{"id":"2bba70ca-f9ec-46b8-be1c-a8d044802f2d","arxiv_id":"2505.09450","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MrTrack combines Mamba-based feature extraction with a register bank for temporal memory, reaching 73.9 FPS and best accuracy on aspiration biopsy needle tracking.","lead":"MrTrack is a new deep-learning tracker that follows a biopsy needle moving rapidly back and forth in ultrasound images. It uses a memory bank of past frames to keep tracking even when the needle blurs or disappears briefly, achieving state-of-the-art results.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ground-truth mapping from optical-tracker 3D coordinates to 2D ultrasound pixels is undocumented, and every accuracy claim depends on it.","rationale":"The paper's core contribution is a new temporal-context mechanism for ultrasound needle tracking, and the empirical evaluation is extensive: 239 robotic trials, 16 manual videos, seven baselines, and ablations. The architecture and loss are plausible, and the ablations isolate the proposed components. However, the entire quantitative comparison rests on a ground-truth pipeline that is described in only one sentence. The optical tracker measures 3D positions, while the tracker outputs 2D image coordinates. The missing calibration is not a minor implementation detail; it is the link that converts physical positions into the 2D ground truth used by every metric. If that link is biased or noisy, the reported margins, such as the Err improvement from 2.93 to 2.29 mm on robotic IPM, could change or disappear. The paper also does not report run-to-run variance or statistical significance, so we cannot tell whether the gaps exceed noise. The reader's weakest assumption identified exactly this gap, and the conditional verdict is appropriate. I do not see an additional load-bearing concern that would push toward rejection; the method itself is coherent and the experiments are carefully designed aside from the calibration documentation. Therefore I recommend keeping the verdict unchanged: conditional acceptance pending the calibration details and a validation of the 3D-to-2D registration.","tokens_in":8813,"tokens_out":2559,"duration_ms":26379,"concrete_test":"Publish the calibration protocol and perform a phantom validation: place a grid of US-visible fiducials (e.g., metal wires) with known 3D optical-tracker coordinates, record simultaneous US and optical data, and compute the 3D-to-2D projection including temporal synchronization. Report the reprojection RMSE in both mm and pixels. Then recompute Table 1 using only frames where the reprojection error is below a stated threshold (e.g., <0.5 mm) and check whether MrTrack still outperforms ODTrack, STMTrack, and TrackingMamba on Err and AUC. If the ranking changes or the margins shrink below significance, the headline claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 states that ground truth was collected with a ClaroNav MicronTracker 3 with an RMSE of 0.189 mm in our setup, but it never specifies how the 3D tip position is transformed into 2D ultrasound image coordinates. All reported metrics (Err, AUC, P) compare the tracker's 2D output against this 2D ground truth. If the 3D-to-2D projection is biased or noisy, the claimed improvements over the second-best method (21.8% Err reduction on robotic, 8.2% on manual) could be artifacts. The problem is especially acute during rapid reciprocating motion: at 30 mm/s, a temporal misalignment of one US frame (~33 ms) shifts the ground truth by about 1 mm, which is comparable to the reported mean errors (2.29 mm). The paper reports no calibration procedure, no reprojection error, and no temporal synchronization method. Without this, the central quantitative claim is not independently checkable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MrTrack, a Mamba-based visual tracker with learnable registers for needle-tip tracking in ultrasound-guided fine needle aspiration (FNA) biopsy. A Mamba-based register extractor condenses each historical search map into compact register tokens stored in a register bank; a Mamba-based register retriever injects these temporal prompts into the template to handle rapid reciprocating motion and transient image degradation. A self-supervised register diversify (RD) loss is introduced to prevent feature collapse. Experiments on a robotic dataset (239 trials, 54,428 frames) and a manual dataset (16 videos used only for testing) compare MrTrack against seven baselines, showing lower localization error, higher AUC/precision, and higher inference speed (73.9 FPS). Ablations examine register-bank length, register scale, transformer-based alternatives, removal of the register mechanism, and removal of the RD loss.","tokens_in":8986,"tokens_out":3594,"duration_ms":39781,"significance":"If the quantitative claims hold, MrTrack would be the first tracker explicitly designed for the rapid reciprocating needle motion of FNA biopsy, and its efficiency advantage is practically relevant for real-time ultrasound-guided intervention. The paper provides a broad comparison with seven baselines under shared training settings, an ablation study of the main components, and generalization testing on a manual dataset that was not used for training. However, the evaluation protocol has two load-bearing gaps: the optical-tracker-to-ultrasound coordinate mapping is not described, and no run-to-run or trial-level statistical variability is reported. The RD-loss equation also has an ambiguity that prevents exact reproduction. These issues are fixable and should be addressed before the central accuracy claims can be