{"id":"1903059c-9472-4bf6-a51f-03f63261964b","arxiv_id":"2501.04515","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A transformer network that outputs B-spline control points for guidewire shape enables a real robot to cannulate the brachiocephalic artery autonomously with a 50% success rate.","lead":"SplineFormer predicts the shape of a medical guidewire inside blood vessels from X-ray images, then feeds that shape into a robot that steers the wire on its own. In physical phantom tests, the robot successfully reached one target artery in 50% of trials, showing progress toward fully automated endovascular procedures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The autonomous-navigation claim lacks a described training objective for the PolicyConv action head; Eq. (8) has no action term, so the reported 50% success rate is not derivable from the presented method.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing gap: Equation (8) contains no action term, so the PolicyConv action head has no described training signal. I agree with the REJECT verdict because the central claim of full autonomy cannot be derived from the method as presented. The shape-prediction component is plausible and the physical robot setup is described concretely, but the policy head is the missing link between spline prediction and navigation. The fix is straightforward—explicitly report the action loss, training procedure, and success-rate statistics with confidence intervals—but until then the headline claim is unsupported.","tokens_in":12154,"tokens_out":3368,"duration_ms":35562,"concrete_test":"Inspect the released code or request the training script, then instrument the backward pass to confirm which loss terms produce gradients reaching the PolicyConv action head. If the only training loss is Eq. (8), PolicyConv weights receive no gradient, proving the actions were not learned by the presented method; if an additional action loss exists, it must be reported in the paper. As a complementary check, re-train SplineFormer exactly as specified, freeze the shape-prediction head, and test whether PolicyConv produces any meaningful commands; random or constant outputs would indicate the navigation claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that SplineFormer 'is able to perform endovascular navigation autonomously and achieves a 50% success rate when cannulating the brachiocephalic artery on the real robot.' That claim requires a policy that maps image or spline features to translation and rotation commands. The architecture figure shows a 'PolicyConv' head outputting translation and rotation, and Section V.A describes collecting state-action pairs and deploying SplineFormer after training, but no training objective for this head is specified anywhere in the text. The only loss, Eq. (8) in Section IV.C, consists of MSE on predicted spline parameters and knots, BCE on end-of-sequence predictions, and a curvature-consistency term; there is no action loss, no reinforcement-learning objective, and no behavior-cloning loss. This is not an internal contradiction so much as a missing component, and it is load-bearing: if the PolicyConv head was trained by an unreported objective, the paper omits a central part of the method; if it was not trained at all, the autonomous navigation result cannot be reproduced or evaluated from the paper as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SplineFormer, a transformer-based architecture that predicts B-spline control points and knots representing the guidewire directly from X-ray images, thereby bypassing pixel-wise segmentation. The predicted spline is fed into a PolicyConv head that outputs translation and rotation commands for a physical endovascular robot. Experiments on a vascular phantom with a leader-follower robot report a 50% success rate in cannulating the brachiocephalic artery, compared with 5.6% for a behavior-cloning baseline, and the paper also presents qualitative shape-prediction results and attention visualizations.","tokens_in":12407,"tokens_out":6957,"duration_ms":69851,"significance":"If the experimental claims are reproducible, the paper would make a useful contribution: direct spline prediction is a plausible way to avoid the fragmentation problems of segmentation, and a fully autonomous 50% cannulation rate on a physical phantom is a nontrivial result. Strengths include the use of real bi-planar X-ray data (8,746 annotated images), a physical robot evaluation, and a comparison with an existing autonomous baseline. However, the claims are currently contingent on a missing description of the policy training objective, the absence of quantitative shape-prediction metrics, and weak statistical reporting. I do not see a circularity problem, because the evaluation uses external baselines on a real robot; the spline representation is not reducing to fitted constants in a circular way. The 'explainable' claim rests only on attention visualizations and has no quantitative support.","major_comments":[{"comment":"The autonomous navigation claim rests on the PolicyConv action head, but no training objective for this head is specified. The global loss in Eq. (8) contains only the spline-parameter MSE, the end-of-sequence cross-entropy, and the curvature-consistency term; there is no