DUSTrack: Ultrasound Point Tracking Framework

• DUSTrack is a deep learning and optical flow-based toolkit for tracking arbitrary points in B-mode ultrasound videos, effectively handling speckle noise, low edge contrast, and out-of-plane motion.
• It integrates a user-friendly graphical interface for sequential annotation with iterative model refinement and bidirectional Lucas–Kanade optical flow to reduce jitter while preserving rapid tissue motion.
• The framework demonstrates superior accuracy compared to zero-shot trackers and is open-sourced for practical use in clinical echocardiography and biomechanics research.

Searching arXiv for the exact DUSTrack paper and closely related tracking work. DUSTrack is a semi-automated framework for tracking arbitrary points in B-mode ultrasound videos. It combines per-frame deep learning with optical-flow-based refinement and filtering, and it is designed for use across clinical echocardiography and biomechanics where speckle noise, low edge contrast, and out-of-plane motion complicate longitudinal point tracking (Namburi et al., 18 Jul 2025). In the formulation presented in "DUSTrack: Semi-automated point tracking in ultrasound videos" (Namburi et al., 18 Jul 2025), the system integrates a graphical user interface for annotation and iterative model refinement, fine-tunes DeepLabCut models for framewise coordinate prediction, and applies a bidirectional Lucas–Kanade procedure with reverse sigmoid tracking correction together with a transposed sliding-window averaging scheme to reduce frame-to-frame jitter while preserving rapid tissue motion (Namburi et al., 18 Jul 2025).

1. Definition and scope

DUSTrack is expanded as Deep learning and optical flow-based toolkit for UltraSound Tracking. It is described as a semi-automated framework purpose-built for tracking arbitrary points in B-mode ultrasound videos, with the stated goal of meeting the accuracy and generalization needs that arise across clinical echocardiography and biomechanics while remaining lightweight and practical to use (Namburi et al., 18 Jul 2025).

The framework is motivated by three persistent difficulties of B-mode ultrasound point tracking: speckle noise, low edge contrast, and out-of-plane motion. The first produces granular textures that can obscure edges and confuse trackers; the second reduces the reliability of edge- or template-based methods; the third reflects the fact that tissues move in 3D while conventional probes capture 2D slices, leading to apparent changes in anatomy or disappearance of landmarks (Namburi et al., 18 Jul 2025). DUSTrack addresses these issues by treating arbitrary point tracking itself as the central primitive from which downstream measurements such as wall thickness, chamber diameters, fascicle length, pennation angles, and localized strain can be derived (Namburi et al., 18 Jul 2025).

The system is positioned as a general tracking toolkit rather than a narrowly specialized measurement algorithm. The paper states that DUSTrack demonstrates superior accuracy compared to contemporary zero-shot point trackers and performs on par with specialized methods, while also being released as an open-source solution (Namburi et al., 18 Jul 2025).

2. Architecture and processing pipeline

DUSTrack is organized around three components. The first is dataset creation and semi-automated annotation through a graphical user interface. The second is deep learning tracking through DeepLabCut. The third is optical-flow refinement and a flow-based filtering stage that suppresses frame-to-frame jitter without erasing fast motion (Namburi et al., 18 Jul 2025).

The GUI is designed around sequential annotation, which is explicitly contrasted with selecting visually dissimilar frames. This reflects the way sonographers interpret ultrasound and is intended to reduce labeling errors caused by speckle and non-linear tissue deformation (Namburi et al., 18 Jul 2025). The interface can also interpolate labels between manually annotated frames using Lucas–Kanade optical flow with reverse sigmoid tracking correction, producing augmented labels that can be verified and corrected by the operator (Namburi et al., 18 Jul 2025). The paper further reports support for verification and correction of interpolated labels, layer management, copy and trim tools, and interactive navigation (Namburi et al., 18 Jul 2025).

The deep learning stage fine-tunes DeepLabCut models, including ResNet-50, MobileNet, and EfficientNet backbones. DeepLabCut processes frames independently and outputs 2D coordinates per frame for each point label. This framewise inference is used specifically to avoid long-term drift accumulation characteristic of sequential algorithms (Namburi et al., 18 Jul 2025).

The refinement stage uses bidirectional Lucas–Kanade optical flow with reverse sigmoid tracking correction, followed by a transposed sliding-window filter. In this design, DUSTrack constructs short, overlapping tracklets using optical flow and averages them to suppress high-frequency noise while preserving rapid tissue motion. The paper presents this as a core contribution because conventional low-pass filtering either leaves residual jitter or over-attenuates true motion (Namburi et al., 18 Jul 2025).

