{"id":"fe3c1453-aba6-4a89-9884-efeb73f98611","arxiv_id":"2507.14368","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DUSTrack integrates per-frame deep learning point detection with LK-RSTC optical flow tracklet filtering to achieve accurate, low-jitter tracking of arbitrary points in B-mode ultrasound videos.","lead":"DUSTrack is a semi-automated toolkit that combines deep learning and optical flow to track arbitrary points in ultrasound videos with low jitter. It could give clinicians and biomechanics researchers a general tool for measuring tissue motion without building custom software.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground truth for the main accuracy/jitter comparisons is itself generated with LK-RSTC, the same algorithm used in the filter, so the reported 'superior accuracy' and PSD match may be partly self-fulfilling.","rationale":"The reader's weakest assumption points to the same load-bearing issue: the ground truth is partly produced by the very optical-flow algorithm being evaluated. This is the single most important threat to the paper's central claims, because it affects the main accuracy comparison (Fig. 3a), the jitter/frequency analysis (Fig. 4e-f), and the label-augmentation experiment (Fig. 5). I also considered the fascicle-tracking comparison (n=4, two of three comparators excluded, Fig. 8) and the framing of fine-tuned versus zero-shot models as 'superior accuracy.' These are real weaknesses, but they are secondary: the circular ground-truth issue undermines the quantitative foundation of the paper's headline claims. The proposed test—restricting evaluation to independently manually annotated frames—is feasible with the existing public code and would either confirm or resolve the concern. Therefore the conditional-accept verdict remains appropriate, with the request for an independent ground-truth subset or sensitivity analysis.","tokens_in":16111,"tokens_out":4552,"duration_ms":49889,"concrete_test":"Recompute the Fig. 3a RMSE and Fig. 4e PSD comparisons using only the manually annotated frames (the ~20 per video) as ground truth, excluding all LK-RSTC-interpolated frames. If the fine-tuned ResNet-50 advantage over CoTracker3 and the PSD overlap between filtered output and ground truth are not substantially reproduced on this independent subset, the reported accuracy and jitter improvements are artifacts of the shared LK-RSTC reference. A complementary check: run the same evaluation on a synthetic ultrasound phantom with known ground-truth deformation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claims—ResNet-50 fine-tuned on 25 frames significantly outperforming zero-shot TAP models (Fig. 3a) and the LK-RSTC filter matching ground-truth frequency content (Fig. 4e-f)—are both measured against a reference described as 'DUSTrack-assisted human annotations' (Results; Methods). The Methods state that only ~20 of ~600 frames are manually labeled and the rest are interpolated with the LK-RSTC algorithm, which is the same algorithm used in the jitter-reduction filter. If the manual corrections on the interpolated frames were sparse, the ground-truth trajectory is largely an LK-RSTC output. Then (i) comparing the LK-RSTC filter's PSD to this ground truth is circular, since both share the same smoothing/weighting characteristics, and (ii) zero-shot models are penalized for not matching LK-RSTC's specific noise profile, inflating the reported gap between fine-tuned and zero-shot accuracy. The paper does not report how many interpolated frames were manually corrected, nor any independent validation (e.g., sonomicrometry or synthetic phantoms). This concern is load-bearing because the abstract's 'superior accuracy' and 'reduces high-frequency frame-to-frame noise while preserving rapid tissue motion' rest on this reference.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DUSTrack, a semi-automated point-tracking toolkit for B-mode ultrasound videos. The workflow combines DeepLabCut-based deep learning models (e.g., ResNet-50) for drift-free point detection with an LK-RSTC optical-flow filter for jitter reduction. The authors claim that a ResNet-50 fine-tuned on about 25 manually annotated frames significantly outperforms zero-shot TAP models (CoTracker3, BootsTAP, LocoTrack, PIPs++) in tracking accuracy, and that the LK-RSTC filter reduces high-frequency noise while preserving rapid tissue motion. The toolkit is demonstrated on cardiac wall motion tracking, upper-arm muscle deformation analysis, and medial gastrocnemius fascicle tracking, with the fascicle comparison