{"id":"ac4dd19f-7ef6-4405-a9fb-3009da798f05","arxiv_id":"2506.21631","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":10,"one_line_summary":"A CTA-guided inverse projection framework for real-time 3D guidewire reconstruction from single-view DSA is proposed, but the depth formula and evaluation do not establish true 3D accuracy.","lead":"This paper reports a system that combines 3D scans of blood vessels with 2D X-ray movies to reconstruct the 3D shape of a guidewire during robot-assisted surgery, in real time. It claims high accuracy and speed, but the depth recovery step is mathematically questionable and the reported accuracy metric does not actually test 3D correctness.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth formula in Section II-D is not a valid monoplane triangulation; a single ray plus a known rigid transform leaves depth unconstrained, so the 3D reconstruction claim is unsupported.","rationale":"Reading the paper charitably, the claimed contribution is monoplane 3D guidewire reconstruction that uses the CTA-to-DSA transform to lift 2D pixels into 3D. That would require an additional constraint, such as a second view, a known vessel-surface prior, or a learned depth model. The paper does not supply one: the derivation in Section II-D produces a depth from a cross-product formula involving only t, R, K, and g, while Section III-C concedes that no 3D ground truth exists. This is internally inconsistent with the claim of accurate 3D reconstruction, not merely a disagreement with current consensus. The projection-consistency metric is self-confirming because any lifted point on the same camera ray reprojects to its source pixel. There is no machine-checked proof or released code that could independently support the geometry. I therefore agree with the reader's weakest-assumption analysis and with the REJECT verdict; the final verdict should remain unchanged.","tokens_in":11207,"tokens_out":10043,"duration_ms":124765,"concrete_test":"Set up a synthetic calibrated monoplane camera with known K, R, t. Pick a pixel g and choose two true depths lambda1 != lambda2 along the same ray, defining ground-truth points P_i = T^{-1}(lambda_i K^{-1}[g;1]); both project to g. Run the Section II-D depth formula for g and show it returns the same d_m in both cases, so the reconstructed point is identical and cannot equal both ground-truth points. Then repeat for two full 3D guidewire curves that share the same DSA projection but have different depth maps, and verify that the reconstructed 3D curves coincide while the projection error stays near zero. This decisively separates the claimed depth recovery from the depth-unconstrained geometry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the closed-form depth in Section II-D. For a calibrated pinhole camera with known T and K, a guidewire pixel g defines a single ray r = K^{-1}[g;1] in camera coordinates. Every point P(lambda) = T^{-1}(lambda r), lambda > 0, projects exactly to g; the actual guidewire depth is one specific lambda fixed by the 3D scene. The proposed d_m is computed only from t, R, K and g, without querying the CTA centerline C_3D or using any vessel-surface prior. Algebraically it is the lambda that makes d_m r lie on the line through t in direction R^T r; nothing in the imaging geometry forces the guidewire to lie on that line. Thus the method returns one arbitrary point per ray, and any number of physically different 3D guidewire configurations sharing the same 2D segmentation would be assigned identical reconstructed coordinates. The evaluation cannot rescue this: Section III-C explicitly states that there is no 3D ground truth, and the projection-consistency metric of Eq. (5) is insensitive to depth because, by construction, every reconstructed point T^{-1}(d r) reprojects to the source pixel. A projection error near 1.76 px therefore reflects segmentation and numerical effects, not 3D accuracy. The central claim of real-time accurate 3D spatial information lacks a valid geometric basis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for reconstructing the 3D shape of a guidewire from intraoperative monoplane 2D DSA images, guided by preoperative 3D CTA. The pipeline consists of segmentation of the guidewire in DSA, extraction of vascular centerlines in both modalities, deformable 3D-2D registration on SE(3), and a so-called