{"id":"53bdf495-bd5f-4463-b3cb-f4281923cdd8","arxiv_id":"2411.08395","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MambaXCTrack introduces SSM-based cross-correlation with interleaved scanning and an implicit motion descriptor, and reports state-of-the-art ultrasound needle tracking accuracy on its own phantom and tissue dataset.","lead":"This paper presents MambaXCTrack, a Mamba-based tracker for ultrasound needle tip tracking that combines a state-space cross-correlation module with a motion prompt. It reports lower tracking error than eight prior trackers on a private phantom and tissue dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The test-time source of the motion descriptor is unspecified; if the motion queue uses ground-truth boxes during evaluation, the reported SOTA gains and the motion ablation are inflated.","rationale":"The paper's central claim is that MambaXCTrack outperforms eight prior trackers on a private ultrasound needle dataset and that each proposed component contributes, as shown by the ablation. The single most load-bearing assumption is the source of the bounding boxes used to build the motion descriptor at test time. Section II-D defines m_t from B = {beta_1, ..., beta_t} but never states whether B is the tracker's own output history or ground-truth annotations. This assumption is structurally independent of the architecture: the Mamba scan, SSMX-Corr, and CIS could all work as described, yet if ground-truth boxes are fed into the motion queue during evaluation, the comparison against trackers without such leakage is unfair and the v4 ablation overattributes gains to the motion descriptor. The paper's ablation pattern is internally consistent, but that consistency does not resolve the leakage question because all ablations share the same motion-queue mechanism. The EM ground-truth RMSE of 0.76 mm versus the reported phantom error of 0.22 +/- 0.14 mm is a related red flag; if the reported error is computed against EM-derived labels, a tracker cannot ordinarily beat the label noise floor unless it has access to those labels or the error definition differs. The SSM convolution kernel derivation in Eq. (5) is not strictly valid for selective SSMs, but that is a theoretical framing issue rather than the decisive correctness risk. The absence of Mamba-based baselines weakens the novelty claim but does not threaten the internal validity of the reported comparison. Because the decisive issue is an unspecified but checkable protocol, the appropriate outcome is the same conditional verdict the reader reached: the paper should be accepted only if code or an explicit inference-time protocol confirms that the motion queue uses predicted boxes only. No change to the reader's verdict is needed, but the concern should be stated as a required condition for acceptance.","tokens_in":14399,"tokens_out":5223,"duration_ms":58202,"concrete_test":"Inspect the released inference code and trace the motion queue construction: at frame t, verify that the queue is updated only with the tracker's own output box from frame t-1, never with the EM-derived ground-truth box. Then rerun the Table II evaluation under two conditions: (i) motion queue populated from ground-truth boxes, and (ii) motion queue populated from previous tracker predictions only. If the gap between (i) and (ii) exceeds a small threshold (e.g., 5% AUC or 0.1 mm mean error), the reported motion-prompt advantage is partly an artifact of label leakage. As a secondary check, recompute the phantom error after accounting for the stated 0.76 mm EM calibration RMSE to confirm that 0.22 mm is not below the label-noise floor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II-D defines the motion descriptor from historical bounding boxes B = {beta_1, ..., beta_t} and computes m_t from beta_t - beta_{t-1}, but it never states whether, at inference, B contains the tracker's own previous predictions or ground-truth boxes. The phrase 'historical bounding boxes' and the notation beta_t (current frame) read naturally as labels, and the training-time augmentation of the motion sequence does not clarify the test-time queue. If ground-truth boxes populate the motion queue at test time, the tracker receives the current target displacement as an auxiliary input. That is a label leak: none of the eight compared trackers receive such a prompt, so the headline comparison would be inflated, and the v4 ablation that credits the motion descriptor would no longer measure what it claims. The phantom error of 0.22 +/- 0.14 mm also sits below the stated 0.76 mm EM ground-truth RMSE; the paper does not reconcile this discrepancy, which is at least consistent with some form of label leakage or a different error definition. The central SOTA claim is therefore not falsifiable from the text alone until