accepted.","major_comments":[{"comment":"The ground-truth mapping from optical-tracker 3D coordinates to 2D ultrasound image coordinates is not described. Section 4.1 reports only that the ClaroNav MicronTracker 3 has an RMSE of 0.189 mm in the authors' setup; it does not specify the calibration procedure, the rigid transformation between tracker and ultrasound coordinate frames, the sonographic pixel-to-millimeter scaling, or the temporal synchronization between the optical tracker and the 30 FPS ultrasound frames. All reported metrics (Err, AUC, P) compare tracker outputs against this 2D ground truth, so a biased or mis-synchronized projection directly affects every quantitative conclusion. At a needle speed of 30 mm/s, a one-frame temporal misalignment shifts the ground truth by about 1 mm, comparable to the reported mean errors. The authors should provide the calibration protocol, reprojection error, temporal synchronization method, and a sensitivity analysis of the main metrics to plausible calibration errors.","section":"Section 4.1 / Table 1"},{"comment":"Only frame-level standard deviation of the localization error is reported; no run-to-run variance, number of repeated training runs, random seeds, or trial-level confidence intervals are given. The headline improvements over the second-best method (21.8% Err reduction on the robotic dataset and 8.2% on the manual dataset) could be within training or trial variability. I ask the authors to report results over at least three training runs (or otherwise justify that the comparisons are stable), and to provide per-trial or per-video error statistics and appropriate significance tests for the manual dataset, which contains only 16 videos.","section":"Table 1 / Section 4.2"},{"comment":"The cross-register diversify term L_div is not well-defined as written. The text defines each descriptor r_i as an element of R^{k×d}, so r_{i,p} is a k-dimensional vector, and the expression [(r_{i,p} - \\bar{r})(r_{j,p} - \\bar{r})]^2 is ambiguous: it is unclear whether the product is an inner product, an outer product, or an elementwise operation, and whether the square is a scalar or a matrix norm. Since L_div is a core component of the proposed RD loss, this equation must be rewritten with the exact tensor operations and the resulting scalar form, and the implementation should be clarified to match.","section":"Equation (4), Section 3.4"}],"minor_comments":[{"comment":"There is a typo in the Related Work section: 'learning-based methods have bee proposed' should read 'have been proposed'.","section":"Section 2"},{"comment":"The name 'MrT rack' appears with an unusual space in several places (e.g., 'MrT rack, an ultrasound needle tracker'); this should be normalized to 'MrTrack' throughout.","section":"Abstract and Section 1"},{"comment":"The sentence 'The robustness of L_RD is also showed by avoiding tuning the value of hyper-parameters α and β' is not supported by any sensitivity analysis; the paper fixes α=β=0.01 but does not show that nearby values give similar performance. Either add such an analysis or rephrase the claim.","section":"Section 4.3"},{"comment":"The table caption does not define IPS/IPM; the terms are explained in Section 4.1 but should also be defined in the caption for readability.","section":"Table 1"},{"comment":"The paper does not state whether the same optical-tracker-to-ultrasound calibration was used for both the robotic and manual datasets; please clarify whether the manual dataset uses the same geometric setup and synchronization, since this affects the interpretation of the generalization results.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is plausible and the empirical scope is reasonable, but the missing ground-truth calibration and synchronization details are a serious reproducibility issue for a medical imaging paper. The authors appear honest about current limitations (no template update mechanism, small manual dataset, future clinical validation). If the calibration and statistical-variance gaps can be closed, the paper would likely be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MrTrack is worth taking seriously: it addresses a real gap—no dedicated needle tracker for FNA's rapid reciprocating motion—and the empirical work is generally careful. But the paper leaves a load-bearing hole: it never explains how the MicronTracker's 3D coordinates become 2D ultrasound annotations. Every accuracy number in Table 1 depends on that mapping, and the text only gives the optical tracker's RMSE (0.189 mm) in \"our setup.\" The stress-test's point about temporal misalignment is fair: at 30 mm/s, one US frame is about 1 mm, comparable to the reported mean errors. That doesn't make the results wrong, but it makes them uncheckable as written. A referee can ask for the calibration and synchronization protocol (and reprojection error) and expect a straightforward answer; until then, the central quantitative claim is conditional.\n\nWhat's new: MrTrack is the first tracker targeted at aspiration-biopsy reciprocating motion. The Mamba-as-register-extractor/retriever design is an incremental combination of existing ideas—Mamba, ViT registers, prior Mamba trackers including the authors' own Mambaxctrack—but the application and the RD loss are new enough. The RD loss (variance plus cross-register decorrelation) is a reasonable regularizer and the ablations show it matters. The experiments are more thorough than usual for this area: seven baselines, both robotic and manual acquisition, manual data held out for generalization, and ablations that test bank length, register scale, transformer-vs-Mamba, and removal of the loss.