inverse projection step that assigns a depth to each 2D guidewire point. The authors report a projection error of 1.76 ± 0.08 px, a length deviation of 2.93 ± 0.15%, and a processing speed of 39.3 ± 1.5 FPS, and claim that the method provides accurate real-time 3D spatial information for robot-assisted endovascular interventions.","tokens_in":11578,"tokens_out":4378,"duration_ms":58070,"significance":"If the 3D reconstruction claim were valid, a monoplane DSA + CTA fusion running at 39 FPS would be a useful contribution to robotic endovascular navigation, where depth information is otherwise missing. The paper has some credible components: a clinical DSA training set with expert annotations, a comparative segmentation study, and a realistic vascular phantom setup. However, the central geometric derivation is not sound: the depth formula in Section II-D does not follow from the imaging model, and the reported accuracy metrics do not measure true 3D accuracy. The central claim of accurate 3D spatial information is therefore unsupported by the evidence presented.","major_comments":[{"comment":"The closed-form depth d_m is not a valid monoplane triangulation. For a calibrated pinhole camera, a pixel g_m defines a ray {λ K^{-1}[g_m;1] : λ > 0} in camera coordinates, and every λ yields a 3D point P = T^{-1}(λ K^{-1}[g_m;1]) that projects exactly to g_m. The proposed d_m selects the λ that makes d_m K^{-1}[g_m;1] lie on the line t + μ R^T K^{-1}[g_m;1], but nothing in the imaging geometry forces the guidewire to lie on that line, and the CTA centerline C_3D does not appear in the formula. Thus the reconstructed coordinates are an arbitrary choice per ray, not a measurement of the actual guidewire depth.","section":"Section II-D, depth formula"},{"comment":"The depth formula is singular at the reported initialization. In Section III-A the registration is initialized with R = I and t = [0,0,1000]^T in pixel coordinates. For these values, a = R^T K^{-1}[g;1] equals b = K^{-1}[g;1], so the denominator a × b is zero and d_m is undefined for every guidewire pixel. The formula therefore cannot be evaluated at the exact configuration from which the reported experiments start.","section":"Section II-D and Section III-A"},{"comment":"The projection consistency error is a self-consistency metric, not a 3D accuracy metric. Since each reconstructed point is defined as P_m = T^{-1}(d_m K^{-1}[g_m;1]) using the same pixel g_m, applying the projection operator necessarily returns g_m (up to numerical precision). The reported 1.76 ± 0.08 px therefore reflects only segmentation and numerical effects, and would be essentially unchanged for arbitrarily wrong depths. The paper itself states in Section III-C that there is no 3D ground truth, so this metric cannot support the claim of accurate 3D reconstruction.","section":"Section III-C, Eq. (5)"},{"comment":"The length deviation metric is not defined. It is unclear whether the reconstructed 3D arc length is compared to the 2D segmentation length, to a known physical guidewire length, or to some other reference. If it is compared to the 2D length, the metric is scale-dependent and can be made small by choosing depth values appropriately; if it is compared to a physical length, that ground truth is not described. A precise definition and an independent 3D ground truth are required before any accuracy claim can be made.","section":"Table II and Section III-C"}],"minor_comments":[{"comment":"The heading contains a typo: \"Experimal Setup\" should be \"Experimental Setup\".","section":"Section III-A"},{"comment":"The text refers to Fig. 4(c) for both segmentation and reconstruction results, but the caption describes panels (a)-(d); the figure references should be corrected and made consistent with the text.","section":"Figure 4 and Section III-D"},{"comment":"The covariance matrices Σ_j and Λ_i are introduced without definitions, and the dimensions of Λ_i are not specified consistently with the residual term T^{-1}(π^{-1}(q_j)) - p_i, which is a 3D vector.","section":"Section II-C, Eq. (2)"},{"comment":"The graph Laplacian and spectral decomposition developed in Section II-B are not used anywhere in the registration or reconstruction pipeline; this section is disconnected from the rest of the method.","section":"Section II-B"},{"comment":"The abstract and discussion describe the