3. Optical-flow formulation and filtering method

The optical-flow component is grounded in the standard Lucas–Kanade brightness-constancy relation

$I_x u + I_y v + I_t = 0,$

where $I_x$ and $I_y$ are spatial gradients, $I_t$ is the temporal gradient, and $(u,v)$ is the optical flow estimated within a local neighborhood (Namburi et al., 18 Jul 2025). For point propagation, the canonical update is given as

$p_{t+1} = p_t + v_t \,\Delta t,$

with $\Delta t = 1$ frame in the practical implementation context (Namburi et al., 18 Jul 2025).

Within a window spanning frames $[t_s,t_e]$, DUSTrack computes a forward tracklet $T^{f}(t)$ propagated from $t_s$ to $I_x$0 and a reverse tracklet $I_x$1 propagated from $I_x$2 to $I_x$3. These are blended by a smooth sigmoid weight $I_x$4 over normalized time:

$I_x$5

The paper emphasizes the directional interpretation of this weighting rather than a specific analytic sigmoid form: the weight smoothly transfers preference from forward to reverse tracking across the window (Namburi et al., 18 Jul 2025).

The final filtering stage aggregates all overlapping tracklets that include a frame $I_x$6:

$I_x$7

The paper states that the implemented method uses uniform averaging across overlapping tracklets, and identifies the novelty as the transposed sliding window together with bidirectional blending, rather than a specialized confidence weighting law (Namburi et al., 18 Jul 2025). This filtering design is presented as the mechanism by which DUSTrack suppresses high-frequency frame-to-frame noise while preserving fast physiological motion.

A key practical configuration reported in the paper is a 0.6-second window, corresponding to 30 frames at 50 Hz, which yields 28 intermediate frame estimates because two frames serve as anchor endpoints (Namburi et al., 18 Jul 2025).

4. Annotation workflow, training protocol, and implementation

The recommended annotation regime is intentionally sparse and sequential. The paper reports that approximately 25 frames within 1–2 motion cycles provides a strong accuracy-effort trade-off, and describes a typical ~600-frame clip in which users manually label ~20–25 frames before optional LK-RSTC augmentation expands the training set (Namburi et al., 18 Jul 2025). After initial training, users inspect predictions, add targeted labels in failure segments, optionally interpolate or correct those labels with Lucas–Kanade flow, and retrain. The process is repeated until the outputs are satisfactory, after which the final LK-RSTC filtering is applied (Namburi et al., 18 Jul 2025).

The GUI is implemented using the authors’ datanavigator and pysampled, with pandas for DeepLabCut I/O, decord for video reading, and OpenCV for Lucas–Kanade optical flow. Annotation layers are saved to JSON with frame numbers and 2D coordinates per label (Namburi et al., 18 Jul 2025). The paper does not report despeckling or contrast-enhancement preprocessing; robustness is instead attributed to task-specific fine-tuning and the post hoc flow-based filter (Namburi et al., 18 Jul 2025).

The paper indicates a layer name dlc_iteration-0_100000 in the GUI, suggesting training runs on the order of 100k iterations, but explicitly notes that optimizer details, learning-rate schedule, and batch size are not specified and instead follow DeepLabCut defaults (Namburi et al., 18 Jul 2025). The framework is described as favoring computational accessibility, including training on widely available 8 GB GPUs, and this accessibility is presented as part of the rationale for choosing DeepLabCut rather than heavier video foundation models (Namburi et al., 18 Jul 2025).

5. Evaluation, use cases, and reported results

The evaluation spans three application domains: upper-arm transverse B-mode ultrasound during reaching, medial gastrocnemius fascicle tracking, and echocardiography (Namburi et al., 18 Jul 2025). Ground truth is based on DUSTrack-assisted human annotations refined iteratively and manually corrected, typically across 2–4 refinement rounds (Namburi et al., 18 Jul 2025). Metrics include position RMSE, power spectral density (PSD) for jitter analysis, paired t-tests, and binomial tests for blind perceptual comparisons, with Bonferroni correction where applicable (Namburi et al., 18 Jul 2025).

On the upper-arm dataset with 36 healthy adults, a ResNet-50 fine-tuned with 25 labeled frames significantly outperformed several state-of-the-art zero-shot trackers. The reported paired t-test results are: CoTracker3 with $I_x$8, $I_x$9; BootsTAP with $I_y$0, $I_y$1; LocoTrack with $I_y$2, $I_y$3; and PIPs++ with $I_y$4, $I_y$5 (Namburi et al., 18 Jul 2025). The PSD analysis showed that CoTracker3 had jitter closest to ground truth among the zero-shot models, but still worse localization accuracy than the fine-tuned DeepLabCut model (Namburi et al., 18 Jul 2025).