performed against an independent public dataset.","tokens_in":16265,"tokens_out":3699,"duration_ms":40053,"significance":"If the central claims hold, DUSTrack would provide a valuable open-source, general-purpose point-tracking framework for ultrasound, with a GUI that supports annotation, augmentation, and iterative model refinement. The fascicle-tracking comparison against an independent public dataset (ground truth from multiple expert raters) is a clear strength, as is the release of the toolkit and code. However, the main accuracy and jitter claims rest on a ground-truth construction that is partly generated by the same LK-RSTC algorithm used in the filter, which, if not resolved, would substantially weaken the reported performance gaps and the filter's evaluation.","major_comments":[{"comment":"The ground truth for the main accuracy and jitter comparisons is generated with the DUSTrack workflow itself: the Methods state that only approximately 20 of 600 frames are manually annotated and the rest are interpolated with the LK-RSTC algorithm, and the Results describe the ground truth as 'DUSTrack-assisted human annotations ... refined over several iterations and manually corrected.' The same LK-RSTC algorithm is then used as the jitter-reduction filter whose performance is evaluated against this ground truth (Figure 4e-f). This creates a self-referential evaluation: the filtered output is compared to a reference that is itself an LK-RSTC product, so the strong PSD overlap in Figure 4e (the black ground-truth trace is described as 'not visible due to significant overlap with the green trace') is expected and does not by itself validate the filter. Furthermore, the comparison between the fine-tuned ResNet-50 and zero-shot TAP models in Figure 3a may be biased, because zero-shot models are evaluated against a reference that has LK-RSTC-like smoothing characteristics. The manuscript does not report how many of the interpolated frames were manually corrected, nor does it provide any independent validation (e.g., sonomicrometry, synthetic phantoms, or a separate manual-only ground truth). This is load-bearing for the central claims of 'superior accuracy' and 'reduces high-frequency frame-to-frame noise while preserving rapid tissue motion.'","section":"Methods (annotation step) and Results (first paragraph)"},{"comment":"The claim that the LK-RSTC filter 'reduces jitter' is supported only by qualitative traces and power spectral density plots (Figure 4e-i); no quantitative RMSE or other numerical error metric is reported comparing the filtered output to the unfiltered model output. Given that the filter is a central contribution, the absence of a quantitative before/after accuracy measure is a significant gap. Reporting, for example, RMSE of filtered versus unfiltered outputs against the ground truth would allow readers to assess the actual improvement.","section":"Results, Figure 4 (LK-RSTC filtering)"},{"comment":"The fascicle tracking comparison against UltraTimTrack is based on a very small sample: the paired t-tests are reported with t3 (d.f. = 3), implying only four participants are included, despite the Methods describing five participants in the public dataset. The claim of 'comparable accuracy' is therefore weakly supported. The paper should clarify the actual number of videos used and consider reporting effect sizes or confidence intervals, or acknowledge the limited statistical power of this comparison.","section":"Results, Figure 8 (fascicle tracking comparison)"},{"comment":"The 'transposed sliding window filter' is described only in prose, without a precise mathematical formulation, pseudocode, or explicit parameter settings (window length, stride, weighting function, etc.). The sliding-window length is listed as a free parameter, but the paper does not report how it was chosen or whether results are sensitive to it. For a methods paper, this hinders reproducibility and makes it difficult for readers to re-implement the filter or assess its robustness.","section":"Methods, LK-RSTC filtering description"}],"minor_comments":[{"comment":"There are several typographical errors: 'speckle-trackig' in the Discussion, a duplicated 'non-invasive' in the Introduction, 'a tracket' in the Figure 4a caption, and 'the a ResNet-50' in the Figure 5a caption. These should be corrected.","section":"General"},{"comment":"The binomial test result is reported in the text as 'Fig. 5e', but Figure 5e shows the power spectral density analysis, while the preference results appear to be in Figure 5f. Please fix the