reconstruction as \"physics-constrained,\" but no physics-based constraint appears in Section II-D; the depth assignment is purely algebraic and unconstrained.","section":"Section II-D"}],"recommendation":"reject","confidential_remarks":"The central flaw is not a matter of presentation: the depth formula in Section II-D is geometrically invalid, and the evaluation metrics in Section III-C do not measure 3D accuracy. These are load-bearing issues that cannot be fixed by local edits; a correct method would need to derive depth from the CTA vessel geometry or a second view, with validation against independent 3D ground truth. I recommend rejection in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked for my read on arXiv:2506.21631. The punchline is that the 3D reconstruction claim does not hold up. Section II-D gives a closed-form depth formula that supposedly recovers the guidewire's depth from a single pixel and the known rigid transform. But a single calibrated ray leaves depth completely free; the formula does not use the CTA geometry or any vessel-surface constraint. It is not a triangulation. With the stated initialization (R=I, t=[0,0,1000]), the denominator actually vanishes, so the formula is undefined for the authors' own setup. That is a load-bearing flaw, not a minor one.\n\nWhat the paper does well: the segmentation pipeline is solid. The UNet results on a clinical dataset are credible, and the phantom experiments with a real robotic setup show the integration works end-to-end at 39 FPS. The authors are also honest enough to state there is no 3D ground truth. That honesty makes the evaluation section more troubling, because they then rely on a projection consistency error that is circular: the reconstructed 3D points were built from the same 2D segmentation they are reprojected onto, so low error is guaranteed by construction. The length deviation metric is also questionable without a true 3D reference.\n\nOther soft spots: the abstract says \"deformable image registration,\" but the method uses a rigid SE(3) transform. That is an internal contradiction. The novelty is incremental, as the paper itself cites prior monoplane guidewire reconstruction and PnP-based vascular registration. The contribution is the integration, but the integration's central step is invalid.\n\nMy verdict: this should not be accepted as is. The depth estimation needs to be replaced with something that actually uses the CTA prior, e.g., back-projecting onto the vessel centerline or surface. The evaluation needs a real 3D ground truth or a physically realistic simulation with known ground truth.\n\nWould I send it to peer review? Yes, if the venue is willing to give the authors a chance to fix the geometry. The system is real, the segmentation is useful, and the failure mode is instructive. But at this stage it is a reject-and-resubmit, not a conditional accept. I wouldn't cite the 3D reconstruction results, though the segmentation comparison could be useful. It's a good reading-group example of circular metrics in image-guided intervention papers.","headline":"The paper builds a real phantom system and reports good segmentation, but its central 3D reconstruction step is geometrically unfounded and the evaluation metrics don't measure what they claim.","tokens_in":12081,"tokens_out":2851,"would_cite":false,"duration_ms":29622,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By registering a preoperative 3D CT angiography model to live 2D DSA images and lifting guidewire pixels along back-projected rays with a closed-form depth formula, this paper claims real-time 3D guidewire reconstruction at 39.3 frames…","keywords":["guidewire reconstruction","2D-3D registration","digital subtraction angiography","CT angiography","endovascular robotics","inverse projection","real-time shape sensing"],"falsifier":"Image a phantom with a guidewire whose true 3D shape is known from a CT scan, using one DSA view, run the pipeline, and compare the reconstructed depth (distance from the X-ray source) of each point with the known value; large depth errors alongside small 2D projection errors would show that the single-view depth formula is not measuring true 3D position.","tokens_in":11027,"feed_emoji":"🩻","tokens_out":14317,"duration_ms":128466,"temperature":0.7,"pith_summary":"This paper sets out to recover the full 3D shape of an interventional