\n\nSoft spots beyond calibration: no run-to-run variance, only frame-level SD; datasets are not public; and the gains over the second-best method are not huge (21.8% on robotic Err, 8.2% on manual), so the significance depends on the calibration being clean. The novelty is modest—it's a strong engineering paper, not a conceptual shift. The self-citations are not a problem; the cited works are relevant and the method stands on its own.\n\nFor a reader in interventional ultrasound or medical tracking, this is useful. For a general vision audience, it's a competent application paper. I'd send it to review—conditional acceptance if the authors disclose the optical-to-US calibration, temporal sync, and annotation pipeline, and ideally host the data or at least a validation subset. It deserves referee time.","headline":"Solid empirical tracker for a real clinical gap; the missing optical-to-ultrasound calibration protocol is a load-bearing gap in the paper's central claim.","tokens_in":9508,"tokens_out":2767,"would_cite":false,"duration_ms":29666,"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":"MrTrack is proposed as the first ultrasound needle tracker for rapid reciprocating aspiration motion: a Mamba-based register extractor stores compact per-frame contexts, and a register retriever folds them into the template when current…","keywords":["ultrasound-guided fine needle aspiration","needle tip tracking","Mamba","state space models","temporal context","register tokens","visual object tracking","register diversify loss"],"falsifier":"Measure the reprojection error of a known 3D point from the optical tracker into the ultrasound image using a calibrated probe geometry; if this calibration error is comparable to or larger than MrTrack's reported mean errors (2.29 mm robotic, 2.80 mm manual), the claimed accuracy advantage would not be distinguishable from registration noise.","tokens_in":8635,"feed_emoji":"🎯","tokens_out":7075,"duration_ms":66626,"temperature":0.7,"pith_summary":"Ultrasound-guided fine-needle aspiration deliberately jabs the needle back and forth, which makes its tip blur or vanish between frames, and no existing needle tracker was designed for that motion. The paper proposes MrTrack, a tracker built around a Mamba-based register mechanism: each historical search frame is condensed into a compact register token, stored in a bank of the last 300 frames, and later retrieved to supplement the template when the current image is degraded. A self-supervised register diversify loss keeps these stored tokens from collapsing into identical features. On robotic and manual aspiration datasets, the paper reports that MrTrack is more accurate and more robust than seven comparison trackers while running at 73.9 frames per second, and that it generalizes to manual insertions it was never trained on.","feed_headline":"Needle tracker keeps up with rapid biopsy jabs at 73.9 FPS","feed_subtitle":"It stores 300 frames of needle history and retrieves them when ultrasound images fail.","key_machinery":"The central mechanism is a learnable register token paired with two Mamba (state-space sequence model) modules. In the extractor, the register is interleaved evenly behind each image segment of the search map before the selective-scan SSM, which condenses that frame's global appearance into $k=8$ compact descriptors; these are stored in the register bank, keeping the last $L=300$ frames. In the retriever, the historical descriptors are inserted before each segment of the template map and processed by another SSM, so temporal cues from the bank are melted into a dynamic template. The register diversify loss is the regularization that prevents collapse: it applies a softplus penalty to dimensions whose variance falls below $\\tau=1$ and pushes cross-register covariances toward zero across different frames, encouraging dimension independence within the register bank.","core_discovery":"On its own terms, the paper establishes that temporal context in needle tracking can be stored frame-by-frame as compact registers and retrieved only when needed, instead of being propagated through the network as a running token. The Mamba-based register extractor inserts a learnable register of $k=8$ tokens into the sequence of image tokens before a selective-scan state-space model, so each frame's global context is distilled into descriptors that are appended to a register bank of length $L=300$. The Mamba-based register retriever performs the reverse insertion, placing historical descriptors before segments of the template sequence, producing a dynamic template that combines the initial template with cues from the last 300 frames. The register diversify loss combines a variance-promotion term and a cross-register covariance decorrelation term, both scaled by $0.01$, to keep the registers informative. The empirical claim is that this design outperforms the comparison trackers on AUC, precision, mean error and error standard deviation for robotic and manual aspiration biopsy, with the largest relative gains in mean error and standard deviation.","pith_inferences":["Because the mechanism only depends on a history of search-image features, the same extractor-retriever design could be carried over to other interventional ultrasound tasks where a target periodically disappears, such as catheter or guidewire tracking; the paper does not test this.","The reported millimeter-level accuracies rest on an unverified projection of the optical tracker's 3D ground truth into ultrasound image coordinates, so the fairest next experiment is a full calibration-error analysis rather than further tracking-metric