serve as a perception module for autonomous guidewire navigation without retraining on the robotic camera.","The maximum clinical error of 1.736 mm bounds the worst-case localization drift, suggesting the tracker remains clinically plausible even during transient artifacts."],"supporting_citations":[{"why":"Supplies the adaptive Kalman filter baseline the proposed Siamese tracker is compared against in Table 1.","marker":"[18]"},{"why":"Supplies the extended Kalman filter baseline used to benchmark mean localization error.","marker":"[26]"},{"why":"Supplies the particle-filter baseline and the real-time exoscope tracking context the method must beat.","marker":"[27]"},{"why":"Supplies the unscented Kalman filter baseline in the quantitative comparison table.","marker":"[28]"},{"why":"Represents the segmentation-dependent multi-task guidewire localization approach the paper argues is fragile in low-contrast DSA.","marker":"[22]"},{"why":"Establishes the Siamese relation-network template-matching paradigm that the proposed architecture builds on.","marker":"[24]"},{"why":"Prior continuum guidewire tracking under fluoroscopy that motivates the segmentation-free visual approach.","marker":"[19]"},{"why":"Documents the DSA artifacts, anatomical variability, and noise that motivate the attention-based design.","marker":"[20]"}],"fun_headline_variants":["Siamese net tracks guidewire tip to 0.42 mm at 57 fps","Real-time sub-mm guidewire tracking with dual-attention Siamese","Guidewire tip located to 0.42 mm in 57 fps DSA tracking","Dual-attention Siamese tracker runs at 57 fps for endovascular","Sub-mm guidewire tip tracking in real time for interventions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every millimeter-level accuracy figure assumes an unstated pixel-to-millimeter calibration of the DSA sequences, because ground-truth boxes are described as roughly 12×12 pixels while errors are reported in millimeters; if the pixel spacing is not fixed and known, none of the reported errors can be verified or compared across methods.","fun_headline_variants_meta":{"raw":{"variants":["Siamese net tracks guidewire tip to 0.42 mm at 57 fps","Real-time sub-mm guidewire tracking with dual-attention Siamese","Guidewire tip located to 0.42 mm in 57 fps DSA tracking","Dual-attention Siamese tracker runs at 57 fps for endovascular","Sub-mm guidewire tip tracking in real time for interventions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00081,"raw_usage":{"total_tokens":3574,"prompt_tokens":983,"completion_tokens":2591,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":2492}},"tokens_in":599,"tokens_out":2591,"duration_ms":19844,"temperature":1.0,"reasoning_tokens":2492,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:57:05.664032+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the three test DSA sequences, read the pixel spacing from their DICOM headers, and recompute center errors from the predicted and ground-truth box centers; if the spacing varies by sequence or is absent, the claimed $0.421$ mm mean cannot be reproduced. A second check is annotating a public DSA dataset with millimeter ground truth and running the released model; the method would be falsified if the observed mean error exceeds the reported value by more than the standard deviation.","supporting_citations":[{"cited_title":"Needle tip tracking in 2d ultrasound based on improved compressive tracking and adaptive kalman filter,","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive Kalman filter baseline the proposed Siamese tracker is compared against in Table 1."},{"cited_title":"Uniting attitude estimation with global asymptotic stability,","cited_arxiv_id":null,"evidence_quote":"Supplies the extended Kalman filter baseline used to benchmark mean localization error."},{"cited_title":"Hybrid tracking module for real-time tool tracking for an autonomous exoscope,","cited_arxiv_id":null,"evidence_quote":"Supplies the particle-filter baseline and the real-time exoscope tracking context the method must beat."},{"cited_title":"Active-model-based precise twist steering for autonomous robotic flexible endoscope,","cited_arxiv_id":null,"evidence_quote":"Supplies the unscented Kalman filter baseline in the quantitative comparison table."},{"cited_title":"A real-time multi-task framework for guidewire segmentation and endpoint localization in endovascular interventions,","cited_arxiv_id":null,"evidence_quote":"Represents the segmentation-dependent multi-task guidewire localization approach the paper argues is fragile in low-contrast DSA."},{"cited_title":"Learning to filter: Siamese rela- tion network for robust tracking,","cited_arxiv_id":null,"evidence_quote":"Establishes the Siamese relation-network template-matching paradigm that the proposed architecture builds on."},{"cited_title":"Real-time pose tracking for a contin- uum guidewire robot under fluoroscopic imaging,","cited_arxiv_id":null,"evidence_quote":"Prior continuum guidewire tracking under fluoroscopy that motivates the segmentation-free visual approach."},{"cited_title":"Enhancing percutaneous coronary interven- tion with heuristic path planning and deep-learning-based vascular segmentation,","cited_arxiv_id":null,"evidence_quote":"Documents the DSA artifacts, anatomical variability, and noise that motivate the attention-based design."}],"review_version":1}