action loss, behavior-cloning loss, or reinforcement-learning objective. Section V.A states that the learned policy is evaluated on state-action pairs and that SplineFormer is deployed after training, but the training procedure that maps spline or image features to translation and rotation commands is never described. Without this component, the reported 50% success rate cannot be reproduced or even evaluated from the manuscript. The authors must state how PolicyConv is trained, including the objective, the data used, and any action-space preprocessing.","section":"Section IV.C, Eq. (8); Fig. 4"},{"comment":"The headline result is reported as a 50% success rate over 20 trials, but the manuscript does not define what constitutes a successful cannulation, does not report the number of successes (10/20 is only inferred), and gives no confidence interval or statistical comparison. It is also unclear whether unsuccessful trials are included in the mean completion time of 150 ± 45.6 s. The comparison to Behavior Cloning (5.6%) and GAIL-PPO (69.4% semi-autonomous) has no significance testing and mixes fully autonomous with semi-autonomous conditions. Finally, Table I leaves the LCCA entry for SplineFormer empty even though the text states that the method could not cannulate the LCCA; if that target was attempted, the success rate and trial count should be reported. Please define the success metric, report per-target trial counts, and provide confidence intervals or raw outcomes.","section":"Section V.A, Table I"},{"comment":"The shape-prediction component, which is the core of the proposed representation, is evaluated only qualitatively. No quantitative error metric (e.g., endpoint error, control-point error, or curve distance) is reported on a held-out test set, and there is no ablation for the loss weights or the number of control points. Because the autonomous navigation system depends on the accuracy of the predicted spline, the paper needs a quantitative evaluation of shape prediction and an analysis of how prediction errors affect navigation outcomes.","section":"Section V.B, Fig. 7"},{"comment":"Several architecture and training hyperparameters required for reproduction are missing. The patch size |P|, the number of encoder and decoder layers N_e and N_D, the number of heads H, the loss weights λ_a, λ_b, λ_c, the curvature sample count n, and the number of control points in the output sequence are not given; Eq. (8) also uses N without defining it. Including these values, along with the learning-rate schedule and any data augmentation, is necessary for the experimental claims to be reproducible.","section":"Section IV.B, IV.C"}],"minor_comments":[{"comment":"The driver part number 'A4899' appears to be a typo for 'A4988'; please correct it.","section":"Section III"},{"comment":"The B-spline recursion is undefined when knot values coincide; the standard 0/0 = 0 convention should be stated explicitly.","section":"Section IV.A, Eq. (3)"},{"comment":"The action space is described as a_t ∈ [−1, 1]^2 corresponding to a maximum translation of 2 mm and a rotation of 15°, but the text does not explain how the continuous PolicyConv output is mapped to these limits or whether the actions are relative or absolute.","section":"Section V.A"},{"comment":"The 'discard factor' used in the attention-map fusion is not defined, which makes the attention visualization procedure non-reproducible.","section":"Section V.C"},{"comment":"Citation [43] is cited for 'advancements in autonomous driving,' but [43] is a guidewire tracking paper; this citation appears mismatched.","section":"Section II"},{"comment":"There are several grammatical errors, including 'most studies were conducted' (should be 'most studies being conducted' or similar) and 'deep learning models face challenges often struggle' (should be 'deep learning models often struggle').","section":"Sections I and II"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the unstated training objective for PolicyConv. If the authors cannot provide a precise description of how the action head is trained, the autonomous navigation claim should be withdrawn and the paper limited to shape prediction. I would also encourage the editor to require the authors to release code or detailed experimental logs, because the current text does not allow the reader to determine how actions are generated from the predicted spline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece is the spline prediction head: predicting B-spline control points and knots directly from X-ray images rather than segmenting a mask. The loss in Eq. (8) for shape, end-of-sequence, and curvature consistency is well-defined, and the attention visualization suggests the model does focus on the guidewire tip. The physical phantom with a bi-planar X-ray system is a real step up from simulation. If the shape prediction works as described, that portion is a solid contribution.