cross-reference.","section":"Results, Figure 5"},{"comment":"The Methods state that a Bonferroni correction was applied for multiple testing in the pairwise comparisons between the fine-tuned ResNet-50 and the zero-shot models, but the p-values reported in Figure 3a appear to be uncorrected (e.g., p = 2.71e-16). The paper should clarify whether the reported p-values are adjusted or unadjusted, and if unadjusted, note that all remain significant after correction.","section":"Statistical Analysis and Figure 3a"},{"comment":"The statement that the fine-tuned ResNet-50 accuracy 'at approximately 200 µm (Fig. 3a)' approaches the perceptual threshold of approximately 100 µm is not directly visible from the figure; please indicate where in the figure these values are shown or provide them in the text.","section":"Discussion"},{"comment":"The description of the 'transposed sliding window' in the filtering step says that estimates are averaged 'across the same frame,' which is confusing because the window is temporal. Clarify whether the averaging is across overlapping window positions at the same time index or something else.","section":"Methods, GUI description"}],"recommendation":"major_revision","confidential_remarks":"The circular-ground-truth issue is the most serious concern and should be addressed head-on. If the authors can show that the results are robust when only manually annotated frames are used as ground truth, or provide an independent validation, the paper would be much stronger. The fascicle comparison, though small, provides some external validation, so the paper is not entirely without merit. I recommend major revision rather than rejection, as the issue is fixable in principle, but the current version does not support the abstract's strongest claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is the transposed sliding-window tracklet filter: instead of smoothing a single trajectory, DUSTrack computes multiple overlapping LK-RSTC tracklets and averages the estimates at each frame. That is a real idea, clearly explained, and the integration with DeepLabCut into a general-purpose ultrasound tracking workflow is useful. The code and GUI are public, which is a concrete contribution. The augmentation study is also honestly done—a null result with a sensible perceptual-threshold estimate, reported without overclaiming.\n\nBut the central accuracy and jitter comparisons have a load-bearing circularity problem. The ground truth is described as \"DUSTrack-assisted human annotations\" refined over a few iterations. Roughly 20 of 600 frames are manually labeled; the rest are interpolated with LK-RSTC, the same algorithm that powers the filtering step. So when Figure 4e shows the filtered output's power spectrum matching ground truth, both signals share the same smoothing backbone. And when Figure 3a shows fine-tuned ResNet-50 beating zero-shot TAP models, the zero-shot models are being scored against a reference that already contains LK-RSTC's specific noise and bias profile. The paper never reports how many interpolated frames were manually corrected, and there is no independent validation (e.g., synthetic phantoms, sonomicrometry). This matters because the abstract's \"superior accuracy\" claim rests on it.\n\nThe fascicle comparison is also thinner than the text implies: four participants, and two of three specialized comparators were excluded due to failure cases. The statistical test is fine, but the parity claim rests on a very small sample. The abstract also conflates fine-tuning with the DUSTrack methodology—the zero-shot models are not given any per-video adaptation, so \"DUSTrack demonstrates superior accuracy\" is really \"fine-tuning on the test video beats zero-shot.\" That framing should be corrected.\n\nI would send this to peer review—the toolkit is useful, the filter idea is worth examining, and the problems are addressable—but the referee should ask for an independent ground truth or at least a sensitivity analysis that shows which frames were manually corrected and how the results change if those corrections are removed. Without that, the headline accuracy numbers are not yet trustworthy.","headline":"Useful open-source ultrasound point-tracking toolkit, but the main accuracy claims rest on a ground truth generated with the same optical flow algorithm used in the filter.","tokens_in":16917,"tokens_out":1452,"would_cite":true,"duration_ms":18590,"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":"DUSTrack claims a fine-tuned per-frame model plus an optical-flow filter can track arbitrary ultrasound