the inference-time motion queue protocol is specified and verified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents MambaXCTrack, a tracking architecture for ultrasound needle tip localization built on Mamba/SSM blocks. The key components are SSMX-Corr, which replaces convolutional cross-correlation with an SSM-based operation intended to perform global long-range matching between template and search features; CIS, a four-directional interleaved scan that keeps template and search pixels locally adjacent; and an implicit motion descriptor computed from historical bounding-box displacements and injected as a prompt. The authors collect a motorized-insertion dataset (57 phantom and 51 tissue videos) with EM ground truth, retrain eight existing trackers under a common protocol, and report that MambaXCTrack achieves the best mean AUC, precision, normalized precision, and millimeter error in phantom, tissue, and pooled results, with ablations v1-v7 supporting each component.","tokens_in":14515,"tokens_out":7990,"duration_ms":82370,"significance":"The contribution is potentially significant for the medical robotics and ultrasound tracking community: it is the first Mamba-based needle tracker, the proposed SSMX-Corr and CIS are architecturally novel, and the dataset with video-level splits is a useful test bed. The paper also provides a systematic ablation and compares against eight trackers, including recent transformer and motion-prompt baselines. If the evaluation protocol is clean, the reported margins over MixFormerV2 and SwinTrack would be a credible advance. However, the central empirical claim currently rests on an underspecified test-time motion queue, and the theoretical description of SSM cross-correlation contains a formal imprecision; these points must be resolved before the results can be taken at face value. No code is provided, and no repeated-seed statistics are reported, so independent confirmation is not yet possible.","major_comments":[{"comment":"The paper does not state whether the historical bounding boxes B used to form m_t are the tracker's own previous outputs or ground-truth boxes at inference. Training uses ground-truth boxes with added Gaussian noise (Section III-A), but the evaluation protocol is not described. If ground-truth boxes populate the motion queue at test time, the tracker receives current-frame displacement information unavailable to the eight baselines, inflating both the headline gains and the v4 ablation. Please specify the inference-time protocol; if ground-truth boxes are used, rerun the experiments with self-generated boxes and re-report Table II and the v4 row of Table III.","section":"Section II-D and Section III-A"},{"comment":"The phantom tracking error 0.22 +/- 0.14 mm and tissue error 0.49 +/- 1.01 mm are lower than the stated EM ground-truth RMSE of 0.76 mm. This is not impossible if the EM error is predominantly a constant bias, but the manuscript does not explain the discrepancy. The authors should report how the EM RMSE was measured, whether it is systematic or random, and how the millimeter error was computed relative to the EM time series; otherwise the sub-millimeter claims cannot be interpreted.","section":"Section III-A and Table II"},{"comment":"Eq. (5) uses the convolution equivalence y = K_bar * u for a selective SSM in which A_bar, B_bar, C, and Delta are functions of the input. For a time-varying or selective system, the unrolled recurrence is not a convolution with a fixed kernel, and K_bar is not constant across the input. The sentence 'Note that K_bar comes from the parameterization of input...' does not restore the equality. Please replace the convolution claim with the actual selective-scan recurrence, or explicitly label Eq. (5) as an LTI idealization used only for motivation.","section":"Section II-C, Eq. (5)"},{"comment":"All comparisons and ablations are reported as single point estimates with no repeated-initialization variance, confidence intervals, or significance tests. Some of the key differences are small, such as the v1 versus baseline comparison in Table III and several tissue-environment entries in Table I. Please report mean +/- standard deviation over at least three seeds, or a paired significance test, for the main comparisons and for the v4 ablation that isolates the motion descriptor.","section":"Tables I-III"}],"minor_comments":[{"comment":"The definition beta_t = (w_t, h_t, cx_t, cy_t) calls (cx_t, cy_t) the 'top-left corner' but uses those coordinates for displacement as if they represent a consistent reference point; please clarify the coordinate convention of the bounding-box tuple.","section":"Section II-D"},{"comment":"The path of the motion descriptor m into each Mamba head is described in the text but is not clearly visible in the overview figure; adding