guidewire during robot-assisted endovascular procedures, where the only live images are single-view 2D X-ray angiograms (DSA) that carry no depth. The proposed framework first segments the guidewire in the 2D image, then registers the pre-scanned 3D vessel model (CTA) to the live 2D vessel projection by optimizing a rigid transform on SE(3), and finally lifts each 2D guidewire pixel to a 3D point using an inverse-projection formula whose per-point depth is given in closed form. The authors report that this runs in real time at 39.3 frames per second, with a projection error of 1.76 pixels and a length deviation of 2.93 percent, outperforming EPnP and B-spline baselines. If correct, the method would give robotic controllers continuous 3D shape feedback from standard imaging hardware, without adding sensors to the thin guidewire.","feed_headline":"3D guidewire shape recovered live from one 2D X-ray view","feed_subtitle":"Fusing pre-op CT angiography with live DSA images hits under 2-pixel error at 39 FPS.","key_machinery":"The load-bearing object is the inverse-projection step that converts 2D guidewire pixels into 3D coordinates: $G_{3D} = \\{[R^{\\top} \\mid -R^{\\top}t]\\,[d_m K^{-1}g_m;1]\\}_{m=1}^M$, with the depth $d_m$ given by the closed-form linear-triangulation quotient. This is what turns a segmented 2D skeleton into a 3D curve, and it is the mechanism that makes the reconstruction real-time because each point costs $\\mathcal{O}(1)$ arithmetic. The other essential component is the registration stage that supplies $T$: a weighted perspective-$n$-point problem solved on the Lie algebra $\\mathfrak{se}(3)$ by manifold-aware Levenberg-Marquardt with a pseudo-Huber robust kernel, which aligns the CTA centerline with the DSA centerline and thus provides the spatial reference frame in which the inverse projection operates.","core_discovery":"The central claim is that, once the preoperative 3D CTA centerline is registered to the intraoperative 2D DSA centerline, the 3D guidewire is recovered by a direct linear mapping: each 2D guidewire point $g_m$ is back-projected along its camera ray and placed at depth $d_m = \\frac{\\|t \\times R^{\\top}K^{-1}[g_m;1]\\|}{\\|(R^{\\top}K^{-1}[g_m;1]) \\times (K^{-1}[g_m;1])\\|}$, giving $P_m = T^{-1}(d_m K^{-1}[g_m;1])$ with $T=(R,t)$ the registered rigid pose and $K$ the intrinsic matrix. The authors call this a physics-constrained inverse projection and a form of linear triangulation; it assigns each point a depth in $\\mathcal{O}(1)$ time using only the pose, the intrinsics, and the pixel coordinate. On their phantom experiments the reconstruction reaches a projection error of $1.76\\pm0.08$ pixels, a length deviation of $2.93\\pm0.15\\%$, and a frame rate of $39.3\\pm1.5$ FPS, and the paper interprets these numbers as evidence that the method provides accurate, real-time 3D spatial information for robotic navigation.","pith_inferences":["The paper states in its evaluation section that no 3D ground truth exists for this problem; projection error and length deviation therefore verify 2D consistency and arc-length preservation rather than true depth accuracy, and a phantom with known 3D positions would be the direct check.","The closed-form depth uses only the registered pose and the pixel coordinate, so the vessel model enters the reconstruction only through the alignment $T$; a testable consequence is that depth accuracy should degrade when the live vessel shape deviates from the preoperative CTA.","Because each reconstructed point costs $\\mathcal{O}(1)$ arithmetic, the pipeline could plausibly be extended to multiple instruments or to dense catheter body reconstruction without a redesign, an extension the paper does not make.","Adding an occasional second X-ray view would provide a direct calibration of the monoplane depths and would test whether the single-view assumption is sufficient."],"forward_implications":["A robotic guidewire manipulator would receive continuous 3D shape feedback at about 39 frames per second from the standard single-view DSA images already used in the operating room, with no sensors added to the guidewire itself.","At vascular bifurcations the reconstructed 3D configuration would supply the spatial information needed to decide which branch the wire enters, addressing a decision that 2D images leave ambiguous.","The reported projection