\n\nThe problem is the headline claim. Section V says SplineFormer 'is able to perform endovascular navigation autonomously' with a 50% success rate, but the paper never states how the PolicyConv action head is trained. Eq. (8) contains no action loss—no behavior cloning, no RL objective, no supervised policy term. The text says 'we evaluate the learned policy on the state-action pairs,' but no learning procedure for the policy is described anywhere. That omission is load-bearing: without it, the navigation result is not reproducible or fully interpretable. The 20 trials also yield a wide confidence interval (10/20 successes), and there is no error analysis. The comparison to GAIL-PPO and Behavior Cloning from [22] uses numbers from a different study, which may not be comparable under identical conditions.\n\nThese are fixable problems, not insurmountable ones. The shape prediction contribution stands on its own, and the navigation result might be salvageable if the policy details are added. But as written, the central claim is missing its method, so I would not rely on the navigation result.\n\nI would send this to peer review. Reviewers can push for the missing policy training details, hyperparameters, and a more careful statistical treatment. It is not a desk reject—the physical setup and the spline representation deserve referee time, and the gap is the kind that a revision can close. I would not cite the navigation result in my own work until the policy is clarified.","headline":"The spline-shape prediction is a genuinely useful idea, but the autonomous navigation claim is not yet supported because the policy head has no described training objective.","tokens_in":12928,"tokens_out":4315,"would_cite":false,"duration_ms":40733,"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":"SplineFormer predicts a guidewire's B-spline shape from X-ray images and uses it to drive a robot that cannulates the brachiocephalic artery autonomously 50% of the time.","keywords":["Endovascular navigation","Guidewire shape prediction","B-spline representation","Transformer","Autonomous navigation","Explainable AI","Fluoroscopy","Robot-assisted intervention"],"falsifier":"Inspect the training code and learning curves for the PolicyConv action head: if no loss or gradient ever updates that head (no behavior-cloning term, no RL return, no auxiliary action loss), then the reported 50% robot success rate cannot be produced by the described method.","tokens_in":11967,"feed_emoji":"🤖","tokens_out":4844,"duration_ms":42159,"temperature":0.7,"pith_summary":"SplineFormer claims that a transformer can infer a guidewire's continuous shape directly from fluoroscopic X-ray images, representing it as a B-spline curve rather than a segmentation mask. This condensed geometric representation is then used as the state for an end-to-end robot control pipeline, and the paper reports that the resulting system cannulates the brachiocephalic artery in a physical vascular phantom fully autonomously in 50% of 20 trials. The authors argue this beats a fully autonomous behavior-cloning baseline (5.6% success) and offers an explainable, fragmentation-free alternative to segmentation-based methods, at the cost of lower success than a semi-autonomous GAIL-PPO approach. If true, the result suggests that direct spline-parameter prediction is a viable state representation for autonomous endovascular navigation.","feed_headline":"SplineFormer drives a robot to cannulate an artery 50% of the time","feed_subtitle":"Predicting the guidewire as a smooth B-spline lets a robot navigate fully autonomously to the brachiocephalic artery.","key_machinery":"The central machinery is the B-spline curve model $C(t) = \\sum_{i=0}^{n} P_i B_{i,p}(t)$, which encodes the guidewire as a small set of control points $P_i$ and knot values $t_i$, giving a smooth, continuous curve with local control. A visual-transformer encoder processes X-ray image patches; a transformer decoder with masked self-attention generates the coefficient–knot pairs autoregressively, starting from a separately predicted tip point; a curvature-consistency loss samples the predicted curve to keep it smooth; and a PolicyConv head maps the decoder's condensed state to translation/rotation commands for the robot.","core_discovery":"On the paper's own terms, the central discovery is that a transformer-based network can predict the B-spline coefficients and knot sequence that define a guidewire's shape in fluoroscopy, and that this spline, together with a learned action head, carries enough information for a physical robot to navigate a guidewire to a target vessel without human intervention. The network is trained with a composite loss on spline-parameter prediction, end-of-sequence classification, and sampled-curve curvature consistency; the robot's translation and rotation actions come from a PolicyConv head attached to the decoder output. In evaluation, SplineFormer reaches the brachiocephalic artery in 50% of 20 fully autonomous trials, with a mean completion time of 2.5 minutes, and cannot reach the left common carotid artery, a limitation the authors attribute to the more complex 3D anatomy of that branch.","pith_inferences":["The reported comparison is uneven: SplineFormer is fully autonomous while GAIL-PPO receives teleoperator assistance; a fair test of the representation would pit the spline state against a mask-based state inside the same fully autonomous policy.","The absence of a policy-loss term in the write-up makes it likely that the PolicyConv action head is trained by some implicit gradient from the composite loss or by a procedure the paper omits; pinning this down is the fastest way to