an explicit annotation or legend for m would improve readability.","section":"Figure 1 and Figure 2"},{"comment":"The caption says the best and second-best results are noted in red and cyan, but these colors are not visible in monochrome print; please use bold or superscript markers as well.","section":"Table I"},{"comment":"The authors motivate SSMX-Corr by comparing with segmentation-based needle trackers [17], [18] but do not include any segmentation-based method in the experiments; a short sentence explaining why these methods are excluded from the comparison would help the reader judge the baseline coverage.","section":"Section III-B"}],"recommendation":"major_revision","confidential_remarks":"The load-bearing issue is the test-time motion protocol. If the authors confirm that the motion queue uses ground-truth boxes at inference, the SOTA comparison and the motion ablation (v4) would be invalid; if they use their own predictions, the paper is much closer to acceptance. I would ask the authors to state the protocol clearly and, if necessary, rerun the experiments. The paper also needs a correction to Eq. (5) and to the interpretation of the sub-millimeter errors relative to the EM ground-truth accuracy. With those changes, the manuscript could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a plausible Mamba-based tracker for US needle tracking with a clever cross-correlation scheme, but the headline numbers cannot be assessed from the text because the test-time source of the motion descriptor is never specified. If the motion queue is populated with ground-truth boxes, the reported gains and the motion ablation are inflated.\n\nWhat's actually new: first Mamba application in US needle tracking; SSMX-Corr with CIS is a reasonable way to combine long-range SSM modeling with local pixel interaction; the implicit low-level motion descriptor is a sensible alternative to explicit motion prediction. The experiments are extensive: phantom and tissue, multiple insertion cases, eight baselines retrained on the same data, and a seven-way ablation. The architecture details are described in enough detail to reimplement.\n\nThe soft spots are serious. Section II-D defines m_t from historical bounding boxes B = {β_1,...,β_t} but never says whether B is the tracker's own output or ground truth. During training, Gaussian noise is added to the motion sequence, which suggests ground-truth boxes are used; at test time, using ground truth would leak the current displacement into the model. None of the compared trackers receive such a prompt, so the comparison would be unfair. The reported phantom error of 0.22 ± 0.14 mm is also below the stated EM ground-truth RMSE of 0.76 mm, which is hard to reconcile unless the error metric is different or the tracker is using label information. The paper needs to state the inference-time motion protocol clearly and, ideally, run an ablation with a motion queue built from predicted boxes.\n\nThere are smaller weaknesses: no comparison with other Mamba-based trackers (e.g., TrackingMamba), no repeated-seed variance or statistical significance tests, and Eq. (5) presents a convolution equivalence for selective SSMs that is only valid for LTI systems (though the text does note the kernel is input-dependent). Code and data are withheld.\n\nWho this is for: people working in ultrasound-guided intervention and visual tracking. The method is interesting and the problem matters, but the empirical claim needs verification. I'd send it to peer review with a request for the missing details; the architecture is worth serious examination. My own verdict is skeptical until the motion leak issue is resolved.\n\nRecommendation: accept for peer review, not desk reject.","headline":"A plausible Mamba-based US needle tracker with strong reported numbers, but the unsourced test-time motion descriptor makes the central claim unverifiable and possibly inflated.","tokens_in":15148,"tokens_out":3545,"would_cite":false,"duration_ms":34555,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MambaXCTrack claims that a Mamba-based tracker with SSM cross-correlation and an implicit motion prompt tracks ultrasound needle tips more accurately than eight prior CNN and transformer trackers, with a mean error of 0.34 mm across…","keywords":["ultrasound-guided needle tracking","Mamba","structured state space models","SSM cross-correlation","motion prompt","visual tracking","medical robotics","computer vision"],"falsifier":"Recompute the v4-versus-baseline ablation under two motion-queue conditions during testing: one fed with ground-truth bounding boxes and one fed with the tracker's own predicted boxes. If the AUC gap between v4 and the baseline narrows to near zero when using self-generated boxes, the motion-prompt gain is explained by label leakage