points accurately enough to replace task-specific trackers.","keywords":["ultrasound point tracking","B-mode ultrasound","optical flow","Lucas-Kanade","jitter reduction","track-any-point models","muscle fascicle tracking","echocardiography"],"falsifier":"Have several experts manually label every frame, or use a phantom with known displacements, in the same upper-arm and gastrocnemius videos, then measure DUSTrack's filtered error against that independent ground truth; if the fine-tuned model no longer beats the zero-shot trackers or the filter no longer reduces jitter, the central claim fails.","tokens_in":15812,"feed_emoji":"🎯","tokens_out":7978,"duration_ms":559270,"temperature":0.7,"pith_summary":"DUSTrack is a semi-automated toolkit for tracking arbitrary points in B-mode ultrasound videos. The paper's central claim is that a model fine-tuned on about 25 hand-annotated frames tracks tissue points more accurately than current zero-shot point trackers, and that a bidirectional optical-flow filter removes frame-to-frame jitter without suppressing fast tissue motion. If this is right, one general workflow can replace task-specific methods for measuring cardiac wall motion, muscle deformation, and fascicle length and pennation angle, while also generating high-quality training data for future automated trackers.","feed_headline":"Fine-tuned tracker beats zero-shot models on ultrasound points","feed_subtitle":"DUSTrack couples deep learning with optical flow to track cardiac, muscle, and fascicle motion from few labeled frames.","key_machinery":"The load-bearing object is the LK-RSTC filter used in a transposed sliding-window configuration: Lucas-Kanade optical flow with reverse sigmoid tracking correction builds short tracklets between anchor estimates, and each frame's final position is the average of all overlapping tracklet estimates passing through that frame. Because the tracklets are short, they inherit Lucas-Kanade's low short-term error; because they are anchored to deep-learning estimates, the average stays globally accurate. The same optical-flow routine also generates interpolated labels for training, augments the training set, and refines the ground-truth annotations.","core_discovery":"DUSTrack combines per-frame deep-learning localization with a transposed sliding-window optical-flow filter. The deep model sees each frame independently, so errors do not accumulate into drift; the filter generates many short overlapping tracklets with a Lucas-Kanade optical-flow routine that tracks forward and backward with reverse sigmoid correction, then averages the tracklet estimates at each frame to suppress jitter. The paper reports that a ResNet-50 fine-tuned with 25 labeled frames has significantly lower root-mean-square tracking error than four zero-shot TAP models, that the filtered trajectories match the ground-truth power spectrum better than low-pass filters, and that DUSTrack matches a specialized fascicle tracker on fascicle length and pennation angle in a public dataset. It also demonstrates cardiac wall-thickness and chamber-diameter measurements across four cardiac cycles and muscle-deformation patterns during a reaching task.","pith_inferences":["Because the same optical-flow routine used in the filter also helps build the ground-truth labels, the accuracy comparison would be stronger with independent manual annotations or a synthetic phantom with known displacements; the reported gap could partly reflect shared algorithm bias.","A direct extension would be to fine-tune a video-based point tracker, since the best zero-shot model already showed the lowest jitter but lower accuracy, on DUSTrack-style annotations to combine global accuracy with temporal smoothness; the paper notes hardware costs but does not test this.","The transposed tracklet-averaging scheme is a general recipe for any jittery per-frame keypoint estimator, not only ultrasound, whenever local optical flow is reliable enough to bridge short intervals.","Clinical validation in patient populations remains untested; pathological tissue or atypical anatomy may violate the assumption that 25 frames in healthy volunteers generalize."],"forward_implications":["A single fine-tuned point tracker can yield clinical and biomechanical measurements that today require specialized pipelines: fascicle length, pennation angle, cardiac wall thickness, chamber diameter, and muscle strain.","Annotating about 25 frames in one or two motion cycles is enough to reach near-asymptotic accuracy; error drops sharply from 5 to 15 frames and then plateaus.","The