test the navigation claim.","The same B-spline regression idea could transfer to other thin deformable instruments, such as catheters, ablation catheters, or endoscopes, and to other imaging modalities, since the representation is modality-agnostic once the tool is visible.","A testable extension would be to run SplineFormer with a frozen spline predictor and a randomly initialized action head; if the 50% success persists, the navigation result is not attributable to shape prediction."],"forward_implications":["If SplineFormer is correct, segmentation-free guidewire shape prediction can feed a robot controller directly, avoiding the discontinuous masks that plague U-Net-style segmentation.","Fully autonomous cannulation of the brachiocephalic artery at 50% success on a real robot, against a 5.6% behavior-cloning baseline, indicates the spline state carries enough information for navigation decisions.","The method's explainability, with attention concentrated near the guidewire tip and target branches, could make robot decisions easier for clinicians to verify during procedures.","The inability to cannulate the left common carotid artery shows the approach needs stronger handling of complex 3D anatomy before clinical translation.","Because the spline is a compact parameterization, the approach may scale to real-time use with modest compute, as demonstrated on an RTX 4080 GPU."],"supporting_citations":[{"why":"Provides the GAIL-PPO and Behavior Cloning baselines that SplineFormer is compared against; its Behavior Cloning result (5.6%) is the fully autonomous reference point.","marker":"[22]"},{"why":"Introduces the B-spline tube model for catheter/guidewire tracking, the representational basis SplineFormer adopts for shape prediction.","marker":"[56]"},{"why":"Supplies the vision-transformer encoder design (patch embedding, multi-head self-attention) used to process fluoroscopic images.","marker":"[57]"},{"why":"Supplies the transformer decoder architecture with masked self-attention and cross-attention that the autoregressive spline generation is built on.","marker":"[58]"},{"why":"Provides the template for autoregressive coefficient generation in transformers, which SplineFormer adapts to spline control points and knots.","marker":"[59]"},{"why":"The U-Net segmentation baseline used to illustrate the discontinuity and tip-missing failure modes that spline prediction is meant to avoid.","marker":"[40]"}],"fun_headline_variants":["SplineFormer predicts guidewire spline, robot cannulates artery 50% of time","Robot guidewire navigation hits 50% success using spline-predicting transformer","Autonomous artery cannulation robot succeeds 50% with spline transformer","Transformer learns guidewire shape spline enabling half-successful autonomous cannulation","B-spline guidewire prediction lets robot reach target artery in half of trials"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The autonomous navigation claim rests on the unstated assumption that the PolicyConv action head is trained by a well-defined objective; the only loss the paper reports covers spline parameters, end-of-sequence, and curvature, with no action term, so the 50% success rate is not derivable from the described training procedure.","fun_headline_variants_meta":{"raw":{"variants":["SplineFormer predicts guidewire spline, robot cannulates artery 50% of time","Robot guidewire navigation hits 50% success using spline-predicting transformer","Autonomous artery cannulation robot succeeds 50% with spline transformer","Transformer learns guidewire shape spline enabling half-successful autonomous cannulation","B-spline guidewire prediction lets robot reach target artery in half of trials"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001241,"raw_usage":{"total_tokens":5078,"prompt_tokens":917,"completion_tokens":4161,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":4058}},"tokens_in":533,"tokens_out":4161,"duration_ms":30984,"temperature":1.0,"reasoning_tokens":4058,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:30:27.527589+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the training code and learning curves for the PolicyConv action head: if no loss or gradient ever updates that head (no behavior-cloning term, no RL return, no auxiliary action loss), then the reported 50% robot success rate cannot be produced by the described method.","supporting_citations":[{"cited_title":"Collaborative robot-assisted endovascular catheterization with generative adversarial imitation learning,","cited_arxiv_id":null,"evidence_quote":"Provides the GAIL-PPO and Behavior Cloning baselines that SplineFormer is compared against; its Behavior Cloning result (5.6%) is the fully autonomous reference point."},{"cited_title":"Robust catheter and guidewire tracking using b-spline tube model and pixel-wise posteriors,","cited_arxiv_id":null,"evidence_quote":"Introduces the B-spline tube model for catheter/guidewire tracking, the representational basis SplineFormer adopts for shape prediction."},{"cited_title":"MultiSpeech: Multi-Speaker Text to Speech with Transformer","cited_arxiv_id":"2006.04664","evidence_quote":"Provides the template for autoregressive coefficient generation in transformers, which SplineFormer adapts to spline control points and knots."}],"review_version":1}