rather than by robust tracking.","tokens_in":14085,"feed_emoji":"🩺","tokens_out":5793,"duration_ms":52604,"temperature":0.7,"pith_summary":"This paper tries to show that a Mamba-based tracker, MambaXCTrack, can locate the tip of an ultrasound-guided needle more accurately than existing CNN and transformer trackers, despite noise, artifacts, and intermittent tip visibility. It replaces the convolutional cross-correlation used in Siamese trackers with an SSM-based cross-correlation (SSMX-Corr) that searches whole maps for distant semantic cues, and it feeds a history of simple tip displacements into the network as an implicit motion prompt. On a dataset of 108 motorized insertions in phantom and tissue, the tracker reports a mean error of 0.34 ± 0.55 mm averaged across environments, 0.22 ± 0.14 mm in phantom and 0.49 ± 1.01 mm in tissue, outperforming eight prior trackers on the reported average AUC, precision, and normalized precision. The ablation study attributes the gains to SSMX-Corr, the cross-map interleaved scan, and the motion descriptor. If the results hold, Mamba-style state space models are a viable backbone for real-time medical ultrasound tracking where local visual features are unreliable.","feed_headline":"Mamba tracker pins ultrasound needle tip to 0.34 mm","feed_subtitle":"State-space-model cross-correlation plus a motion prompt beats eight prior trackers on phantom and tissue ultrasound data.","key_machinery":"The load-bearing object is the SSMX-Corr operation: SSM cross-correlation, which rewrites the standard convolutional cross-correlation $z * x$ as a convolution $y = \\bar{K} * u$ between the interleaved scan of template, search submap, and motion descriptor and the input-dependent SSM kernel $\\bar{K} = (C\\bar{A}^j\\bar{B})_{j \\in [L]}$. This formulation gives global search over the whole map while retaining the convolution-like form of X-Corr. Two supporting mechanisms feed it: the cross-map interleaved scan (CIS), which alternates template and search pixels along four directions so local pixel-wise interaction and positional inductive bias survive SSM scanning, and the implicit low-level motion descriptor $m_t = (\\Delta cx_t, \\Delta cy_t)$, a queue of raw tip displacements that acts as a non-visual prompt during invisibility.","core_discovery":"On its own terms, the paper's central claim is that cross-correlation for visual tracking can be implemented with structured state space models instead of convolution, and that this combined with an implicit motion prompt solves the ultrasound needle tracking problem. SSMX-Corr treats the interleaved template, search, and motion sequence as input to a selective SSM, whose closed form is a convolution with an input-dependent kernel; because the kernel is sequence-wide, the operation searches globally rather than in a local window. The motion descriptor is just the pixel displacement between consecutive historical bounding boxes, stored in a queue and concatenated into the scan. The paper reports that this design yields a mean tracking error below the needle radius in most frames, and that each component contributes in ablation.","pith_inferences":["Beyond the paper: the evaluation protocol does not state whether the historical bounding boxes feeding the motion queue are the tracker's own previous outputs or ground-truth boxes; if ground truth is used at test time, the motion prompt can leak future tip positions and the ablation gain would not transfer to real deployment.","Beyond the paper: the EM ground truth has a reported RMSE of 0.76 mm, yet the phantom tracking error is 0.22 mm, so the sub-0.76 mm accuracy figures should be interpreted with that measurement floor in mind.","Beyond the paper: SSMX-Corr is architecture-agnostic enough that a direct test would replace ConvX-Corr in an existing Siamese tracker with SSMX-Corr and compare on natural-image benchmarks to see whether the global-search benefit generalizes beyond ultrasound."],"forward_implications":["If the reported numbers are right, SSMX-Corr gives an end-to-end tracker with a long-range search range that convolutional X-Corr lacks, so similar state space cross-correlation could improve other tracking tasks with degraded local features.","An implicit low-level motion prompt can keep a target tracked when it disappears for up to the queue duration (T = 60 frames, about 2 seconds), without an explicit motion predictor.","The method runs at 34.9 FPS on a single GPU, so it meets real-time clinical rates, but only with a high-end GPU; the authors state this limits deployment on low-end hardware.","The ablation in Table III reports that replacing SSMX-Corr with ConvX-Corr lowers mean AUC by 4.7 points, removing CIS lowers it by 1.1 points, and removing the motion descriptor lowers it by 5.1 points, supporting each component's