For filtering, the paper compares the LK-RSTC scheme against low-pass filters at 10 Hz and 5 Hz. It reports that the LK-RSTC filtered trajectories are nearly indistinguishable from ground truth in the frequency domain, whereas the 10 Hz filter fails to suppress jitter adequately and the 5 Hz filter over-attenuates true motion (Namburi et al., 18 Jul 2025).

The effect of annotation augmentation is more limited. LK-RSTC augmentation reduces jitter in initial DeepLabCut outputs, with a paired t-test of $I_y$6, $I_y$7, but after final LK-RSTC filtering the PSDs are nearly identical with or without augmentation (Namburi et al., 18 Jul 2025). In blind human evaluation, the vote count was 127 for augmentation versus 109 without, with binomial test statistic 0.538 and $I_y$8, and 32.7% of responses labeled the outputs the same (Namburi et al., 18 Jul 2025). The paper further notes an average perceptual threshold for visible differences of approximately 100 μm, illustrated by reference lines at about 92 μm and 140 μm (Namburi et al., 18 Jul 2025).

On the public medial gastrocnemius dataset, using 5 participants, 18 annotated frames each, and 9 expert ratings per frame averaged as ground truth, DUSTrack was reported as comparable to UltraTimTrack. Specifically, there was no significant difference in fascicle length RMSE, with paired t-test $I_y$9, $I_t$0, and a trend to lower RMSE for pennation angle, with paired t-test $I_t$1, $I_t$2 (Namburi et al., 18 Jul 2025). HybridTrack and DLTrack were not included in that comparison because of severe failures on the dataset (Namburi et al., 18 Jul 2025).

The use-case demonstrations include cardiac wall motion tracking in a parasternal long-axis view from EchoNet-LVH, cross-sectional muscle deformation analysis during reaching, and fascicle tracking during ankle plantarflexion. In these examples, DUSTrack is used to derive quantities such as interventricular septum thickness, left ventricular internal diameter, muscle area change, fascicle length, and pennation angle from tracked point trajectories (Namburi et al., 18 Jul 2025).

6. Interpretation, limitations, and relation to nearby tracking paradigms

A central technical interpretation advanced by the paper is that DUSTrack exploits a complementary division of labor between deep learning and optical flow. DeepLabCut provides globally stable, per-frame coordinate estimates that do not accumulate drift, while Lucas–Kanade optical flow provides short-term local smoothness. The transposed sliding-window averaging then repeatedly reuses short-term optical-flow consistency to suppress jitter without inheriting long-term drift (Namburi et al., 18 Jul 2025). This suggests that DUSTrack should be understood less as a purely learned tracker than as a hybrid point-tracking framework in which learned anchors and local motion constraints are explicitly coupled.

The paper also makes clear that DUSTrack is semi-automated, not fully automated. Accuracy depends on label quality and on human perceptual limits, which the study places at roughly 100 μm under its conditions (Namburi et al., 18 Jul 2025). The optical-flow refinement is sensitive to frame rate, with lower frame rates reducing the method’s ability to capture fast motion (Namburi et al., 18 Jul 2025). The system does not explicitly model out-of-plane motion; instead, the recommended response is reinitialization and local relabeling where 3D motion disrupts the 2D slice (Namburi et al., 18 Jul 2025). Robustness in pathological clinical populations is also noted as not yet validated (Namburi et al., 18 Jul 2025).

A common misconception is that DUSTrack is simply an ultrasound-specific wrapper around DeepLabCut. The paper’s workflow argues against that simplification. The GUI-centered sequential annotation protocol, Lucas–Kanade interpolation for augmented labels, reverse sigmoid tracking correction, and the transposed sliding-window filter are all presented as essential to the framework’s reported performance (Namburi et al., 18 Jul 2025). Another plausible misconception is that conventional low-pass smoothing could substitute for the filtering stage. The frequency-domain comparison in the paper suggests otherwise: standard cutoff filters either leave residual jitter or suppress true high-frequency tissue motion (Namburi et al., 18 Jul 2025).

The name itself is potentially ambiguous in the broader literature. Other papers and implementation-oriented summaries in the supplied corpus use “DUSTrack” as a descriptive label for dust-particle tracking, aerosol motion estimation, or radar tracking pipelines, but the formally titled system "DUSTrack: Semi-automated point tracking in ultrasound videos" refers specifically to the ultrasound tracking toolkit described above (Namburi et al., 18 Jul 2025). This suggests a naming overlap rather than a shared technical lineage.

The toolkit is reported as open source, with code available at https://github.com/praneethnamburi/DUSTrack (Namburi et al., 18 Jul 2025). As presented, it functions both as a practical software system for current ultrasound research and as a foundation for future work on more automated ultrasound tracking, including possible integration of more advanced optical-flow variants such as robust M-estimators, sparsity constraints, or learned motion dictionaries (Namburi et al., 18 Jul 2025).