comparisons.","The ablation with an unbounded register bank ($L=\\infty$) slightly improves robotic accuracy but hurts manual accuracy and drops speed by 15.9 FPS, which suggests that a learned forgetting or importance-weighting mechanism over the register bank could beat the fixed $L=300$ window.","For clinical translation, training on manual and real clinical video may shift the optimal bank length and register size, since manual insertions contain more abrupt imaging disturbances than the robotic dataset."],"forward_implications":["If the reported results hold, MrTrack would be the first needle tracker capable of following the rapid reciprocating motion of fine-needle aspiration, a motion regime that breaks trackers designed for stable insertions.","The per-frame register storage is presented as a cure for error accumulation in temporal token propagation: replacing the Mamba-based extractor and retriever with a transformer variant raises robotic mean error by $0.81$ mm, and removing the register entirely raises it by $0.20$ mm.","The tracker runs at 73.9 FPS on the reported hardware with a $384\\times384$ search map and $192\\times192$ template, fast enough for real-time use in high-frame-rate ultrasound.","The register diversify loss contributes a measurable gain: without it, robotic mean AUC drops 1.3 points and manual AUC drops 2.6 points, supporting the claim that decorrelated registers carry more useful temporal information.","The method generalizes without training to manual insertions with two to three needle oscillations per second, since the manual dataset was used only for testing."],"supporting_citations":[{"why":"Supplies the Mamba selective-scan state-space model used by both extractor and retriever for efficient temporal modeling.","marker":"[9]"},{"why":"Introduces the learnable register token as a compact global representation, the core idea MrTrack adapts to temporal needle context.","marker":"[4]"},{"why":"Shows that Mamba benefits from registers and suppresses artifacts, motivating the cross-map scanning insertion used in MrTrack.","marker":"[25]"},{"why":"Provides the space-time memory temporal baseline that MrTrack compares against and improves on with compressed registers.","marker":"[7]"},{"why":"Provides the online dense temporal token learning baseline whose token-propagation error accumulation motivates per-frame register storage.","marker":"[36]"},{"why":"Provides the vanilla Mamba tracking baseline, demonstrating that Mamba without explicit temporal context modeling performs worse.","marker":"[26]"},{"why":"Supplies the variance-covariance regularization idea that the register diversify loss adapts to decorrelate registers across frames.","marker":"[1]"},{"why":"Provides a motion-prediction needle tip tracker, a prior motion-integration approach that MrTrack's nonlinear motion setting challenges.","marker":"[29]"}],"fun_headline_variants":["Register bank rescues needle tracking from rapid biopsy jabs","Needle tracker recalls 300 frames to stay steady during fast jabs","Mamba-based tracker retrieves past cues when ultrasound images fail","Needle tracker banks temporal cues to beat rapid biopsy jabs","Ultrasound needle tracker uses register bank to outlast fast jabs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the optical tracker's 3D ground-truth needle tip positions are accurately registered into the 2D ultrasound image coordinates, and the paper does not describe or verify that registration; if the mapping is wrong, all reported millimeter-level errors inherit that error.","fun_headline_variants_meta":{"raw":{"variants":["Register bank rescues needle tracking from rapid biopsy jabs","Needle tracker recalls 300 frames to stay steady during fast jabs","Mamba-based tracker retrieves past cues when ultrasound images fail","Needle tracker banks temporal cues to beat rapid biopsy jabs","Ultrasound needle tracker uses register bank to outlast fast jabs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000672,"raw_usage":{"total_tokens":3055,"prompt_tokens":936,"completion_tokens":2119,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2030}},"tokens_in":552,"tokens_out":2119,"duration_ms":16031,"temperature":1.0,"reasoning_tokens":2030,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:30:14.109886+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the reprojection error of a known 3D point from the optical tracker into the ultrasound image using a calibrated probe geometry; if this calibration error is comparable to or larger than MrTrack's reported mean errors (2.29 mm robotic, 2.80 mm manual), the claimed accuracy advantage would not be distinguishable from registration noise.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Provides the space-time memory temporal baseline that MrTrack compares against and improves on with compressed registers."},{"cited_title":"In:ProceedingsoftheAAAIConferenceonArtificialIntelligence.vol.38,pp.7588– 7596 (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the online dense temporal token learning baseline whose token-propagation error accumulation motivates per-frame register storage."},{"cited_title":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sens- ing (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the vanilla Mamba tracking baseline, demonstrating that Mamba without explicit temporal context modeling performs worse."},{"cited_title":"Medical Image Analysis88, 102847 (2023)","cited_arxiv_id":null,"evidence_quote":"Provides a motion-prediction needle tip tracker, a prior motion-integration approach that MrTrack's nonlinear motion setting challenges."}],"review_version":1}