final LK-RSTC filtering step makes the optional label-augmentation step imperceptible in output trajectories, so augmentation's main remaining value is helping annotators see and verify motion.","DUSTrack's measured accuracy of roughly 200 micrometers before filtering sits near the paper's estimated human perceptual threshold of roughly 100 micrometers, giving future zero-shot ultrasound trackers a quantitative target to beat."],"supporting_citations":[{"why":"Introduces the LK-RSTC algorithm with reverse sigmoid tracking correction that DUSTrack uses for tracklet generation, filtering, and label augmentation.","marker":"[18]"},{"why":"Provides the specialized fascicle tracker against which DUSTrack's fascicle length and pennation angle accuracy are compared.","marker":"[20]"},{"why":"Supplies the Lucas-Kanade optical flow algorithm whose bidirectional form underlies the LK-RSTC filter.","marker":"[23]"},{"why":"Supplies the fine-tuning workflow through which the ResNet-50 tracking model is trained and iteratively refined.","marker":"[24]"},{"why":"Defines the ResNet-50 backbone architecture used for per-frame point localization in DUSTrack.","marker":"[26]"},{"why":"Supplies the public gastrocnemius dataset and expert-annotation ground truth used in the fascicle tracking comparison.","marker":"[30]"},{"why":"Serves as the best-performing zero-shot baseline that the fine-tuned DUSTrack model is compared against in tracking accuracy.","marker":"[32]"},{"why":"Provides a second zero-shot baseline for the tracking accuracy comparison.","marker":"[33]"},{"why":"Provides a third zero-shot baseline for the tracking accuracy comparison.","marker":"[34]"},{"why":"Provides the fourth zero-shot baseline for the tracking accuracy comparison.","marker":"[35]"}],"fun_headline_variants":["DUSTrack: Semi-automatic ultrasound tracking beats zero-shot models","Deep learning + optical flow for robust ultrasound point tracking","Few labeled frames: DUSTrack beats zero-shot trackers in US","DUSTrack: Accurate cardiac and muscle tracking from few labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy numbers treat DUSTrack-assisted human annotations, which the same optical-flow routine helps interpolate and refine, as unbiased ground truth; if those labels are not independent of the tracker, the reported accuracy and jitter improvements could be partly self-fulfilling.","fun_headline_variants_meta":{"raw":{"variants":["DUSTrack: Semi-automatic ultrasound tracking beats zero-shot models","Deep learning + optical flow for robust ultrasound point tracking","Few labeled frames: DUSTrack beats zero-shot trackers in US","DUSTrack: Accurate cardiac and muscle tracking from few labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3255,"prompt_tokens":1002,"completion_tokens":2253,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":2193}},"tokens_in":618,"tokens_out":2253,"duration_ms":16399,"temperature":1.0,"reasoning_tokens":2193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:58:06.281453+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have several experts manually label every frame, or use a phantom with known displacements, in the same upper-arm and gastrocnemius videos, then measure DUSTrack's filtered error against that independent ground truth; if the fine-tuned model no longer beats the zero-shot trackers or the filter no longer reduces jitter, the central claim fails.","supporting_citations":[{"cited_title":"& Anthony, B","cited_arxiv_id":null,"evidence_quote":"Introduces the LK-RSTC algorithm with reverse sigmoid tracking correction that DUSTrack uses for tracklet generation, filtering, and label augmentation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the specialized fascicle tracker against which DUSTrack's fascicle length and pennation angle accuracy are compared."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Lucas-Kanade optical flow algorithm whose bidirectional form underlies the LK-RSTC filter."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the fine-tuning workflow through which the ResNet-50 tracking model is trained and iteratively refined."},{"cited_title":"F., Hullfish, T","cited_arxiv_id":null,"evidence_quote":"Supplies the public gastrocnemius dataset and expert-annotation ground truth used in the fascicle tracking comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides a second zero-shot baseline for the tracking accuracy comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides a third zero-shot baseline for the tracking accuracy comparison."}],"review_version":1}