contribution."],"supporting_citations":[{"why":"Supplies the selective-scan SSM that SSMX-Corr builds on.","marker":"[24]"},{"why":"Supplies the convolutional cross-correlation that SSMX-Corr replaces and the SiamRPN++ baseline.","marker":"[10]"},{"why":"The strongest transformer baseline and the closest second-best comparison in the tissue experiments.","marker":"[11]"},{"why":"Prior ultrasound needle tracker using explicit motion prediction that the implicit motion descriptor is designed to outperform.","marker":"[6]"},{"why":"Baseline that stores historical visual templates, contrasting with the paper's motion-based memory approach.","marker":"[33]"},{"why":"Transformer tracker that fuses raw motion, serving as a contrast for the low-level motion descriptor.","marker":"[34]"},{"why":"Shows that SSM output equals convolution with the SSM kernel, the identity at the core of SSMX-Corr.","marker":"[32]"},{"why":"Defines the AUC success-plot metric used throughout the evaluation.","marker":"[35]"},{"why":"Defines the precision and normalized precision metrics used for center-location-error evaluation.","marker":"[36]"}],"fun_headline_variants":["First Mamba tracker for ultrasound needles hits 0.34 mm error","Mamba SSM cross-correlation out-tracks eight prior needle trackers","Motion prompt and Mamba beat noise in needle tip tracking","Ultrasound needle tracking with Mamba achieves sub-millimeter accuracy","SSM cross-correlation upgrades needle tracking to global search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"During evaluation, the historical bounding boxes that feed the motion descriptor are assumed to be the tracker's own previous outputs; if instead ground-truth boxes are used, the motion prompt could leak future label information and the reported ablation gain would be inflated.","fun_headline_variants_meta":{"raw":{"variants":["First Mamba tracker for ultrasound needles hits 0.34 mm error","Mamba SSM cross-correlation out-tracks eight prior needle trackers","Motion prompt and Mamba beat noise in needle tip tracking","Ultrasound needle tracking with Mamba achieves sub-millimeter accuracy","SSM cross-correlation upgrades needle tracking to global search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1282,"prompt_tokens":923,"completion_tokens":359,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":268}},"tokens_in":539,"tokens_out":359,"duration_ms":4092,"temperature":1.0,"reasoning_tokens":268,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:38:16.026418+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the v4-versus-baseline ablation under two motion-queue conditions during testing: one fed with ground-truth bounding boxes and one fed with the tracker's own predicted boxes. If the AUC gap between v4 and the baseline narrows to near zero when using self-generated boxes, the motion-prompt gain is explained by label leakage rather than by robust tracking.","supporting_citations":[{"cited_title":"Siamrpn++: Evolution of siamese visual tracking with very deep networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the convolutional cross-correlation that SSMX-Corr replaces and the SiamRPN++ baseline."},{"cited_title":"Mixformerv2: Efficient fully transformer tracking,","cited_arxiv_id":null,"evidence_quote":"The strongest transformer baseline and the closest second-best comparison in the tissue experiments."},{"cited_title":"Learning-based needle tip tracking in 2d ultrasound by fusing visual tracking and motion prediction,","cited_arxiv_id":null,"evidence_quote":"Prior ultrasound needle tracker using explicit motion prediction that the implicit motion descriptor is designed to outperform."},{"cited_title":"Stmtrack: Template-free visual tracking with space-time memory networks,","cited_arxiv_id":null,"evidence_quote":"Baseline that stores historical visual templates, contrasting with the paper's motion-based memory approach."},{"cited_title":"Swintrack: A simple and strong baseline for transformer tracking,","cited_arxiv_id":null,"evidence_quote":"Transformer tracker that fuses raw motion, serving as a contrast for the low-level motion descriptor."},{"cited_title":"Combining recurrent, convolutional, and continuous-time models with linear state space layers,","cited_arxiv_id":null,"evidence_quote":"Shows that SSM output equals convolution with the SSM kernel, the identity at the core of SSMX-Corr."},{"cited_title":"Online object tracking: A benchmark,","cited_arxiv_id":null,"evidence_quote":"Defines the AUC success-plot metric used throughout the evaluation."},{"cited_title":"Trackingnet: A large-scale dataset and benchmark for object tracking in the wild,","cited_arxiv_id":null,"evidence_quote":"Defines the precision and normalized precision metrics used for center-location-error evaluation."}],"review_version":1}