error ($1.76\\pm0.08$ px) and length deviation ($2.93\\pm0.15\\%$) are lower than the compared EPnP and B-spline baselines, so the framework is claimed to be both more accurate and faster than those methods.","Because the reconstruction is a closed-form linear mapping from pixels plus a registered pose, it can be embedded into planning, simulation, and reinforcement-learning pipelines, including the digital-twin direction named in the paper's future work."],"supporting_citations":[{"why":"Establishes the monoplane fluoroscopic 3D guidewire reconstruction problem that this work extends with CTA assistance.","marker":"[22]"},{"why":"Supplies the iterative PnP 3D-2D vascular image registration approach on which the registration stage builds.","marker":"[25]"},{"why":"Provides the B-spline keypoint reconstruction baseline and spline model used for comparison.","marker":"[20]"},{"why":"Provides the EPnP-based real-time 3D pose reconstruction baseline used for comparison.","marker":"[28]"},{"why":"Motivates the covariance-weighted 2D/3D registration with uncertainty priors used in the energy function.","marker":"[27]"},{"why":"Documents why embedded shape sensors such as FBG and EM are impractical for thin guidewires, motivating the vision-based approach.","marker":"[12]"}],"fun_headline_variants":["Live 3D guidewire from single 2D X-ray at 39 FPS","Real-time guidewire 3D shape from DSA and CTA fusion","Sub-2-pixel 3D guidewire tracking in real time","Robotic navigation gets live 3D guidewire from 2D DSA","Fusing CTA and DSA yields guidewire 3D in real time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a single 2D X-ray picture, together with the known position and angle of the X-ray source and the alignment between the pre-scanned vessel model and the live picture, uniquely fixes how far each guidewire point is from the X-ray source; without a second picture or a constraint from the vessel wall, that distance is free.","fun_headline_variants_meta":{"raw":{"variants":["Live 3D guidewire from single 2D X-ray at 39 FPS","Real-time guidewire 3D shape from DSA and CTA fusion","Sub-2-pixel 3D guidewire tracking in real time","Robotic navigation gets live 3D guidewire from 2D DSA","Fusing CTA and DSA yields guidewire 3D in real time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1692,"prompt_tokens":1103,"completion_tokens":589,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":719,"completion_tokens_details":{"reasoning_tokens":498}},"tokens_in":719,"tokens_out":589,"duration_ms":5699,"temperature":1.0,"reasoning_tokens":498,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:57:36.416967+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Image a phantom with a guidewire whose true 3D shape is known from a CT scan, using one DSA view, run the pipeline, and compare the reconstructed depth (distance from the X-ray source) of each point with the known value; large depth errors alongside small 2D projection errors would show that the single-view depth formula is not measuring true 3D position.","supporting_citations":[{"cited_title":"3d guidewire shape reconstruction from monoplane fluoroscopic images,","cited_arxiv_id":null,"evidence_quote":"Establishes the monoplane fluoroscopic 3D guidewire reconstruction problem that this work extends with CTA assistance."},{"cited_title":"Iterative pnp and its application in 3d-2d vascular image registration for robot navigation,","cited_arxiv_id":null,"evidence_quote":"Supplies the iterative PnP 3D-2D vascular image registration approach on which the registration stage builds."},{"cited_title":"Suture thread spline reconstruction from endoscopic images for robotic surgery with reliability-driven keypoint detection,","cited_arxiv_id":null,"evidence_quote":"Provides the B-spline keypoint reconstruction baseline and spline model used for comparison."},{"cited_title":"A reinforce- ment learning approach for real-time articulated surgical instrument 3d pose reconstruction,","cited_arxiv_id":null,"evidence_quote":"Provides the EPnP-based real-time 3D pose reconstruction baseline used for comparison."},{"cited_title":"2d/3d registration with a statistical deformation model prior using deep learning,","cited_arxiv_id":null,"evidence_quote":"Motivates the covariance-weighted 2D/3D registration with uncertainty priors used in the energy function."}],"review_version":1}